PREDICTIONS ON THE RECORD
Dated and falsifiable — full record at /api/episode/<id>.
Dylan Patel · by end of 2027 · 3:50
By the end of 2027, half of the world's incremental new compute goes to just OpenAI and Anthropic.
“it's really by the end of next year, it's already half of the incremental compute is going to Anthropic and OpenAI.”
Dylan Patel · by end of 2028 · 5:52
By the end of 2028, OpenAI and Anthropic together control most of the world's usable flops.
“by the time you're in 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.”
Dylan Patel · by end of 2027 · 22:31
Most compute will still transact at under $20 billion per gigawatt at the end of 2027, despite the headline $40B SpaceX deal.
“I think most compute will still continue to transact at sub $20 billion a gigawatt.”
Dylan Patel · by end of 2027 · 25:55
By the end of 2027 a leading lab's blended revenue per megawatt reaches $70-80M or higher.
“it's like 70, 80 million dollars a megawatt blended across the company, if not higher”
Dylan Patel · in 2029 · 33:45
Global incremental AI compute additions reach 90-100 gigawatts in 2029.
“30 this year, 50 next year, 70 and 28 29 should be on the order of 90 to 100”
Dylan Patel · in 2028 · 34:34
China will have roughly 30 gigawatts of AI compute or less in 2028.
“it's pretty easy to say China will have like 30 gigawatts of AI compute or less.”
Dylan Patel · in 2028 · 35:46
Chinese domestic fabs (SMIC, CXMT) add 5-10 gigawatts of domestically produced compute in 2028 alone.
“now they're incrementally adding 5, 10 gigawatts in just 2028 of domestically produced chips.”
Dylan Patel · in 2029 · 37:49
China can add 50 gigawatts of incremental compute in 2029, though quality-weighted that is worth closer to 20 gigawatts of American chips.
“it's completely reasonable that China in 2029 can do 50 gigs.”
Dwarkesh Patel · by 2030 · 44:10
Incremental global AI CapEx approaches $10 trillion a year by the end of the decade.
“So very plausibly, incremental capex every year is getting close to 10 trillion dollars.”
Dylan Patel · 2024-2029 · 54:35
AI CapEx from 2024 to 2029 totals about $11 trillion, roughly $6T cash-funded and $5T debt-funded.
“In the modeling that we do, we have about $11 trillion of CapEx from 2024 to 2029”
Dylan Patel · by 2029 · 56:48
Meta's borrowing cost rises to around 8%, up from the 5-6% it raised at recently, dragging the whole economy's cost of credit up with it.
“I don't see why they wouldn't pay 8 percent.”
Dwarkesh Patel · unspecified · 58:44
A second Volcker shock repeats the 1980s pattern, defaulting a wave of heavily-indebted developing countries.
“he made this point that we'll see a second Volcker shock ... And I think that will probably happen again”
Dwarkesh Patel · unspecified · 51:31
Heavily-indebted, low-tax-revenue countries such as Pakistan and Nigeria are hit hardest by the new interest rate regime.
“Those countries like Pakistan or Nigeria or something, I think are just going to be very fucked in this new interest rate regime.”
Dwarkesh Patel · in the 2030s · 59:56
In the 2030s the rate of interest reaches tens of percent as an automated economy potentially doubles every year.
“So then we'll go into a world, I think in the 2030s, where the rate of interest is like tens of percent.”
Dwarkesh Patel · by end of the decade · 1:07:36
By the end of this decade, a single frontier lab holds more effective AI labor than there are people on earth.
“And I think that's like a thing that is very plausible by the end of this decade.”
WHO CLAIMED WHAT
Every position taken in the episode, attributed to its holder, with conviction (0–1) and the verbatim receipt.
Dylan Patel — 20 claims
assertionconviction 0.853:50
The centralization of compute into OpenAI and Anthropic is accelerating rather than slowing.
“And this centralization doesn't look like it's slowing down or stopping. In fact, it looks like it's only accelerating.”
predictionconviction 0.755:52
By the end of 2028 OpenAI and Anthropic will control most of the world's usable flops on their own, because their share of incremental compute is also the newest and most efficient.
“you've got them just controlling most of the usable flops in the world on their own.”
assertionconviction 0.810:04
The supply chain will expand, but with a long bullwhip lag — the price signal takes years to reach suppliers like Carl Zeiss, so the constraint holds in the near term.
“yes, capitalism will cause these things to expand, but it's a whip, right? It takes a long time for the whip signal to get to the tail end.”
assertionconviction 0.7516:03
The labs cannot reach ~100GW by 2028 at today's compute prices — they would have to pay $25-50M per megawatt to take 70% of the world's compute.
“They do have to start paying 25, 30, 50 million dollars a megawatt to really gobble up 70% of the world's compute”
assertionconviction 0.816:03
The safety regulation the US labs advocate for slows them down far more than it slows open-source Chinese models.
“This regulation that they advocate for is actually slowing down the labs a lot more than it slows down sort of the open source Chinese language models.”
assertionconviction 0.7516:03
Because the labs are withholding their best models, revenue per megawatt can stall or decline as competing models catch up.
