What this covers

Dwarkesh Patel interviews Dylan Patel and Jon Y on how semiconductors are actually manufactured, designed, and controlled—the full stack from physics to geopolitics. The conversation moves between technical depth (how process nodes shrink, why memory stalled while logic continued, the multiplicative yield problem in cutting-edge fabs) and strategic questions (whether export controls can slow China's chip ambitions, why Pascal's wager logic drives AI spending regardless of ROI, what happens if Taiwan's fabs vanish). The hosts treat semiconductors not as a commodity market but as humanity's largest optimization problem: a search space of nearly infinite arrangements across 100-220 billion transistors, where the target is intelligence per picojoule.

The substance splits between the technical and the geopolitical. On the technical side: why Moore's law became partly self-fulfilling through industry coordination; how hardware-software co-design means the optimal AI architecture depends on whether you're running on Google's TPUs or OpenAI's NVIDIA gear; why memory is now the most under-improved layer of the stack, and why breaking that wall requires manufacturing scale few entrepreneurs can access. On geopolitics: how TSMC's dominance was won not through superior technology alone but through organizational intensity (Morris Chang's "Nightingale Army" running round-the-clock R&D when Samsung threatened); how master-apprentice tacit knowledge transfer caused cutting-edge process know-how to migrate from America to Taiwan as apprentices followed jobs; and why the jigsaw puzzle of export controls signals China exactly which capabilities to develop domestically rather than stopping progress. The conversation argues that semiconductors operate under entirely different constraints than AI—requiring advanced degrees, paywalled knowledge, and years of apprenticeship—which makes it nearly impossible for outsiders to break in the way they do in AI.

Sharpest takeaway

China could plausibly match or exceed any single US AI lab's compute for a frontier training run by 2026-2028 if it chose to centralize its chips, while export controls remain a piecemeal 'jigsaw puzzle' that slows but cannot stop China's semiconductor progress; meanwhile the AI buildout is a Pascalian bet that does not require near-term revenue because the cost of falling behind dwarfs the cost of over-investing.

  • China lacks centralization, not chips - if Xi became scale-pilled, foreign and domestic chips alone could fund a frontier-scale single training run
  • Export controls are an 'art' of removing jigsaw pieces; each removal pushes China to build that capability domestically rather than halting progress
  • AI capex precedes revenue by years, and Pascal's wager logic (cost of under-investing > cost of over-investing) drives continued spending regardless of current ROI

The claims · ranked64 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.58

Moore's Law became a coordinating 'word of God' for the industry: Gordon Moore's observation, predicting multiple orders of magnitude of transistor increases 7-8 years out, drove tens of millions of engineers across every abstraction layer to converge on hitting each 2x-density-every-two-years shrink, making the law partly self-fulfilling rather than a pure technical inevitability.

causalcontestednovelty 3/4durability 4/4· Dylan Patel

“It's like God came and proclaimed it: "We will shrink density 2x every two years." Gordon Morris made an observation. It didn't go nowhere. It went way further than he ever expected”

0.55

AI lets cracked outsiders (high-school dropouts, ex-McKinsey consultants like Sholto) enter and move across the stack because state-of-the-art is public on arXiv, whereas semiconductors require master's/PhDs (≈50,000 of TSMC's 75,000 workers have advanced degrees), papers are paywalled and siloed, knowledge is apprentice-master tacit intuition, and no one can pick it up or move between layers - making it impossible for cracked generalists to break in the way they do in AI.

factualestablishednovelty 3/4durability 3/4· Dylan Patel

“You can't get a job at a tool company without at least a master's in chemistry, probably a PhD. Of the 75,000 TSMC workers, like 50,000 have a PhD or something insane.”

0.55

The acceleration points in China's semiconductor industry came from Taiwanese talent poaching: Richard Chang recruited a 'conga line' of Taiwanese engineers to Shanghai, and Liang Mong Song, after losing a power game at TSMC, took talent to Samsung (briefly making it beat TSMC) and then to SMIC, which caught up rapidly under him.

factualestablishednovelty 3/4durability 3/4· Jon Y

“Generally that's actually true of a lot of the acceleration points within China's semiconductor industry. It's from talent flowing from Taiwan.”

0.55

When Samsung temporarily surpassed TSMC and won Apple business, Morris Chang created the 'Nightingale Army' running round-the-clock R&D shifts (called 'burning your liver'), finished FinFET, blew past Samsung, and simultaneously sued Liang Mong Song for trade-secret theft, illustrating how organizational intensity rather than just technology decided leadership.

factualestablishednovelty 3/4durability 3/4· Jon Y

“He sets up what is called the Nightingale Army. They split a bunch of people and they say, "You are working R&D night shift. There is no rest at the TSMC fab."”

