Scott Galloway: 95% of enterprise AI spend connects to no return a CFO can name

The man who called the dot-com crash from a faculty meeting is making the same call on AI valuations

Jun 04, 2026

The man who called the dot-com crash at a faculty meeting while his colleagues were still buying Pets.com is making the same argument about AI valuations today.

I watched the full Prof G Markets live show in San Francisco so you do not have to.

Here are the 10 biggest takeaways:

📢 A quick word before we get into it.

Galloway’s whole argument is that enterprise AI runs ahead of the controls that should govern it. The cost side breaks first. The access side breaks next, and quietly.

Opal’s 2026 Identity Governance Report just showed that 80% of systems are exposed through stale access. Access that should have been removed and was not…

Now layer AI agents on top, with permissions that compound every week, and the blast radius becomes impossible to contain.

Opal maps how AI-ready teams fixed the access layer before agents made the problem permanent.

Read the 2026 Report !

The CFOs running the AI cost audit Galloway describes will run the access audit next. The teams ready for both are on the right side of every line in this breakdown

1. Uber burned its entire 2026 AI budget before April. The COO said the costs are getting hard to justify

The flagship promise of enterprise AI was the simplest one. AI does the work, the bill goes down.

“The technology is actually more expensive than the humans it is supposed to replace.”

AI eliminated nearly 50,000 jobs this year. That approaches the full 2025 total. The layoffs arrived on schedule. The savings did not.

Uber exhausted its entire 2026 AI budget before the first quarter ended. Microsoft canceled cloud code licenses across multiple divisions. An NVIDIA executive confirmed publicly that compute runs far beyond the cost of employees.

The Uber COO did not hide it. CFOs say costs are hard to justify only after the internal conversation has already gotten uncomfortable.

Three things broke the model at once:

1️⃣ Compute costs were modeled against 2023 pricing. Usage assumptions were wrong by a factor.

2️⃣ Enterprise deployment requires far more inference volume than any projection captured.

3️⃣ Nobody budgeted for what happens when 10,000 employees get an AI tool with no usage ceiling.

The entire ROI case for enterprise AI depends on the cost equation flipping. Every CFO eventually faces a line item they cannot explain to a board. That moment is already here at Uber, Microsoft, and Salesforce.

For the honest version of which jobs AI is actually replacing right now, see no one is safe from AI and the SaaS defense playbook for the AI era.

2. Stripe spends $100,000 a day on AI tokens. The person who spends the most gets the trophy

The cost problem goes beyond size.

“They’ll create leaderboards at these companies, like Meta, like Amazon, where they will track how many AI tokens you’re using. The people who are using the most tokens are the most AI-deployed. Those are the ones who get recognized. Maybe they’ll get a promotion.”

Stripe’s technical staff spends nearly $100,000 on AI tokens every day. Salesforce is on track for $300 million in Anthropic spend this year. One Anthropic employee ran up a Claude Code bill of $150,000 in a single month. ServiceNow exhausted its entire Anthropic budget. Shopify reported earnings partially offset by increased LLM costs.

Meta. Spotify. Pinterest. Same story. Same quarter.

Companies built recognition systems around consumption. Employees rationally responded by consuming as much as possible. The incentive and the outcome are completely disconnected.

Four things broke in parallel:

1️⃣ A KPI measuring activity, instead of output

2️⃣ A leaderboard that rewards cost generation, instead of value creation

3️⃣ A promotion system tied to spend, instead of return

4️⃣ A budget with no anchor to any identifiable result

Token leaderboards are a company car with no fuel budget. Hand someone a tool, measure them on how much they use it, and the bill becomes the metric.

For the operator playbook on which AI workflows actually deliver ROI a CFO can name, see 25 Claude Skills that give your startup a marketing team it cannot afford yet, build your own stock analyst with Claude, and the 5-agent sales team you can build this weekend.

3. An MIT professor found 95% of AI projects connect to no return. The market has not priced that in

The number is the thesis.

“About 5% of the projects that people are using tokens for, CFOs can connect to some sort of return. They’re really intoxicated. Using AI as much as you can and talking about it in your earnings calls, it’s like adding .com back in the 90s.”

