Dylan Patel

Patel urges Nvidia to learn from AWS Trainium's fine-grained profiling hardware for its future Feynman GPUs.

Since AWS is a partner of Nvidia, Jensen should tell his engineers to book a meeting with the AWS Trainium team on the hardware features needed to do accurate fine-grained profiling such that Nvidia can learn how to add this capability to Feynman or Feynman-Next GPUs.

Patel says Neuron compiler's long compile times are a meaningful friction for researchers, though AWS is working to cut them by orders of magnitude.

The part where it was slightly disappointing was the Neuron compiler took a long time to compile, leading to a long wait in the first step.

Patel argues Meta is logically worth far more than ~$1.5 trillion given its cash flows and hoarded compute it can sell at high prices.

In a sense, Meta trading at… I think they’re like a $1.5 trillion company. It’s like, what? Silly. They’re worth way more than that, at least in a logical sense.

Patel warns that every force in AI points toward centralization of compute and power, which he finds deeply frightening.

What world do you see, Dwarkesh, where everything is not centralized? Because it seems to me that every force is screeching towards centralization. And that’s scary as hell.

Patel argues the constraint on AGI is societal tolerance — regulation, interest rates, data center bans — rather than research speed.

So the limiter on AGI is not how fast the research engineers, like our roommate Sholto, can crank the gears. It’s actually just how much does the rest of the world let that happen?

Patel suggests hyperscalers like Meta would happily borrow at 8% given huge compute returns, pushing up borrowing costs economy-wide.

This is going to be extremely vibed out, but recently Meta’s raised at 5 to 6%. I don’t see why they wouldn’t pay 8%. They would happily pay 8% because the return from the compute that they’re going to build is humongous.

Patel projects ~$11 trillion of AI CapEx through 2029, with $5 trillion debt-funded, which will push interest rates up.

No, I’m just saying till 2029 there’s something on the order of $11 trillion of CapEx. $6 trillion of that is funded with cash, and $5 trillion of that is funded with debt. If that’s the case, $5 trillion of debt being raised across the whole ecosystem does make interest rates go up.

Patel says Anthropic's Mythos pre-training used under 200 megawatts for about two months, despite the lab having multiple gigawatts.

But ultimately when they do the training run, when Anthropic trains Mythos, it’s sub-200 megawatts.

Patel breaks down lab compute as roughly 50% research, 10% development (final training runs), and 40% inference.

Actually 50% of the compute is research, 10% of the compute is development, and then 40% is inference. What I mean by research and development is: researchers are generating ideas, testing new architectures, testing new data mixes, testing new hyperparameters, new attention techniques, blah, blah, blah.

Patel warns China's semiconductor subsidies exceed the rest of the world combined, so a slower AI takeoff lets China catch up drastically.

So the Chinese semiconductor industry has significantly more subsidies than the rest of the world’s semiconductor industries combined. 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

Patel expects China's domestic AI compute to hockey-stick from 2028 as SMIC and CXMT fabs ramp, possibly adding 50 GW in 2029.

But ultimately, China is definitely going to hockey stick. If there’s anything China’s really good at, it’s scaling manufacturing really, really quickly.

Patel says 70% of new AI data center watts are deployed in the US and under 10% in China, down from roughly 30-35% in 2022.

So today, 70% of watts are being deployed in America. China is really a very small number. Sub-10% of watts being deployed for data center AI compute is in China.

Patel warns state-level data center restrictions in New York, Texas and Ohio will reduce supply and raise costs.

But more and more, the method of regulation is New York’s banning data centers. Texas is holding moratoriums. Ohio’s saying, or at least trying to say, you have to pay everyone’s property tax in a certain radius. These sorts of things are going to decrease supply and increase cost.

Patel holds the non-consensus view that labs will allocate a shrinking share of compute to inference and more to training and research.

The labs are going to allocate less and less compute to inference over time. I think that’s very non-consensus. The standard belief of most people is, "Oh, most compute will go to inference."

Patel expects lab revenue per megawatt to exceed $50 million by end-2027, contingent on being allowed to release top models.

I think it’s highly dependent on who has the best model, if they’re allowed to keep releasing their best models. But I don’t see why it wouldn’t be $50-plus million a megawatt.

Patel expects supply chain prices to keep rising, with memory and substrate makers raising prices quickly while TSMC moves slowly.

But over time, the supply chain will rebalance and things will cost more and more. To get that incremental capacity, you sort of have to. So TSMC raising prices very slowly, but memory companies raising prices very quickly. Substrate companies raising prices very quickly.

