Two Labs Are About to Own the World's Compute — and Your Business Is Already a Tenant

Two Labs Are About to Own the World's Compute — and Your Business Is Already a Tenant

Let me tell you something that's been sitting with me since yesterday. One of the smartest people in the AI hardware game sat down with Dwarkesh Patel — the guy who interviews everyone who matters — and laid out a future that should scare the hell out of anyone whose business depends on AI. Which, let's be honest, is most of us now.

His name is Dylan Patel. He runs SemiAnalysis, the research firm that's been right about the AI buildout more times than any bank I can name. And his message was simple: Anthropic and OpenAI are on track to control most of the world's compute by 2028. Not a lot of it. Most of it. And if you think that doesn't affect you because you're not in the AI business, you're wrong. You're already a tenant. You just don't know it yet.

The Setup — What the SemiAnalysis Founder Just Said

Let me give you the numbers, because I don't do vibes. At the start of this year, OpenAI had about 2 gigawatts of compute running. Anthropic had less than 2. By the end of this year, both of them are above 5 gigawatts. That's a 3-4x increase in twelve months.

Here's the part that matters. Roughly a third of all new compute coming online this year is going to those two labs. Next year, based on contracts already signed, it's 40 to 50 percent. By the end of 2027, half of the world's incremental new compute goes to just two companies. And by the end of 2028 — if nothing changes, and Dylan says he sees nothing stopping it — they control most of the usable compute on the planet.

Let that sink in. World compute roughly doubles every year. Frontier lab compute triples. Two companies outrunning the entire planet, every single year, forever.

Why the Duopoly Wins — The Economics Nobody's Arguing With

You might think this is some regulatory capture conspiracy. It's not. It's colder than that. It's arithmetic.

Here's the core economics. A megawatt of compute capacity costs $10 million to $15 million. When OpenAI served GPT-4 on Hopper GPUs, that math was negative — they lost money on every token. Now? When OpenAI serves GPT-5.6 or Anthropic serves Opus 5 or Fable 5, revenue per megawatt blows way past the cost. Anthropic has seen revenue as high as $50 million per megawatt. Spend $10 on inference, generate $50, reinvest the profit into training. That's the flywheel.

That's why Anthropic turned a profit in Q2. That's why OpenAI is expected to turn a profit in Q3, powered by Codex and the 5.6 line. And that's the whole game: the labs monetize compute better than anyone else, so they can outbid anyone else. Every time a bank, a hedge fund, an AI company, or a hosting provider tries to buy compute, they're bidding against two companies with the deepest pockets in the history of capitalism. Guess who wins. Ent?

And because they outbid everyone, compute prices go up. That's not a prediction — it's the mechanism. When two buyers can pay more per megawatt than everyone else combined, the price of every megawatt follows them.

The Two Readings — A Rational Flywheel, or a Duopoly With a Balance Sheet

I've been running hosting infrastructure for over a decade, and I've learned to hold two readings of this industry in my head at the same time. Both are true. That's the uncomfortable part.

Reading one: capitalism working exactly as designed. They ran the fab-side numbers. About $6 billion of fab capex produces a gigawatt of compute every year, and that gigawatt generates around $100 billion in revenue a year. Over five years, $6 billion at the fab level creates over a trillion dollars in end AI revenue. Even after you take away half for the middlemen — data centers, power, R&D — you've got a 100x gap between fab capex and end revenue. When you can turn a dollar into a hundred, you build. That's not a bubble. That's a stampede.

Reading two: the same flywheel is a duopoly forming in real time. Two companies on track to own most of the world's compute. Two companies whose models increasingly are the workforce — effective AI labor at the frontier is growing 10x a year, from millions of AI workers this year to hundreds of millions next to a billion the year after. Two companies that, by the end of the decade, could each have more effective population than there are people on Earth.

I don't care how you feel about the singularity. As a business owner, I care about single points of failure. And this is the mother of all single points of failure. If your company's intelligence runs through two APIs, your company's intelligence is a lease. And the landlord has no competition.

The Secondary Bottleneck Nobody's Talking About — The $5 Trillion Credit Wedge

Here's the part that should worry you more than the compute numbers. Every force in this industry is screaming toward centralization, but one force could bend the curve: the credit market.

SemiAnalysis models about $11 trillion in AI capex from 2024 to 2029. Maybe $6 trillion gets funded by cash flow. The rest — north of $5 trillion — has to be borrowed. That's $5 trillion of new credit into a world economy already drowning in debt. And when you raise that much money, interest rates move. Meta recently raised debt at 5 to 6 percent. Dylan's point: they'd happily pay 8, because the compute returns far more. But when hyperscalers pay 8 percent, everyone else pays 250 basis points more. Banks scream — their credit costs reprice faster than their assets. And every non-AI equity gets crushed on the discount rate.

This is the second Volcker shock argument. In the 1980s, Paul Volcker raised rates and something like 40 countries, mostly in Latin America, defaulted. Dylan's version: countries with heavy debt and thin tax revenue — Pakistan, Nigeria — get wrecked in a rising-rate world. He didn't mince words, and neither will I: the AI buildout's financing could take down economies that have nothing to do with AI.

So the real governor on this buildout isn't a competing lab. It's the bond market. It's regulators. It's communities saying no to data centers. Dylan said it himself: the limiter on AI isn't just research — it's how much the rest of the world lets the buildout happen. Interest rates, regulations, backlash. Those bend the curve. And right now, that's the only thing standing between us and the pure concentration math.

What This Means for Independent Hosting Providers

If you run an independent hosting business — and I know a lot of you reading this do — here's what I'd do today:

First, stop betting your business on a single lab's API. The moment your stack depends on one frontier model provider, your margin is their pricing power. Build multi-model redundancy into everything you sell. When the duopoly raises prices — and they will, because they can — the providers who can route around them keep their customers.

Second, lock your compute and power costs now. If two labs are outbidding the world for megawatts, every GPU lease, every colo contract, every power agreement in your portfolio reprices. Lock terms early. The window is closing.

Third, serve the long tail the labs don't care about. The labs want frontier megawatts for frontier models. They don't want the boring middle — regulated industries, mid-market inference, edge workloads, the 99 percent of businesses that need AI that's good enough and private enough. That's your market. That's where independent hosting always survives.

Fourth, position as the neutral ground. When two companies control the compute and the workforce, trust and independence become a product feature. Your customers are starting to fear the duopoly — that's your moat. Be the operator who isn't owned by anyone.

Fifth, watch the credit market like it's your own balance sheet. The $5 trillion debt wedge is the thing most likely to bend this curve. When hyperscaler borrowing starts moving rates, it moves your financing costs too. Model your pricing against 8 percent money, because you may be living with it.

The Bottom Line

I've been in this industry long enough to watch consolidation stories come and go. This one is different. This isn't two companies merging. It's two companies on track to own the world's compute, the world's models, and increasingly the world's workforce — with economics so brutal nobody can compete with them.

The only counterweight isn't a competitor. It's the credit market and the political system, and both are already pushing back. As a founder, I don't bet on either being rational. I plan for the world where the duopoly wins, and I make sure my business can survive as a tenant — while owning the one thing the labs can't: the relationship with the customer.

Because in a world where two companies own the machines, the people who own the customers are the only independents left.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Dwarkesh Podcast — "Dylan Patel – Two labs will soon control most of the world's workforce" (Aug 25, 2026); SemiAnalysis.