# "Two Companies Will Own Most of the World's Compute" — Notes on Dylan Patel's 2028 Map > Listening notes on the Dwarkesh Podcast conversation with SemiAnalysis founder Dylan Patel: how compute is concentrating into two labs, why revenue-per-megawatt is the ruler that matters, and how $5 trillion of credit reaches every stock you own. Educational; not investment advice, no tickers, no price targets. Published: 2026-08-26 Locale: en Tags: AI compute, capex, interest rates, bottlenecks, investing frameworks ![A night-time data center construction site: a crated precision optical component sits under a work lamp on a truck bed, with half-built halls and transmission towers receding into haze behind it](/covers/dwarkesh-2026-08-25-dylan-patel-anthropic-openai-will-have-most-of-the-cover.png) > Wealth lies in method, not in toil; profit lies in position, not in ploughing harder. > > — Huan Kuan, *Discourses on Salt and Iron*, "Tong You" (Western Han, c. 81 BCE; my translation) ## What This Episode Is About The 25 August 2026 episode of the Dwarkesh Podcast brings back Dylan Patel of SemiAnalysis. The two open by joking that their once-a-year recording has become a kind of family Thanksgiving dinner — "but we're not actually related. Don't tell the people this, it will destroy the myth." The dinner conversation is about one hard question: where compute is going. It starts with this year's actual numbers, runs out to 2028 and 2029, and then takes a turn I did not expect — into interest rates, sovereign debt, and why a company that makes medical devices or runs a railway might get cheaper because someone else is building a data center. What I took away wasn't the scary numbers. It was a clean demonstration of how to reason: find a unit you can measure, trace how that unit changes everyone's incentives, then say out loud which things would break the whole chain. What follows is my own synthesis, not a transcript. The figures and judgments come from the conversation; the extensions are mine. ## The Points Worth Keeping **1. The model layer just flipped from losing money to making it — and the ruler is revenue per megawatt.** The unit used throughout is unusual and useful: not revenue growth, but how much revenue one megawatt of compute generates per year. The base cost of compute runs roughly $10–15 million per megawatt. A year and a half ago the tokens sold off that megawatt were worth less than that — negative gross margin, funded by venture capital. That has inverted. Frontier labs now generate far more than cost per megawatt; Anthropic is described as having reached the $50 million range. Turning $10 into $50 is the engine behind everything else in the episode. **2. Concentration is accelerating, and effective compute is more concentrated than the headline share.** To name the "two companies" in the title outright: Anthropic and OpenAI. Roughly a third of this year's incremental compute goes to these two labs. Next year, based on contracts already signed, it's 40–50%. By 2028 the projection is 70–80%. There's a multiplier that's easy to miss: each new chip generation delivers three to five times the performance per watt of the last. So taking "half the new watts" means taking considerably more than half of the usable compute. The percentage understates the reality. **3. A genuinely non-consensus call: the inference share will fall, not rise.** Most people assume the future is mostly inference — compute deployed to serve customers and earn revenue. Patel argues the opposite. If a megawatt earns $70 million on inference, a board still has to choose between paying that out and spending it on research toward something bigger. He thinks that choice has an obvious answer, and that it's already happening: monthly compute additions keep climbing while revenue additions have plateaued. The machines didn't disappear. They changed jobs. **4. The real bottleneck sits at the very top of the chain, and the signal takes years to travel.** My favourite passage. The labs are effectively saying: we could make a trillion dollars right now, and we're bottlenecked on the mirrors that go into ASML machines. Mirrors. Crack the whip and the tip moves late — Carl Zeiss didn't think it needed that scale of EUV output six months ago and now does, and by the underlying economics it should probably be more still. Patel throws out a half-joking arbitrage: if you had $400 million and could persuade ASML to sell you a tool, buy it, wait a year, and sell it for north of a billion. People are already doing exactly this with gas turbines. **5. Where the money comes from: about $11 trillion of capex through 2029, roughly $5 trillion of it borrowed.** The model in the episode is $6 trillion funded from cash flow and $5 trillion from credit. One detail gets consistently overlooked: when people say "AI capex" they usually mean servers, networking and optics. That excludes the buildings and the power plants — and the power plants have to be built years ahead and are thirty-year assets. The real capital requirement is larger than the already absurd number. **6. Interest rates are the brake, and they press on people who have nothing to do with AI.** If Meta will happily borrow at 8% because the returns on the compute are enormous, everyone else in the economy pays more too. That's crowding out: governments, consumers, telecoms, banks and consumer goods companies all draw from the same pool. The episode cites an economist friend's view that this could be a second Volcker shock — the 1980s rate rise that pushed some forty countries, mostly in Latin America, into default. **7. Most of the value currently sits with users — and the person saying so doesn't quite believe it lasts.** Jane Street makes far more from the tokens it buys than the model company makes selling them. So does Meta, optimizing ad algorithms. Patel calls this the one saving grace — and then immediately labels it: "this is my cope." Because the logic he laid out earlier (tokens are worth more used internally) eventually reclaims those gains. It is rare for a forecaster to point at the exact spot where his argument collides with his hope. ## Going Further ### "The news is full of trillion-dollar capex. Should I be buying this stuff?" The episode doesn't answer that, and it gives you something more useful instead: a map of how the rent-collecting position moves. Look at where the margin has actually sat over three years. Around 2023, memory makers were delivering enormous value and capturing almost none of it, while chips and foundry took most of the gross margin, the model companies sold tokens below cost, and hyperscalers built infrastructure without knowing whether it would pay. Today the model layer has flipped to large positive margins, memory has become the fastest and most aggressive price riser, and foundry raises prices only slowly. So the question isn't "will AI do well" — the episode treats that as a premise. The question is: **which layer will the money sit in next?