Frontier labs are growing their compute 3x a year, while the rest of the world grows at 2x.
Run that gap forward 2 years and you get a sentence that should stop you cold:
“You’ve got them just controlling most of the usable flops in the world on their own.”
Not a slice of the world’s compute, but most of it, held by 2 companies.
Dylan Patel runs SemiAnalysis, which tracks lab compute, chip supply, and data center buildouts closer than almost anyone outside the labs themselves. He sat down with Dwarkesh Patel for the most honest read yet on what AI is doing to the economy.
I watched the full interview so you can skip it.
Here are the 10 takeaways that matter.
Together with Hard Skill Exchange:
If Patel is right, the compute layer is settled and out of your hands. What stays in your hands is everything you build on top of it: your data, your workflows, your decision rights, your agents.
That’s the sovereign AI question, and it’s the premise of the Agentic Harness Summit, Sept 8-10:
“Today, most companies are built on business processes. In the future, most companies will be built on harnesses.” Jensen Huang
3 days on how enterprises secure agent identity and actions while keeping their own business alpha, with James Currier (NFX), Mike Maples Jr. (Floodgate), Wade Foster (Zapier), and 100+ CXOs, VCs, and analysts.
Free and virtual.
1. 2 labs are on pace to control most of the world's usable compute by 2028
“By the time you’re towards the end of 2028, if this trend continues, which I see nothing that’s stopping it, you’ve got them just controlling most of the usable flops in the world on their own.”
Patel’s math starts simple. World compute in gigawatts roughly doubles every year, while compute at the 2 frontier labs grows 3x, so the frontier goes from 2 gigawatts at the start of this year to 6 by year end, 18 by the end of 2027, and 54 by the end of 2028.
Each new watt is also more efficient than the last, since GB300, TPU v7, and Trainium 3 chips deliver 3x to 5x more performance per watt than the generation before them.
“A new watt deployed this year is significantly more efficient than the watts deployed two years ago.”
Stack the growth rate on top of the efficiency gain and the compounding turns 2 labs into the dominant buyer of the planet’s usable compute inside 3 years, over the 10 most forecasts assume.
2. 1 lab could employ more workers than exist on Earth before 2030
“Then pretty soon, even if compute scaling slows down, it doesn’t take many more years before each company individually has more labor equivalents than there are people on Earth.”
Frontier compute in FLOP terms is growing 4x to 5x a year, while the compute needed to hit a given capability level falls roughly 3x a year. Multiply the 2 curves and the effective AI population at OpenAI and Anthropic compounds 10x year over year, before any recursive self-improvement enters the picture.
Patel’s estimate: OpenAI goes from roughly 10 million AI laborers this year, to 100 million next year, to 1 billion the year after. Even if scaling slows later, the curve crosses 8 billion, the population of the planet, within a few more years.
That turns a compute story into a labor story concentrated inside 2 companies, which Patel argues the world underrates relative to how much attention nationalization and state power get. If those systems end up misaligned, most of the world’s effective labor is misaligned with them.
3. The business flipped from losing money to $50 million a megawatt in under 2 years
“Anthropic started turning a profit in Q2. It’s believed at some point in Q3, OpenAI could potentially start turning a profit even with the big rise of Codex and 5.6 and all this. But if we go back a year ago, everything that they, all the money they had was venture funded losses, right?”
Serving GPT-4 on Nvidia Hopper chips generated negative gross margin for OpenAI, and base compute still costs $10 million to $15 million per megawatt. Anthropic’s revenue per megawatt has since climbed to $50 million.
Spend $10 on inference capacity, generate $50 of revenue, then plow the profit back into training. That flywheel funds the entire buildout, and it exists only because the gap between compute cost and compute revenue widened this fast.
The results already show it: Anthropic turned profitable in Q2, and Patel expects OpenAI to follow in Q3 even while absorbing the cost of scaling Codex and GPT-5.6, a swing from venture-funded losses that were still running at the start of this year.
4. Why Labs Would Rather Build AGI Than Bank $70 Million a Megawatt
“If you now get to generating $60, $70 million per megawatt, do you still allocate 40% to inference and generate all this profit and then do dividends and share buybacks? Or do you go build AGI? And I think the obvious answer from Anthropic and OpenAI, and not just at the executive level, but also their board, is go build AGI because it’s way more profitable.”
Labs have historically split compute roughly 60% training and 40% inference, and Patel expects that split to keep tilting toward training, a view he flags as non-consensus.
The standard assumption is that most new compute eventually serves paying users. Patel’s read runs the other way: as revenue per megawatt climbs, the labs would rather generate the same profit off a smaller inference footprint and pour the freed-up compute into research.
Every megawatt spent on inference today is a megawatt not spent compounding next year’s model.
That decision sits at board level, over engineering, and Patel says the shift is already measurable, with Anthropic’s incremental compute skewing toward research for 3 straight months while older capacity kept revenue climbing.