Companies added “.com” to their names in 1999 to signal relevance. It worked for about 18 months. Mentioning AI on earnings calls serves the same function today. A market communication strategy. Not a proof of value.

Galloway’s specific prediction: the first major CEO who announces a significant AI pullback due to missing ROI is the trigger event. NVIDIA’s first earnings miss follows. That miss takes the broader market down 5% to 10%.

NVIDIA has beaten estimates 15 consecutive quarters. The streak is the pressure.

The profile of the trigger CEO:

▫️ A traditional Fortune 500 company, not a tech native

▫️ Someone with credibility and no evangelism invested in the AI trade

▫️ A CFO-driven decision announced quietly on an earnings call

▫️ A statement that starts with “efficiencies” and ends with “scaling back”

When the first credible CEO says it out loud, the repricing happens the day of the announcement. NVIDIA is the tell.

For the investor lens on this exact moment, see Coatue’s 18-chart AI report, where VC money is going in AI, and the most valuable VC-backed startups in the world.

4. Chinese AI models are 10-30x cheaper. 80% of American startups are already using them

The part most enterprise AI conversations skip.

“Every developer in the world has heard of them because 80% of American AI startups are now using Chinese models.”

DeepSeek. Kimi K2. Xipu. GLM. The pricing gap runs 10x in some comparisons. 30x in others.

Two structural reasons explain it. Chinese AI companies receive direct government subsidies. They also engage in what the industry calls distillation: industrially harvesting outputs from American frontier models and using those outputs to train their own systems. Galloway names it plainly. Sophisticated theft.

Cost-pressured companies unable to justify their Anthropic or OpenAI bills migrate to whatever delivers similar outputs at a fraction of the price. The economics make the decision for them.

The migration runs without coordination:

1️⃣ CFO flags AI spend as unjustifiable

2️⃣ Procurement gets involved

3️⃣ Technical team identifies cheaper alternatives

4️⃣ DeepSeek performs comparably at 10-30x lower cost

5️⃣ Migration happens without announcement

American AI companies compete against models that are cheaper partly because they were built on stolen outputs. Markets alone will not correct that.

For the AGI race across labs and what each one is betting on, see Demis Hassabis named his AGI year, Anthropic just passed OpenAI in revenue spending 4x less, and Sam Altman watched 900 million people talk to one personality every week.

5. Galloway's precise prediction: AI valuations fall 50-70% in 24 months.

One number forces a choice between two futures.

“One of two things is going to happen. Either the valuations of AI are going to come down by 50% or 70%, or we’re going to have labor chaos in these industries. I absolutely think it’s the former.”

Walk the numbers. 155 million working Americans. Half are AI-vulnerable. That is 75 million people. At $100K average labor cost, that is a $5 trillion pool of potential savings. To justify current AI valuations through labor replacement alone, you need 5-7 million layoffs in 2-3 years. Roughly 10% labor destruction.

That has not happened. The layoffs this year are real. They are far below what the valuation math requires.

The AI fear machine runs on catastrophizing as a business model. Fear is the product. Capital is the outcome.

Mass disruption at the speed the valuation requires means visible chaos, political crisis, regulatory response. Disruption at the speed the evidence supports means valuations compress 50-70%. Quiet. Certain. Galloway thinks the second outcome is where this ends.

Being right about the technology and wrong about the timeline produces the same financial result as being wrong entirely.

For the VC version of this argument, see what top-tier VCs actually look for in 2026, what top VCs check in due diligence before writing checks, and $80 billion in 3 months: Q1 2026’s record-breaking fundraising.

6. Xi Jinping can flood the U.S. with cheap AI. 90-day window before Trump acts

The optimal play from Beijing is already obvious.

“AI is the only thing that feels like it’s propping up the economy right now. The Trump administration has too much to lose.”

A White House at 34% approval has one economic prop: the S&P and NASDAQ. Those indices have one primary support mechanism, which is the AI trade. Flood the U.S. market with cheap Chinese LLMs, go directly to CFOs already sick of their AI bills, and the trade begins to unwind.