“They're clearly not releasing their best models, and in which case, the revenue per megawatt stalls or even can start to decline again because other models are competitive again.”
assertionconviction 0.8518:23
Most of the value AI models generate is captured by their users, not by OpenAI and Anthropic.
“most of the value that these models generate does not get given to OpenAI and Anthropic.”
predictionconviction 0.7528:31
State-level data-center restrictions will cut supply, raise costs, and pass those costs through — making AI regulation inflationary.
“These sorts of things are going to decrease supply and increase cost and that's going to get passed on as well.”
assertionconviction 0.728:31
In a takeoff scenario Anthropic would keep its best model roughly six months ahead of anything released externally, for both safety and competitive reasons.
“why would Anthropic not have their best model six months ahead of what is externally available because of safety and regulation, but also the competitive advantage.”
opinionconviction 0.830:30
The labs will allocate a declining share of compute to inference over time — a view he flags as explicitly non-consensus — because building AGI beats taking the profit.
“this is sort of what I personally believe that the labs are going to allocate less and less compute to inference over time, which I think is very non-consensus”
assertionconviction 0.832:40
The labs are already shifting marginal compute toward R&D rather than inference — visible in revenue adds plateauing while compute keeps growing.
“they are factually increasing their compute towards R&D today.”
predictionconviction 0.837:01
China will eventually hockey-stick its compute build because scaling manufacturing quickly is what it is best at.
“China is definitely going to hockey stick if there's anything China's really good at is scaling manufacturing really, really quickly.”
predictionconviction 0.739:25
If takeoff is slower than the fast-RSI scenario implies, Chinese subsidies mean China catches up drastically on semiconductors and therefore on compute.
“if takeoff is not as fast as you're implying, but actually takes longer, then ultimately China will catch up drastically on the semiconductor side”
assertionconviction 0.7552:25
The crowding-out effect on credit — not physics — is what stops an unbounded compute buildout.
“So, this crowding out effect is actually the thing that I like, is the reason it's not like YOLO 1 billion gigawatts.”
opinionconviction 0.71:01:28
If you take AI seriously, every non-AI asset should trade at two or three times earnings — meaning the stock market should crash.
“if you're really AI-pilled, everything in the economy should trade it like two or three times earnings.”
assertionconviction 0.851:01:28
The limiter on AGI is not researcher throughput but how much the rest of the world permits the reallocation of capital toward it.
“the limiter on AGI is not how fast can the research engineers like our roommate Sholto can crank the gears. It's actually just like, how much does the rest of the world let that happen?”
opinionconviction 0.61:03:00
A slow takeoff remains possible even with capable models, because economic and regulatory friction — not capability — is the binding constraint.
“But slow takeoff is, at least in my hope, possible because everything in the economy and regulatory world”
assertionconviction 0.851:12:22
Every structural force in the market — economies of scale, continual learning, RSI — points to centralization of compute and power.
“All of these things point to centralization.”
opinionconviction 0.91:12:47
Neither the government nor the lab CEOs should be trusted with centralized control of AI.
“I don't trust the government and I don't trust Dario and I don't trust Sam.”
assertionconviction 0.71:13:58
The one saving grace of the current arrangement is that Anthropic does not capture most of the value it creates — the rest of the economy does.
“the one positive thing here is that today, Anthropic does not capture most of the value.”
Dwarkesh Patel — 8 claims
assertionconviction 0.79:11
The gap between fab CapEx and end AI revenue is so large (roughly 100x) that capitalism should force a much faster supply expansion than is happening.
“That still means there's 100X discrepancy between fab capex and end revenue generated.”
predictionconviction 0.618:18
Full AGI would generate many hundreds of billions of dollars of value per gigawatt, far above today's ~$100B.
“But it would be many hundreds of billions of dollars if you get like full AGI per gigawatt.”
opinionconviction 0.6538:37
If China's compute stock really is that far behind by the time automated research arrives, US export controls will have been a notable policy success.
“So, if that ends up being the case, that would have worked. I think that's actually a notable success.”
opinionconviction 0.747:27
The US will be fine fiscally because it can tax the data centers, provided they are built in America.
“I think the US will be fine at the end of the day, because if the data centers are built in America, we can fundamentally just tax the data centers.”
predictionconviction 0.757:55
A higher discount rate from AI-driven interest rates will crater the valuation of stable-cash-flow, non-AI equities even if the index holds up.
“if interest rates rise, the discount rate increases, which means that the discount cash flows of all equities crater”
predictionconviction 0.71:07:36
By the end of this decade a single lab could plausibly hold more effective AI labor than there are people on earth.
“That there's more AI labor, more effective population within a single lab, than there are people on earth.”
assertionconviction 0.71:07:36
Because most of the world's cognitive work would sit inside two labs, misalignment there is effectively misalignment of most of the world.
“if these AIs are misaligned, then most of the world is misaligned basically, because most of the world's minds are there.”
assertionconviction 0.81:11:20
AI training has huge economies of scale — training effort amortizes across billions of sessions — which structurally compounds whoever is ahead.
“I think the fundamental problem is that lab, AI training has huge economies of scale.”