Feb 5, 2026 · 1h 16m · mad
Dylan Patel: NVIDIA's New Moat & Why China is "Semiconductor Pilled”
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews SemiAnalysis founder Dylan Patel on the macroeconomics, technological shifts, and geopolitical forces shaping the artificial intelligence revolution. Their discussion covers Nvidia's hardware and software defensive moats, US-China semiconductor rivalries and trade controls, data center power grid constraints, and the transformative economic impact of AI coding automation.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 10.3% of the talking time here. How this is scored →
Matt as informed peer 3.7 Guest teaching 5.5 Guest disagreement 2.3 Matt pushing back 2.1
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:03 Rejection of circular debt narrative
Dylan forcefully rejects Matt's suggestion that debt guarantees and circular financing in AI infrastructure create systemic fragility, calling public anxiety a narrative where none should exist.
Hardest push from Matt ▶ 1:03 Host challenges circular deal fragilityMatt directly challenges Dylan on the unnerving financial fragility created by large tech players guaranteeing debt and forming circular supply commitments across compute providers.
Biggest teaching moment ▶ 0:58 Burger metrics water debunkingDylan uses granular research to reframe data center water usage myths, revealing that Elon Musk's giant Colossus cluster consumes no more water than 2.5 In-N-Out burger locations.
Matt holds his own ▶ 0:50 Model progress as lagging CapEx indicatorMatt demonstrates sharp economic insight by framing AI model capability leaps as lagging indicators of previous hardware CapEx cycles rather than real-time achievements.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| The MAD Podcast Title Card | 3 | 6 | 1 | 1 | Matt sets up the discussion by asking about Nvidia's acquisition and licensing deal with Grok and what it signals about GPU strategy. Dylan explains autoregressive model architectures, decode vs pre-fill trade-offs, and Nvidia's move to cover broad surface area. | |
| Antitrust Concerns and Startup Acquisition Structures | 4 | 5 | 2 | 2 | Matt asks if structuring startup acquisitions as licensing deals is good for the market given antitrust scrutiny. Dylan discusses regulatory limbo, Jensen Huang's Andy Grove mentality, and competing AI chip startups. | |
| The CUDA Moat, Networking, and Open Source Frameworks | 4 | 7 | 2 | 1 | Matt inquires whether CUDA and Mellanox networking remain durable moats. Dylan educates Matt on open-source frameworks like VLLM and SGLang, explaining how PyTorch compilers and KV cache management are redefining software moats. | |
| AMD's Competitive Position and Market Limits | 3 | 6 | 2 | 1 | Matt asks if AMD can catch up to Nvidia. Dylan explains the hardware leapfrogging cycle, software lag, single-digit market share caps, and the extreme risk profile for hardware startups. | |
| Multi-Silicon Workloads and Custom ASICs | 4 | 5 | 1 | 1 | Matt synthesizes the concept of a multi-silicon world where customers use different specialized chips. Dylan expands with concrete examples like Anthropic's language focus versus Midjourney's video compute requirements. | |
| Impact of US Export Restrictions on Nvidia and Chinese Workarounds | 4 | 6 | 2 | 2 | Matt asks about Nvidia's declining revenue share in China and whether Huawei or US restrictions are responsible. Dylan details ByteDance renting huge clusters in Malaysia as a workaround for US export controls. | |
| China's Semiconductor-Pilled Culture and Local Municipal Competition | 3 | 5 | 2 | 2 | Dylan describes Chinese cultural enthusiasm for semiconductors, including TV romance dramas set in fabs. Matt jokes about Western pop culture comparison and contrasts US vs Chinese municipal subsidy dynamics. | |
| Global Semiconductor Specialization and Supply Chain Interdependence | 3 | 7 | 2 | 1 | Dylan details the global supply chain interdependence in chips, invoking the 'I, Pencil' analogy and noting how smaller countries like Austria hold critical bottlenecks. Matt listens while adding examples like guitar manufacturing cities. | |
