AI runs on tokens, tiny chunks of text you pay for going in and coming out. They cost tenths of a cent each, cheaper than the rounding error on a coffee run.

So why is your CFO losing sleep? Companies are burning a year’s AI budget by spring. Nvidia’s CEO said he would be alarmed unless his top engineers were each torching a quarter-million in tokens.

Here is the paradox: tokens keep getting cheaper, fast, and the bills keep climbing anyway. The same task can cost ten times more from one run to the next, and the bill barely explains why.

The products that win hide the token machinery and sell you the result instead. Some already do.

together with Lovable:

Lovable is one of those products. You describe the feature, and the AI ships inside your app, with zero API keys and the token machinery handled for you.

50M projects already run on it, 4 in 5 built by non-technical founders:

▫️ Hundreds of connectors for context: CRMs, databases, calendars, comms

▫️ A new Settings tab with per-feature credit usage, so the bill finally tells you something

10% off for AI Corner readers:

Start building with Lovable !

Table of Contents

1. The Bills Exploded Because Tokens Got Cheap

2. The Free-Money Era Just Ended

3. Signal and Noise Wear the Same Uniform

4. Tokens Are the New Headcount

5. The Gold Rush to Measure Tokens

6. The Hidden Winner Sells Outcomes, Not Tokens

1. The Bills Exploded Because Tokens Got Cheap

Here is the fact that breaks the popular story. Tokens did not get expensive. They got radically cheaper, and the spending went up anyway.

That reads like a contradiction until you remember a 160-year-old observation about coal, which turns out to describe AI better than almost anything written about AI this year.

Cheaper Every Year, by an Order of Magnitude

The price of running a model has been falling roughly tenfold a year, a slide some investors call “LLMflation.”

A unit of inference that cost around sixty dollars per million tokens when the GPT-3 API launched in 2020 now goes for about a dime on the economy tier. That is a 600-fold drop in under six years, faster than Moore’s Law ever moved silicon.

Nvidia’s Jensen Huang named the dynamic from the stage. Cost per token keeps falling while total AI spend keeps climbing, because cheaper inference pulls in demand that did not exist before.

He started filing tokens under cost of goods sold, the same budget conversation as energy and payroll.

When the man selling the shovels says the gold is getting cheaper to mine and the spending is still going vertical, that is worth sitting with.

The Architecture Is What Changed

Cheap tokens did not just mean more of the same usage. They made an entirely more expensive way of computing viable.

When inference is costly, you write a prompt and take an answer. When it is cheap, you build an agent that loops.

It retrieves, reasons, calls a tool, checks the result, retries, escalates, and only then responds. Every one of those steps burns tokens, and the loop can run dozens of times for a single request.

Researchers call this the structural version of the old paradox. As prices fall, companies reach for more elaborate architectures, and the token multiplier from that complexity swamps the per-unit savings.

The budgets broke for a subtle reason. Cheapness made expensive behavior rational, and everyone adopted it in the same quarter.

The frontier makes the squeeze worse, with the cost of running top-tier models rising several times a year even as the price of a single token keeps dropping. Two trends pointed in opposite directions, and the buyer feels both at once.

2. The Free-Money Era Just Ended

For two years the corporate instruction was simple. Use more AI.

Flood every team with tools, reward experiments, worry about the bill later. Later arrived.

The same executives who pushed adoption are now installing tiered access, mandatory efficiency reviews, and hard caps on who is allowed to spend.

The Receipts Are Public Now

Uber drained its annual AI coding budget in roughly four months, and its operating chief said the spending is getting hard to justify against the returns he can actually see.

Amazon pulled an internal leaderboard that had turned AI usage into a game among engineers.

Microsoft canceled a swath of internal coding subscriptions.

One consultant described a client that ran up around half a billion dollars of model usage in a single month before anyone noticed.

A senior technology executive at a major financial firm put the mood in a single line to the Wall Street Journal, speaking anonymously. The free-money period for AI is over.

The squeeze hit hardest at firms that locked multi-year deals before anyone understood their real usage patterns.

Nobody says that at the top of an adoption curve. It is what people say once the meter has started to scare them.

The Inversion Nobody Priced In

The pitch for all this spending was substitution. Inference replaces labor at lower cost, so you trade headcount for compute and pocket the difference.

A Nvidia vice president flagged the problem with that math out loud. In some organizations, compute costs have already passed human labor costs.

Read that slowly, because it turns the whole thesis upside down.

A company that cut staff on the promise that AI would absorb the work, then watched its model bill climb past the salaries it eliminated, sits in the worst possible spot. Fewer people. A bigger bill. No obvious route back to the efficiency that justified the cuts.

That is a strategy coming apart in real time, not a line item that needs trimming.

3. Signal and Noise Wear the Same Uniform

SaaS trained a generation of operators to read usage as a stand-in for value. More queries, more seats, more logins meant the thing was working.

AI severs that reflex at the root, and the reason is mechanical rather than philosophical.

The Meter Cannot Tell You What Happened

The same workflow on the same input can consume five to ten times more or fewer tokens depending on the prompt, the context retrieved, the model chosen, the tools called, and how often the agent had to retry.

The unit on the invoice holds steady. The amount of real work it stands for does not.

So a rising bill is, at the same time, the evidence that valuable work got done and the evidence that compute leaked into bad prompts, bloated context, and redundant reasoning. The number alone cannot separate the two.

Two companies with identical token bills can be running completely different operations underneath. One is converting inference into outcomes. The other is paying for expensive thrash that looks exactly the same on the line item.

To be fair to the optimists, plenty of that spend is real work. Blackstone said model spending across its portfolio companies rose fifteenfold in a single quarter year over year.

About 11 percent of the live backend code Uber ships now comes from AI agents handling ride matching, pricing, and bug fixes.

The point is not that the spend is wasted. The point is that the invoice cannot tell you which half is which.

What the Survivors Did Differently

The most-cited evidence here is MIT’s NANDA study, and its real finding cuts sharper than the headline.

Across hundreds of deployments, 95 percent produced no measurable profit impact. The gap was not model quality. It was a learning gap, systems that never adapted to the workflow they were dropped into.

The 5 percent that worked shared a discipline. They aimed at back-office operations rather than splashy front-office demos, judged the work by business outcomes instead of benchmark scores, and refused to scale anything that did not already pay for itself.

The lesson buried in that failure rate is plain. Value showed up only where someone defined the outcome before turning on the meter.

4. Tokens Are the New Headcount

The budget fight inside these companies looks like a finance argument.

It is a power struggle wearing a finance costume, and missing that is the fastest way to misread what is happening in boardrooms right now.

Headcount Was the Old Power Marker

For thirty years, the visible marker of a senior executive was the size of the organization they ran. Directs, skip levels, total headcount. Scope equaled status.

When intelligence becomes the scarce resource inside a company, that marker moves. The new measure of seniority becomes how much of that intelligence you direct.

This is why the allocation fight runs so hot. Whoever controls the token budget controls the AI equivalent of org scope, and a thirty-year instinct does not surrender quietly.

The wars in the phrase are literal. They name a contest over who owns the most valuable resource the company now buys, dressed up as a debate about cloud bills.

For an operator, the move is uncomfortable and obvious. The career hedge stopped being the size of the team you protect.

It became the ability to point a swarm of agents at a hard problem and come back with a result the business can actually price.

The managers who treat the token budget as someone else’s spreadsheet will wake up reporting to the ones who learned to allocate it.

Why the Fight Lands on Outsourcing First

When AI spend competes with labor, executives need a baseline, and outsourced work offers the cleanest one.

A business process contract is already priced in finished units. Cost per ticket, per claim, per invoice, per reviewed contract. That makes it the easiest place to stage the comparison between a human and an agent.

Internal labor resists the same scrutiny, because employees do a hundred fuzzy things and nobody volunteers their own team for the chopping block.

And the honest numbers, when they finally surface, run humbler than the marketing.

On coding tools, the celebrated returns shrink to roughly 1.6 times once you count the full token cost instead of the seat license alone.

Still positive.

Nowhere near the tenfold story that sold the budget upstairs.

5. The Gold Rush to Measure Tokens

A new category is forming at speed, and it is the predictable reaction to a measurement problem.

If the meter is noisy, sell people a better meter. Capital is already pouring into the idea that the cure for bad token data is more token data.

The Category Is Real and Crowded

At the FinOps industry’s flagship 2026 event, the dominant theme got branded the Great Token Panic.

Vendors shipped tools that trace every model call down to the user, the agent, and the workflow that triggered it.

One platform demonstrated catching a single employee’s surprise 76,000 dollar token spend within minutes of it happening.

The established cost management players are racing the same direction, bolting token attribution onto dashboards they already sell.

The reframing they are all chasing is the one executives now demand. The question stopped being what did we spend on AI. It turned into what did we get for it.

The tell is who shows up to buy. This is no longer a developer tools story.

It is finance, sitting in the same chair it took over cloud spend from a decade ago, walking in the moment a cost line gets big enough to need a referee.

Why Counting Harder Will Not End the War

Here is the limit nobody selling a dashboard wants to say out loud. The value of a token is irreducibly contextual.

The same trace is signal in one workflow and waste in another, and no amount of attribution fully resolves that, because the thing being measured, the business value of a unit of inference, stays partly unknowable before the work runs.

Better instrumentation catches anomalies and trims obvious waste. Useful, real, worth paying for.

But anomaly detection is a smoke alarm. It tells you something is burning. It cannot tell you whether the fire was worth lighting.

It does not make spend legible enough to settle the allocation fight, because perfect attribution fails for the very reason the meter is noisy in the first place.

The measurement layer is scaffolding for whatever replaces the token entirely, not the building.

6. The Hidden Winner Sells Outcomes, Not Tokens

Watch what the model providers are doing as they walk toward the public markets, because it tells you where this ends.

OpenAI is weighing deep cuts to what it charges per token, expecting Anthropic to follow.

Both are filing for the largest tech listings in years against revenue run rates that have gone near vertical.

Run that through the coal paradox and a price cut stops looking like a gift. Cheaper tokens pull more consumption, and the demand is elastic enough to count, with roughly a 0.29 percent usage drop for every 1 percent the price goes up.

Cut the price, grow the volume, and reset every customer’s blown budget right before handing the SEC a filing that needs to show durable demand. The favor to the buyer is the land grab for the seller.

The endgame is plainer than the attribution gold rush makes it look. The winner will not be whoever measures token-to-outcome best.

It will be whoever stops selling tokens and starts selling the outcome itself. A resolved ticket, a reviewed contract, a processed claim, at a fixed price, eating the token variance on their own books.

That move solves the customer’s measurement problem by swallowing it whole, and only the labs control cost-to-serve well enough to price it.

The token was never going to survive as the thing you buy. It was the meter on the way to a market that sells finished work, and the companies printing the tokens will be the first to stop counting them.

And it is still short.

CFO Sarah Friar runs the most expensive build ever attempted. In one hour she laid out how OpenAI thinks about compute, margins, pricing, and the Jony Ive device, and answered every question with numbers.

I watched the full hour so you can skip it.

Here are the 10 takeaways that matter, and the one thread under all of them:

The personal version: the people who pull genuine productivity out of AI climb the curve fastest. That starts with knowing how to drive the tools.

together with Outskill:

Most people use 5% of Claude. Master the rest in one weekend.

Join the world’s first live Claude-Athon, a 2-Day Claude AI Mastery Workshop that condenses 800+ hours of research into 16 hours, completely free:

▫️ Hands-on with Skills, Connectors, Cowork, and MCP

▫️ Build, scale, and deploy with 10+ AI tools, LLMs, and workflows

▫️ Walk out shipping AI workflows you keep

Live Saturday and Sunday, 10AM to 7PM EST

Block your FREE Seat

1. The gigawatt equation: 1 GW of compute equals $10B a year

“One gigawatt is roughly equivalent to about $10 billion a year of revenue to OpenAI.”

One number explains the whole capital strategy. More compute, more tokens, more revenue capacity. The conversion runs direct.

The same formula governs Anthropic, Google DeepMind, and every model provider you build on. Compute is the one input that maps straight to addressable market.

Why it matters: your vendor’s compute ceiling is your product ceiling. The scarcity is a multi-year cap on what any AI company ships, yours included.

2. The compute crunch is already here

“The landscape right now, in 26, if you want to buy more compute, good luck to you. Tell me, because I don’t know where else to find it. In 27, it’s pretty limited as well, frankly.”

Data centers breaking ground today produce usable compute in late 2027 at the earliest. OpenAI is already buying for 2028. The shortage that worries her starts in 2030.

Her named choke points: energy and powered land, permitting speed, memory costs spiking now, the talent pipeline, and community trust. Plan your AI products inside that constraint.

Why it matters: if you ship AI in the next three years, you build inside a constrained token environment that touches every layer downstream.

3. A 97% cost drop in two years

“From 4 to 5.4, the deprecation cost was something like 97%. That happened in like two years.”

OpenAI priced GPT-5.5 above 5.4, and you still get a 20 to 30 percent cost reduction per token, because the model runs more efficient. The price went up while the cost per unit of intelligence went down.

That compounds: each new release lowers the cost per token on older runs, the multi-chip strategy keeps OpenAI on efficient hardware, and 900 million weekly users drive volume that lowers per-unit cost further. The mechanics sit in any serious margin model.

Why it matters: price on today’s costs and you misprice tomorrow’s value. Lead the cost curve, the way the best pricing playbooks tell you to.

4. One cube became a quintillion configurations

“Today we sit on top of every CSP: Oracle, CoreWeave, Microsoft, GCP, AWS, and a bunch of small neoscalers.”

Two years ago OpenAI had one cloud, one chip, one product, one price. Today the infrastructure is a Rubik’s cube with a quintillion settings. NVIDIA leads the chip roster, the next training run is on Vera Rubins, AMD is next, OpenAI and Broadcom are building a chip together, and Cerebras is live for low-latency coding.

The strategic point: CSPs convert CapEx into OpEx, so you pay as revenue arrives, rather than upfront. That flexibility is the moat.

Why it matters: lock yourself to one cloud and one chip and you fall off the frontier. Optionality is a financing strategy, rather than indecision.

5. The $2,000-a-month bet looked absurd a year ago

“We think they will pay upwards of maybe $2,000 a month for it, which is kind of laughable in hindsight. But nobody believed. They were like, I don’t even know what she’s talking about.”

She heard the same skepticism about ChatGPT Pro at $200. The market underprices what people pay for genuine intelligence, every time. Free users ask 7 questions a day; Pro users engage 11x that.

Once someone tastes the productivity, they climb the commitment curve. The agentic tier is the same pattern, one rung higher, the kind of leverage the five-agent build makes concrete.

Why it matters: if you price an AI-native product on conventional SaaS logic, you are too conservative. Price for what the tool does in 18 months.

6. If Google and Meta had a baby, it would be ChatGPT

“If Google and Meta had a baby, it would be ChatGPT. You have Google search’s high intent. You have Meta’s demographic targeting. And we have memory on top of both.”

Google knows what you want now. Meta knows who you are. ChatGPT holds both, plus persistent memory across every conversation. OpenAI already holds at least 11 percent of the search market, and that figure undercounts it, since one long conversation counts as a single query.

Her ad commitments: sponsored results stay subordinate to model output, an ad-free tier persists, and ads reach the free tier.

Why it matters: if you build ad-dependent consumer products, this is your serious medium-term competition. Your SEO and distribution strategy needs a rethink toward ranking when AI answers.

7. The layer closest to the customer captures the margin

“Where everyone is trying to make sure they reside is the layer that is closest to the customer, where usually you take the largest portion of the profits of the ecosystem. No one wants to find themselves abstracted away.”

NVIDIA makes chips, and now ships models. Google runs cloud, and now makes chips and models. OpenAI builds models, owns the consumer interface, and is building chips with Broadcom. Every layer is climbing toward your customer.

LLM commoditization stalled, because the agentic layer pushed the other way. Memory and context create switching costs that compound, which is the moat the SaaS defense playbook is built around.

Why it matters: every layer below you wants your customer. Your moat is the context you accumulate, rather than the model you rent.

8. The Jony Ive device ships this year

“By the end of this year, we will unveil it. Early next year, you’ll be able to buy it. I have seen it. I’ve tried it. What Johnny and team are really good at is bringing humanity to devices. It feels very natural, but it feels very lovable.”