0.55

In a large GPU cluster's total cost of ownership (multi-year, including debt and operations), roughly 80% is the GPUs/servers, ~10% the data center, and ~10% (under 15%, actually under 10%) the power - so the marginal cost of power is nearly irrelevant; the real constraint is the ability to generate and deliver new power, not its price, which is why clusters won't run only in daytime even with cheap solar.

factualestablishednovelty 3/4durability 3/4· Dylan Patel

“Power is sub 15% of the cost. It's sub 10% of the cost actually. The biggest, like 75-80% of the cost, is just the servers.”

0.52

Because current GPUs and TPUs are built for very high arithmetic intensity and transformers run well on them, all research concentrates on transformers, creating a self-fulfilling prophecy: a theoretically superior architecture that only achieves half the usable FLOPs on existing chips is worthless even with a 30% compute-efficiency win, because it runs half as fast on the hardware.

causalestablishednovelty 3/4durability 3/4· Dylan Patel

“All the research is in transformers that are very high arithmetic intensity because the hardware we have is very high arithmetic intensity and transformers run really well on GPUs and TPUs. You sort of have a self-fulfilling prophecy.”

0.52

Memory followed Moore's Law until 2012, when it stalled into only incremental gains, while logic has continued (still >10-15% CAGR in density/cost), so memory is now the most under-improved part of the stack; breaking the memory wall - integrating memory with accelerators - would change the world, though it's nearly impossible for an entrepreneur because it requires absurd manufacturing scale or custom-memory capability the industry doesn't allow.

factualestablishednovelty 3/4durability 3/4· Dylan Patel

“The memory industry followed Moore's law up until 2012 when it stopped. Tt became very incremental gains since then. Whereas logic has continued”

0.52

The relationship between chips and models is not one-way; hardware heavily influences which model architecture is optimal, so the best model to run on Google's TPUs given fixed dollars and compute is architecturally different from the optimal one for OpenAI on NVIDIA - down to networking and data-center design choices - meaning architectures diverge based on the underlying hardware constraints.

causalestablishednovelty 3/4durability 3/4· Dylan Patel

“Hardware has a huge influence on the model architecture that's optimal. It's not a one-way street that better chip equals... The optimal model for Google to run on TPUs ... is different architecturally than what it is for OpenAI with NVIDIA stuff.”

0.52

Semiconductor manufacturing and design is the largest search space of any human problem because leading-edge chips have ~100-220 billion transistors as on-off switches, connected by ~15 metal layers in nearly infinite possible arrangements - a search space larger than any other humans know of, where the optimization target is useful compute (intelligence per picojoule).

factualestablishednovelty 3/4durability 3/4· Dylan Patel

“semiconductor manufacturing and design is the largest search space of any problem that humans do because it is the most complicated industry that humans do.”

0.51

The stated goal of export controls is to stop China from having better AI than the US, which requires controlling chips, which requires allowing China at least some chip level above what it can build domestically; but current restrictions are flipped - China can build better chips domestically than what NVIDIA/AMD/Intel are allowed to sell them, and the equipment China can still buy lets it build chips better than the export-allowed ones.

causalcontestednovelty 4/4durability 2/4· Dylan Patel

“currently the restrictions are flipped the other way. They can build better chips in China than what we restrict them from in terms of chips that NVIDIA or AMD or Intel can sell to China.”

0.50

The semiconductor tech stack is so hyper-optimized yet under-modernized that most fab equipment runs on Windows XP servers and chip-design tools run on CentOS 6, and people are afraid to touch this 'unholy amalgamation' that miraculously works - signaling huge unexploited opportunity to bring AI and modern tooling to under-optimized layers.

factualestablishednovelty 3/4durability 2/4· Dylan Patel

“Most of the equipment in semiconductor fabs runs on Windows XP. Each tool has a Windows XP server on it. All the chip design tools have CentOS version 6, which is old as hell.”

0.50

America lost the ability to do processes like etch with hydrofluoric acid because the knowledge is transmitted master-to-apprentice like Sith lineage; as apprentices stopped coming in the US and moved to Taiwan (e.g., via National Tsing Hua University), the tacit, secret knowledge migrated with them.

causalcontestednovelty 3/4durability 3/4· Jon Y

“It's very master-apprentice. You know like in Star Wars with the Sith, there's only one, right? Master apprentice, master apprentice.”