** The test offered here is practical: who can say no when demand doubles? Turbines get flipped for profit because they gate the data center. An EUV tool is worth holding for a year because mirror capacity cannot be conjured in twelve months. Foundry raises prices slowly precisely because its expansion path is known and already underway. One layer deeper: **being sold out is not the same as having pricing power.** A supplier can be overwhelmed with orders and earn nothing, as long as its capacity can be rebuilt within a year or two — customers will simply wait rather than bid. The rent goes to whatever cannot be expanded and cannot be routed around. The way to check isn't reading about shortages; it's checking whether gross margin rose along with the shortage. If it didn't, that company is merely busy. The failure conditions matter too. This map misleads in two cases: if demand growth actually stops, the upstream bottleneck falls hardest, because its order visibility rests entirely on customers' expansion plans; and if a bottleneck gets designed around — a process change, an architecture change, or buyers building it themselves. Historically that happens more often than people expect. ### "I don't own any of these companies. Why should I care?" You should, and the transmission path isn't the one most people picture. It doesn't run through "the AI bubble pops and drags the market down." It runs through the denominator. A company valued on stable cash flows — the episode names the Johnson & Johnson and railway archetypes — is worth the next thirty years of cash discounted back to today. The discount rate is the denominator. Roughly: if cash flows grow 3% a year and you require 5%, the business is worth about fifty times cash flow. Require 8% instead and the same cash flows are worth twenty times. **Nothing about the company changed and 60% of the value is gone.** That is the counterintuitive core of the episode. If AI capex genuinely requires $5 trillion of credit, it raises the cost of capital not just in its own supply chain but everywhere — and the assets most exposed are the ones most dependent on distant cash flows. The steadiest, safest-looking holdings are the most denominator-sensitive. There's a blunt line in the conversation to the same effect: if you truly believe AI reshapes the economy, then everything in the economy should trade at two or three times earnings — including the memory stocks that currently look cheap. The practical takeaway for an ordinary investor: **don't treat "unrelated to AI" as shelter.** Whether an asset is exposed depends not only on its revenue but on how much of its valuation rests on long-dated discounting. Mortgages, corporate credit and emerging-market sovereign debt all run through the same pipe. This is a causal chain, not a prediction — whether it happens depends on whether that $5 trillion actually gets borrowed. ### "He sounds very sure. How much of this should I believe?" This section isn't about whether the episode is right. It's about how to read forecasts of this kind. Notice a small move: pressed on global compute growth, Patel adds that the number is the upper bound — "the I'm-so-bullish version." Labelling your own figure as a ceiling rather than a midpoint is a completely different kind of content. Asked about 2029 and beyond, he says plainly that anything more than four years out is hard to predict. My approach is to split any forecast into two parts: **mechanism** and **extrapolation**. Mechanism travels. Revenue per megawatt exceeding cost per megawatt produces bidding for compute. Bidding hands pricing power to whoever holds unsold compute built on their own balance sheet. Upstream expansion has a physical time floor, so price increases propagate through the chain segment by segment rather than all at once. These are checkable causal claims; take them to next quarter's earnings and you learn something either way. Extrapolation does not travel. Triple, triple, triple, therefore fifty-four gigawatts in 2028. Numbers like that help you feel the scale; they are not a plan. The episode itself names three things that would break the chain: regulation preventing labs from releasing their best models (which caps revenue growth and therefore their ability to outbid), whether credit markets supply the $5 trillion, and local political resistance to data centers — New York, Texas and Ohio all come up. One dynamic deserves attention because it cuts both ways. Restrictions on releasing models suppress external revenue and slow concentration. But they may also create a harder problem: internal progress continuing while the outside world falls behind. The genuinely unsettling scenario in this conversation isn't "AI moves too fast." It's a six-month gap during which nobody outside can see what is happening inside. A rule of thumb to carry: **a forecast whose author cannot tell you what would prove them wrong isn't a forecast, it's a position.** This one passes that test, which is why it was worth the time to write up. ## Where to Read More - The Dwarkesh Podcast conversation with Dylan Patel, 25 August 2026 (the source for this piece) - SemiAnalysis's public research on semiconductor supply chains, data centers and fab equipment - Economic histories of Volcker's 1980s rate rises and the Latin American debt crisis, useful for judging the "second Volcker shock" claim - For how discount rates drive valuation, any textbook chapter on the Gordon Growth model is enough — no special tools required ## One Thing to Take With You **One idea: scarcity moves, and the rent-collecting position is worth more than the growing position.** Every part of this episode is a facet of the same thing. When demand explodes, the money doesn't go to the hardest-working link in the chain. It goes to the link that cannot be expanded quickly and cannot be bypassed. The mirror polisher isn't working harder than the data center builder; it's just standing in that spot. And the spot moves — three years ago it was chips, now it's drifting toward memory and power, next it will be somewhere else. The real skill isn't identifying who collects rent today. It's building the habit of re-asking, on a regular basis, where the constraint is now. **One exercise you can do today (no investing required):** Pick something you did at least twice this week — making the kids' breakfast, writing a routine report, running a workout, handling a batch of customer queries. Anything. Break it into five steps on paper. Then ask one question: **if you had to do this ten times tomorrow, which step jams first?** Not "which step takes longest." Which step breaks when the volume rises. They're often not the same step. You'll find some steps cost nearly the same at ten repetitions as at one, and others collapse past three. The one that collapses is where your effort belongs — even if it's dull, even if your best skill lives somewhere else. Most people pour effort into the step they're best at, because that step feels good. Capacity doesn't grow there. It grows at the constraint you'd rather not look at.