A CFO made the biggest product reveal of the interview, and described it in emotion over specs. The context she set: the last generation trained you to talk with your thumbs, and this device exists to solve exactly that. Multimodal, instant, screen-free.

Why it matters: when it ships, every screen-first assumption you designed around faces a direct challenge. The interaction layer you build for today differs from the one that dominates in 18 months.

9. Enterprise and consumer revenue are now 50-50

“Right now, our revenue is getting pretty balanced, about 50-50. People are really moving on AI right now.”

OpenAI is a consumer and enterprise company at equal scale. The value proposition to enterprise runs past cost reduction: the model acquiring the intuition of a business.

Her example: the data says the stock rises post-earnings, but the trader knows a fund is forced to sell, so it falls. AI wired to memory and context starts to replicate that institutional read.

Why it matters: treat the model as a chat box and you miss the value. It compounds in the institutional memory layer, so start building it now.

10. The IPO is a milestone. Durability is the goal.

“In the end, the market is a weighing machine, not a popularity machine. No one remembers who went first, Google or Yahoo, Lyft or Uber.”

OpenAI raised $122 billion so the IPO question would stop mattering. The previous fundraise record was Saudi Aramco near $30 billion. Anthropic filed its S-1 confidentially mid-interview, the moderators tried to make it a race, and she declined.

Q1 2026 alone saw $80 billion raised. The capital environment rewards durability over timing.

Why it matters: treat an IPO as a destination and you optimize for the wrong things. Treat it as financing and you optimize for durability, the only strategy that survives a decade-long build.

What this means for you

Friar ran an hour-long masterclass in capital, infrastructure, and positioning for an unprecedented era. The numbers are large, the timelines long, the logic transferable.

• Founders: the pricing lesson is the most actionable. Your users pay more than you think for genuine intelligence. Stop pricing for today and price for what the tool does in 18 months, the way top operators do.

• Investors: the compute-scarcity thesis is underpriced across most portfolios. If your companies depend on token access and skipped a constrained 2026 to 2028 model, they skipped the actual risk. Audit your portfolio’s token dependency now.

• Operators: enterprise value lives in the institutional memory layer, rather than the chat box. A search box or a summarizer is the easy half. Value compounds when the model holds your company’s intuitions, so build toward that layer.

The 5 principles to steal

1. Invest ahead of demand. OpenAI bought compute when it looked crazy and prescient 18 months later. Your version of that call is happening now.

2. Optionality is a financial strategy. Maximum flexibility is how you survive a decade-long build on unpredictable demand.

3. The layer closest to the customer captures the margin. Your moat is the memory and context you accumulate, rather than the model you rent.

4. Pricing leads the cost curve. Price on today’s costs and you misprice tomorrow’s value.

5. Durability wins. The weighing machine measures what you built, rather than when you listed.

The compute shortage is here. The pricing ceiling sits higher than you think. The intelligence layer is worth fighting for.

Keep reading

The AI build

▫️ AI tools and models library

▫️ AI agents library

▫️ The SaaS defense playbook for the AI era

▫️ Anthropic passed OpenAI at $30B ARR

Capital, pricing, and the market

▫️ Q1 2026 US fund activity and record fundraising

▫️ The startup pricing journey

▫️ What top VCs look for in 2026

▫️ The investor list of lists

Full interview:

Only about 4% of a GPU’s silicon did the actual AI math. The rest was graphics machinery, carried along out of habit. Andrew Feldman’s read: the hard part of deep learning is moving data around, rather than doing the multiplication. So the right machine gets built around data movement, from a blank sheet.

That idea lived in a Series A deck marked working name, from a company so early its identity was still a working title and its vocabulary predated the word transformer. Most people who saw it thought it was crazy.

In May 2026, that company, Cerebras, IPO’d at a valuation near $56 billion, the biggest US tech IPO since Uber.

Here is the part worth keeping even if you stop here. Two questions decide a bet like this, and the deck answers both:

1. Is the demand curve about to bend? They saw AI compute going vertical while the broader industry still treated AI as a science project.

2. Is the incumbent tool accidentally good, or deliberately good? The GPU was accidentally good at AI, a graphics chip pressed into service. That gap is the opening.

When demand is bending and the incumbent is only accidentally good, you have the setup for a generational company. That lens works on any hardware bet you evaluate.

Now the concrete part. We got the confidential Series A deck, and the numbers behind the climb. Below the line:

▫️ The confidential Series A deck, all 20 slides, with what they pitched versus what they shipped

▫️ The 135x bridge, the engineering slide that splits the headline claim into four honest multipliers

▫️ The revenue climb, $25M to $510M, with the growth rates and the swing to profit

▫️ Every round, from the seed off this deck to the $23B private mark to the $56B IPO

▫️ The IPO play-by-play, the pricing, the 68% pop, and the near-halving six weeks later

▫️ The risk file, the customer concentration and OpenAI dependency the headlines skip

▫️ The operator lessons, what this deck teaches anyone building in AI right now

1 subscription unlocks every teardown

This is one breakdown in a growing library. Premium opens all of them:

▫️ The AI tools and models library

▫️ The business and investing library

▫️ The Claude and Anthropic library

Plus a fresh teardown every week. One deck like this, read before the room reads it, pays the subscription back many times over.

📂 The Cerebras Series A Deck

The full confidential deck, the slide-by-slide teardown, the revenue and funding history, the IPO play-by-play, and the risk file, in one place.

Get The Cerebras Series A Deck below 👇

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

Most people pour effort into Claude every week and watch it vanish by the next session. Re-explaining, re-prompting, re-pasting. It feels like work. By the next week it is gone.

A small group flipped that. They spend one hour on a setup that pays back on every task after, then they delegate the rest. This is that setup, copy-paste ready.

Master it once, and Claude stops being a chatbot you wrestle and becomes the teammate who already knows how you work.

Here is the full build :

▫️ The leverage map, the one rule that sorts every Claude habit into compounds or evaporates

▫️ The 10 moves that move the needle, each with the copy-paste asset

▫️ The one context file that lifts output more than switching models, with the template

▫️ The skill you build once that retires your whole prompt library

▫️ The verification loop, the single highest-ROI habit, and the prompt that runs it

▫️ Where to spend your effort and your tokens, the budget that splits power users from busy ones

▫️ The traps to drop, the work that only looks productive

▫️ The 1-hour setup and 7-day plan, plus a self-audit scorecard

One subscription unlocks every build kit

This is one kit in a growing library. Premium opens all of them:

▫️ The Claude best practices power-user guide

▫️ Loop engineering for coding agents

▫️ Claude managed agents guide

▫️ The Granola and Claude second-brain stack

▫️ The one-person startup operating system

Plus a fresh build kit every week. The one hour of setup below pays the subscription back the same afternoon.

🧭 The Claude Leverage Kit

The full system in one place: the leverage map, the 10 moves with copy-paste assets, the context file template, the skill builder, the verification loop, the effort-and-token budget, the traps to drop, and the 7-day plan.

Get The Claude Leverage Kit below 👇

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

In 2024, a junior paralegal reviewed contracts by hand. In 2026, the same desk auto-flags 328 issues across 12,842 documents. That job was eliminated, rather than upgraded.

Of 1.2 million layoffs in 2025, about 4.5% cited AI. Yet Stanford found a 13% employment drop for workers aged 22 to 25 in the most AI-exposed roles. AI leaves senior careers standing and removes the first rung into them. The $150K paralegal keeps her job. The junior role that led to it has vanished.

Here are 6 places it is happening right now, with the numbers.

together with Gelt:

▫️ PTE elections missed in Q2

▫️ S-corp savings you needed an entity in place to claim

▫️ Retirement contributions left unmentioned

The moves that lower your tax bill happen now. Gelt gives you a team of tax strategists who reach out to you, backed by AI, built for founders with genuine income complexity.

See your savings with Gelt’s calculators, then get 10% off as an AI Corner reader:

Build your tax strategy (10% off)

Table of Contents

1. Paralegals: The Document Review Floor Dropped Out

2. Customer Service: Klarna Already Did the Math

3. Freelance Writers: The Numbers Nobody Wants to Cite

4. Junior Developers: The On-Ramp Closed in 2023

5. Bank Back-Office: 200,000 Seats Already Marked for Removal

6. The Replacement Roles: What Comes Next Pays Less

1. Paralegals: The Document Review Floor Dropped Out

In 2024, the paralegal market had a clean structure. Junior paralegals handled document review and legal research. Senior paralegals handled judgment-heavy work. The junior role was the training ground. The senior role was the destination.

That structure is now broken at the bottom.

The Numbers Behind the Shift

Salaries for paralegal roles in document review and legal research have declined 10 to 15%. That decline is not spread evenly. It is concentrated at the entry level, in the positions doing the work that AI now handles faster and for less.

At the same time, AI-augmented senior associates are seeing compensation rise 20 to 30%, because their judgment is now amplified rather than replaced. One senior associate with the right AI stack does what a senior associate plus two junior paralegals did before. The junior headcount does not get hired. The senior gets better. The firm holds margins with a smaller team.

What the Missing Rung Actually Cost

What is gone is not the paralegal profession. What is gone is the entry point into it. The junior role was where you learned to read contracts, flag inconsistencies, build the pattern recognition that eventually made you valuable. That learning now happens somewhere else, or it does not happen at all.

The senior roles do not disappear. They just become harder to reach.

2. Customer Service: Klarna Already Did the Math

Most AI job displacement stories are predictions. The Klarna story is not.

In 2025, Klarna’s CEO Sebastian Siemiatkowski said what most executives are thinking but not saying: their AI assistant now handles the work that 700 customer service agents used to do. Klarna’s headcount, which peaked above 5,000, had dropped to around 4,000 by early 2026. The direction of travel was explicit, not implied.

What Survived and What Did Not

The AI handles high-volume, repeatable transactions. Returns, refund status, account questions, password resets. The human roles that remained are the ones dealing with escalations, edge cases, fraud disputes, and situations requiring judgment that cannot be scripted in advance.

That is exactly the task-versus-job distinction that matters here. The tasks are gone. The jobs that were mostly tasks are gone with them.

For founders and operators: customer service is no longer a headcount scaling story. It is an exception management story. The companies that have not restructured around that insight are carrying cost they no longer need to carry, and the companies that have restructured already have a cost structure you will eventually need to match.

3. Freelance Writers: The Numbers Nobody Wants to Cite

The research from Washington University is the most specific dataset on this topic, and it is cited far less than it should be.

Freelance writers on Upwork lost approximately 2% of contracts and 5.2% of earnings following the widespread deployment of ChatGPT. Those numbers sound modest until you multiply them across a platform with millions of active freelancers. A 5% earnings decline at that scale is not noise. It is a structural repricing of an entire category of work.

Where the Damage Is Concentrated

Image freelancers were hit harder than writers. The study found a more significant decline in that segment. Image generation tools produce a finished output from a two-sentence prompt. The task is eliminated more completely, more immediately.

The decline is real and it is concentrated in the most commoditized work. Listicles, product descriptions, SEO filler, template-driven content at scale. The writers doing that work were always in a race with efficiency.

That race ended.

4. Junior Developers: The On-Ramp Closed in 2023

Every article about AI and software development focuses on the senior engineer. The senior engineer is fine. The question nobody is asking is what happens to the person who has not yet become a senior engineer.

Stanford’s Digital Economy Lab found a 13% relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations. This is not coming from senior roles being cut. It is coming from the entry-level positions that organizations used to hire for routine coding tasks, QA, basic feature work, and ticket resolution.

The Compounding Silence in Hiring Data

A senior engineer with Claude or Cursor handles the backlog that used to require two junior hires. The team ships faster with the same headcount. The junior position does not get created. Three years from now that reads as a talent development decision, not a productivity win.

The profession is not threatened. The on-ramp is.

And the problem will not appear on anyone’s radar until 2028, when the organizations that stopped hiring juniors in 2025 discover they have no mid-level pipeline. Every senior engineer on your team was once a junior. The senior engineers you need in five years are the people who are not getting hired right now.

5. Bank Back-Office: 200,000 Seats Already Marked for Removal

Wall Street banks have been transparent about this in a way other industries have not. The numbers are large enough to require disclosure, and the largest institutions have stopped treating this as an internal strategy question.

Approximately 200,000 jobs across US banking will be removed over the next three to five years, concentrated in entry-level and back-office functions. This is not a leaked internal memo. It is the public-facing strategic direction of the largest financial institutions in the world.

The Work That Actually Goes Away

Back-office banking is structurally almost entirely tasks. Trade settlement confirmation. Regulatory compliance reporting. Loan document processing. Know-your-customer data entry. High-volume, rules-based, structured-input operations. These are the conditions under which AI performs best and where human labor was most expensive relative to the actual complexity of the work.

The relationship manager who knows a client’s real risk tolerance, the credit analyst making a judgment call on a complicated deal, those people are not going anywhere. The 22-year-old who used to spend two years learning the business from inside a compliance spreadsheet is not getting that job.

The question is who enters financial services in five years, having never had that training, and what that means for the quality of senior judgment a decade from now.

6. The Replacement Roles: What Comes Next Pays Less

Here is the number that closes the loop on everything above.

The jobs being created to replace the ones described in this article are real. AI Operations Coordinator, a role that did not formally exist in 2023, now appears widely across major job boards at $50,000 to $70,000 annually. Data Quality Analyst, the human who reviews and corrects AI outputs, runs $55,000 to $75,000. These are not bad jobs. They are real positions with real employers paying real salaries.

They just pay 20 to 35% less than the roles they are absorbing displaced workers from.

The Transition Tax Nobody Is Counting

The worker who was a junior paralegal earning $68,000 and is now an AI Compliance Reviewer earning $52,000 did not lose their job in the way the doom narrative describes. They transitioned. The transition cost them $16,000 a year. That is real money affecting a real person’s rent, retirement contribution, and long-term career trajectory. Multiply it across a workforce and it is a structural income shift that does not show up in unemployment statistics.

The doom camp counts bodies and finds few. The optimist camp counts new job titles and finds opportunity. Neither is looking at what the jobs pay, where they lead, or what the 10-year earnings trajectory looks like for someone who started their career in 2026 instead of 2022.

Both camps are answering the wrong question.

The Hiring Problem Nobody Has Calendared

The AI Corner audience is founders, operators, and investors. Most of you reading this are not the one losing a job to AI.

But most of you have also quietly stopped hiring at the entry level over the last two years. Some did it deliberately because AI handles the output. Some did it by default because headcount pressure made it easy to justify. The result is the same either way.

The Bill Arrives in 2029

The problem with stopping the junior pipeline is not 2026. The problem is 2029.

Every mid-level operator on your team was once a junior. Every senior engineer spent two years writing code nobody shipped to production. Every account manager who now closes complex enterprise deals started by answering support tickets and learning the product from the complaint queue up. The person making the hard judgment call today was once the person doing the document review, running the compliance spreadsheet, triaging the support queue.

That learning took time. It required the entry-level role to exist.

When AI handles the tasks that junior roles were hired for, those roles stop getting created. The short-term cost savings are real and measurable. The long-term talent pipeline consequence is invisible for three to four years and then suddenly very visible all at once.

The entry-level role was never primarily about the work it produced. It was about the person it built.

That is what AI cannot replicate. And right now, most organizations are eliminating it without any line on their strategic planning agenda for what replaces it.

The $150K role is not gone. The $60K role that was the only path to it is. And nobody has put that on their three-year hiring plan yet.

I built that system into a free AI Creator Kit: one Notion workspace with AI agents, templates, a creator CRM, and partner credits worth thousands. Below, how it works and the data on why it matters now.

together with Outskill:

Most people use 5% of Claude. Master the rest in one weekend.

This live 2-Day Claude AI Mastery Workshop condenses 800+ hours of research into 16 hours, completely free.

▫️ Hands-on with Skills, Connectors, Cowork, and MCP

▫️ Build, scale, and deploy with 10+ AI tools, LLMs, and workflows

▫️ Live Saturday and Sunday, 10AM to 7PM EST

Block your seat FREE

Here is what changed under all of it. Buyers now research you through AI before a single conversation happens.