0.50

Chip manufacturing is like 3D chess or a jigsaw puzzle; removing one piece at a time via incremental export restrictions just signals to China exactly which capability to build domestically, so a half-measure approach forces China to develop its own industry node-by-node - you must either smash the whole table at once or let them do whatever they want, but the incremental approach is the worst of both worlds.

causalcontestednovelty 3/4durability 3/4· Dylan Patel

“It's like, "Let's take one jigsaw puzzle piece out. Walk away. Oh shit. Let's take two more out. Oh shit." You either have to go full "bat to the fricking table/wall" or chill out and let them do whatever they want.”

0.50

AI investment follows Pascal's wager logic for big tech: the risk of under-investing is worse than the risk of over-investing (as Satya Nadella stated word-for-word, and Sundar Pichai and Zuckerberg implied), so each CEO must believe they are AGI-pilled because if they are not and a competitor is right, they are absolutely finished - this psychology, plus capital holders (UAE, Saudi) buying in, drives the buildout regardless of whether it later deflates.

causalcontestednovelty 3/4durability 3/4· Dylan Patel

“Satya said it on his earnings call. The risk of under-investing is worse than the risk of over-investing. He has said this word for word. This is Pascal's wager.”

0.49

The semiconductor industry started vertically integrated (everyone made their own equipment in the 1960s-70s), but engineers spun off to build better dedicated tools and sell them back; companies that accepted outside equipment got better products and outcompeted the fully integrated ones, creating today's hyper-stratified industry where each layer has only one to three competitors.

causalestablishednovelty 2/4durability 3/4· Jon Y

“What happened was that the companies that accepted outside products and equipment got better stuff and did better. There were companies that were totally vertically integrated for decades. They are still good, but nowhere near competitive.”

0.49

In rapidly growing AI infrastructure, the 'one bottleneck' narrative is misleading because the moment you unlock one constraint, the next becomes binding - from CoWoS in 2023, to HBM and CoWoS-L, to data centers, transformers, substations, power generation, batteries, UPSs, and water cooling - so the picture is permanently cloudy with shifting limiters rather than a single chokepoint.

factualestablishednovelty 2/4durability 3/4· Dylan Patel

“Always when you grow super rapidly, people want to say that's the one bottleneck. ... the moment you unlock one… Only 10% higher, the next one is the thing.”

0.48

Moving a process node no longer doubles density or power efficiency (Dennard scaling ended; SRAM doesn't scale, logic scales ~30%, only ~20% power saving per transistor), but AI gains far more than the per-transistor benefit because higher density lets you fit more of a computation (e.g., a large matrix multiply) on-chip, reducing costly off-chip data movement to memory - so data locality, not transistor power, drives AI's hunger for new nodes.

causalestablishednovelty 3/4durability 3/4· Dylan Patel

“because of data locality and movement of data, you actually get a much larger improvement in power efficiency by moving to the next node than just the individual transistors' power efficiency benefit.”

0.48

If Taiwan stopped shipping (invasion or earthquake), day one the market crashes because the Magnificent Seven are 60-75% of the S&P 500 and rely entirely on chips; within six months car production halts because every car contains 2,000+ chips (even ICE vehicles), and trailing-edge dependence means no fridges, appliances, or display drivers - because 80% of 28nm and effectively 100% of older 130nm production is in Taiwan.

forecastestablishednovelty 2/4durability 2/4· Dylan Patel

“The six or seven biggest companies, the Magnificent Seven, are like 60-75% of the S&P 500 and their entire business relies on chips.”

0.48

Investors operate in a world where, per Keynes (often misattributed to Buffett), the market can remain delusional/irrational longer than you can remain solvent - so whether the AI buildout eventually deflates is irrelevant in the near term because you must operate within the reality that the capital is flowing in now.

factualestablishednovelty 1/4durability 4/4· Dylan Patel

“The one about how the market is delusional for longer than you can remain solvent”

0.47

Even at 98-99% success per step, a process with ~1,000-10,000 steps multiplies down to roughly 40% yield (or worse), so bad yields in cutting-edge Chinese 7nm are inherent to a brand-new process run without EUV; yield is a multiplicative effect requiring six-sigma perfection at every step.

factualestablishednovelty 2/4durability 4/4· Jon Y

“Even if you're 98-99% for each, in the end you'll still get a 40% yield.”