| Huawei's Technological Resurgence and Nvidia's CoreWeave Investment | 4 | 6 | 2 | 2 | Matt asks if China is missing key supply chain links. Dylan outlines lithography lags, Huawei's vertical integration strength, and Nvidia's strategic $2 billion CoreWeave investment in response to Google TPUs. | |
| Global Market Competition, Open-Source Software, and CUDA's Moat | 3 | 6 | 2 | 1 | Matt asks if Nvidia and Huawei are competing directly in neutral international markets. Dylan outlines domestic Chinese supply constraints, open-source software feedback loops, and the AI economic cold war. | |
| Evaluating the US CHIPS Act and Reshoring Complexities | 5 | 6 | 3 | 5 | Matt challenges Dylan on whether US CHIPS Act reshoring is delayed or hopeless given scale differences. Dylan pushes back positively, contrasting $50B US subsidies with $500B+ Taiwan investments while emphasizing tariff leverage. | |
| Dylan Patel's Controversial Social Media Post | 3 | 5 | 3 | 2 | Matt brings up Dylan's controversial Twitter joke about ICE. Dylan pivots to how public backlash against AI is intensifying due to power grid strain, Waymo deployments, and deepfakes. | |
| Evaluating AI Infrastructure CapEx and Bubble Concerns | 4 | 7 | 2 | 2 | Matt prompts Dylan on whether AI CapEx is in a bubble. Dylan breaks down unit economics ($100B revenue run rate, 50% gross margins, 5-year hardware depreciation) to argue current spending is justified. | |
| Lagging Indicators: How Past CapEx Drives Current Model Progress | 5 | 6 | 1 | 3 | Matt articulates that AI model performance is a lagging indicator of hardware CapEx. Dylan agrees and illustrates with a story about his roommate spending $10k in Claude tokens in one week to build a complete RTS game. | |
| The Financial Dynamics and Payback Period of AI Infrastructure | 4 | 6 | 3 | 3 | Matt questions the alignment between supply-side buildout and enterprise demand timing. Dylan explains 5-year capital payback math, then addresses grid capacity constraints and gas turbine alternatives. | |
| Debunking Data Center Water Consumption Myths via Burger Metrics | 3 | 7 | 3 | 1 | Matt brings up Dylan's recent analysis on data center water consumption. Dylan debunks environmental myths using 'hamburger metrics', showing Elon Musk's Colossus data center uses as much water as 2.5 In-N-Out restaurants. | |
| Investment Opportunities in IPPs, Gas Power, and Coal Plant Restarts | 5 | 6 | 3 | 3 | Matt asks directly about energy investment plays like Constellation nuclear vs Vistra IPPs. Dylan dismisses nuclear as too slow and shares a case study of a client restarting a coal plant for hyperscale off-take. | |
| Positive Externalities and Economic Benefits for Skilled Trades | 6 | 6 | 6 | 7 | Matt aggressively challenges Dylan on the perceived circularity and financial debt fragility of AI infra financing deals. Dylan forcefully rejects the premise, explaining that customer guarantees and compute asset backstops are standard financial structures. | |
| Tracking AI Model Progress and Coining 'Tokenomics' in Tech | 4 | 6 | 4 | 3 | Matt asks about software trends and teasingly accuses Dylan of stealing the term 'tokenomics' from crypto. Dylan jokingly expresses his disdain for crypto and details how Claude Code is replacing junior analyst workloads. | |
| Comparing Frontier AI Models and the Paradigm Shift of Claude Code | 4 | 6 | 2 | 2 | Dylan compares the technical stacks of OpenAI, Anthropic, and Google across RL and pre-training. Matt adds context from his previous interview with Boris Cherny regarding Claude Code writing its own product features. | |
| San Francisco Roommate Lore: Sholto and Dwarakash | 3 | 2 | 2 | 2 | Matt and Dylan share lighthearted San Francisco tech lore about living with AI personalities Sholto Douglas and Dwarakash Patel. Matt defines Dwarakash as the quintessential 'podcaster's podcaster'. | |
| Interview Outro and Farewells | 0 | 0 | 0 | 0 | Matt delivers a brief monologue concluding the episode and encouraging listeners to subscribe and leave reviews. |
Statements from this episode (66)
Patel: AI workload scale enables 10x performance gains from specialized chips
“But now the workload is so large that there is room for specialization that will give you 10 X increases in certain domains, right?”