As of the 2026 6sense Buyer Experience Report, 94% of B2B buyers use large language models to synthesize their research, and 95% of deals go to a vendor already on the buyer’s day-one shortlist. The work of getting on that shortlist happens long before anyone fills in a form.

That is why I built a free AI Creator Kit that gives you the infrastructure in one Notion workspace: AI agents, ready-to-use templates, a content calendar, a creator CRM, and partner credits worth thousands.

Get the free AI Creator Kit

What you get

A content system that runs itself:

▫️ An AI-powered social media calendar built in Notion

▫️ A creator CRM to track partnerships and outreach

▫️ A memory-persistent AI content agent that holds your brand voice across every post, thread, and article

▫️ Slack, email, and meeting notes syncing automatically into your Notion databases

▫️ Influencer and creator partnership guides for your first collaborations

Free credits included:

▫️ Notion Business, 3 months free, a $6,000 value

▫️ CapCut, 7 days free for short-form video

▫️ Lovable, a discount to build custom content tools

▫️ Secret, a free trial for 600+ startup deals

How the AI content engine works

Most AI content tools fail founders for one reason: they start from a blank box and forget everything the moment you close the tab. This kit runs the opposite way, on the same agent pattern behind any reliable AI agent.

The content agent reads the context already living in your Notion workspace, your positioning, past posts, product notes, customer language, and writes from that. It stays on voice because the voice is on record, the core idea behind good context engineering.

The engine sits on four layers:

▫️ Context. Your workspace holds the brand voice, the offers, and the audience, so every draft starts informed over generic.

▫️ Capture. Slack, email, and meeting notes sync into your databases on their own, the same Notion-and-MCP pattern in the second-brain stack, so every idea reaches the system on its own.

▫️ Creation. The agent plans the calendar and drafts posts, threads, and articles from that context, the way the managed agents pattern lets one agent run a recurring job.

▫️ Compounding. A creator CRM tracks every partnership and collaboration, so reach builds on reach rather than resetting each month.

You set it up once, and it keeps producing while you run the company. That is the same logic behind the one-person startup operating system, pointed at distribution.

Why this matters now

Investors search you before the meeting. Candidates decide whether to apply based on what they find. Customers choose between you and a competitor partly on who they have seen more often. The numbers behind that are blunt:

▫️ 54% of buyers drop a vendor with a thin social presence before the first meeting, per the 2026 B2B social benchmarks.

▫️ 78% of teams that practice social selling outperform the teams that skip it, per LinkedIn.

▫️ AI-edited short-form video under 90 seconds earns 5.7x higher comment rates than long-form on the same topic, per the 2026 Vidyard and LinkedIn study, which is why the kit bundles CapCut.

Content is distribution. Distribution is leverage. The founders who understood that early are raising at better terms, hiring faster, and closing customers who arrived inbound. Being the answer an AI surfaces is the 2026 version of that edge, which is exactly what consistent, on-record content earns you. The business and investing library goes deeper on founder-led growth.

Set it up this weekend

The kit takes a weekend to set up. The compounding starts the week after.

Everything is pre-built and ready to duplicate directly into any existing Notion workspace: templates, agents, and workflows, plus the partner credits.

Duplicate it into your workspace

Free, about 5 minutes to begin, and built for founders who want the system and the results that follow.

FAQ

Q: What is included in the free AI Creator Kit? An AI-powered social media calendar in Notion, a creator CRM for partnerships and outreach, a memory-persistent AI content agent that holds your brand voice, automatic syncing of Slack, email, and meeting notes into Notion, influencer partnership guides, and credits including 3 months of Notion Business worth $6,000, plus CapCut, Lovable, and Secret.

Q: Who is the AI Creator Kit for? Founders and operators who want a consistent content presence that reaches investors, attracts candidates, and closes inbound customers, run with AI in place of a dedicated content team.

Q: How does the AI content agent work? It pulls context from your existing Notion workspace to plan and generate content that stays consistent with your brand voice, and it plans your content calendar using the organizational context already stored in your workspace.

Q: How long does setup take? About a weekend. The templates, agents, and workflows are pre-built and duplicate directly into any existing Notion workspace.

Q: Is the AI Creator Kit really free? Yes. It is free at ntn.so/vcc-creator, including templates, AI agents, workflow guides, and partner credits, free of charge.

Keep reading

▫️ AI Agents library

▫️ Prompting and Context Engineering library

▫️ Claude managed agents guide

▫️ The Granola and Claude second-brain stack

▫️ The one-person startup operating system

▫️ How to rank when AI answers the query

Within 48 hours, a Reddit post titled “Opus 4.7 is not an upgrade but a serious regression” had 2,300 upvotes. A post on X claiming zero improvement over 4.6 hit 14,000 likes.

VentureBeat ran a piece asking if Anthropic was nerfing Claude. The Register quoted AMD’s AI director calling Claude Code “dumber and lazier.”

Six weeks later, Opus 4.8 shipped with the same framing repeated almost word for word: every benchmark improved, no downside, upgrade now.

together with Attio:

The model underneath you keeps changing. The system that drives your revenue should hold steady.

Attio is the CRM that drives revenue around the clock, turning every signal, from emails and meetings to agent activity, into one live picture of every account.

▫️ Put agents on every account to research, qualify, and move every deal forward in your pipeline

▫️ Ask Attio anything about your business and get instant answers and actions from one chat thread

It is the CRM for the new way of going to market. Join 90,000+ teams already on Attio:

start Free today !

“Better benchmarks don’t mean better for your production system, and the industry has now lived through this pattern three times in a row.”

Here are the 6 things you need to understand before the next model drops:

Table of Contents

1. The Gap Nobody Benchmarks

2. Why It Breaks: Models Get Literal

3. The Second Failure Mode: Giving Up Early

4. The Compaction Problem Nobody Talks About

5. The Upgrade Tax

6. The Exodus: Where the Frustrated Developers Went

1. The Gap Nobody Benchmarks

Opus 4.7 was not a bad model. Cursor’s CEO Michael Truell confirmed it lifted resolution by 13% over Opus 4.6 on Cursor’s internal 93-task benchmark.

It solved tasks that neither 4.6 nor Sonnet 4.6 could touch. On the metrics that get published in launch posts, it was a clear step forward.

And thousands of developers spent the next 48 hours describing it as broken.

Benchmarks Measure the Wrong Thing

A benchmark measures performance on a fixed set of tasks designed to be representative. Your production system is not a representative task.

It’s a specific set of prompts, tuned over months, carrying assumptions about how the model fills gaps you never spelled out. A benchmark has no way to see any of that.

When Anthropic’s model card says every benchmark improved, that’s true. It’s just an answer to a question you didn’t ask.

The question that matters is whether the model got better at the thing you built, the way you built it. That gap is where production breaks.

The benchmark and your prompt are answering two different questions, and only one of them is about you.

2. Why It Breaks: Models Get Literal

The most consistent complaint across Reddit, Hacker News, and developer Discords wasn’t worse code. It was prompts that suddenly produced shorter, terser, sometimes flatly different results.

Some API calls started returning 400 errors with no obvious explanation.

The Model Stopped Filling In the Gaps

Opus 4.7 does what you write. Not what you meant. Not what 4.6 would have inferred from context.

If your prompt left a gap that 4.6 used to fill with a sensible default, 4.7 doesn’t fill it anymore. It does less, or does something narrower, matching the literal instruction exactly.

Think about a prompt you wrote eighteen months ago and never touched again. It worked, so you stopped looking at it.

But “worked” was a relationship between your prompt and a specific model’s habit of completing it. Every upgrade shifts that habit a little, and your prompt didn’t get the memo.

Your prompt didn’t get worse. The thing silently finishing it for you changed jobs without telling you.

3. The Second Failure Mode: Giving Up Early

There’s a separate problem that’s harder to benchmark and showed up constantly in developer reports.

Opus 4.7 gives up early.

The Persistence That Made 4.6 Valuable Degraded

Ask the model to take test coverage from 55% to 80%. It writes a few tests, declares victory at 58%, and asks if you want it to continue.

You say yes. It writes two more, declares victory at 60%, and asks again.

The persistence that made Opus 4.6 genuinely valuable for long agentic sessions degraded. This wasn’t an isolated anecdote.

It was one of three specific complaints developers kept naming on Reddit and Hacker News, alongside the literalism problem above and a tokenizer change that inflated token usage by 20 to 35% on identical inputs.

If your workflow depends on a model grinding through a checklist unsupervised, this regression doesn’t show up in a benchmark and absolutely shows up in your week.

4. The Compaction Problem Nobody Talks About

There’s a third failure mode that’s less about model behavior and more about how context gets managed during long sessions.

It sounds like a footnote until it eats your afternoon.

The Window on the Spec Sheet Isn’t the Window You Get

Anthropic advertised a 1M-token context window for Claude Code. In practice, a documented case showed quality dropping at around 20% of that window’s usage.

The auto-compaction routine fired at roughly 76,000 tokens into a 1M-token session, discarding history while most of the window sat empty.

Engineers running long-horizon agents at companies like Vercel and Replit flagged the same pattern independently. The number on the spec sheet told them almost nothing about what the model could still see.

A bigger context window on paper doesn’t mean a bigger working memory in practice.

5. The Upgrade Tax

Put the pieces together and you get something worth naming. Once you name it, you can design around it instead of getting surprised by it every six weeks.

Call it the Upgrade Tax. It has six parts.

Six Costs, One Cycle

Benchmarks improve while production breaks, because nobody benchmarks your specific prompts but you.

Implicit-to-explicit drift means models get more literal over time, so prompts that relied on the old model filling gaps quietly stop working.

The compaction and context problem means a bigger advertised window doesn’t guarantee more usable memory before quality drops.

The no-rollback problem means your old prompt versions, tuned for the previous model, may not even be accessible once you’ve moved on.

Anthropic’s own changelog format compounds this. Version notes describe model behavior in aggregate, never in terms of your specific workflow.

And the compounding cost is the QA layer itself. Version control for prompts, side-by-side testing, regression detection that catches drift early.

Real protection against the Upgrade Tax is an entire QA discipline, and most teams building on these models don’t have one yet.

6. The Exodus: Where the Frustrated Developers Went

The rough Opus 4.7 launch landed seven days before OpenAI shipped GPT-5.5. The timing turned a bad week into a referendum.

Developers who had built workflows around Claude Code started testing Codex side by side, and some posted before-and-after comparisons with receipts.

A Six-Week Round Trip

Cursor and Windsurf both added faster model-switching as a direct response, letting users route around a single provider’s bad week. RouteLLM-style setups, once a niche optimization, became a hedge.

Six weeks later, Opus 4.8 landed with the benchmarks back up and some of that traffic returned. But the switching infrastructure stayed.

That’s the real long-term effect of this six-week cycle. Not that anyone abandoned Claude. That nobody fully trusts a single provider’s release notes anymore.

Every upgrade is a migration, not a toggle, and the teams that learned that in April aren’t going back.

So What Now

Pull your highest-traffic prompts this week. Write down what you expect them to produce, in plain language, before the next model drops.

Run those prompts against both versions side by side. Not after rollout. As the gate before it.

The teams that got burned in April weren’t slow to upgrade. They never wrote down what “working” meant in the first place.

That document is the only thing standing between you and the next Reddit thread.

That is the opening move in a 90-minute argument about productivity, AI, space, and what “fair share” means.

I watched the entire Squawk Box interview from Blue Origin’s Rocket Park so you can skip it.

Here are the 8 ideas worth stealing.

Together with Outskill:

Bezos said it best: AI elevates the people who learn to wield it

Here’s the catch. Most people use barely 5% of Claude, while Skills, Connectors and Cowork keep shipping every week.

Outskill packed 800+ hours into a free 2-day Claude Mastery Workshop: Claude plus 10+ AI tools and workflows, live this weekend:

save your Free spot !

The bulldozer is in your hands. Time to use it.

1. The Tale of Two Economies

“You have a bunch of people doing really well, and a bunch struggling to pay rent and groceries. Politicians use this age-old technique of picking a villain and pointing fingers. That doesn’t solve anything.”

One group thrives. One group drowns. The political class picks a villain, charges at it, and collects applause.

Bezos prefers the five whys: trace a broken thing to its root cause and fix it there, so the fix holds. Root-cause work is expensive. Finger-pointing is free.

Steal this: when a “fix” feels good for ten seconds and changes nothing, that is the tell. The same discipline separates the founders top VCs actually back from the ones who pitch villains.

2. The Nurse in Queens Pays $12,000 (Bezos Wants Zero)

“A nurse in Queens making $75,000 pays more than $12,000 in taxes. How about we start by having her pay nothing?”

He wants the federal income tax burden on the bottom 50% set to zero. That cohort generates 3% of federal revenue: small for the government, large for the household. $1,000 a month is rent in many cities, or a year of groceries.

Steal this: if you price for lower-income customers, run the same math. Your price ceiling sits below what your finance team modeled.

3. America Already Runs the World's Most Progressive Tax System

“We don’t have a revenue problem. The top 1% pay 40% of all tax revenue. The bottom half pay 3%. We have a spending problem.”

He flips the debate from revenue to spending. His receipt: New York City schools at $44,000 per student, 30% above Chicago, LA and Boston, triple Miami and Houston, with worse outcomes than all of them. Same country, same curriculum, triple the cost, weaker results. A company run that way dies; a school system survives because the market can’t reach it, and the waste stacks in the middle layers between the dollar and the classroom.

Steal this: before any budget cycle, ask whether existing money is spent well. Founders who treat unit economics as a daily dashboard find the answer fast, and so do the investors reading their board decks.

4. AI Hands the Engineer a Bulldozer

“There will be no more radiologists, no more software engineers because AI does it better. These people are wrong. It is going to elevate all of them.”

You dug a basement with a shovel. Someone handed you a bulldozer. The job moves up an altitude and the work survives. This cuts straight against the AI-eats-white-collar-jobs thesis running the discourse.

Steal this: audit what your engineers do, then split it into identifying problems versus typing code. The best AI coding tools already cover the second half.

5. The Productivity Surge Brings a Labor Shortage

“One earner in many two-income households will drop out by choice. Productivity gains mean people can afford it. I predict deflation in certain core categories.”

The contrarian call: families opt down to one income, so labor gets scarce rather than redundant. Food and housing construction get cheaper, assuming the technology stays loosely regulated. It reads a lot like the calls Mark Cuban keeps getting right.

Steal this: build a watchlist of three categories where AI deflation lands first (construction, food, software). Track unit costs quarterly. The signal reaches sharp investors before it reaches the headlines.

6. The Building Permit Is an AI Agent Problem

“Why does a permit take six months to five years? I should get a yes or no in 10 seconds, then the authority to build.”

Permitting is rules-checking: zoning, setbacks, height, use. A well-trained model returns a decision in seconds and lists the reasons on a no. Process time drops from years to seconds, with no policy fight required.

Steal this: municipal middleware sits wildly underbuilt. If you want a wedge, ship the permit-decision agent cities adopt. Start with how to build an AI agent and 100 agent ideas with implementation notes.

7. The Engineer's Job Was Never Typing Code

“What does a good software engineer really do? We identify problems and help solve them. The code is execution.”

Most engineers measure value in code shipped. Bezos says that metric was always off. The problem-identification layer survives automation, while syntax expertise loses ground each quarter. Engineers who move into product thinking and systems design pull ahead.

Steal this: list the last five problems you solved in code. For each, did you find the problem, or did someone hand it to you? That ratio tells you who survives the transition.

8. For-Profit Value Beats Charity

“If I do my job right, the value to society from my for-profit companies will be far larger than my charitable giving.”

His giving rule is independence over dependence: he funds family shelters with sub-180-day stays that teach interviewing, skills, and financial management. The deeper claim is one tech intuits and rarely says aloud. A company solving a problem for hundreds of millions creates more aggregate welfare than a large foundation, because exchange stays voluntary. Customers pay when value beats price, the same logic behind every method used to value a company.

Steal this: if your company delivers value, treat it as your primary moral act. The giving is the company.

What to Steal Today

The doctrine distills to six moves: stop pointing fingers, fix spending, untax the bottom half, treat AI as productivity, build companies that create value, and trust root-cause analysis.

Founders: the bulldozer is in your hands. The work was never typing, it was finding problems worth solving. Map your week against that ratio. If you are raising on the back of it, study the decks that built unicorns and what top VCs check in diligence.