0.46

Because the vast majority of an H100's power goes to moving data (networking and memory) rather than compute, minimizing data movement and maximizing compute could yield 100x efficiency gains from architecture alone - even if process shrinking stopped entirely - via more efficient ALUs, better memory technologies, power delivery, and chip-to-chip networking.

forecastcontestednovelty 3/4durability 3/4· Dylan Patel

“There are 100x gains possible from architecture. Even if we literally stopped shrinking, we could have 100x gains from architectural advancements.”

0.45

xAI built a ~100K GPU cluster fast in Memphis by buying a factory, parking ~14 mobile generators (normally reserved for natural disasters) outside, adding a Tesla battery pack, drawing grid power, and tapping a natural-gas line to a nearby gigawatt plant - costing ~$4-5 billion rather than $1 billion - demonstrating that improvised power sourcing is the way to build the largest clusters quickly.

factualestablishednovelty 3/4durability 1/4· Dylan Patel

“xAI is like, "Oh shit, we're going to buy a Memphis factory, put a bunch of mobile generators usually reserved for natural disasters outside, add a Tesla battery pack, drive as much power as we can from the grid, tap the natural gas line ... and just send it."”

0.45

Improving a process node is not linear research but a massive multivariable search: engineers run experiments turning knobs (pressure, temperature, chemical recipes) on each sequential tool, optimizing simultaneously for yield, performance, and power rather than a single 'better/worse' metric, relying heavily on deep intuition from PhD specialists.

definitionestablishednovelty 2/4durability 3/4· Dylan Patel

“It's basically a multivariable problem. Every single tool sequentially you're processing the whole thing. You turn knobs up and down on every single tool.”

0.45

Huawei out-competes Western firms with two hands tied behind its back - making a phone as good as a year-old Qualcomm phone on a process node equivalent to one three-to-four years old, while banned from selling to the US/Europe and lacking TSMC access - which proves engineering capability rather than mere espionage explains its success.

factualcontestednovelty 3/4durability 2/4· Dylan Patel

“The new phone is as good as a year-old Qualcomm phone on a process node that's equivalent to something three or four years old. They actually out-engineered us with the worst process node.”

0.45

Counting training FLOPs as active-parameters x tokens x six is now stupid because the paradigm includes synthetic data, RL, post-training, data verification, search, and inference-time compute - so the actual pre-training FLOPs may be smaller (likely under 1e30), while the total summed FLOPs across pre-training, post-training, synthetic-data generation, and inference-time compute is what reaches ~1e30.

factualcontestednovelty 3/4durability 2/4· Dylan Patel

“the way you count flops on a training run is really stupid. You can't just do like active parameters x tokens x six.”

0.45

China's barrier to a frontier-scale AI training run is centralization, not chip supply: it still imports over a million NVIDIA H20/Hopper-class GPUs a year plus under a million domestic chips, but no single Chinese entity has aggregated them into a 100,000-GPU cluster the way OpenAI or xAI have, because the ecosystem remains decentralized across Alibaba, Baidu, ByteDance, DeepSeek, etc.

causalcontestednovelty 3/4durability 2/4· Dylan Patel

“If you are Xi Jinping and scale-pilled, you must now centralize the compute resources. Because you have sanctions on how many NVIDIA GPUs you can get in now. They're still north of a million a year”

0.45

Even though America as a whole imports millions of GPUs a year, no single US company gets more than hundreds of thousands, so no individual US training run can be as large as one China could mount if it pooled its national chip supply into a single coordinated run.

factualcontestednovelty 3/4durability 2/4· Dwarkesh Patel

“the fact that any one company is only getting hundreds of thousands or fewer means that there's no one person who can do a single training run as big in America as if China as a whole decides to do one together.”

0.45

As nodes advance, development costs rise so much that you must aggregate demand from all customers to fund the next node, concentrating profit into fewer fabs; this means N2 (2nm) is economically a 'monstrosity' that should not exist without the concentrated spend of only about five remaining market players, and at some point the next node will not be economically possible at all - unless new demand like AI makes A16/1nm viable.

factualcontestednovelty 3/4durability 2/4· Jon Y

“now there's no room in the market for an N2 or N3. You could argue that economically, N2 is a monstrosity that doesn't make sense. It should not exist without the immense concentrated spend of like five players in the market.”

0.45

Funding the next process node would no longer be economically viable - given that Apple's mobile profit, formerly the engine of leading-edge demand, is petering out and Samsung/Intel fund their nodes through memory, monopoly profits, or debt - were it not for AI taking off and generating humongous demand for the most leading-edge chips.

causalcontestednovelty 3/4durability 2/4· Dylan Patel

“There's a strong argument that funding the next node wouldn't be economically viable anymore if it weren't for AI taking off and generating humongous demand for the most leading-edge chip.”