Patel says Meta and SpaceX, by building compute on their balance sheets without end customers, are the only plausible third players and can rent it to labs at high margins.

So there’s this completely different power structure where Meta is effectively hoarding compute. Them and SpaceX are the only plausible #3, because they’re hoarding all this compute.

Patel notes most value generated by AI models accrues to users like Jane Street and Meta rather than the labs.

Most of the value that these models generate does not get given to OpenAI and Anthropic. Thankfully, so far it is mostly just being given to the users.

Patel expects compute prices to rise sharply, with labs needing to pay $25-50 million per megawatt to absorb 70% of world compute by 2028.

They will continue to gobble up more of the compute. But ultimately they can’t do it at current pricing or anywhere close to it. They do have to start paying $25, $30, $50 million a megawatt to really gobble up 70% of the world’s compute in 2028

Patel says the safety regulation the labs advocate is slowing US labs far more than Chinese open-source models.

This regulation that they advocate for is actually slowing down the labs a lot more than it slows down the open-source Chinese language models. OpenAI not releasing Astra. OpenAI stopping training for two weeks.

Patel estimates ASML will produce about 100 EUV tools a year by 2030 at current expansion pace, unless massive capital is injected into suppliers like Zeiss.

The way the supply chain is expanding currently, 100 is roughly still the right number. For 2030? 100 ASML tools for 2030. But if you said, "Carl Zeiss, here’s $10 billion. Please fucking just expand production," that would change things.

Patel argues EUV tools are so bottlenecked they would resell for over $1 billion, and supply chain expansion lags demand due to a bullwhip effect.

I think if anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one, wait, and then sell it for north of a billion dollars. But ultimately, yes, capitalism will cause these things to expand. But it’s a whip.

Patel expects that by end-2028 OpenAI and Anthropic will control most of the world's usable FLOPs.

By the time you’re towards the end of 2028 — if this trend continues, and 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 predicts Anthropic and OpenAI will take 40-50% of new compute next year and that compute centralization is accelerating.

Anthropic and OpenAI are taking as much as 40% to 50% of compute next year. This centralization doesn’t look like it’s slowing down or stopping. In fact, it looks like it’s only accelerating.

Patel says frontier labs' revenue per megawatt now far exceeds compute cost, with Anthropic reaching about $50 million per megawatt versus $10-15 million cost.

But now, when OpenAI serves GPT-5.6 or Anthropic serves Opus 5 or Fable 5, their revenue generation has passed well beyond the incremental $10-15 million per megawatt. In the case of Anthropic, the revenue has gone as high as $50 million per megawatt.

Patel says Anthropic turned profitable in Q2 and OpenAI may turn profitable in Q3, driven by Codex and GPT-5.6.

Anthropic started turning a profit in Q2. It’s believed at some point in Q3, OpenAI could start turning a profit even, with the bigger rise of Codex and 5.6 and all this.

Patel says AI CapEx is just over $1 trillion this year, will exceed $2 trillion by 2028, and the labs are taking a growing share of it.

We’re at a little bit over a trillion dollars of CapEx this year. As we go out into ’28, it’s going to be more than $2 trillion. The labs are also taking an increasing percentage of this.

Patel rebuts datacenter skeptics such as Brad Gerstner, arguing that giant AI datacenter sites are being built and many are on schedule despite labor, supply chain and capital constraints.

But the fact is simple: gigantic sites are simply getting built, and many are on time.

Patel forecasts roughly 50GW of new datacenter capacity in 2027, equal to the entire cumulative global market through 2024, even after adjusting for constraints.

To be fair, +50GW is essentially the global cumulative datacenter market up through 2024 being built in one year! But our Datacenter Industry Model adjusts for all of that and still points to a material acceleration in datacenter capacity in 2027.

Patel argues the buildout will bring cheaper, smarter AI, faster model progress and revenue acceleration for AI labs, but also enormous capex that will push interest rates up.

The natural implication of the acceleration will be significantly more compute supply. That means cheaper and smarter AI, faster pace of AI model progress, and likely material revenue acceleration for leading AI labs and other token vendors. That also means unthinkable capex numbers. Reflexively rates will have to shoot up to enable the buildouts from a capital demand perspective.

Patel warns that memory and logic chips may become the top constraint on AI buildout in 2027-28, surpassing datacenters and power.

As we explained in The Great Silicon Shortage, datacenters & power may be the tip of the iceberg, with memory and logic possibly becoming the new #1 constraint in 2027-28.