Investors: the productivity gains point to labor shortages and core deflation. Build a watchlist around cheaper construction, food and code, and keep the investor lists close for the rounds that follow.

Operators: the five whys is free. Use it on every broken process this month. Fund causes, not costumes.

For the AI thread running under all of this, The AI Corner tracks AI agents, tools and models, and where the jobs debate goes next.

If this saved you 70 minutes, send it to one founder or investor who needs it.

Full podcast:

The gap is rarely talent. It is knowing which steps matter and which to skip.

This is the path production agents actually run, stripped to the parts that ship. Copy the code, follow the seven steps, and your agent goes live today.

A year from now, building an agent will feel as routine as writing a script. The people who learn it this week run the teams that lean on it later.

Here is the full build:

▫️ The 7-step path, from picking the problem to a running agent, each step with the copy-paste asset

▫️ The minimal Agent SDK starter, a working agent in about 15 lines, plus the custom-tool pattern

▫️ The reusable system-prompt template that makes an agent behave the same way twice

▫️ 5 first agents worth building this week, each with the tools it needs and the hours it gives you back

▫️ The build-or-script decision that saves two weeks of wasted effort

▫️ Memory done right, the three levels, and the one most beginners over-build

▫️ A full worked agent, an inbox-triage agent wired end to end and ready to clone

▫️ The guardrails, hooks and permissions that keep an autonomous agent safe

▫️ The cost controls, model routing and caching that keep a 24/7 agent cheap

▫️ The debugging checklist for the moment it breaks on a live task

▫️ The framework picker, Claude Agent SDK versus LangGraph, CrewAI, and Managed Agents

🤖 The First Agent Build Kit

The full system in one place: the 7-step path, the copy-paste Agent SDK starter, the system-prompt template, 5 ready-to-build agents, memory done right, a full worked agent, the guardrails, the cost controls, the debugging checklist, and the framework picker.

One subscription unlocks every build kit

This is one kit in a growing library. Premium opens all of them:

▫️ Loop engineering for coding agents

▫️ Build a five-agent sales team in a weekend

▫️ The Claude managed agents guide

▫️ The Claude best practices power-user guide

▫️ The Codex background workflows playbook

▫️ The one-person startup operating system

Plus a fresh build kit every week. One agent shipped from these guides pays for the year in the first afternoon.

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

His playbook is teachable. That is the whole opportunity.

You can encode an operator’s rules into agents that run the motion for you: source, enrich, sequence, forecast, expand. Your people keep the part that closes deals, the judgment and the room.

Most teams bolt AI onto a broken funnel. This wires operator-grade GTM into a system that runs itself.

Here is the full build. Behind the paywall:

▫️ The 8-agent GTM stack, each one mapped to a proven operator rule, with the copy-paste prompt

▫️ The operator rules as guardrails, so your agents inherit the discipline that scaled revenue, rather than generic best practices

▫️ The 3 workflows that chain them, the morning loop, the outbound loop, and the forecast loop

▫️ The brutal-negativity forecast agent that keeps your pipeline honest and your board trust intact

▫️ The land-and-expand agent that turns one small contract into a whole-account motion

▫️ The tool stack, Clay, HubSpot, Claude, and the connectors that let agents act inside your systems

▫️ The keep-it-human map, the GTM work to protect from automation

▫️ The 30-day rollout, from your first agent to a self-running engine

Pair it with the deeper AI Corner library:

▫️ The AI Agents library for orchestration and reliability

▫️ The Prompting and Context Engineering library for the agent prompts

▫️ The AI Tools and Models library for the stack

▫️ The Business and Investing library for the revenue math

Related builds worth reading next: the five-agent sales team, loop engineering, the how to build an AI agent guide, and the VC founder playbook.

🛠️ The GTM Agent Stack

The full system in one place: the 8 agents with copy-paste prompts, the operator guardrails, the 3 workflows, the tool stack, the keep-it-human map, and the 30-day rollout.

Get The GTM Agent Stack below 👇

Try premium free for 7 days. Or get 50% off this week only.

Get 50% off forever

Read more

Has your friend already built an AI agent? He probably did, and described it excitedly over coffee:

“My AI agent watches my inbox, tags each email, drafts a reply, sends it!”

Four steps, same order, every single time, nothing about it deciding anything on its own.

The problem is… that is not really an AI agent. It is a workflow wearing a more expensive name.

The word has done real damage this year. Companies are buying autonomy they will never use, skipping the work that would have made it pay, and filing the result under “AI doesn’t work for us.”

In reality, it works fine. They just never figured out which of two very different things they were building.

That difference decides whether the money compounds or evaporates. It is also the easiest thing in AI to get right, the moment someone hands you the question that actually sorts it.

Something worth knowing before we get into it:

The people actually shipping AI systems in 2026 are overwhelmingly from non-technical backgrounds.

Lovable just published a data study on this, and the numbers are worth sitting with for a second:

The barrier between idea and running product has essentially collapsed.

This is the story nobody is telling loudly enough. Software creation used to be a narrow technical discipline. It is becoming something much broader. Operators, founders, marketers, and domain experts are shipping real products by themselves, without a dev team, without a technical co-founder, without months of runway burned on engineers.

The barrier between “I have an idea” and “I have a running product” has essentially collapsed. If you have been sitting on something because you assumed you needed an engineer first, that assumption is no longer accurate.

Start building on Lovable

Exclusive 10% off for readers.

Everyone wants the shiny AI agent. Most teams just need the boring workflow. And now, anyone can build either one themselves.

Table of Contents

1. The Only Difference That Counts

2. Why Most of What You Want Is a Workflow

3. You Can’t Buy Your Way There

4. The Filter Nobody Bothers to Apply

5. Where to Actually Point It

6. The Part You Won’t Want to Hear

1. The Only Difference That Counts

One question splits every AI system into a workflow or an agent. And it all boils down to:

Who decides the steps?

A workflow runs your script

A workflow walks a path someone already drew.

The sequence is locked before a single email arrives, and the model only fills the slots that need a brain.

Back to that inbox. Sort by topic, pull the order, draft from a template, send.

The model does the sorting and the writing, but the road never bends. It cannot read a strange message and decide today calls for a detour.

A recipe is the closest thing to it. The cook is skilled and the knife work is real, but the steps and their order belong to whoever wrote the card.

Underneath, almost every workflow turns out to be one of three plain forms. Chain the steps so each output feeds the next.

Classify the request up front, then send it down its own lane. Or fire several checks at once and gather what comes back.

Dress it up however you like, that is most of business automation.

An agent writes its own

An agent starts from a goal and a toolbox, then builds the road as it walks.

First move, second move, whether to circle back, when to quit. None of that lives in a script.

Same inbox, different animal. The agent reads a complaint, decides on its own to open the order system, spots that the package shipped cracked, checks what the refund policy permits, sends the money back, and writes the customer a note.

Nobody choreographed any of it. The thing reasoned its way through in real time.

That improvisation is the entire appeal and the entire hazard. A workflow keeps your plan. An agent throws the plan away and bets on itself.

Every real argument about cost, risk and control grows straight out of that one fact.

2. Why Most of What You Want Is a Workflow

The instinct runs backward. Founders reach for the harder, smarter-sounding option, certain their problem deserves it, when the plain one fits almost every time.

At the end of the day, the agent is the exception and not the opening move.

Agents feel like the real thing

An agent looks like arrival. Autonomous, reasoning, deciding without anyone holding its hand.

Building one feels like graduating from plumbing to intelligence, and that feeling has sold an enormous amount of software nobody needed.

Look at the actual work, though. Tagging tickets. Summarizing a contract. Drafting a first pass. Lifting figures off an invoice. Tidying a CRM record.

Every one of those has a known form, and a workflow handles known forms cheaper, faster, and without a surprise at 2 a.m.

Autonomy is never free. An agent can loop on one step until the credits run out, talk itself into something absurd, or stall in a corner you never thought to test, and because the reasoning happens in the dark, you may never learn why.

The bill for that freedom arrives as cost, as debugging time, and as the runs that simply go wrong.

The honest test

A ten-second test ends most of these debates. This is it:

Could you draw the flowchart yourself?

Yes means build the workflow. The path is knowable, so write it down. What comes back is steady, the kind of thing a junior engineer can fix on a Tuesday without paging anyone senior.

A real no, where the right steps genuinely hinge on what the system uncovers and reshuffle on every run, is the one place an agent earns its keep.

Even there, letting it roam loose is the rookie error. So put structure around the autonomy.

The agent owns the open-ended part, but it runs inside a process with checkpoints, a log of everything it touched, and a human gate at the moments that carry money. The flexibility survives and the blindness dies.

That hybrid has a name, the “agentic workflow”, and for the handful of problems that truly need a mind of their own, it beats a naked agent nearly every time.

3. You Can’t Buy Your Way There

Choosing the right architecture clears half the problem. The other half kills more projects than any model ever has, and it is the quiet belief that value ships in a box you can buy and bolt onto this morning’s process.

The lesson from the last big rewiring

Think of it like this:

Electricity reached the factory floor decades before it paid a dime. Owners ripped out the steam engine, dropped an electric motor in the same spot, and changed nothing else. Output barely moved.

The payoff waited on a smaller idea. Motors could be cheap and tiny, so every machine could carry its own, so the floor no longer had to crowd around a single shaft of power. Arrange the machines in the order the work actually flows, and you have invented the assembly line.

Nobody got rich buying the motor. The money lived in redrawing the whole building around what a cheap motor suddenly made possible.

Redesign beats a shopping spree

AI repeats the pattern beat for beat. A license here, a seat there, the process underneath frozen exactly where it stood.

The tools see a little use, the results land flat, and a meeting goes on the calendar to ask where the magic went.

A system that does not understand how the work truly happens gives back nothing worth keeping. The people who own that work and were never consulted will route around it on instinct, and adoption stays thin no matter how clean the demo looked.

The honest move is to take one process and cut its work into three piles.

1. Rule-bound steps get scripted.

2. Judgment steps go to AI inside a workflow.

3. The heavy calls stay with a human who has to answer for them.

Most of the value hides in the first two piles, not in some oracle that does everything at once.

Going live is not a switch, either. Run it in a sandbox, then in shadow mode beside the people still doing the job by hand, then in supervised production once it has earned a little trust.

Log every action and every correction so the thing keeps sharpening instead of freezing at the accuracy it had on launch day.

4. The Filter Nobody Bothers to Apply

Workflow or agent is the second question. The first one, the one almost everyone walks straight past, is whether the task deserves automating at all.

Skip it and a full month can vanish into something that never held a payoff.

Four things worth checking first

Volume leads. The task should run hundreds or thousands of times a month, or touch enough money that improving it lands on a real line of the books.

Pattern follows. Not carbon-copy identical, but regular enough that rules and old examples still bite.

Scatter belongs on the list too. The more a person bounces between Gmail, Slack, a spreadsheet and the CRM to finish one task, the more an automation has to pull together.

Last, the pain has to sit still long enough to measure. Hours burned, errors logged, deals stalled, a number you can name today and check again later.

No measurement, no way to know you won.

Run a candidate against all four bars and watch most of them trip on at least one, which is the entire point of owning a filter.

Write it down before you wire it up

One step decides the result, and it touches no software at all. Write the process out by hand. Every move, every odd exception, in the words a person would actually use to explain it.

Output quality sits downstream of input clarity, almost completely. Feed the model fog and it hands fog right back. The slop everyone pins on AI is usually slop somebody fed in first.

An afternoon of honest documentation, then handed to the model to draft a plan and pick fights with your logic, is the highest-return hour in the whole build. It happens once per process and never again.

5. Where to Actually Point It

Knowing the difference is dead weight without a place to aim it. A short list of functions keeps surfacing as the right first ground, because the work there runs nonstop, sprawls across systems, and bores everyone unlucky enough to do it by hand.

Start in the back office

Accounts payable is the cleanest opening move in most companies. Invoice intake and purchase-order matching are rules in a costume, so script them.

Coding entries to the right account asks for a little judgment, so let AI take the first swing and let a person check the odd ones. Cheap to baseline, miserable by hand, easy to love.

Procurement lives on context buried inside contracts, portals and inboxes. Onboarding a vendor or testing compliance means dragging facts out of six open tabs.

Automate the dragging and leave the sourcing and the negotiating to people who can read a room.

Operations is a machine built for exceptions, and routing them is repetitive, patterned, automatable work. Returns and allocation blend hard rules with judgment, so a workflow with a human watching the expensive calls beats setting a loose agent free in the warehouse.

The pain, the delays and the misroutes, sits in plain sight and counts itself.

Then the revenue side

Sales buries fortunes in clerical work. One large deal can cross six teams and eleven handoffs before anyone signs. Routing it, enriching the CRM, running commission math, all of it devours hours a rep should spend across the table from a buyer.

Lift it off their plate and you hand back the only resource they cannot make more of.

Marketing is the most tempting ground to automate and the easiest to wreck. Lead scoring and routing is high-volume and pattern-rich, a strong early bet, as long as a human still studies the borderline names instead of trusting the score blind.

Botch this one and the cost is not wasted minutes, but most likely dead deals.

Content is the piece everyone sprints at and limps away from. Treat it as a button that prints posts and you get exactly what that sounds like. The win is feeding it something that already worked and reworking that across formats, so the model recasts proven ideas instead of conjuring hollow ones from a cold prompt.

Campaign reporting earns its automation the second the volume starts to ache. Hauling numbers out of ad platforms, email tools and analytics into one straight answer is repetitive and scattered, the precise sweet spot.

One channel trips the volume bar. Ten channels pays you back every Monday you stop losing to the rollup.

Customer research rewards you in proportion to how far people have to dig today. A rep assembling the company site, fresh news, old tickets and CRM history before every call is doing exactly the hunt an automation erases. A clean single source of customer truth shrinks the prize, so slide it down the list.

Pipeline hygiene is the dull one worth doing first anyway. Records kept current, stale deals flagged, blank fields chased down. It never stops, reps loathe it, the cost is trivial to measure, and a wrong move costs nothing to undo.

One line ties the function together. Give the busywork to the machine so people pour their hours into judgment and relationships, where revenue is genuinely won and lost. Automate the friction around the human work, never the human work itself.

Outreach mass-produced with nobody in the loop reads as exactly that, and it spends the trust you were trying to earn.

6. The Part You Won’t Want to Hear

Failed AI projects rarely die of bad models. The models are fine.

AI projects die because someone bought autonomy the business would never touch and skipped the unglamorous work that turns any of this into money.

That unglamorous work is the entire game.

Map a process by hand, script the dull stretches, point AI at the judgment, keep the weighty calls human, then measure what actually changed. None of it looks good on a slide. All of it shows up on the P&L.

The companies pulling ahead are not the ones with the flashiest agents. They are the ones who ran that boring loop once, watched it pay, then ran it again on the next process, and the next, until the whole place had rebuilt itself around work that compounds.

The agent can wait. The honest first step is admitting the thing you needed was a workflow, and that chasing the cooler toy is what slowed you down.

That refusal to flinch is the whole story.

I watched the full Circuit documentary so you can skip it. Here are the 10 takeaways that matter.

One thread runs through them: once AI takes the execution, the edge moves to judgment and go-to-market craft. Fitting, given who powers this issue.

Together with Attio:

Dario’s point cuts to one question: what do you do that AI struggles to replicate?

GTM Atlas Volume II is the operators answering it, with six new entries live:

▫️ What only a human can deliver by Travis Bryant, Anthropic’s Head of US Mid-Market GTM

▫️ Hire for slope, not for pedigree by Cristina Cordova, COO of Linear (ex-Stripe, Notion)

▫️ Your product is the pitch by , Head of Growth at Lovable

Plus a perk stack from Wispr Flow, Granola, Notion, and Anthropic.

Steal the systems the best operators run right now:

Explore the Free GTM Atlas

Now, to Dario, starting with the half of the automation curve almost everyone skips.

1. The automation curve has a brutal second half

“You automate 90% of the job, great. People are 10 times more productive in the other 10%. But eventually it gets close to 100%. Now, the sequel to that is, well, then you have to find something else for them to do.”

Most analysis stops at 90 percent and calls it a productivity revolution. Dario keeps going. For some workers, AI already replaces rather than augments, and augmentation and replacement turn out to be the same curve, the same people, separated by time.