0.45

Apple, historically ~25% of TSMC and ~100% of each newest node, does not need the 2nm/A16/A14 capacity TSMC is aggressively building; if AI scaling continues to deliver capability gains, AI could plausibly consume 60-80% of those leading-edge nodes by 2028, replacing Apple's mobile demand as the funder of leading-edge silicon.

forecastcontestednovelty 3/4durability 2/4· Dylan Patel

“In that world—not of 5 nanometer, but of 2 nanometer, A16, A14, these are the nodes that'll be in that timeframe of 2028—used for AI, I could see it being like 60-80% of it. No problem.”

0.45

Near-term revenue does not need to support AI spending because there is a multi-year lag between when capex is committed/raised and when revenue flows: GPT-4 cost ~$500M to train and generated billions in recurring revenue, so the question is only whether each prior huge chunk of capex eventually makes sense or whether enough people can be convinced to raise the next round - and most of a data center's cost (the chips) is deferred, so companies commit capacity now without spending most of the money.

causalcontestednovelty 3/4durability 2/4· Dylan Patel

“You don't necessarily need the revenue in 2025 to support this. You don't need the revenue in 2026 to support this.”

0.45

The AI investment surge (~$55-60 billion of private capital this year) is nowhere near the dot-com bubble, when private industry flew through ~$150 billion a year, and unlike the debt-financed 1990s bubble, today's AI buildout is funded by the most profitable companies humanity has created from their own balance sheets - so even if it is a bubble, there is more room to grow and (per the dot-com fiber analogy) it would still build durable infrastructure.

factualcontestednovelty 3/4durability 2/4· Dylan Patel

“during the dot-com bubble, the private industry flew through like $150 billion a year. We're nowhere close to that yet.”

0.44

The export sanctions have, to a significant extent, made China believe in the strategic primacy of its semiconductor industry more than anything else, reinforcing rather than deterring national investment.

causalcontestednovelty 2/4durability 3/4· Dylan Patel

“To some extent, the sanctions have made China believe in the importance of the semiconductor industry more than anything else.”

0.44

Huawei's competitive intensity comes from a militaristic, PLA-adjacent 'struggle' culture and constant existential paranoia (the belief they will fight the US in five years, that their country depends on every engineering gain), embodying the Andy Grove 'only the paranoid survive' mindset that lets paranoid companies like Google resurge.

causalcontestednovelty 2/4durability 3/4· Dylan Patel

“It's the Andy Grove-ian mindset. Shout out, the based Intel of Andy Grove. Only the paranoid survive. Paranoid Western companies do well.”

0.44

East Germany and Soviet industry ran on a conveyor belt of stolen Western secrets but were never good at mass-producing them, whereas China both acquires secrets and can mass-produce - distinguishing China's threat from the historical Soviet one.

factualcontestednovelty 2/4durability 3/4· Jon Y

“East Germany and the Soviet industry was basically a conveyor belt of secrets coming in ... But the Soviets were never good at it. They could never mass produce it. But now you have China.”

0.44

Even though the dot-com bubble burst and many companies went bankrupt, the spending laid out the infrastructure that enabled the web; analogously, many AI/foundation-model companies may go bankrupt but their buildout could enable the singularity - though the fiber-bubble counterexample (MEMS/optical investment that never revitalized after 2002) shows the infrastructure overhang can also strand for decades.

factualcontestednovelty 2/4durability 3/4· Dwarkesh Patel

“Even though the dot-com bubble obviously burst and a lot of companies went bankrupt, they in fact did lay out the infrastructure that enabled the web and everything.”

0.44

The cost of intelligence is tanking analogously to Moore's Law: GPT-4-level output cost roughly $120 per million tokens in February 2023 and is now around $10, driven partly by cheaper compute and partly by models' compute-efficiency wins, and this falling cost will drive adoption in a scale-up-and-cheapen feedback loop.

factualestablishednovelty 2/4durability 2/4· Dylan Patel

“OpenAI and GPT-4 in February 2023, was roughly $120 per million tokens. Now it's like $10. The cost of intelligence is tanking”

0.44

AI is already delivering single-to-low-double-digit gains in chip design (NVIDIA and Google AI layout tools, RL circuit-placement, simulation) using even 5-8-year-old RL techniques, while manufacturing has seen little AI adoption so far (only possibly inverse lithography) - so generative AI could revolutionize the industry's under-optimized layers, though a massive data problem remains because knowledge is siloed and tacit.

factualestablishednovelty 2/4durability 2/4· Dylan Patel

“Some examples of where this is already shown true are some of NVIDIA's AI layout tools, and Google as well, laying out the circuits within a small blob of the chip with AI.”