Patel: Groq chips cannot cost-effectively perform general-purpose large model inference
“In a general purpose workload, crock. Grok doesn't work, right? You know, it can't train, it can't, you know, it can't inference really, really large models cost efficiently, right? You can't serve many, many, many users, but what it can do is it can go block,…”
Patel advised Enfabrica during its transaction with Nvidia
“This happened for a company. I was an advisor for NVIDIA acquired in Fabrica just maybe a few months before they did Grok and similar style of deal, right?”
Patel: Microsoft's custom Maia AI chip is not a credible competitor
“And then, you know, Microsoft's Maya is not credible, but like, you know, maybe it will be one day, right?”
Patel: Groq missed revenue significantly before being acquired
“In fact, they missed revenue last year significantly and yet they got bought, right? Because the value of the IP was there and the value of the team.”
Patel: Most AI chips will be consumed via open-source engines, not direct programming
“I think most AI chips will not be consumed by people programming anything for it. They will download an open source inference engine, they will download an open source model, and then they will put it on the, and it's really simple to download VLM and, like, m…”
Patel: AI coding startup costs are driven by pre-fill context, not decode tokens
“If you look at coding as an application and you like look at these coding companies and how much they're paying for prefill versus decode, actually majority of their cost is prefill tokens, not decode tokens, because the context is just so large and it's switc…”
Patel: All major non-Nvidia AI chips will fully support vLLM by mid-2026
“All of them will have a very good UX for download model, run model on VLM by The middle of the year, I think, right? Certainly AMD is already there by the end of this quarter.”
Patel: SemiAnalysis runs InferenceMax benchmark on $60M of donated GPUs
“It's called inferencemax. It's open source. All the code is and the results are, but we run across, I think, sixty million dollars of GPUs, which are donated to us by companies like Nvidia, AMD, OpenAI, Microsoft, Amazon, Crusoe, CoreWeave, Together AI.”
Dylan Patel: AMD will remain in single-digit percentage AI market share
“I don't think they'll Go beyond, like, I think they'll stay in single digits market share, single digit percentage market share.”
Patel: Chip startups cannot beat Nvidia playing Nvidia's game
“You're never going to beat Nvidia at their own game, right? They're going to have the supply chain on lock. They're going to get to the newest memory technology or process technology or whatever packaging technology, whatever it is, sooner than you.”
Patel: First-wave AI chip startups all bet on on-chip memory
“The first wave of AI hardware bets, Graphcore, Cerebris, Samanova Grok, where they all made the same bet on memory and putting the memory on the chip.”
Patel: New AI chip startups like Etched have under 1% success chance
“I don't know what a venture capitalist views as like likely chances of succeeding, but I think all of them are less than one percent.”
Patel: Meta operates two MTIA chip lines for recommendations and Gen AI
“Meta actually has two lines of AI chips. They're MTIA. There's a line that's focused on recommendation systems, and then there's a line that's focused on Gen AI.”
Patel: ByteDance develops custom chips for recommendation systems rather than Gen AI
“ByteDance also has a recommendation system line of chips, and it's not really focused on Gen AI”
China missed its 2020 and 2025 domestic semiconductor production targets
“So in 2015, they made these five-year plans for two, 20 20 and 20 20 five, where they set the percentage of semiconductors they wanted domestically produced. And they've missed the goal both times.”
OpenAI surpassed ByteDance as the largest GPU renter globally
“And so when you look at who rents the most GPUs in the world, it's three companies, right? So one of them is obviously OpenAI. Second one, actually they were bigger than OpenAI. They are bigger than OpenAI today, or no, they were bigger than OpenAI than OpenAI…”
ByteDance is taking over 1 GW of Oracle data center capacity in Malaysia
“It's instead being built in Malaysia, and Oracle has over a gigawatt of capacity in Malaysia that ByteDance is gonna take, right?”
Patel: Chinese TV dramas feature semiconductor fab and solar cell researchers
“There are dramas where people fall in love in the fab or dramas where people fall in love and they're photovoltaic, like solar cell researchers and engineers.”