At Anthropic, Claude writes almost all the code today. Engineers feel superhuman, which is exactly why the endpoint hides. The window to plan for the second half is open now, while it still reads as good news.

Track before the turn:

▫️ Which roles you augment today

▫️ Which roles raise output while headcount holds flat

▫️ Whether your workforce plan accounts for the second half at all

Why it matters: the companies that survive the second half built for it during the first. Plan for the jobs turn ahead of the curve announcing it.

2. The super weapon Anthropic held back

“Some of the early companies that we gave this to said things like, this is a super weapon. You should have to own a gun license to use it. Please don’t release this.”

A project called Mythos sits under every public Anthropic story. It found thousands of cybersecurity vulnerabilities and mapped exploitable flaws across every major operating system. Anthropic’s own read: released openly, it could hack banks, reach state secrets, and hit critical infrastructure.

Dario calls it “a particularly large jump,” one that arrived ahead of any internal schedule. The response was Project Glasswing: controlled access for a few trusted organizations, federal agencies included. It is the first live test of whether a private company can hold that much power responsibly.

Why it matters: enterprise and government buyers now bet on which vendors hold the governance to handle a Mythos-level event. The labs with pre-set release criteria set the precedent the rest follow.

3. One engineer, zero lines of code, 17x growth

“Now I talk to my Claude and it writes the code. And then while it does that, I talk to the next Claude and it writes some code. And at any point I have either a few Claudes running and up to a few thousand Claudes running, doing things.”

Boris Cherny, the engineer behind Claude Code, spent six months coding through Claude rather than typing syntax. This is parallelized AI labor: dozens to thousands of instances running at once on complex tasks. Claude Code’s internal volume at Anthropic is up 17x year over year, and one Code with Claude attendee put it plainly: “I have the confidence of a 22-year-old with VC money.”

The job changed. The winners orchestrate the most parallel agents against the hardest problems:

▫️ Understanding what is worth building

▫️ Judgment on direction and priority

▫️ Directing agents toward the right goal

▫️ Knowing when to override the output

Why it matters: treat this as a tool and you go 10x faster. Treat it as a replacement for thinking and you become replaceable. Judgment survives, execution commoditizes.

4. A 25% chance of collapse, and the airline test

“If there was a 25% chance of an airplane crashing, you wouldn’t get on that plane. That’s right. 25% is too high. We’re trying to make that probability much, much lower. That is the goal.”

He puts collapse odds at 10 to 25 percent and holds the number under pressure. His logic: the technology exists, many actors will build it regardless, and collective action is the hardest problem class in history, so one player opting out shifts little. The better move is to sit at the frontier with the most safety-focused team.

The person building the thing is telling you the risk is genuine. That is the data point. Reducing that number is half of Anthropic’s operating budget, and the same person who calls 25 percent too high builds harder, faster.

Why it matters: a 10 to 25 percent risk estimate belongs in scenario analysis and board conversations as a planning input, from the CEO of the most valued AI company by revenue.

5. Why he left OpenAI, and what actually drove it

“When you feel that you can’t trust someone, when you feel that their values are not what they say they are, when you feel that they’re not honest, that makes it very hard to continue to work with the company.”

Coverage compresses Dario’s exit into safety disagreements. The deeper driver was trust. Safety frameworks invite honest disagreement among smart people. A gap between stated values and actual behavior is a different category, and once it opens, the argument ends.

The number that says the rest: every Anthropic co-founder still works there, which at this scale and under this much Pentagon and competitor pressure is rare. That cohesion is a competitive asset.

Why it matters: trust is the load-bearing wall of every partn

6. The Pentagon blacklisted them, and the lines held

“What does winning this fight actually look like? I won’t even call it a fight. This is more a debate about what the proper use of AI by the government is.”

Anthropic wrote its red lines before the meeting: bans on mass surveillance, human oversight required on any targeting. The Pentagon pushed to strip them, Anthropic held, and the blacklist followed. Lawsuits came, Trump called Dario “an ideological lunatic,” and the ban later lifted with the lines intact.

The lesson sits in the sequence: specificity before the meeting beats philosophy after it.

Why it matters: the companies that built Washington credibility during calm periods gain leverage when the next capability reveal forces a policy response. Define your red lines before the room.

7. A school was hit, and Claude was in the chain

“What we’ve seen here is Claude assists, but a human makes the final call. So a human made that final call, not Claude. Imagine if you had a world in which the AI model just makes the decision and the human never sees it. That’s what we were standing up for.”

A US missile hit a girls’ school in Iran, more than 150 people died, and Claude may have been in the targeting chain. Dario calls the strike terrible and separates two questions most coverage fuses: was the outcome acceptable, and was the decision architecture acceptable.

The counterfactual is the hard part. The alternative to Claude assisting under human oversight is AI from less careful actors making lethal calls with a human fully outside the loop.

Why it matters: “human in the loop” is the load-bearing phrase in high-stakes AI. Every org using AI in consequential calls needs a written answer to one question: who makes the final call, and how do you verify it happened?

8. The bet everyone called crazy

“At that point in time, I know it sounds crazy now looking back, not a lot of people believed, hey, scale-up is the way that these models are going to get smarter and better. That was sort of an unusual, countercultural, scientific perspective.”

Before ChatGPT made scaling obvious, Dario argued it would work: add compute and data to large models and they improve, predictably, on the same algorithm. In 2016 that read as fringe, and the founding team held it anyway. They were right.

The rarer thing is the conviction held under social pressure for years until it paid out. To spot the 2025 version: find the belief researchers hold privately, consistent for two-plus years under pressure, where the holder stays put when the room pushes back.

Why it matters: Anthropic’s whole position traces to one technical conviction held under resistance. The next decade-long return hides inside a belief today’s consensus still calls wrong.

9. $285 billion gone in a day, and why the pie still grows

“The existing incumbents may be smaller in relative terms. Some of them may go down in value. Some of them may even go out of business if they don’t adapt. But I would guess that the software industry gets larger, not smaller, although there will be some big losers.”

Claude Cowork launched, and $285 billion in software market cap vanished in a day. Traders called it the SaaSpocalypse. Dario’s actual view is more useful: total software grows while distribution shifts hard. Those are two statements, and fusing them is how sharp operators misread the moment.

Which side you land on:

▫️ You can name what your product does that AI struggles to replicate

▫️ That answer is specific, defensible, and hard to build around

▫️ Your users would lose something irreplaceable if you vanished

▫️ Your moat lives in the workflow, the data, the relationship, or the judgment layer

Try this: complete the sentence “AI can do everything we do except ___.” A specific, defensible ending is your moat. A blank is the work to finish before the next launch ships.

10. Silicon Valley yo-yoed on regulation. Dario held the third path

“They started with a position of, you know, even having transparency around this technology is all just totally apocalyptically destroy our potential. And then as soon as they see the first real danger, there’s all this talk of nationalization and the government should just seize it. Come on, folks.”

Silicon Valley swung from “any regulation kills us” to “nationalize AI” inside one capability reveal. Dario stayed put. His position is specific: the technology rides a smooth exponential, it keeps getting more capable, and governance belongs in place ahead of the crisis.

His mechanism:

▫️ Pre-release testing and auditing before any major model ships

▫️ Accountability that skips government ownership

▫️ Transparency enough that outsiders verify what labs claim

▫️ A line between “the government should know” and “the government should own”

Why it matters: the policy fight stays open, and the next reveal forces another response. Show up with consistent, specific positions. The ones who held them help write the rules.

What this means for you

Dario runs the most valued AI company by revenue, believes civilization faces a meaningful chance of collapse, and builds harder anyway, because handing the lead to someone less careful is the worse outcome.

Three principles to carry forward:

1. The automation hump reads as productivity until it turns. Build for both halves while it still feels like good news.

2. Trust is the actual reason partnerships break. Values that drift from behavior end things faster than any strategy fight.

3. The smooth exponential feels like stillness until it goes vertical. Most people are standing in that flat stretch right now.

Founders: build something that compounds on the right side of the curve. Physical world, human judgment, or directing AI at scale: pick one lane deep enough that the next Claude launch makes you harder to replace. Write your moat answer down before the next reveal.

Investors: Anthropic went underdog to trillion-dollar in four years on two contrarian bets. The next decade-long return hides in a belief the current consensus still calls fringe. Back the person who holds it and stays put.

Operators: the SaaSpocalypse repeats. Map where your users lose irreplaceable workflow context if you vanished. That is your moat.

Quiet, quiet, quiet.

And then zoom.

Plan for the zoom.

Keep reading

Anthropic, Claude, and the power fight

▫️ Claude and Anthropic library

▫️ Inside the Fable 5 and Mythos government ban

▫️ Anthropic passed OpenAI at $30B ARR

▫️ Dario Amodei on safe AI and AGI

Agents, Claude Code, and the new job

▫️ The Claude Code system that replaces your dev loop

▫️ Loop engineering for coding agents

▫️ Claude managed agents guide

▫️ The Anthropic AI jobs report

Moats, markets, and the SaaS shift

▫️ SaaS defense playbook for the AI era

▫️ Marc Andreessen on the AI moat

▫️ The biggest VC-backed startups

▫️ The investor list of lists

Full documentary

If this breakdown saved you 47 minutes, share it with one founder or investor who needs to see it. They will thank you later.

One of the most powerful companies in history almost slept through the biggest platform shift in a decade.

Google held every ingredient: the talent, the data, the compute, and a decade of AI research sitting in the building. Then ChatGPT shipped, and for a few months the giant looked flat-footed while a startup set the pace.

What Sundar Pichai did next is the part worth studying. He rebuilt the entire company in 18 months, in a deliberate order most analysts read right past.

He walked through it with Nilay Patel on Decoder after Google I/O 2025. I went through the full 45 minutes so you can keep yours.

Here are the 10 moves that matter, and the thread running through every one:

Raw capability counts for little until you organize to act on it.

Founders feel that sharpest at go-to-market, where a beautiful plan and an empty pipeline weigh exactly the same.

together with HubSpot for Startups:

Pichai’s lesson is execution over planning. Here is the go-to-market version of it.

On June 18, HubSpot for Startups and Clay run a live build-along for seed and pre-seed founders. You build it in the room with them:

▫️ Define your ICP from genuine buying signals

▫️ Build and enrich a prospect list in Clay

▫️ Launch AI-personalized outbound in HubSpot

You leave with pipeline, rather than a plan.

Apply to HubSpot for Startups and you become eligible for up to 90% off HubSpot, plus a shot at $1,000 in Clay credits.

📅 June 18, 11 AM ET (4 PM CET). 45 to 60 minutes, live:

Save your FREE Seat

With that on your calendar, back to Pichai, starting with the diagnosis almost everyone missed:

1. Pichai read the ChatGPT shock as permission

“I always internalized that moment. It was tough to convey it outside. But I felt like the Overton window had changed. People were adopting these technologies faster than we had expected. And so to me, it was a way to go and actually express ourselves through our products.”

Most teams read a rival’s breakthrough as a gap to close. Pichai read it as clearance to move.

Google held the ingredients. What it missed was the organizational readiness to deploy them, so the ChatGPT launch exposed a structural deficit rather than a capability one. He diagnosed that first, and the order is the whole difference between a strategic response and a reactive one.

The useful question after a competitor’s big launch skips past “are we behind.” It is “do we hold what we need to move, and are we organized to use it.”

Why it matters: a shock stays existential only while you lack the assets to answer it. Hold them, and the shock becomes permission. The teams still standing after the wave diagnosed the bottleneck before touching the org chart.

2. Merging Brain and DeepMind: the 5-step sequence that followed

“One AI team, we had world-class research teams in Brain and DeepMind. Bringing that together as Google DeepMind was harder than it sounds because it’s like saying, go put Stanford and MIT together and create a department out of it.”

From outside, Google DeepMind looked like a branding call. Pichai calls it the hardest part of the entire restructuring: two elite, proud, culturally distinct research houses fused into one model capability.

Merging two functions that both succeed and both believe they are right is harder than shutting one down, and most companies stop short of the will to do it. The moves that followed were sequenced on purpose:

▫️ Unified AI research under Google DeepMind

▫️ Centralized infrastructure under Amin Vahdat

▫️ Created a Chief AI Architect role, filled by Koray

▫️ Consolidated Search under single leadership

▫️ Weekly AI product reviews run by Pichai himself

Why it matters: each move set up the next. Research, then infrastructure, then accountability, then decision speed. Founders restructuring under pressure should map dependencies the same way, since sequence decides whether a restructuring compounds or collapses.

3. The 90-day filter that sets your decision speed

“A big part of my framework is over time understanding that there are very, very few decisions which are really consequential, and most decisions aren’t. So what matters much more is that you make the decision, because that’s what determines the velocity of an organization.”

This line runs counter to how most founders were trained. Treating every decision as heavy creates the same jam as treating every decision as light.

The distinction Pichai draws is simple: some choices look big yet reverse inside 90 days, and others lock in. Combining Brain and DeepMind was the second kind. Most product launches are the first, and experience is what lets you sort them fast.

Why it matters: organizational velocity comes from decision throughput on the reversible calls. Before your next decision stalls, ask one thing: can we fix this in 90 days if we are wrong? A yes means make the call and move. That single filter buys more speed than any process overhaul.

4. 13 products, one infrastructure layer, a moat that compounds

“For the first time, we have such a common infrastructure powering all of them with our Gemini models and the underlying AI infrastructure. So we are more able to, with intent, do things which cut across things.”

The fragmentation critique of Google is as old as the company, and the 2025 answer is structurally new. Thirteen products, each with a billion users, shared distribution for most of Google’s history yet ran on separate technology.

Gemini changes that. One model and one intelligence layer now power Search, Maps, Gmail, and Docs from a single base. Personal Intelligence is the proof: one infrastructure build, one user toggle, every product upgrades at once.

Why it matters: platforms with common infrastructure compound, because each improvement reaches every surface at once. Google’s moat here is architectural depth multiplied by distribution scale, and matching one side leaves the other gap wide open. Few rivals close both inside 36 months.

5. Agents finally deliver Google's 10-year assistant dream

“This long-running vision of Assistant we’ve all had and worked through myriad forms of it, and failing to fully do it well. I think we are closer than ever before to deliver on that promise. We haven’t delivered it yet.”

That admission is worth isolating. Google Assistant, Duplex, a decade of attempts, all fell short. Agents are the first technology able to execute the original vision, on four building blocks Pichai sees finally in place:

▫️ Reasoning across a problem

▫️ Tool use and code execution

▫️ Planning over multiple steps

▫️ Long-running task management

The same stack underpins what teams now ship with Claude Code and agent loops.

Why it matters: the agent paradigm is durable because it pays off a user expectation that already existed for a decade. Products that deliver an old promise grow faster than products that ask users to believe in a new one, so the adoption curve is shorter than it looks.

6. Pichai called a live Search result wrong on camera

“I think it’s probably more opinionated than it should be for the particular query you showed me. That’s my reaction as a user. I think that’s the scope for improvement.”

The most candid exchange in the interview. Patel ran a live search for “best Chromebook,” the AI overview gave one answer, Reddit gave another, the Times gave a third, and Pichai owned the problem on the spot.

The tension underneath is structural. The more AI overviews answer directly, the more useful Search feels, and the less it surfaces the organic content underneath. Google’s stated check is measurement: 25 years of satisfaction data, bounce rates, session length, return visits.

Why it matters: watch AI overview accuracy against click-through to organic results over the next 18 months. That gap is the cleanest signal of whether Search strengthens or erodes, and it shapes how you should rank when AI answers the query.

7. Conde Nast told its teams to plan for zero search traffic

Conde Nast CEO Roger Lynch: “Every year our search traffic was down more than we had forecast. So last year I told our teams, assume there is no search. You have to have your businesses planned as if search is zero.”

Google Zero moved from a publishing meme to a formal planning assumption at one of the largest media companies on earth. Pichai’s reply pointed to surfacing quality content, filtering low-quality clicks, and growing information overall. Each holds up, and each sidesteps the core worry.

The worry: AI overviews answer the question so well that the click vanishes. A great answer can produce zero traffic, because intent and outcome are separate things.

Why it matters: every business living on search referral should already hold its own version of the Conde Nast plan. Planning around the loss of a channel you fail to control is the rational baseline, so build the GEO playbook and your AEO backlink sources now, ahead of the curve forcing it.