Xi does not need to win the AI race. He needs to win the pricing war in procurement meetings.

80% of smaller American startups have already shifted to Chinese LLMs on cost alone. With a strategic push, it accelerates.

Galloway’s prediction unfolds in four steps:

1️⃣ Mass Chinese LLM adoption surfaces in earnings calls

2️⃣ American AI companies lobby for protection using distillation claims

3️⃣ The administration frames it as national security

4️⃣ The Trump administration bans Chinese LLMs within 90 days, using the BYD vehicle ban as the template

A Chinese LLM ban is a geopolitical event with direct market consequences. It validates the threat, protects American AI incumbents from below-cost competition, and installs the first real regulatory layer on a sector operating without one.

For the geopolitical and vertical-integration thesis, see Elon Musk and the outer limit of vertical integration and Dario Amodei named “the zeroth world”.

7. China ran this playbook on German industry for 30 years. AI is next.

“What China has done to Europe economically, they’re going to try to do to the AI market what they tried to do to the steel market here in the 80s and 90s.”

China attracted Volkswagen, Daimler, and Siemens with favorable economics. Chinese production and R&D became so profitable that when Germany tried to restrict IP theft and product dumping, its largest companies lobbied against the policy from the inside. They were already dependent on the arbitrage.

The steel playbook, now running on AI:

1️⃣ Subsidize model development to below-market pricing

2️⃣ Attract American developers with cost savings

3️⃣ Harvest outputs via distillation to improve models further

4️⃣ Create economic dependency before policy can respond

5️⃣ Use dependent American companies as lobbying assets against any proposed ban

Industrial strategy disguised as market competition.

For the strategic context of who positions for the next decade, see Mark Cuban on the AI thesis, Marc Andreessen on why the AI moat is not the model, and the AI agent that thinks like Jensen Huang, Elon Musk, and Dario Amodei.

8. SpaceX is the greatest business built in 50 years. The IPO bundles it with a division losing $2.5B per quarter.

Separate the asset from the structure.

“If you want to hang out with Snow White, SpaceX, you have to also invest in this money furnace called xAI.”

SpaceX by the numbers: 90% global launch capacity. Two-thirds of all low-Earth orbit satellites. $16 billion in revenue. $8 billion in operating profit. 30% annual growth. The deepest competitive moats of any business operating today.

At $700-800 billion, that company is priced aggressively but defensibly. Double it to $1.6 trillion for the Elon premium. That case can be made.

At $2 trillion, at 107 times sales, growing revenue at 15% annually, losing $20 billion per year, the numbers collapse.

The comparison that ends the argument: Google went public growing at 240% annually at 10 times revenue. SpaceX goes out growing at 24% at 100 times revenue. The ratio is roughly 1,000 to 1 in Google’s favor.

The Snow White is real and worth owning. The furnace is real too. Buying both without separating them is the mechanical path to overpaying by several hundred billion dollars.

For the VC playbook on separating asset from structure, see the most valuable VC-backed startups in the world, the ultimate investor list of lists, and Elon Musk and the outer limit of vertical integration.

9. SpaceX, OpenAI, and Anthropic IPO at $4T combined. None profitable. Retail investors are the last buyers.

The smartest people in the room are selling to you. That is not a metaphor.

“When these companies go public, it’s effectively the smartest people in the room saying: we’ve squeezed as much juice out of this as we can. We need to find people stupider than us to invest at this valuation.”

$4 trillion combined. That exceeds every dot-com IPO adjusted for inflation. It equals roughly half the combined value of every IPO in the prior 50 years. None of the three are profitable.

NASDAQ changed its rules specifically for these companies. Previously: 12 months of public trading before joining passive index funds. Now: 15 days for mega-caps. These three will represent approximately 6% of global public equity markets on listing day, before reporting a single public quarter.

The exit sequence is mechanical:

1️⃣ VCs and early investors know the financials better than anyone outside

2️⃣ Private capital can fund rounds of $10-15 billion without a public listing

3️⃣ Going public means insiders believe they have extracted maximum private value

4️⃣ The NASDAQ rule change accelerates passive fund inclusion before the public has one quarter of data

5️⃣ Retail investors become the last buyers in the chain

Galloway’s advice: sell immediately. Regardless of VC pressure, manager pressure, or any long-term vision framing.