0.43

Centralizing all compute under one entity yields immense gains from scaling laws but hurts compute efficiency, because fewer independent teams can experiment and lead competing efforts; a single bureaucrat at the top creates a single point of failure, so there is a genuine balancing act between scale and innovation diversity.

causalcontestednovelty 3/4durability 3/4· Dylan Patel

“there are obviously immense gains from being centralized because of the scaling laws. The flip side is compute efficiency which is obviously going to be hurt because you can't experiment and have different people lead and try their efforts as much if you're more centralized.”

0.43

The semiconductor industry is so stratified that nobody understands the full stack - experts are each deep in a single 'well' (etch, lithography, packaging) and assume others know more, a Gell-Mann amnesia dynamic - so the industry coordinates the path to the next node not by central knowledge but by a social, gossip-like consensus process where, like SEMATECH historically, people argue until a 'blue moon' alignment forms around a direction.

factualcontestednovelty 3/4durability 3/4· Jon Y

“They're all in this single well. They're digging deep for what they're getting at. They don't know the other stuff well enough. In some ways, nobody knows the whole stack.”

0.41

The US sanctioned SMIC at a specific Shanghai location rather than the whole company, so SMIC keeps buying tools for 7nm/6nm/5.5nm (e.g., for the 910C) as long as it's not in Shanghai; with ~45-50 high-end immersion lithography tools it could theoretically produce up to 60,000 7nm wafers/month, though real yielded output is far lower because using DUV instead of EUV on a brand-new process gives bad yields.

factualcontestednovelty 3/4durability 1/4· Dylan Patel

“The way the US has sanctioned SMIC is really stupid. They've sanctioned a specific spot rather than the entire company.”

0.41

If China centralized, it could release a 1e27 FLOP model by 2026 and have a 1e28 model in the works using only foreign (NVIDIA) chip supply, matching any individual US lab's compute in 2025-2026; the only obstacle is the decision to centralize, not chip availability.

forecastcontestednovelty 3/4durability 1/4· Dylan Patel

“They could have just as many flops in 2025 and 2026 if they decided they were scale pilled. Just from foreign chips, for an individual model.”

0.41

The largest single cluster is ~100K GPUs in 2024, rising to 300-700K next year (≈1M H100-equivalents given 2-3x faster GPUs), with single gigawatt sites in 2026 and aggregate buildouts north of 2-3 gigawatts; if globally produced FLOPs grow 10-30x year-on-year and multi-site linking improves, total compute delivered to a single model could reach ~1e30 by 2028-2029 - about 100,000x (six orders of magnitude) more than GPT-4.

forecastcontestednovelty 3/4durability 1/4· Dylan Patel

“It's 100k GPUs this year for the biggest cluster. Next year, 300-500k depending on whether it's one site or many. ... on an H100 equivalent basis, you're at a million chips next year in one cluster by the end of the year.”

0.41

By 2028-2029 China could deliver more FLOPs to a single model than the US - possibly reaching 1e30 first - because the US is not even attempting 10-gigawatt single data centers and instead must link multiple sites, whereas China has no data-center-scale constraint and would likely centralize sooner if it became convinced AI is critical.

forecastcontestednovelty 3/4durability 1/4· Dylan Patel

“It's possible they actually have 1e30 before the US because data center isn't as big an issue. A 10 gigawatt data center… I don't think anyone is even trying to build that today in the US”

0.41

Microsoft and OpenAI have figured out how to train across multiple geographically separate data centers, evidenced by Microsoft signing over $10 billion in fiber deals (with Lumen, Zayo, and others) and filed permits showing digging between data centers; analysis suggests five regions comprising many data centers are being connected, with total power easily north of a gigawatt.

factualcontestednovelty 3/4durability 1/4· Dylan Patel

“Microsoft has signed deals north of $10 billion with fiber companies to connect their data centers together. There are some permits already filed to show people are digging between certain data centers.”