8. Pichai calls AI anxiety rational, and that reshapes consumer AI

“No, I don’t think so. That’s the point I’m making. I’m, in fact, arguing against it. People are standing and telling about how AI could make a lot of jobs go away. Why wouldn’t you feel a sense of anxiety about it? I think those are deeper issues which we have to tackle as a society.”

Several executives frame public distrust of AI as a messaging miss that better demos fix. Pichai rejects that outright. In his telling the anxiety is grounded: job displacement fears hold weight, and energy concerns are legitimate.

The gap between polling (young people broadly dislike AI) and usage (nearly a billion users) reflects genuine tension rather than confusion. Closing public trust means solving the underlying problems, and reframing them leaves the gap open.

Why it matters: founders building consumer AI should plan for a trust deficit that marketing fails to fix. Products showing immediate, personal value close it, while products that ask users to bank on a future benefit stall. A sitting CEO calling this societal sets the terms for the next wave.

9. "Foothills of the singularity" has a precise definition

“I think for him, the advent of AGI is what he thinks of as the singularity. There is a harder definition of AGI, which is that it has to more comprehensively do the wide range of tasks, including cognitive tasks, in a way that’s comparable. So I think we’ll, at some point, actually put it out as a company.”

Demis Hassabis closed Google I/O with one of the year’s most-quoted lines, and Pichai grounds it in a specific bar: AI performing the full range of human cognitive tasks at a comparable level. All of them, rather than some.

Current systems sit below that bar, so “foothills” does precise work: closer than consensus expects, with road left. Google plans to publish a formal company definition of AGI, and that document rewards a close read when it lands.

Why it matters: when major labs publish operational AGI definitions, those definitions become product roadmaps, and the capabilities that close the gap attract the most capital. Track the definition and you track the priorities.

10. Skip the AGI label, build for the 36-month window

“I think that timeline doesn’t matter because the rate of progress means you’re dealing with ever more intelligent systems in a profound way. Three years from now, whether you and I call it AGI or not, doesn’t matter because it’ll be very, very powerful and we have to prepare for it.”

A precise answer dressed as an evasion. Pichai treats the label as a distraction from preparation that already needs to start. Three years out, the systems become powerful enough that societal readiness outweighs the name.

Why it matters: founders building today should model roadmaps against systems much more capable in 36 months. The strategic risk is a capability jump landing ahead of any declaration. Build for the present environment, and design the architecture that survives the next one. Both belong on the roadmap now.

What this means for you

The Decoder interview reads as a strategic document rather than a product update. The company that almost missed generative AI reorganized around one thesis: unified research, common infrastructure, faster decisions, and AI converging across every product surface. The structure is in place, and the execution runs on.

For founders. Structural readiness decides whether you can answer a competitive shock, because capability that lacks organization stays latent. Pichai’s order applies at any scale: unify research, centralize infrastructure, create single accountability, consolidate decisions. Get the capability layer unified before you build product on top, the same discipline the one-person operating system runs on.

For investors. Google’s moat is the pairing of distribution and infrastructure rivals struggle to copy quickly. The threat to it is a better distribution channel rather than a better model, so watch which AI companies build consumer reach off Google surfaces. The biggest startups and the VC money flows point to where that reach is forming.

For operators. Google Zero is a planning assumption now. Quality content still gets indexed, and traffic-dependent revenue needs diversification that holds even as referral volume falls. The Conde Nast plan is the template, and the GEO and AEO playbooks give you the build.

For anyone building in AI. The agent paradigm is durable because it pays off an expectation a decade in the making. Winners deliver the promise everyone already believed, since fulfilled expectations beat novel use cases in a market where the underlying model swaps out.

The three principles to carry forward:

1. Structural readiness decides your response to a shock. The bottleneck is almost always structure ahead of technology.

2. Most decisions are reversible. Speed on the routine ones buys the room to deliberate on the few that lock in.

3. The AGI label is a distraction, the trajectory is the signal. Model your business against systems much more powerful in 36 months and build accordingly.

The window to prepare is shorter than the debate about the timeline suggests. The companies that treat this as a platform shift rather than a product category set the terms. The ones waiting for clarity inherit the conditions the first group creates.

If this breakdown saved you 45 minutes, share it with one founder or investor who should see it.

Keep reading

Strategy, moats, and restructuring

▫️ Business and Investing library

▫️ Founder mental models for the AI era

▫️ Marc Andreessen on the AI moat

▫️ The one-person startup operating system

▫️ What top VCs look for in 2026

Agents and the build stack

▫️ AI Agents library

▫️ The Claude Code system that replaces your dev loop

▫️ Loop engineering for coding agents

▫️ How to build an AI agent in 2026

Search, GEO, and the traffic shift

▫️ How to rank when AI answers the query

▫️ RIP SEO: the GEO playbook

▫️ AEO and GEO backlink sources for founders

AGI, jobs, and the market

▫️ Claude and Anthropic library

▫️ Dario Amodei on safe AI and AGI

▫️ The Anthropic AI jobs report

▫️ The biggest VC-backed startups

▫️ Where VC money is going in AI

Full podcast:

If this breakdown saved you 45 minutes, share it with one founder or investor who needs to see it. They will thank you later.

The gap traces back to one thing most teams overlook. Every time a model answers, two separate operations run on the GPU, and each one fights a different battle. The first reads your entire prompt in a single burst, and its speed rides on raw compute. The second writes the answer one token at a time, and its speed rides on memory bandwidth.

That split sets your latency and your bill, and inference engineering is the craft of bending it in your favor. Three years ago the work stayed locked inside frontier labs. Today every team running serious AI workloads leans on it, because the payoff is concrete: a latency target you reliably hit, and an inference bill that falls by most of its size once your volume earns the work.

Here is the full system:

▫️ The prefill and decode split, explained so the entire field organizes itself in your head, with the two metrics that matter

▫️ All 6 optimization techniques, mapped to the exact phase each one speeds up, with the tradeoff each forces

▫️ The prompt-structure rule that turns prefix caching from zero savings into most of your prefill cost gone

▫️ The 2026 serving stack, vLLM versus SGLang, and which one fits your workload

▫️ The build-versus-buy crossover, the honest math on when self-hosting open models wins and when the API stays cheaper forever

▫️ The 3 signals that tell you the moment to leave off-the-shelf APIs, plus the compliance trigger that overrides the cost math

▫️ The quantization sensitivity map, which layers tolerate compression and which ones poison quality

▫️ The decision framework to pick the right techniques for your product, rather than all of them

Pair it with the deeper AI Corner library (included in the premium subscription):

▫️ The AI Tools and Models library for the model and serving stack

▫️ The AI Agents library for the workloads that stress inference hardest

▫️ The Prompting and Context Engineering library for the prompt structure that drives caching

▫️ The Claude and Anthropic library for caching mechanics and pricing

▫️ The Business and Investing library for where this margin compounds

Related builds worth reading next: the token cost playbook, the AI coding tools guide, the context engineering guide, and loop engineering.

⚙️ The Inference Engineering Playbook

The full system in one place: the prefill and decode split, all 6 techniques mapped to phase and tradeoff, the prompt-structure caching rule, the vLLM versus SGLang choice, the build-versus-buy crossover, and the decision framework.

Access The Inference Engineering Playbook below 👇

Try premium free for 7 days. Or get 50% off this week only.

Read more

Jensen Huang has called deep learning right for 15 years, and he went unfiltered for 1 hour. I watched all of it so you can skip to the moves.

Here are the 10 that matter, with the action under each👇

Together with Outskill:

Move 7 below is the sharpest line in the whole talk: the job now goes to the expert AI user, and the edge comes from built workflows rather than a title.

That edge is learnable in a weekend:

Outskill runs a live 2-day Claude AI Mastery Workshop that drives Claude past the chat box, into the Code and Cowork modes where the agent work actually happens. It condenses 800+ hours of research into 16 hours:

▫️Chat, Cowork, and Code end to end
▫️ Skills, Connectors, and Plugins, plus 10+ tools that keep Claude working while you sleep
▫️ Live Saturday and Sunday, 10AM to 7PM EST

Walk in able to answer the one question a resume hides:

Sign up for $0

Now, back to the roadmap, starting with the inflection point almost everyone missed:

1. The actual inflection point landed 6 months ago, and almost everyone missed it

Agentic AI crossed the line, which means every infrastructure thesis you hold deserves a rebuild rather than an update.

“AI in the last several months became useful. That’s the big idea.”

Agentic systems now understand, reason, plan, and use tools to finish actual work. Jensen names Claude Code as the first system to do this productively, and that benchmark just became the floor every other system gets measured against.

Here is the number that rewrites the infrastructure math: agentic AI demands roughly 1,000 times the compute of generative AI, since agents read, reason, call tools, and generate far more tokens moment to moment. Multiply that by 100 times more users arriving at once, and GPU demand compounds instead of scaling linearly. The chips NVIDIA sold four years ago carry more value today than the day they shipped.

Try this → audit every AI workflow in your operation. Generative workflows are the 2023 answer to a 2026 problem, so find them and rebuild them with the loop engineering and agentic patterns that fit the new floor.

2. The model layer gets the attention, and NVIDIA deploys capital two layers below it

Capital follows value, and right now the value sits two layers below the conversation.

“AI is not just an application. AI actually reinvented the computer industry.”

Jensen frames AI as a five-layer stack: energy, chips, infrastructure, models, applications. The media lives at the model layer, while NVIDIA’s capital lands at infrastructure, specifically CoreWeave, Nebius, and Nscale. One dollar NVIDIA committed unlocked nine more from the institutional market, and the application layer comes next.

Healthcare, transportation, financial services, and retail each sit at the starting line of genuine transformation. That transformation depends on compute infrastructure existing first, before any application runs on top of it.

Try this → find the layer in your space that few are fighting for yet. That is where the next fundable company lives, and the YC requests for startups list is a useful map of the open ground.

3. NVIDIA's capital rule: $1 in must activate $100 out

Jensen writes checks to remove the constraint blocking the next $100, rather than chasing a return on the check itself.

“We invest at $1, it activates AI maybe by $100. If we can make that kind of amplification for the entire ecosystem, it would be tremendous.”

Two years ago the bottleneck was chips. Today it is energy, land, and power. The CoreWeave investment worked as a precisely placed lever: anchor capital that unlocked institutional confidence and nine dollars for every one NVIDIA put in. Every investor who co-invested alongside NVIDIA in CoreWeave, Nebius, and Nscale is seeing strong returns now. Early bottleneck identification is the entire thesis.

Try this → map the constraint in your market. The company removing that constraint is worth more than every company it enables combined, so the founder mental model to internalize is simple: get to the constraint first. The investor list of lists shows who is already funding the layer below.

4. AI-native gross margins flipped positive in 90 days, and the market still owes a reprice

The unit economics stopped being a bet. They are a fact, and the capital chase just changed sha

“Both of these companies and most of the AI native companies have turned. Their gross margins have gone extremely positive.”

OpenAI. Anthropic. Cursor. Gross margins crossed into strongly positive territory in the last three to six months. Token economics work. Demand is there. That is why OpenAI and Anthropic race for capacity right now, not profitability reassurance.

$100 billion went into AI startups last year. The largest single-year startup investment in human history. Software engineering job openings are rising, not falling.

Try this: watch capacity announcements from OpenAI and Anthropic. Capacity commitments are the leading signal for where AI product investment accelerates next.

5. The radiology jobs prediction was wrong, and the reason applies to your industry too

Computer scientists predicted the task correctly and predicted the job incorrectly.

“100% of radiology is now infiltrated by AI. It is completely integrated. And yet, the radiologist job was not wiped out.”

Computer vision reads scans at superhuman accuracy now, and radiologists review more patients, take more cases, and generate more revenue. Radiology departments turned into profit centers, and hospitals started hiring more radiologists. AI compresses the task, purpose expands to absorb the freed capacity, and hiring follows the expanded purpose.

Every knowledge-work vertical AI enters rides the same curve: task compression, then purpose expansion, then more hiring at the level of the expanded purpose. Your industry rides it too, as the AI jobs data keeps showing.

Try this → list the three tasks in your role AI automates today, then list what you actually get paid to accomplish. The gap between those lists is where you build the next two years, a theme the the AI jobs reckoning breakdown runs all the way down.

6. Jensen's task list is fully automated, and he works harder than ever

The CEO of NVIDIA is the sharpest living proof of his own argument.

“The purpose of a job and the task of the job are related, not the same.”

Jensen’s tasks are typing and talking, both superhuman by AI, and he works harder than ever because the purpose scaled up to absorb the freed capacity. A software engineer’s task is writing code, and the purpose is solving problems and building things that existed only as ideas. AI handles more of the task, the purpose expands, and the same logic reaches every role.

The skill gap between AI-native and non-AI-native graduates already shows up in this hiring cycle. The dislocation lives in the current offers, ahead of the next round.

Try this → write your purpose list and your task list as two separate documents. Every automated task is time returned, and every item on the purpose list is where that time goes. The AI engineer roadmap shows what that expanded purpose looks like in practice.

7. Graduating in 2026 with AI fluency wins the job, this hiring cycle

The most specific and consequential warning in the full 46 minutes.

“If you graduate and you’re not an expert AI user, you’re not going to take a job from another kid who is. That’s a dislocation.”

A skill that was optional yesterday became a requirement today, and Jensen named it flat out. The companies competing hardest for talent want people who use AI fluently, who built workflows around it, who direct AI systems toward outcomes. The AI-native candidate runs more productive at every level, from daily compounding rather than a job title, and it shows in the offers right now.

Try this → change your hiring filter. Ask every candidate to walk you through an AI-assisted workflow they built in the last 30 days. The answer reveals everything a resume hides, and the top-1% Claude setup plus Claude best practices define what fluent actually looks like.

8. Jensen's biggest fear is American disengagement, ahead of China getting AI

The competitive risk is cultural rather than external.

“My greatest concern is that we scare United States people to the point where AI is so unpopular they don’t actually engage it. That we lose our lead as a nation.”

Jensen wants everyone to hold access to AI, China included, and something else worries him. Science-fiction narratives, repeated loudly enough, breed a culture of fear, fear breeds disengagement, and disengagement is how you lose a technology race. America led the last industrial revolution by applying the technology faster than anyone, more than by inventing it.

Try this → audit how you talk about AI publicly and internally. A culture of application and a culture of avoidance grow from the same soil, and you choose which one you water. The MIT research on engagement design sharpens how that choice plays out.

9. Jensen's answer to a billion-dollar AI weapon inverts the arms race

The arms-race instinct loses before it starts, and Jensen’s framework flips it entirely.

“The way you defend against a super force is not with another super force. It’s with an abundance of cheap force.”

The threat is specific: a powerful coding AI in a bad actor’s hands scans thousands of endpoints at once and finds every vulnerability in your codebase faster than a human team could surface them. A single more-powerful defensive AI loses that race, because a motivated attacker out-scales any one defender at the top of the stack. The answer is distribution.

Picture open-source models, each small and cheap, each trained on one threat surface: authentication endpoints, API exposure, dependency chains, credential handling. Individually limited, collectively exhaustive. You can guarantee more coverage than they hold attack surface, even while their model might match yours, which is the whole defensive posture for the AI era.

Try this → map your threat surface by category rather than by severity. For each category, find whether an open-source model trained on that surface exists, and start deploying a swarm of small agents where the gap is widest.

10. Research that took months now takes a day, so add a zero to your ambition

The closing argument, and the one most people will hear and still underweight.

“Whatever level of ambition you have, it’s just not high enough. Whatever expectations I have for the company, you’ve got to increase it by about 100x.”

Jensen described his morning before taking the stage: a professor, then a scientist, research that used to take months collapsing to a day. Drug discovery, climate science, energy science, the physical sciences, all of it compressing across every field. He watched it happen this week, in person, ahead of any projection deck. The constraint on progress moved past compute, past models, past talent. The constraint is now ambition.

Try this → take your three-year roadmap and run one scenario: research timelines 30 times shorter, engineers 10 times more productive. Whatever ships in that version is your actual starting point, so rebuild from there.

The AI playbook: five principles to start using today

Every playbook built before agentic AI earns a rebuild. Here is what that rebuild looks like in practice.

