2 Companies Will Control Most of… | The AI Corner Summary

1. Key Themes

Compute is centralizing into a duopoly

Frontier labs are growing compute at 3x/year vs. the world's 2x/year, and running that gap forward produces near-total concentration.

"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."

The economics flipped from burning cash to a self-funding flywheel

Labs went from negative gross margins to massive profitability per unit of compute in under two years, which now funds further scaling.

"Anthropic started turning a profit in Q2... if we go back a year ago, everything that they, all the money they had was venture funded losses." Revenue per megawatt has reportedly hit "$50 million," against base compute costs of "$10 million to $15 million per megawatt."

The real AI buildout cost is far larger than headline capex figures

The commonly cited numbers only capture "critical IT" and ignore power plants and data center shells.

"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."

Debt, not equity or cash flow, will finance the buildout — with macro consequences

Hyperscalers can't self-fund an $11 trillion buildout, forcing massive credit issuance that could ripple through rates and emerging-market debt.

"You still end up with north of $5 trillion of credit that needs to be issued for this $11 trillion plus..." "So in the 80s... Fed Chair Paul Volcker raised interest rates... And that caused some 40 different countries... to default in that decade. And I think that will probably happen again."

Centralization is structural, not just a product of a single breakthrough

Fixed costs amortized over huge user bases, compute scarcity pricing power, and continual learning all compound advantages for whoever is already ahead.

"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."

2. Contrarian Perspectives

Labs will choose to under-monetize inference in favor of training

Conventional wisdom assumes rising revenue per megawatt gets reinvested into serving more paying customers. Patel argues the opposite: labs will intentionally shrink inference share to funnel more compute into research, because building AGI is judged more profitable than harvesting profit today.

"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?... 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."

Regulation, not physics or compute supply, is the real ceiling on AI progress

Rather than chip shortages or power constraints, Patel worries governments pausing or restricting model releases could be the binding constraint by the end of the decade — and a short external delay could mask years of hidden internal capability gains.

"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."

The labor concentration story is underrated relative to nation-state framing

Most public debate focuses on state power and nationalization risk, but Patel argues the more urgent and underappreciated risk is that effective AI "labor" — potentially exceeding Earth's population — will be concentrated inside just two private companies.

"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."

3. Companies Identified

OpenAI — Frontier AI lab. Cited as one of the two labs on pace to dominate world compute and labor equivalents; expected to turn profitable in Q3. Quote: "OpenAI could potentially start turning a profit even with the big rise of Codex and 5.6."
Anthropic — Frontier AI lab. The other of the two dominant labs; already profitable in Q2, revenue per megawatt reached $50M, reportedly restricted access to its next model internally. Quote: "Anthropic started turning a profit in Q2."
SemiAnalysis — Research firm tracking chip supply and data center buildouts, run by Dylan Patel. Source of the core analysis in the article. Quote: "[Dylan Patel] runs SemiAnalysis, which tracks lab compute, chip supply, and data center buildouts closer than almost anyone outside the labs themselves."
Nvidia — Chip maker referenced via GB300 chip generation and historical company profile. Mentioned for efficiency gains driving compute growth ("GB300... deliver 3x to 5x more performance per watt than the generation before").
Google, Microsoft, Amazon, Meta — Hyperscalers funding the AI buildout, now spending all generated cash plus raising debt. Quote: "All 4 now spend everything they generate on capex and still raise debt on top of it."
Cerebras — Chip company mentioned as a case study of an alternative bet against GPUs, cited as a related read: "Cerebras bet against the GPU. It just IPO'd at $56B."
Jane Street — Cited as a value-capture benchmark, extracting more value per unit of compute than AI labs themselves. Quote: "Jane Street reportedly extracts more value per megawatt than Anthropic does."
Zapier — Referenced via founder Wade Foster as a summit speaker on agentic infrastructure (sponsor content, tangential to core analysis).

4. People Identified

Dylan Patel — Founder/analyst at SemiAnalysis. Central figure of the interview; source of nearly all quantitative claims on compute growth, financing, and centralization. Quote: "You've got them just controlling most of the usable flops in the world on their own."
Dwarkesh Patel — Podcast host who conducted the interview. Described as producing "the most honest read yet on what AI is doing to the economy."
Paul Volcker — Former Fed Chair, referenced historically for 1980s rate hikes that triggered developing-world defaults, used as an analogy for future rate risk. Quote: "Fed Chair Paul Volcker raised interest rates like more than 5%... real interest rates 8%."
Basil Halperin (with Trevor Chow and J. Zachary Mazlish) — Economist whose paper connects AI growth expectations to real interest rates; credited as the source of Patel's Volcker framing. Referenced via "Transformative AI, Existential Risk, and Real Interest Rates."
Jensen Huang — Nvidia CEO, quoted in the sponsor section on the shift from business processes to "harnesses." Quote: "Today, most companies are built on business processes. In the future, most companies will be built on harnesses."
James Currier (NFX) and Mike Maples Jr. (Floodgate) — VCs mentioned as speakers at the sponsored Agentic Harness Summit (promotional context, not core analysis).

5. Operating Insights

Track revenue-per-megawatt as the cleanest strength signal. The article frames this metric — not user counts or funding rounds — as the key indicator of lab health: "It's the cleanest signal on lab strength, and that number moved from negative to $50 million in under 2 years."
Benchmark against private, not public, model checkpoints. Teams risk building against stale capability baselines: "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."
Build pricing power into margins now, anticipating rising compute costs. Founders building on frontier models are building on infrastructure controlled by price-setting oligopolists: "Plan for compute costs to keep climbing, and build pricing power into your own margins now."

6. Overlooked Insights

The wafer-level economics of a single gigawatt are staggeringly favorable, implying chip/fab capacity — not just model capability — is the actual chokepoint of the entire industry: "1 gigawatt of Vera Rubin-class compute needs roughly 55,000 N3 wafers, 6,000 N5 wafers, and 170,000 DRAM wafers... $6 billion of CapEx at the fab level will have generated over a trillion dollars of end AI revenue."
Continual learning may be a bigger moat than the model itself, since a model deployed to more users learns from more live usage in a self-reinforcing loop independent of recursive self-improvement — a claim buried in point 9 but potentially more durable than the compute/capital arguments driving the rest of the piece: "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."