0.41

Microsoft is taking on immense credit risk by signing data-center deals (CapEx rising from $50-80B direct plus ~$20B via Oracle/CoreWeave plus ~$10B partners) that it can afford because it truly believes in OpenAI; OpenAI must in turn deliver a far better next model and raise $50-100 billion, which Dylan believes it will because Sam Altman is exceptional at recruiting and raising money.

factualcontestednovelty 3/4durability 1/4· Dylan Patel

“Today Microsoft is taking on immense credit risk. They've signed these deals with all these companies to do this stuff. ... that's affordable for them because they believe they're a great partner to OpenAI.”

0.40

NVIDIA delivers more than 2x performance per dollar generation to generation, and hyperscalers will try to exceed that, driven by combined gains from process nodes, power delivery, system design, chip-to-chip networking, and memory technologies pursued simultaneously.

factualestablishednovelty 2/4durability 2/4· Dylan Patel

“NVIDIA gen to gen to gen will do more than 2x performance per dollar. I think that's very clear.”

0.40

TSMC will build the silicon needed for 1e30-scale AI, but only if continuously convinced through demonstrated growth - it must see NVIDIA and Google volumes growing straight up over and over; if people work silently it won't happen, because TSMC needs repeated visible proof across the industry to commit to aggressive capacity expansion.

normativecontestednovelty 2/4durability 2/4· Dylan Patel

“it's going to be a constant game of convincing them constantly that they must do this. It's not a simple game. If people work silently, it's not going to happen. They have to see the demonstrated growth over and over and over again”

0.40

Big tech companies (Microsoft, Google) cannot do crazy improvised power buildouts like xAI did because of ESG and net-zero commitments, which forces them to work through partners and converted crypto-mining sites; these commitments are likely to be dropped as the scaling picture intensifies.

causalcontestednovelty 2/4durability 2/4· Dylan Patel

“the big tech companies can't do crazy shit like Elon did. ... ESG. They can't just do crazy shit”

0.40

The US/global power and transformer-substation supply chains, weakened by years of flat-to-declining US power demand, can be revitalized much faster than people realize (6-month to 3-year timescales) once tripling demand and AI tooling create urgency, because people innovate and expand production capacity when given the need and the margin incentive.

forecastcontestednovelty 2/4durability 2/4· Dylan Patel

“the US industrial base as well as the Japanese, all across the world can get revitalized much faster than people realize. I truly believe that people can innovate when given the need to.”

0.40

China has been hacking ASML for over five years and likely already possesses its files, but building EUV lithography from those files is extraordinarily difficult - so espionage alone does not confer capability.

factualcontestednovelty 2/4durability 2/4· Dylan Patel

“China has been hacking ASML for over five years and ASML is kind of like, "Oh, it's fine." ... They already have those files in my view. It's just a very difficult thing to build.”

0.39

Data centers today are only 4-5% of total US power consumption, so on a national basis AI power demand is not crazy - the difficulty is localized delivery (generating and routing power to dense sites), especially as coal has been curtailed, not aggregate national supply.

factualestablishednovelty 2/4durability 1/4· Dylan Patel

“Today, data centers are like 4-5% of total US power consumption. ... that's not that much. ... On a localized basis, it is, because it's about the delivery of it.”

0.39

GPU rental has shifted to a buyer's market - H100s that rented for $3-4/hour last year now go for ~$2.15 for six-month deals - yet remains profitable because the natural all-in cost (data center pricing plus debt/cost of capital) is about $1.70/hour, so the good deals (e.g., Microsoft renting CoreWeave at ~$1.90-2.00) still leave margin even as the cost of compute keeps tanking.

factualestablishednovelty 2/4durability 1/4· Dylan Patel

“Last year you would buy H100s for like $4 or $3 an hour for shorter-term or mid-term deals. Right now, if you want a six-month deal, you could get it for like $2.15 or less.”

0.39

China was about 48% of ASML's revenue and ~45% of Applied Materials' revenue in recent quarters, showing that equipment exports to China are not being controlled effectively even though they could be.

factualestablishednovelty 2/4durability 1/4· Dylan Patel

“China in the most recent quarters was like 48% of ASML's revenue and like 45% of Applied Materials'.”

0.38

Because Chinese chips are memory-constrained (less compute per memory bandwidth) and China invests heavily in compute-in-memory, Chinese models will diverge architecturally from American ones - in sparsity, expert arrangement, attention mechanisms, model width vs depth, and possibly optical on-chip networking - and diverge in use cases too, with China dominating video/image recognition due to surveillance data while US models focus on consumer learning.

forecastcontestednovelty 3/4durability 2/4· Dylan Patel

“it'll start to diverge more and more architecturally because they'd be stupid not to make chips for China.”