For founders. Your moat is domain expertise applied through AI at a speed rivals struggle to match. The first agent you deploy should solve the recurring problem in your operation that goes unowned. Ship that, then find the next one, using the one-person operating system and the five-agent build as templates.

For investors. The $1-activates-$100 rule is the most useful filter in the market right now. Find the constraint in your sector, find the company removing it, and remember infrastructure is producing returns today while applications come next. The window before both turn into consensus is the one you sit in, and the VC playbook plus the investor list of lists help you move inside it.

For builders. AI fluency compounds, and the gap between the AI-native engineer and everyone else widens every month. Build the workflows now, because the hiring cycle moves now.

Five principles that hold regardless of what ships next:

▫️ Agentic AI crossed the line. The 1,000x compute math is permanent.

▫️ Capital belongs at the constraint, beyond the conversation. Find the layer few are funding.

▫️ AI compresses tasks, purpose expands, and hiring follows the expanded purpose.

▫️ The competitive risk is disengagement. Apply the technology.

▫️ Ambition is the binding constraint now. Add a zero and rebuild from that number.

Jensen spent 46 minutes telling you the ceiling moved. Most people will nod and keep their current plans. You now hold the moves to do something else with that information.

If this breakdown saved you 46 minutes, share it with one founder or investor who should see it.

Keep reading

Agentic AI and the build stack

▫️ AI Agents library

▫️ The Claude Code system that replaces your dev loop

▫️ Loop engineering for coding agents

▫️ Codex background workflows playbook

▫️ Build a five-agent sales team in a weekend

▫️ How to build an AI agent in 2026

AI fluency, tools, and models

▫️ AI Tools and Models library

▫️ How to use Claude like the top 1% of users

▫️ Claude best practices power-user guide

▫️ AI coding tools complete guide

▫️ The 2026 AI engineer roadmap

▫️ SaaS defense playbook for the AI era

Founders, capital, and the market

▫️ Business and Investing library

▫️ The one-person startup operating system

▫️ Marc Andreessen on the AI moat

▫️ Anthropic passed OpenAI at $30B ARR

▫️ The investor list of lists

▫️ What top VCs look for in 2026

▫️ Where VC money is going in AI

▫️ The Coatue AI report in 18 charts

Full interview:

If this breakdown saved you 46 minutes, share it with one founder or investor who needs to see it. They will thank you later.

That question is also your invoice. You pay per token, roughly three-quarters of a word each, on the way in and the way out. Most production apps resend the same system prompt, the same tool list, and the same documents on every call, paying full freight thousands of times a day. Prompt caching alone trims repeated input by up to 90% on Claude. Stack the rest of the system and a typical bill drops by half or more, with the output quality held steady.

This is the full system:

▫️ The 4-layer token model that maps every dollar you spend to a lever you control, from the prompt to the agent loop

▫️ Before-and-after prompt rewrites that cut input tokens 30 to 60% while holding output quality, ready to copy

▫️ The prompt-caching setup that delivers Claude’s 90% discount in practice, with the prefix-ordering rule, the cache_control breakpoints, and the hit-rate target

▫️ The retrieval pattern that replaces stuffing whole documents with searching for the chunks that matter

▫️ The agent and tool diet, including the serialization trick that halves the cost of structured data

▫️ The worked ROI math on a realistic agent workload, so you can size your own savings before you touch a line of code

▫️ The 8 failure modes that silently erase your savings, each with the fix

▫️ The 30-day rollout from measuring your spend to a fully optimized stack

Pair it with the deeper AI Corner library (all included in the premium subscription):

▫️ The Prompting and Context Engineering library for the patterns underneath every rewrite

▫️ The AI Tools and Models library for model rates and routing

▫️ The AI Agents library for the agent-loop economics

▫️ The Claude and Anthropic library for caching mechanics and model choice

▫️ The Business and Investing library for where this margin compounds

Related builds worth reading next: the context engineering guide, the 2026 prompt engineering guide, Claude best practices, loop engineering, and the Codex background workflows playbook.

💸 The Token Cost Playbook

The full system in one place: the 4-layer model, the prompt rewrites, the caching setup, the retrieval pattern, the agent and tool diet, the ROI math, the 8 failure-mode fixes, and the 30-day rollout.

Get The Token Cost Playbook below 👇

Get 50% off forever

Read more

This marks the first time a leading AI lab has pulled a publicly deployed frontier model because the federal government told it to. Every other Anthropic model, including Claude Opus 4.8, stayed live. Both sides now fight in public, and the details cut in directions that embarrass almost everyone involved.

Here is what happened, what each side claims, and why the precedent matters far more than the incident.

The timeline, straight

1. June 9. Anthropic launches Fable 5, the first public release from its top-tier Mythos family, with safeguards built to make Mythos-level capability safe for broad use, alongside Mythos 5 for select enterprise partners. Both build on Claude Mythos Preview, the model held back in April under a cybersecurity program called Project Glasswing.

2. June 10. CEO Dario Amodei publishes a policy essay arguing that governments should hold legal authority to block or reverse the release of frontier models that fail independent safety testing, comparing it to the FAA grounding unsafe aircraft.

3. Around June 11. Amazon, one of Anthropic’s largest investors, calls administration officials to share a report claiming its researchers jailbroke Fable and reached portions of Mythos that pose a national security risk, according to Axios. The Information reported that Amazon CEO Andy Jassy was among the tech leaders raising concerns to senior officials that week.

1. June 12, 5:21pm ET. Lutnick’s letter imposes export controls on Mythos 5 and Fable 5 for any location outside the US and any foreign national inside it, including Anthropic’s own non-citizen staff.

2. June 12, ~10pm ET. Faced with a directive that reached its foreign-born employees, Anthropic switches off both models for all users and posts its statement.

The business and investing stakes here run high, and the Claude and Anthropic library tracks the company’s policy posture across the year for context, including its path to a $1 trillion valuation and the moment it passed OpenAI on revenue.

What the government and its allies say

White House AI czar David Sacks laid out the administration’s version in a public post. His framing, paraphrased: Fable is Mythos with guardrails, so a guardrail failure exposes advanced cyber capability to people who should stay far away from it. A trusted partner testing Fable surfaced a jailbreak, the administration asked Amodei to fix it or pull the model, and in Sacks’s telling Amodei declined, so the export control followed reluctantly. Sacks argues the ball now sits in Anthropic’s court and that remediation would lift the controls quickly.

An administration official told Axios the Commerce Department acted after another company claimed it could jailbreak Mythos, and that the government had earlier tried and failed to get Anthropic to pause the release. Sacks also leaned on a pointed irony: Anthropic itself promoted Mythos as something close to a cyberweapon that warranted regulation, so a vulnerability in its consumer wrapper becomes, by Anthropic’s own standard, the company’s job to patch.

What Anthropic says

Anthropic’s statement tells a different story. The company says the letter withheld the specific national security concern, and that its understanding is the government learned of a jailbreak method. Anthropic reviewed a demonstration and characterized it as narrow and non-universal, essentially asking the model to read a specific codebase and fix software flaws, a capability it says ships in other public models including OpenAI’s GPT-5.5 and runs every day in the hands of the defenders who keep systems safe.

The company’s core arguments:

▫️ Testers so far have surfaced only narrow, non-universal bypasses, while the universal kind that broadly unlocks dangerous cyber capability stays unseen

▫️ It runs a defense-in-depth strategy, including 30-day data retention to detect and shut down attacks fast, a policy it says carries genuine commercial cost

▫️ Recalling a model over a narrow finding sets a precedent that, applied industry-wide, would freeze frontier deployments across every lab

▫️ It is complying while disputing the basis, calling the action a misunderstanding and working to restore access

Anthropic officials told Axios they walked the administration through how the alleged Amazon jailbreak was simple, reproducible on other models, and consistent with healthy safety systems. Both accounts agree a narrow technique was demonstrated. They split completely on what it means.

The irony everyone noticed

The framing that spread fastest, including a TechCrunch headline arguing Anthropic’s safety warnings may have backfired, writes itself. Anthropic spent years as the safety-first lab calling for government oversight of powerful models. Amodei published an essay days earlier requesting exactly the authority the government then used on his own flagship. Critics also seized on the 30-day retention policy, casting the safety-first company as one that now logs every prompt and output to a Mythos-class model.

Anthropic’s defenders read it the opposite way: the company asked for a transparent, fair, fact-grounded statutory process, and argues this directive, delivered by letter with the evidence sealed and the harmful result still undisclosed, departs from every one of those principles. Both readings stay live, and which one lands depends largely on how serious you judge the jailbreak to be, which is the part the public still waits to see.

The precedent every builder should track

Strip away the personalities and one question remains: should a government be able to switch off a commercially deployed model on a few hours’ notice, by letter, while the evidence stays sealed? This episode is the first working example of model-layer regulation in practice, and it rewards seeing clearly from both sides.

The case for it runs straight. A frontier model with advanced cyber capability is a national security asset, the government holds a legitimate interest in who can reach it, and a lab that markets its model as dangerous invites the state to treat it as dangerous. When independent testers surface a bypass, pausing the model while it gets patched reads as a reasonable precaution.

The case against it is the one Box CEO Aaron Levie made publicly, and it is the deeper structural worry. Regulating at the model layer hands the government sole discretion over when a model reaches the public, based on risk judgments that stay inherently subjective. Every release becomes a months-long negotiation over what a model can do and how dangerous that is, which produces a backlog, slows the market, and starts to make AI look like any other heavily gated industry. Levie’s sharpest line: had this paradigm existed three years ago, we might still sit near GPT-4. His proposed alternative is to regulate the applied use of AI in cyberattacks, fraud, and biosecurity, leaving the model itself free to ship.

The disagreement reaches past Fable. It is about whether the kill switch belongs at the model or at the use. This week was the first live test, and everyone walked away still arguing.

The AI tools and models library and the Claude and Anthropic library both track how this regulatory question keeps reshaping the release calendar.

The Amazon question

The strangest thread is the trigger. Amazon ranks among Anthropic’s largest investors and a major customer, having poured billions into the company and sold Claude across its cloud. Yet by the reporting, Amazon’s own researchers produced the report that set the takedown in motion, and Andy Jassy carried concerns to the administration.

Why would a major investor strike a blow this disruptive against a company it owns a large stake in? Three readings circulate, each still unconfirmed:

▫️ Genuine safety concern, where Amazon researchers found something alarming and escalated it through the channels available

▫️ Competitive positioning, where a backer with its own model ambitions gains from a rival flagship sitting grounded

▫️ Hedging a complicated relationship, where Amazon balances its Anthropic stake against its government and enterprise commitments

The honest answer is that outsiders still lack the evidence to tell which it is, and the venture dynamics around the big labs make all three plausible at once. The investor list of lists and the biggest VC-backed startups show how tangled the cap tables behind frontier AI have become.

The dual-use paradox at the center of it

Here is the technical knot that makes this hard to resolve cleanly. The capability in dispute, a model reading a codebase and finding or fixing software flaws, is exactly what you want on the defensive side of cybersecurity. The same skill that helps an attacker find a vulnerability helps a defender close it, and defenders run this work daily. By Anthropic’s account, and by Levie’s, the capability already ships broadly across current models.

That is why “is this a cyberweapon” resists a clean verdict either way. A frontier model is dual-use by design, and a rule that grounds any model capable of finding a software flaw would ground most of them. The disagreement between Anthropic and the administration is partly about severity and partly about this deeper question of whether dual-use capability can be governed at the model level while the field keeps moving. The SaaS defense playbook covers how defenders actually put these capabilities to work in production.

The three paths back to release

Anthropic says it is working to restore access. Realistically, Fable returns through one of three doors:

1. Government reversal. Anthropic demonstrates the jailbreak is narrow and reproducible elsewhere, the administration accepts the technical case, and the controls lift. Sacks has signaled this is the outcome the administration wants, which makes it the fastest path once the technical facts land.

2. Negotiated remediation. Anthropic ships a patch and extra safeguards, both sides claim a win, and the model returns with tighter controls. This reads as the likeliest middle path given the prior friction between the two.

3. Court order. Anthropic, already in litigation with the administration over a separate supply chain risk designation, escalates legally. Slower, and it hardens the conflict rather than settling it.

The backdrop matters here. In February, the administration directed federal agencies to stop using Anthropic’s technology after the company objected to certain defense uses, the DOD later labeled Anthropic a supply chain risk, and Anthropic sued. Sacks insists this week’s action runs separate from that history. Others read one continuous conflict. The truth likely sits in between, and the Dario Amodei safety thesis is now colliding with the politics it helped invite, an arc the Davos zeroth-world essay foreshadowed.

What this means for you

Set aside who is right. Three durable lessons hold whichever way the fight ends.

For builders. Treat single-model dependence as a live risk you design around. A model you built a product on can vanish overnight for reasons that trace to forces beyond your control. Design for multi-model resilience: an abstraction layer, a tested fallback to another provider, and an honest answer to “what happens to my product the night this specific model goes dark.” The AI coding tools guide and the power-user setup both cover building across providers.

For investors. Regulatory action now prices in at the frontier as a base case rather than a tail. A model’s release calendar, and therefore a lab’s revenue, can hinge on a letter. That reshapes how you underwrite the frontier labs and lifts the value of companies positioned one layer up from the raw model. The Coatue AI report and the venture map at VC Corner help size where that risk sits, and Mark Cuban’s track record on calling these moments early is worth a read.

For operators. The public, direct-to-audience nature of this fight is the new normal. As Anthony Pompliano observed, both sides skipped the press and posted their own versions, leaving citizens to judge. Read the primary statements, weight the partisan accounts accordingly, and assume every party argues its own book.

While Washington argues, build the leverage

The policy fight will run for weeks. The leverage sits in wielding Claude deeply enough that any single model’s fate matters less to your output. Here is where to go deeper, all in one place:

▫️ Loop engineering for designing the system that prompts your agent for you, in the loop engineering playbook

▫️ The one-person company that runs like a 30-person team, in the one-person startup operating system

▫️ Background automations that run while you sleep, in the Codex background workflows playbook

▫️ The six patterns Anthropic engineers use in the Claude Code dynamic workflows guide

▫️ The Claude Code system that replaces your dev loop, in the full system breakdown

▫️ A five-agent sales team you assemble in a weekend, in the build guide

▫️ A second brain on Claude with Obsidian and the Granola MCP stack

▫️ Your own stock analyst in 12 prompts, in the Claude stock analyst build

▫️ The power-user playbook in Claude best practices and managed agents

▫️ The complete skills library in the Claude Skills startup marketing library and the Claude Skills complete guide

What to watch next

▫️ Anthropic’s promised technical detail, and whether it convincingly shows the jailbreak reproduces widely

▫️ The administration’s evidence, since the sealed specifics sit at the heart of Anthropic’s complaint

▫️ Amazon’s explanation, if any arrives, for triggering an action against a company it heavily backs

▫️ The restoration timeline, and which of the three paths it travels

▫️ Whether other governments copy the move, since a US precedent for model-layer control invites imitation abroad

TL;DR

The US government ordered Anthropic to suspend Fable 5 and Mythos 5 three days after launch, over a jailbreak the administration calls a national security risk and Anthropic calls narrow and widely reproducible. Amazon, a major Anthropic investor, appears to have triggered it. The deeper fight is about model-layer regulation, where the government gates releases directly, and whether that freezes the field or protects it. For builders, the lesson is multi-model resilience plus deep Claude leverage. For investors, regulatory risk now prices in at the frontier. The evidence the public can verify stays thin, so weight every account, including this one, by who gains from it.

Keep reading

Anthropic, Claude, and the policy fight

▫️ Claude and Anthropic library

▫️ Anthropic’s $1 trillion valuation breakdown

▫️ Anthropic passed OpenAI at $30B ARR

▫️ Dario Amodei’s zeroth-world thesis from Davos

▫️ Dario Amodei on safe AI and AGI

▫️ The Claude Code source code leak

Build with Claude while the policy plays out

▫️ Loop engineering for coding agents

▫️ The one-person startup operating system

▫️ Codex background workflows: 10 automations

▫️ Claude Code dynamic workflows: 6 patterns

▫️ The Claude Code system that replaces your dev loop

▫️ Build a five-agent sales team in a weekend

▫️ Obsidian and Claude second brain playbook

▫️ Granola and Claude second brain MCP stack

▫️ Build your own stock analyst with Claude

▫️ Claude best practices power-user guide

▫️ Claude managed agents guide

▫️ Claude Skills startup marketing library

▫️ Claude Skills complete guide

▫️ How to use Claude like the top 1% of users

The capability, security, and market side

▫️ AI Tools and Models library

▫️ SaaS defense playbook for the AI era

▫️ AI coding tools complete guide

▫️ MIT on how ChatGPT is designed to hook you

▫️ Mark Cuban has been right every time

▫️ Business and Investing library

▫️ The investor list of lists

▫️ Where VC money is going in AI

▫️ The Coatue AI report in 18 charts

Peter Steinberger, founder of OpenClaw, put it to 8 million viewers in a day: stop prompting coding agents, start designing the loops that prompt them.