5. A $6 billion fab can generate more than $1 trillion
“$6 billion of CapEx at the fab level will have generated over a trillion dollars of end AI revenue.”
The wafer math behind a single gigawatt is almost absurd. 1 gigawatt of Vera Rubin-class compute needs roughly 55,000 N3 wafers, 6,000 N5 wafers, and 170,000 DRAM wafers. Tooling to produce that gigawatt costs $3 billion to $4 billion a year, and adding clean rooms and shell brings the total near $6 billion.
That $6 billion produces 1 gigawatt every year, and that gigawatt currently produces $100 billion of revenue every year, for roughly 5 years before it needs replacing. Cut the number in half for every middleman along the chain and the ratio between fab capex and downstream AI revenue still runs over 100x.
If the wafer math changes how you think about the chip supply chain, these go deeper:
▫️ Cerebras bet against the GPU. It just IPO’d at $56B
▫️ The AI trade just lost $1 trillion. Here is where the value went
▫️ Jonathan Ross turned a 3-week phone call into a $20 billion NVIDIA deal
6. The Actual Price Tag on AI Buildout Is $7 Trillion to $10 Trillion a Year (Not the Number You’ve Heard)
The figure everyone quotes for AI capex is missing most of the bill.
“If you end up in a world where you're doing 100 gigawatts a year, at current prices that would be 5 trillion of CapEx every single year. And then stack on the fact that you have to build the power plants way before then, slash it's a 30-year asset. You stack on the fact that the data centers are a, you know, 15, 20-year asset and you have to build that then too. So the 5 trillion, you know, you have to account for future years growth. So it's actually going to be more like 7 or 10 trillion. Of CapEx.”
At 100 gigawatts a year, current pricing puts capex near $5 trillion annually, and that figure only covers what Patel calls “critical IT”: servers, networking, fiber, and optical communications.
It skips the power plants, 30-year assets built years ahead of demand, and it skips the data centers themselves, 15 to 20 year assets that also have to be built ahead of the compute they will house.
Stack those together and Patel’s number lands between $7 trillion and $10 trillion a year by the end of the decade, close to a 10th of the world economy.
7. Financing the $11 Trillion Buildout Takes $5 Trillion of Debt
Hyperscalers can’t fund this alone anymore.
“The total, let's say the big... The hyperscalers in total will rate and all the clouds. In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029. Total. And if you do, you know, if you fund a lot of this with cash flows and as much as you can, you still end up with north of $5 trillion of credit that needs to be issued for this $11 trillion plus...”
Google, Microsoft, Amazon, and Meta funded most of AI’s buildout so far. All 4 now spend everything they generate on capex and still raise debt on top of it.
Patel’s model puts total buildout costs at $11 trillion through 2029, split roughly $6 trillion cash and $5 trillion credit, a debt share closer to 45 percent than a clean half.
That debt has to come from somewhere: semiconductor companies, traditional infrastructure investors, and eventually ordinary capital that would otherwise buy homes, government bonds, or Anthropic’s own corporate debt.
If you’re trying to price your own exposure to that debt wave, these go deeper:
▫️ Rate Cuts, 80+ Ways to Fund Without Diluting, AI Is Breaking Valuation Models
▫️ The Venture Capital Liquidity Crisis That Nobody Is Talking About
▫️ Scott Galloway: 95% of enterprise AI spend connects to no return a CFO can name
8. The Original Volcker Shock Broke 40 Countries. Patel Bets on a Sequel
The 1980s debt crisis has a modern echo, and it starts with a data center.
“So in the 80s, to fight inflation, Fed Chair Paul Volcker raised interest rates like more than 5%, or it's like something like 8%, real interest rates 8%. And that caused some 40 different countries, mostly Latin America, to default in that decade. And I think that will probably happen again.”
That 40-country toll is history, not forecast; Patel expects a repeat but doesn’t size it. He credits the framing to economist Basil Halperin, whose paper with Trevor Chow and J. Zachary Mazlish, “Transformative AI, Existential Risk, and Real Interest Rates,” ties AI growth expectations to long-term real rates.
The 8 percent figure Patel cites is his recollection of how far Volcker pushed real rates, not a forecast for the new shock’s size. It checks out: the Fed funds rate peaked at 19.1 percent in June 1981.
If Meta is raising debt at 5 to 6 percent today, Patel doesn’t see why it wouldn’t pay 8 percent. A 250 basis point jump ripples through the market, raising the discount rate on every company’s cash flows, so stable, low-growth stocks take the hit even if the S&P 500 looks fine.
“Every stock that is not an AI stock is worth basically zero, because discounted cash flows are worth nothing.”
Developing countries with heavy debt loads and thin tax bases, Pakistan and Nigeria among them, sit most exposed to that repricing, a point Patel’s broader read on where AI is closer than people think society is ready for returns to often.