Patel: Chinese local governments, not national, banned Nvidia's H20 and H200
“But as far as I understand, the national government has not banned Nvidia's H-twenty or H-two hundred, but the local ones have. Right. A lot of local ones have said, no, you know, you must use China manufactured chips.”
Patel: 15 to 20 countries could single-handedly shut down semiconductor manufacturing
“I would say there's like 15 or 20 countries that can shut down the entire semiconductor industry.”
Patel: China currently has the world's most vertically integrated semiconductor stack
“China has the most vertical stack and semiconductors today.”
Patel: US cannot build a fully independent fab even for 20-year-old tech
“America could not build a fully vertical fab without stuff from elsewhere, even if it's 20 year old tech.”
Patel: China will narrow its lithography gap to five years shortly
“Their lithography is like 10 years behind and I think it'll be five years behind in a couple of years, right?”
Patel: Nvidia is backstopping CoreWeave's land and power capital needs
“NVIDIA invested two billion in CoreWeave, but what's more important is that that's like sort of just like the sticker. What's really relevant is NVIDIA is going to work with CoreWeave to acquire and backstop and all these things, the land, the power, the energ…”
Patel: Huawei is the most vertically integrated company in the world
“Always the most vertical company in the world. No company is more verticalized than Huawei, which then leads to huge innovations.”
Patel: Huawei used shell companies to procure TSMC chips and Korean HBM
“Well, actually they were using shell companies to get chips from TSMC and using Different methods of like sneaking HBM, which is memory from, you know, Korea through Taiwan to China, right?”
Patel: AI software industry could hit $100 billion ARR this year
“I think the industry could hit a hundred billion ARR by the end of this year, like 45, 50 for open AI, like 35, 40 for anthropic.”
Patel: Without AI leadership, China will overtake the US as hegemon
“But without AI, China definitely will rise to be the global hegemony. They're just gonna outrun America.”
Patel: Cumulative semiconductor CapEx in Taiwan exceeds $500 billion
“The collective total amount of, like, CapEx that has been spent in Taiwan is, like, five hundred billion plus, right? Across the industry, across all the companies that are making semiconductors in Taiwan.”
Patel: TSMC is actively manufacturing chips in Arizona for Nvidia, Apple, and AMD
“TSMC is literally making chips for NVIDIA and Apple and AMD and others in Arizona today, right?”
Patel: U.S. CHIPS Act would not have passed without COVID car shortages
“Chips Act did not get passed, only got passed because that happened. And people are like, oh my God, the semiconductors are why cars can't be made. If that didn't happen, we wouldn't even have the Chips Act.”
Patel: Elon Musk is considering building semiconductor fabs due to AI chip shortages
“Even Elon's talking about building fabs now, because he sees the shortages in the world, right? There's a lot of semiconductor-related shortages for building out AI”
Patel: Microsoft Nebius NJ data center didn't drive power price hikes
“That data center has nothing to do with power prices going up. It's super storm standee, like five years ago, knocking or whatever, how many years ago, knocking down the state's electrical infrastructure and then the, then improving all these improvements. And…”
Patel: New Waymo vehicle costs dropped from $300k to $90k-$100k
“Their Waymo's went from like 300 K to like a hundred K or 90 K, the new Waymo car.”
Patel: Hyperscaler capex will reach roughly $500 billion this year
“Hyperscalers capex is going to be like five hundred billion dollars this year or something like this.”
Patel: AI model performance is a lagging indicator of prior hardware CapEx
“Ultimately the capex that Microsoft spent in 2024 for OpenAI is what results in 2025 for OpenAI or CoreWeaver or whoever is what results in their models being so good this year. Same with Anthropic and Amazon Google and their models now being so good now is th…”
Patel: Two percent of global GitHub commits are generated by Claude Code
“But two percent of GitHub commits today are cloud code.”
Patel: Annual global software wages total two trillion dollars
“Two trillion dollars of software wages paid in the world.”
Patel: AI is underearning the economic value it creates by a significant margin
“AI is under earning the value that it's producing in the world, right? By a significant margin already today.”