For two years the move was simple. You wrote a good prompt, fed enough context, read what came back, typed the next thing. The agent was a tool and you held it the whole time, one turn after another. That posture is ending. The new one is a small system that finds the work, hands it out, checks it, records what is done, and decides the next thing, while you watch instead of type.

Here is what changed in 2026, and why this stopped being a weekend bash-script project. The pieces ship inside the products now.

Claude Code added /goal in v2.1.139 on May 11, 2026, a native loop that runs across turns until a condition you wrote is true, with a separate fast model grading the work after every turn.

Codex shipped the same primitive in CLI 0.128.0 on April 30, plus an Automations tab and a triage inbox. Once you see that both tools expose the same five blocks, you stop arguing about which one and start designing a loop that works in either.

This is the full build:

▫️ The five building blocks of a loop mapped command-for-command across Claude Code and Codex, so your loop runs in whichever tool you are sitting in

▫️ Copy-paste /goal conditions that actually stop, plus the four constraints that silently break most of them (the evaluator reads the transcript, not your files)

▫️ A working SKILL.md skeleton and a sub-agent config for the maker-checker split that lets you trust a loop running unattended

▫️ The full reference loop, the morning-triage-to-merged-PR blueprint you design once and never prompt again

▫️ The token economics, where a loop burns money, and how to cap spend whether you are token rich or token poor

▫️ The 8 failure modes that turn an unattended loop into unattended mistakes, each with the fix

▫️ The memory spine, the on-disk state file that survives between runs because the model forgets and the repo does not

▫️ The 30-day rollout from your first /goal to a self-feeding loop with a verifier you trust

Pair it with the deeper AI Corner library:

▫️ The AI Agents library for orchestration and reliability

▫️ The Prompting and Context Engineering library for skills and intent

▫️ The Claude and Anthropic library for Claude Code and Codex internals

▫️ The AI Tools and Models library for the build stack

▫️ The Business and Investing library for where this leverage compounds

Related builds worth reading next: the Claude Code system that replaces your dev loop, the AI agent reliability playbook, the multi-agent code review build, the Claude Skills complete guide, and the context engineering guide.

🔁 The Loop Engineering Build Kit

The full system in one place: the five blocks mapped across both tools, copy-paste /goal conditions, SKILL.md and sub-agent configs, the reference loop, the token math, the 8 failure-mode fixes, and the 30-day rollout.

Get The Loop Engineering Build Kit below 👇

Get 50% off forever

Read more

The reason is a quiet role change. The founder stops being the individual contributor who writes the code and runs the ops, and becomes the orchestrator of agents that carry the work out. Your edge moves from execution to judgment. Anthropic made the same argument in its 36-page Founder’s Playbook, and the math is blunt: validation cycles that took months now take an afternoon, and a single founder can run like a team many times their headcount.

Here is the part most people miss. When building gets this cheap, the often-cited CB Insights number, 42% of startups dying because they built something nobody wanted, does not shrink. It climbs. The bottleneck moved from “can you build it” to “should you,” and that favors people with domain judgment, not just engineers.

This is the full operating system for running that company as one person. Here is what sits behind the paywall:

▫️ The 3-surface decision framework that tells you exactly when to reach for Chat, Cowork, or Code so you stop using a chat box for work that needs an agent

▫️ The 4-stage AI-native build (Idea, MVP, Launch, Scale) with the goal, the exit criteria, and the AI-era trap for each stage

▫️ 30+ copy-paste prompts that turn Claude into your validator, your architect, your security reviewer, your GTM strategist, and your board prep, each one ready to run today

▫️ The 12-function Founder OS, a complete prompt library across validation, business model, fundraising, GTM, product, sales, marketing, growth, ops, finance, customer success, and legal

▫️ The CLAUDE.md context system that keeps an agent building coherent software across months instead of re-deriving your architecture every session

▫️ The 8 AI-era failure modes that kill solo startups, each with the exact prompt-level fix

▫️ The solo tool stack for build, design, video, analytics, and support, so one person covers six departments

▫️ The 30-day rollout from raw idea to a company that runs without you in every loop

Pair it with the deeper AI Corner library:

▫️ The AI Agents library for orchestration and reliability

▫️ The Prompting and Context Engineering library for the prompt patterns underneath every exercise

▫️ The Claude and Anthropic library for the surfaces and skills

▫️ The AI Tools and Models library for the build stack

▫️ The Business and Investing library for fundraising and GTM

Related builds worth reading next: the Claude Skills complete guide, the Claude Code system that replaces your dev loop, the Claude Cowork breakdown, the AI GTM playbook, and the VC Corner investor list of lists.

🛠️ The One-Person Startup OS

The full system in one place: the 4-stage build, the 12-function prompt library, the CLAUDE.md context setup, the 8 failure-mode fixes, the solo tool stack, and the 30-day rollout.

Get The One-Person Startup OS below 👇

Get 50% off forever

Read more

Benedict Evans has called technology shifts right more consistently than almost anyone alive. If this is the first time you are hearing his name, you are about to read something full of gems.

He spent 25 years as one of the sharpest analysts in tech, across equity research, strategy consulting, and a long run as a partner at Andreessen Horowitz. Twice a year he publishes a macro deck on where technology is heading. His 2013 “Mobile Is Eating the World” became required reading for the smartphone wave while it was still forming, and he called mobile before most people knew what mobile meant.

His weekly newsletter reaches around 200,000 readers, and he has presented for Alphabet, Amazon, LVMH, and Goldman Sachs. When Evans publishes, serious people in tech read it carefully and argue about it for months. “AI Eats the World” is his most consequential deck yet, because it does not tell you AI is revolutionary. It forces you to ask which part of the revolution you are in, and whether the part you are betting on is the part that wins.

together with Hard Skill Exchange:

Evans makes you pick which part of the AI shift you are betting on. Here is one that is already arriving: your next customer may not be human.

Hard Skill Exchange is running a FREE virtual AI summit on exactly that. Agent-to-Agent GTM, where the people building machine-native go-to-market show what breaks when software becomes the buyer, the seller, and the decision maker.

▫️ 3 days, fully virtual, free for everyone, with no ticket and no travel

▫️ Agent-native GTM playbooks for a market where buyers skip ads, brands, and funnels

▫️ The operators defining the category before the consensus catches up

It runs live June 23 to 25, and then it is gone. Your next customer may not be human, so it would be a mistake to sit this one out:

Save your Free seat !

Table of Contents

1. Platform Transitions Eat Their Parents

2. The Financial Model Is Quietly Breaking

3. 900 Million Users. Almost Nobody’s Hooked.

4. Models Are a Commodity. Deal With It.

5. The Right Question About Jobs

6. The Uncomfortable Math

1. Platform Transitions Eat Their Parents

Every 10 to 15 years, the tech industry doesn’t just add a new layer. It rewrites the rules of the previous one, and the prior winners often don’t survive the rewrite.

Evans opens the deck here, and it deserves more than a quick nod.

The historical progression runs from Mainframes to PCs to the Web to Smartphones to Generative AI. Each transition didn’t just produce new companies. It made the dominant players of the prior era structurally irrelevant in ways they couldn’t have anticipated or defended against.

Microsoft went from owning essentially all of global computing in 2000 to less than 15% of computing units by 2025. Not because Windows got worse. Because the iPhone redefined what “a computer” was, and Microsoft had no answer for it.

This is what makes the current spending war legible. The CEOs committing hundreds of billions are not doing it because the returns are obvious. They are doing it because they have studied that chart and have no interest in becoming the next example on it.

Sundar Pichai said the risk of under-investing is significantly greater than the risk of over-investing. Zuckerberg said the worst case is that they “just prebuilt for a couple of years.” Those are not confident statements. They are the sound of people who know what happens to the unprepared.

The question nobody is asking loudly enough

What Evans doesn’t say explicitly, but the framing implies: the companies most at risk here are not the ones who refuse to invest. They are the ones who invest in the wrong layer.

Platform transitions have a habit of making yesterday’s moat look like a liability, and there are a lot of very expensive bets being placed right now on layers that may not be where the value ends up.

The companies that dominated the PC era built for PCs. The companies that dominated mobile built for mobile.

Winning the current infrastructure race does not guarantee winning what comes after it. That distinction is what the rest of the deck is really interrogating.

2. The Financial Model Is Actually Breaking

The four biggest tech companies, Meta, Microsoft, Alphabet, and Amazon, are planning a combined $700 billion in capital expenditure in 2026 alone.

The entire global telecom industry spends around $300 billion annually. Big Tech has entered that tier of physical investment almost overnight, and the balance sheets are starting to show it.

For most of their existence, these were asset-light businesses. Software scales without proportional physical investment. High margins, minimal infrastructure, free cash flow that funded buybacks, acquisitions, and everything else. That was the entire premise of their valuations, and it is now under serious pressure.

Meta and Microsoft are both approaching 55% of revenue going to capex in 2026. In 2015, Meta was spending around 10%. These numbers do not look like software companies anymore. They look like the capital structures of the industries they used to disrupt.

When a company’s cost structure changes this radically this fast, something has to give. Either the revenues scale to justify it, the margins compress, or the valuation assumptions get repriced. Right now, markets are betting hard on option one.

The deeper issue is that this investment is not discretionary anymore. You can’t half-build the infrastructure. You can’t decide in two years that you over-invested and sell the data centres back.

This is a one-way door, and the world’s most valuable companies walked through it while telling their shareholders it was the obvious move. Maybe it was. But the size of the bet means there is no graceful exit if the thesis is wrong.

A quick one, courtesy of Hard Skill Exchange:

While Big Tech bets $700B on the infrastructure, the go-to-market layer on top of it is already changing

The free Agent-to-Agent GTM Summit covers what happens when software becomes the buyer. 3 days, fully virtual, free for everyone, live June 23 to 25.

Grab your FREE seat

3. 900 Million Users. Almost Nobody’s Hooked.

ChatGPT has over 900 million weekly active users, but only 5% of them are paying. Evans shows the growth curve and it looks extraordinary, right up until the next slide reframes everything.

OpenAI’s own “Wrapped” data tells a different story. For at least 80% of users, total usage across all of 2025 was under 1,000 prompts. That is fewer than 3 per day. For the vast majority of people who “use” ChatGPT, it is something they open occasionally, the way you open a dictionary.

Evans calls this “a mile wide and an inch deep.”

At the end of the day, this is not product-market fit. Product-market fit is when people use something daily without being reminded it exists, when they complain loudly if it goes down, when they genuinely cannot describe their workflow without it.

The comparison to smartphones cuts in the wrong direction. Within three years of the iPhone, people were sleeping next to their phones. Within three years of ChatGPT, 80% of users are sending fewer prompts per day than they send text messages per hour.

The VC Corner

Model-Market Fit: The New Make-or-Break for AI Startups

The Framework Every AI Founder Is Missing…

Read more

2 months ago · 74 likes · Ruben Dominguez

The enterprise story mirrors the consumer one.

Bain’s data shows that across every major business function, the gap between “in pilot” and “in production” is enormous.

Everyone has a proof of concept. Almost nobody has something running at scale that has actually changed how the business operates.

Pilots are easy to start and hard to kill. They generate internal goodwill, satisfy board questions about AI strategy, and require minimal organisational change.

Getting from pilot to production means redesigning workflows, retraining people, and accepting that the model will sometimes be wrong. Most companies haven’t started that part.

The one real exception is coding. Enterprise AI spending on software development is roughly five times the next biggest category.

Among developers, AI has clearly become a daily essential. But coding is one profession, and the distance between “coding is transformed” and “work is transformed” is still vast.

4. Models Are a Commodity. Deal With It.

This is the part of the deck the industry is not engaging with honestly, and it has direct consequences for how you think about the biggest valuations in private market history.

Look at the benchmark chart for frontier LLMs. OpenAI, Anthropic, Google, Meta, the Chinese labs: scores are converging at the top. For most general use cases, the models are now functionally interchangeable.

And there are no network effects. A model doesn’t get better because more people use it. It doesn’t accumulate data about you that makes switching painful.

Evans draws the comparison to mobile telecom infrastructure, and it is the most pointed analogy in the deck.

Global mobile data traffic exploded over 15 years. Global telco stocks went essentially nowhere over the same period. The companies that built the pipes carried all the value but captured almost none of it.

Infrastructure businesses at scale, serving commodity demand, get priced like utilities. Necessary, steady, and low margin. Sam Altman’s “intelligence as a utility like electricity or water” is more accurate than he probably intends. Utilities are worth owning. They are not worth $850 billion.

Electricity producers don’t capture the value created by the industries running on their power. That value goes to whoever builds the product people actually use every day.

The interesting question is not who trains the best model. It is who builds the thing on top of it that becomes genuinely indispensable.

5. The Right Question About Jobs

Most of the public conversation about AI and employment is asking the wrong question. People want to know which jobs disappear. So the ultimate question according to Evans is:

Is this a job, or is it a task?

A task is a defined operation with a known input and output. A job contains judgment, relationships, and decisions made under genuine ambiguity. That’s a huge distinction when we talk about AI.

In 1950, there were nearly 100,000 elevator operators in the United States. Otis automated the elevator, and the profession was gone within a decade, because the job was just a task.

There was nothing else there. But accountants grew as a share of US employment for decades after software automated their core computational work, because what accountants actually do expanded as the cost of calculation fell.

Before grocery barcodes in 1974, the average US supermarket stocked around 8,000 products. By 2010, that number was 50,000. Automated inventory management didn’t just make stocking cheaper. It made a level of variety that was previously impossible suddenly achievable. The free task unlocked something that couldn’t exist before.

The live test nobody wants to talk about

Evans points to the Philippine outsourcing industry as the most immediate real-world test of this framework.

Two million people, 8% of GDP, built entirely on a skill and income arbitrage that AI is now directly threatening.

Whether those roles survive depends entirely on whether they are tasks or jobs. Some are pure tasks with defined inputs, defined outputs, no judgment required. Those are (or will be) gone.

Others involve enough human context, relationship management, and situational reading that they are genuinely jobs. Those will change, probably dramatically, but they won’t simply disappear.

Anyone telling you they know exactly which category any given role falls into is guessing. The honest answer is that nobody knows yet, and the outcome will vary enormously by industry, company, and how aggressively employers choose to act.

6. The Uncomfortable Math

One slide in Evans’ deck is more unsettling than any of the capex charts.

OpenAI and Anthropic are each valued in private markets at roughly $850 to $900 billion. The total market capitalisation of every US venture-backed IPO from 1995 to 2000, the entire dot-com boom adjusted to today’s dollars, was around $780 billion.

Two companies, each worth more than an entire era of internet investment, with business models nowhere near equilibrium, in a model layer that is commoditising in real time.

There are three ways to make that math work:

1. The valuations are irrational and a correction is coming.

2. The commodity thesis is wrong and one or both labs builds a lasting moat through something not yet invented.

3. The application layer ends up being worth so much that the model providers capture enough of it to justify the numbers.

Evans ends with William Goldman’s line: “No one knows anything.” The internet era produced AOL, Yahoo, and Pointcast before it produced Google. The mobile era produced WAP and Nokia before it produced the iPhone.

Every platform transition has false starts, confident predictions that age terribly, and a long stretch of genuine uncertainty before the real picture emerges. This one will be no different.

The infrastructure is real. The growth is real. The most important chapter of this story hasn’t been written yet.

The accountant who saw VisiCalc in 1979 and understood what it meant built something lasting. The one who called it a curiosity missed everything that followed.

That decision is sitting in front of every operator and founder reading this right now.