9. The 2 Forces That Make the AI Race Centralize (Even Without a Single Breakthrough)
2 forces keep pushing the gains toward whoever is already ahead.
“So that's like one effect. The other effect is if you're slightly ahead in the AI race and compute is in shortage, you can charge a much higher markup because you can better economize this scarce resource. So there's like two effects, which are give more and more to the person who's like ahead in the AI race.”
Training a skill into a model is a fixed cost, amortized across billions of sessions, which rewards whoever already has the most users.
Scarcity does the rest. When compute is short and one lab is ahead, that lab can charge a premium simply because it can better ration a resource everyone else needs too.
That’s also the argument behind why some investors say the model was never the moat, and why the smartest reads on AI Corner increasingly conclude the actual moat lives somewhere else entirely: continual learning adds a third multiplier, since the model deployed to more people learns from more live usage, widening the gap again, even without recursive self-improvement in the picture at all.
10. Why a 6-Month Delay Could Matter More Than Physics
Society, not silicon, is turning out to be the actual limiter here.
“I mean, right now, at least more revenue because the models aren't capable of RSI. But I'm worried about a world where it's 2030 and the government's like, we're going to wait six months before you can release your model to the public.”
Patel already sees regulation biting harder than most assume. OpenAI has paused training runs for stretches, and Anthropic has reportedly restricted access to Model 2, widely believed to be the next version of Mythos, for some employees, echoing a documented episode where the US government switched off Anthropic’s most powerful model days after launch.
If governments keep labs from releasing or fully using their best models, revenue per megawatt stalls, and so does each lab’s ability to outbid everyone else for compute.
Patel’s other fear runs the opposite direction. A 6-month external release delay could let a lab run years of unseen internal progress, widening the gap between what the public gets and what the lab actually has.
If the regulation angle above changes your timeline, start with these:
▫️ Dario Amodei named “the zeroth world.” You probably live in it.
▫️ What If the Real AI Risk Isn’t Superintelligence?
▫️ The Real Reason AI Costs Keep Rising
The compute centralization playbook
2 labs are on pace to out-compute and out-employ most of the planet by 2028, and the open question has moved from whether it’s possible to whether the rest of the world lets it happen.
▫️ Founders: building on frontier models means building on infrastructure 2 companies increasingly ration by price. Plan for compute costs to keep climbing, and build pricing power into your own margins now.
▫️ Investors: the value-capture question is unsettled. Jane Street reportedly extracts more value per megawatt than Anthropic does. Watch where the spread between compute cost and end value lands before betting on any single layer of the stack.
▫️ Operators: if your team isn’t testing internal workflows against the newest checkpoints, start this week. What you’re benchmarking against may already be a generation behind what the labs run privately.
▫️ Everyone else: interest rates, data center siting, and AI regulation stopped being side issues. Watch the permitting fights and the financing terms, over the model releases.
The 5 principles to steal
Revenue per megawatt over hype. It’s the cleanest signal on lab strength, and that number moved from negative to $50 million in under 2 years.
Training now, inference later. Inference funds the flywheel, and the labs believe the bigger prize sits in the next model over this quarter’s serving margin.
Follow the debt markets. An $11 trillion buildout never gets funded on cash flow alone, and where that credit comes from will reshape borrowing costs for everyone.
Physics stopped being the limit. Mirrors for EUV tools and turbines for power plants move slowly, and politics moves slower still, or stops things outright.
Expect 2 price-setters. Whatever you build on frontier models, plan for a market where your supplier gains pricing power every year.
This trajectory hinges on how high revenue per megawatt climbs and how much regulatory room the labs get. 2 labs already look like the market, and the rest of the decade decides how much of it they own.
If this breakdown saved you an hour of trying to model this yourself, send it to one founder or investor who needs it.
Keep reading
The money behind the compute
▫️ Anthropic just passed OpenAI in revenue, spending 4x less
▫️ Sam Altman on where AI is actually going
▫️ Dario Amodei and the long game of safe AI
The people making the bet
▫️ How Jensen Huang turned a green van and 30 years of rejection into Nvidia
▫️ Dario Amodei’s full picture: 10 takeaways that matter
▫️ Jensen Huang: 10 lessons from the CEO building the most important company in history
Where the bet could break
▫️ The venture capital liquidity crisis nobody is talking about
▫️ The US government switched off Anthropic’s most powerful model 3 days after launch
I dont understand, the frontier labs barely own any data centers, how are they securing 3x compute versus everyone else 2x? If they keep spending on training with ever decreasing marginal revenue on new models they go bankrupt and hyperscalers, the true owners of compute will lease it out elsewhere or just lease it out to the highest bidder.
This doesn’t matter if there is 100x more compute. Its a Pareto improvement across the board.