Patel: Engineer built an RTS game using $10K of Claude API
“He used, like, 10,000 dollars of Claude in one week and built an entire RTS from scratch about, like, but instead of, like, being a standard RTS where it's like, oh, Age of Empires where you advance through ages or Starcraft, it is an RTS where it's China vers…”
Patel: Only a few holdouts left writing code manually at Anthropic
“We have an indicator internally at Anthropic where you see how many people actually write code now. There's only a few holdouts left.”
Patel: Much of current AI infrastructure spending yields no immediate financial return
“That's what's happening today is that people are spending all this infra money on infra and there's no return for a lot of it, right? A lot of it is just doing research and like trying to get adoption and it's free users.”
Patel: The US has not built wholesale new power capacity in 50 years
“America's not built power for 50 years really, right? It's like converted from coal to gas and like things like this, but like really just have not built wholesale new power on a large scale.”
Patel: Data center power consumption will grow from 2% to 10% of US grid
“And then you've got data centers now all of a sudden coming online and going from two percent to 10% of the U S grid in just a handful of years.”
Patel: Data centers will use under 1% of US water by 2030
“So the U.S. Grid will get to, like, 10% of power by, like, 28, 27, is data centers. For water consumption, it's not even gonna crack one percent. By the end of the decade.”
Patel: xAI's Colossus uses as much water as 2.5 In-N-Outs
“I think the metric was the entirety of Elon Musk's Colossus data center, right? Uses as much water as two and a half in and outs.”
Patel: Meta's Louisiana data center will reach 4 to 5 gigawatts
“Meta's data center in Louisiana is getting protested because the water It's gonna be the largest data center in the world. It's gonna be like four or five gigawatts, at least announced so far.”
Patel: Independent power producers will profit from pairing power assets with data centers
“I think IPPs will do well. I think IPPs can secure contracts at premiums to what they've previously been able to for new power plants that are either dedicated or grid connected, but come with a pairing Of a grid load, right?”
Patel: Nuclear power is not relevant for short-term AI energy needs
“I'm not necessarily bullish nuclear. Existing nuclear, fine, yeah, it'll, it'll, it can find a higher buyer, higher priced buyer, but majority of it will be gas.”
Patel: SemiAnalysis advised a client on buying and restarting a coal plant
“Have clients would like, had a client buy a coal plant. And we were advising them on the transaction based on, they just like showed up and they're like, yeah, we want to buy power assets.”
Patel: Hyperscalers are funding public transmission grid upgrades
“Hyperscalers are paying for transmission grid upgrades, which people will benefit from.”
Patel: Electrician and plumber wages are skyrocketing from AI expansion
“Electricians wages are skyrocketing, you know, et cetera. Right. Like the plumbers wages are skyrocketing.”
Patel: Concerns over AI circular financing and debt are overblown
“I think it's completely fine, and I think, like, people are, like, freaking out and making narratives where there really shouldn't be one.”
Patel: CoreWeave secured GPU debt without backstops using Microsoft contracts
“In the case of core wave, they were actually able to no backstop, right? They were able to just say, Hey, look, here's our Microsoft contract for this many GPUs.”
Patel: Over 99% of OpenAI's spend is likely compute
“99 plus percent of their spend at the company is probably just compute.”
Patel: NVIDIA CEO Jensen Huang adopted the term 'tokenomics'
“Yeah, yeah, and Jensen's used it now, so I've like, I've convinced him to use the word. He's used it as sovereigns, and so I think we've won.”
Patel: Companies no longer need junior analysts due to AI automation
“This is Claude code now. You don't need junior analysts.”
Patel: Many tech companies have stopped hiring L4 software engineers
“Just like a lot of companies have stopped hiring L four engineers because it's useless.”
Patel: OpenAI's next model will outperform Opus 4.5 around February-March
“OpenAI's new model, I think, will be better than Opus 4.5, and it's coming, like, somewhat soon in March-ish timeframe, maybe February, March-ish, but”
Patel: OpenAI has a better RL stack than Anthropic, but inferior pre-training
“Because OpenAI has a better RL stack than Anthropic today, it's just their pre-trained models suck compared to Anthropic's pre-training, right?”
Patel: Google has better pre-training than OpenAI or Anthropic, but worse RL
“Flip side, Google has a better pre-trained model than Anthropic or OpenAI, but their RL stack sucks.”