0.35

China could build and hide a gigawatt-to-10-gigawatt AI data center because it adds power capacity equal to roughly half of Europe each year; a gigawatt is a drop in the bucket relative to its industrial base, and it could shut down an energy-intensive facility like the world's largest aluminum mill (over 5 GW) near the Three Gorges Dam, repurpose that power to AI, and spin up replacement aluminum capacity, making the switch undetectable.

factualspeaker onlynovelty 3/4durability 2/4· Dylan Patel

“China could just build it in six months, I think, around the Three Gorges Dam or many other places. They have the ability to do the substations. They have the power generation capabilities.”

0.35

The return on invested capital for all the big tech firms (Meta, Microsoft, Amazon, Google) is up since 2022, suggesting investing in AI has been fruitful so far - though this is a general company-level metric confounded by other factors (e.g., Meta's ad efficiency), not a clean AI-specific measure.

factualcontestednovelty 2/4durability 1/4· Dylan Patel

“The return on invested capital for all of the big tech firms is up since 2022. Therefore, it's clear as day that investing in AI has been fruitful so far”

0.35

The entire AI investment case is contingent on the next model (GPT-5) being impressive enough to blow people away; if GPT-5 sucks or fails to impress, the whole financing structure is void - and Jon notes GPT-5 is not here and appears late, making the bet uncertain, while Dylan disagrees and expresses tremendous belief in the GPT-5-area capability jump based on public signs.

forecastcontestednovelty 2/4durability 1/4· Jon Y

“It's all dependent on GPT-5 being good. If GPT-5 sucks, if GPT-5 looks like it doesn't blow people's socks off, this is all void.”

0.31

Reports of Chinese military incompetence (e.g., 'filling water in rockets') may be deliberate deception, because if China wants the West to believe it is incompetent, it can suddenly reveal advanced capabilities like hypersonic missiles and drones manufactured far faster than Raytheon - so apparent incompetence and false propaganda cut both ways.

factualspeaker onlynovelty 2/4durability 2/4· Dylan Patel

“If I'm the Chinese military, I want the Western world to believe I'm completely incompetent because one day, I can just destroy everything with hypersonic missiles and drones.”

0.30

There is more opportunity today than any time in human history because there are a thousand under-optimized abstraction layers far from Pareto-optimal; the right move is not picking one 'correct' spot but finding where you can vibe and work hard - whatever your passion (copper wires, B2B SaaS) - and applying AI to make that layer more efficient, because working hard at something you're interested in plus AI leverage reliably produces success.

normativespeaker onlynovelty 1/4durability 3/4· Dylan Patel

“There's more opportunity today than any other time in human history, in my view. Just go out there and try. What engages you? Because if you're interested in it, you'll work harder.”

0.30

Locating AI data centers in power-rich places like Ethiopia (where the Nile dam generates over a gigawatt but the country uses ~400MW) is blocked not by power or technical feasibility but by genuine security risks - you don't want your IP/AGI somewhere you can't monitor technicians, and an adversary could destroy every GPU by tampering with the grid.

causalspeaker onlynovelty 3/4durability 2/4· Dylan Patel

“There are true security risks. If you're China, or if you're a US lab, to build a data center with all your IP in Ethiopia… You want AGI to be in Ethiopia?”

0.26

Dylan built SemiAnalysis from a Reddit/forum shitposting hobby into a leading research firm by going to ~40 technical conferences a year across every siloed layer of the stack, bridging economics (reading earnings since age 18) with technical detail, hiring specialists across the stack, and selling granular supply-chain data validated by hyperscaler, semiconductor, and investor customers.

factualspeaker onlynovelty 1/4durability 2/4· Dylan Patel

“I go to about 40 conferences a year, not trade show type conferences, but technical ones: chip architecture, photoresist, AI and NeurIPS, ICML, and so on.”

0.24

Asianometry grew from a tourist channel made for his mother into the leading semiconductor/business-history channel simply by covering whatever genuinely interested him (zippers, oil industries, ASML), laboring for three years with ~200 views before it took off in 2018-19, while simultaneously running a full-time job and a textile business.

factualspeaker onlynovelty 1/4durability 2/4· Jon Y

“Asianometry is always just stuff I'm interested in. ... I labored on for the first three years with no one watching. I got like 200 views”

0.13

Jon admits he still feels like a fraud and doesn't personally believe he 'knows' semiconductors, reflecting how the field's depth means even experts only command narrow slices.

factualspeaker onlynovelty 0/4durability 1/4· Jon Y

“I still feel like a fraud. I sit in front of Dylan and I feel like a legit fraud”