Retail investors are the last buyers in a chain that started with the people who know these companies best. Sell on day one.

For the broader macro context on the AI capital cycle, see Coatue’s 18-chart AI report, where VC money is going in AI, and Anthropic is closing in on a $1 trillion valuation.

10. The one career skill AI cannot average its way into.

AI fluency is commoditizing. Relationships are not.

“The only competence that’s really important is storytelling and relationships.”

When Google posts a job, 200 resumes arrive within 60 minutes. The listing closes. Of the top 10 candidates brought in, 70% of the time the person hired had an internal advocate. Not the strongest resume. Not the most AI-proficient. A relationship inside the building.

AI drives output toward the median. The literal mechanism is averaging the previous words to predict the next one. Everything AI produces is a weighted mean of existing thought.

AI cannot produce the person who sends a text saying you were exceptional today. It cannot manufacture the advocate in the room when your name comes up.

Galloway’s four-part framework for anyone under 40:

1️⃣ Be as social as possible. Out of the house, not on a screen. Relationships compound like money.

2️⃣ Invest in storytelling and writing. If someone reads your work and says it sounds like AI, that is the worst professional insult available right now.

3️⃣ Get comfortable with rejection at volume. More no’s now means more advocates later.

4️⃣ Find your group before you need it. The person who gets the job had a friend in the room before the job existed.

Everyone is training to use AI better. Almost no one is training to be more human. The median is crowded. The internal advocate is not.

For the writing voice that distinguishes you from AI output, see your voice is the only AI moat that compounds. For the personal knowledge stack that makes your work sound like yours, see I built a $0 second brain with Obsidian + Claude that compounds for life and I built a second brain in 10 minutes with Granola + Claude.

Full podcast:

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

The AI repricing playbook

The core argument is not that AI fails. AI is mispriced, misbilled, and misunderstood simultaneously, at historic scale.

For founders

The AI cost conversation is coming to your board whether you invite it or not. Map every dollar of AI spend to a specific outcome a CFO can name. Produce one number this week that a board member can connect to a result.

▫️ The SaaS defense playbook for the AI era

▫️ 50 game-changing AI agent startup ideas for 2026

▫️ The AI GTM playbook for 2026

▫️ 70 startup ideas YC wants you to build

▫️ The Claude Code system that replaces a 5-person team

For investors

Separate the technology from the valuation. Unprecedented fundamentals at insane valuations still produce bad returns. Model the technology on its merits. Price it on evidence, instead of narrative.

▫️ What top-tier VCs actually look for in 2026

▫️ Coatue’s 18-chart AI report

▫️ Where VC money is going in AI

▫️ The U.S. VC database most founders never find

▫️ The full investor lists archive

For operators and people in tech

The cheapest route to irrelevance is doubling down on AI fluency alone. AI fluency is commoditizing in real time. Write in a voice that sounds like you. Build one internal advocate this month, before you need it.

▫️ Your voice is the only AI moat that compounds

▫️ The single best productivity decision you can make with Claude right now

▫️ I built a $0 second brain with Obsidian + Claude

▫️ Build your own stock analyst with Claude

▫️ Why ChatGPT and Claude keep disappointing you

The 5 principles to take from Galloway

1️⃣ 95% of AI projects connect to no return a CFO can name. That number will move. Watch when it does.

2️⃣ Valuations correct before labor markets collapse. The apocalypse is not on schedule.

3️⃣ China’s pricing advantage is structural. Policy is the likely U.S. response. The 90-day clock is running.

4️⃣ The IPO window is insiders exiting at peak valuation. Sell on day one.

5️⃣ Relationships compound. Start before you think you need to.

The technology survives the valuation. It always has. The question is what you paid to hold it while the market caught up.

If this saved you 2 hours, share it with one founder or investor who needs to see it.

The AI Corner

Discussion about this post

Ready for more?