Patel says over $1 trillion of datacenter lease value has been signed in 2026, a leading indicator underpinning a massive 2027-28 buildout.

Just in 2026 year-to-date, we’ve identified over a trillion USD of lease value that was signed and closed. That trillion is supporting a gigantic datacenter buildout in 2027-28.

Patel notes Oracle started the era of GW-scale leases in late 2025, and the buyer pool now spans hyperscalers, chipmakers and AI labs.

Since H2 2025, things have dramatically changed. Oracle kicked off the era of gigantic leases, after closing on ~4.5GW of datacenter IT capacity across three sites between August and October 2025. Everyone else followed, and the pool of buyers has never been broader

Patel argues that behind-the-meter generation and new OEMs have broadly solved the power bottleneck for AI datacenters.

From fuel cells to truck engines to refurbished turbines, the power generation problem has broadly been solved.

Patel expects ex-US datacenter buildout to accelerate sharply in 2027 and gain significant share in 2028 due to US labor, power and permitting constraints.

And as shown below, the ex-US buildout is sharply accelerating in 2027, and we expect it to gain a lot of market share in 2028, largely driven by US challenges.

Patel highlights modular, prefabricated datacenter designs cutting construction to under 10 months and MEP labor by up to 63%.

Construction timelines are now under 10 months. Hence our “LEGO Datacenters” title for the modular movement, you literally slot them in and go, and doing so allows for MEP labor (7-11 months of work) to drop up to 63%!

Patel believes ASML is building a D2W hybrid bonding tool as a natural extension of its vertical integration that will strengthen customer ties.

We believe it will start with a die-to-wafer (D2W) hybrid bonding (HB) tool. This is harder to build than wafer-to-wafer but is where ASML could gain an edge. We see the move as a natural extension of ASML’s vertical integration, and one that should strengthen its customer ties.

Patel expects advanced logic to adopt ASML's bonder first, which threatens Besi's logic-heavy hybrid bonding business at TSMC and Intel.

We expect advanced logic, where D2W already runs in volume, to adopt ASML’s D2W HB tool first. This is bad news for Besi. Logic accounts for most of Besi’s HB business today concentrated at TSMC and Intel.

Patel contrarily argues HBM will not adopt D2W hybrid bonding before 2029, if ever.

The market expected HBM to become the big D2W volume driver from 2028. However, our view is different: we believe HBM will not adopt D2W HB before 2029, if at all.

Patel views hybrid bonding as the most transformative semiconductor manufacturing innovation since EUV.

We did a deep dive on hybrid bonding in 2024, and we still see it as the most transformative innovation in semiconductor manufacturing since EUV.

Patel notes D2W hybrid bonding lags because current tools lack sufficient accuracy and throughput, keeping costs high.

Die-to-wafer (D2W) HB is moving slower, because the industry still lacks mature tools that combine satisfying alignment accuracy and throughput (our check indicates upper bound in current applications is only one wafer per hour), which keeps cost per die high.

Patel argues ASML's lithography expertise in alignment and particle control can give it an edge in bonding yield.

We believe ASML could gain an edge in alignment accuracy and particle control for Hybrid Bonding. These are the two factors that set bonding yield. And ASML has expertise in solving the same problems in lithography.

Patel cautions that ASML must still solve metrology, contact, tool adaptation and cleanliness challenges for its track-based bonder.

Despite the advantages, ASML still needs to solve challenges around feed-forward metrology, the die-to-wafer contact, the modification of the non-bonding tools so they can process dies on the track, and keeping the whole loop clean.

Patel says Besi's HBM prospects are weakening due to Korean competitors and delayed adoption.

Besi’s HBM prospects are also weakening. Korean challengers are catching up, in addition to an HBM adoption that we expect to slip to 2029 or beyond.

Patel expects ASML to price its bonder at a premium, which logic customers can afford.

Advanced logic dies are also among the most expensive chips, so logic customers can afford a premium tool. We expect ASML to price its tools at a premium to peers.

Dylan Patel argues that AI infrastructure investment is larger than commonly reported, with US capital expenditure next year expected to reach about $2 trillion, or 5-6% of GDP.

AI Infrastructure Investment Is Larger Than Reported

Dylan Patel said SemiAnalysis' ClusterMAX modelling shows Clockwork's TorchPass cuts training goodput loss from 14% to under 3% for a gold-rated neocloud.

In our ClusterMAX, TorchPass cuts training goodput loss from 14% to under 3% for a gold-rated neocloud,