Ep. 044 - Is NVIDIA Over? EMERGENCY Pod with ChinaTalk | Dylan Patel, Jordan Nanos, Jordan Schneider
Topics
US-China AI Rivalry and Geopolitical Controls
Open Source vs Closed Source AI
Physical Bottlenecks in AI Clusters
Mindset and Psychology of Tech Founders
Pioneers
Dylan Patel
Summary
- The episode frames Jensen Huang’s reported departure as an export-enforcement crisis with consequences beyond NVIDIA. The panel links the crisis to alleged chip diversion, but Nanos distinguishes circumstantial evidence from proof of Huang’s knowledge; Patel qualifies his theory that the board has evidence with “If I had to guess.” Legal overseas rentals and illegal physical smuggling remain separate questions in their account.
- Patel argues that remotely switchable U.S. compute gives Washington valuable leverage over Chinese AI. A blanket cutoff could provoke retaliation against supply chains American servers still depend on; renting through U.S.-controlled clouds preserves the ability to “turn it all off.” Schneider disputes assumed Chinese escalation dominance and thinks Beijing might welcome RASA as pressure for domestic substitution.
- Chinese GPU rentals can increase American profits while keeping Chinese competitors’ compute expensive. Patel estimates $7 million of NVIDIA profit per megawatt plus $7 million–$10 million for the cloud provider at current prices; Meta could potentially operate at one-quarter of ByteDance’s compute cost. He also puts American open-model application ARR “well north of $10 billion,” while Schneider asks whether scarce gigawatts would create more value at U.S. frontier labs.
- MATCH draws the clearest consensus because equipment controls constrain China’s manufacturing scale. The panel favors restricting advanced DUV alongside EUV; Patel’s jigsaw analogy is to “blow up the board” instead of leaving one missing piece China can engineer. Schneider says Ascend is comparable to Hopper and describes Chinese developers supporting domestic accelerators, making the equipment bottleneck increasingly consequential.
- An American open-model champion requires concentrated talent, compute and data beyond a large budget. Patel rates Moonshot “probably just as good as the US labs” skill-wise, with compute and deployment deficits; Schneider adds data. NVIDIA’s Nemotron spending failed to beat Chinese open models, supporting Patel’s answer to the few-billion-dollar fix: “It’s not just money.”
- Schneider gives NVIDIA a high chance of remaining healthy for two to four years, while Patel sees execution risk compounding. Supply-chain dominance and the Rubin roadmap support the near term, but executive retention is Schneider’s one-year test. Patel’s downside case includes slower yields, roadmap delays and a possible loss of chip leadership and supply-chain majority in 2029: “Hardware progresses slower without Jensen at the helm of NVIDIA, period.”
- Succession exposes unresolved questions about technical leadership and state influence. Patel warns of possible “civil war” between finance, technology and sales, with finance in charge. Schneider favors Colette Kress, while Nanos hopes she does well. Schneider’s analogy to NVIDIA’s “nationalization” draws categorical rejection from Patel and Nanos, who characterize the intervention as enforcing export law while leaving lawful chip sales and public advocacy available.
Deep dive
1. NVIDIA’s leadership shock becomes an argument about enforcement
- The speakers take Huang’s departure as their starting point, then separate personal-conduct allegations from alleged chip diversion. For Schneider and Patel, smuggling is the more consequential issue; Nanos says the evidentiary test is Huang actually knowing and being complicit. Patel’s inference stays an inference: “If I had to guess, the board has evidence on this.”
- Schneider describes an alleged Supermicro scheme: a Taiwan employee using a hairdryer to transfer server stickers so fake machines could pass an audit in September of the previous year. Schneider says the employee is cooperating with prosecutors. Nanos asks how to explain NVIDIA’s cessation of Singapore revenue reporting; Schneider maintains that establishing Huang’s knowledge is the necessary next step.
- Patel insists overseas GPU rental is “fully legal,” drawing a boundary between Malaysia-based cloud access and chips physically smuggled into China. Nanos frames NVIDIA’s choices as a conflict between the “letter of the law” and the “spirit of the law.”
- Nanos and Schneider say the Biden-era diffusion framework was left unenforced rather than formally removed; they describe BIS guidance, not a conventional statute. Schneider points to a May 2025 footnote promising replacement rules that never arrived. Patel is surprised: “I thought we would have to repass them.”
2. Patel wants Chinese compute dependence to remain switchable
- Patel’s national-interest framework has two goals: maximize America’s AI lead and control Chinese access. A blanket ban on ownership, deployment and rentals, he argues, could trigger retaliation against components embedded in NVIDIA, Google and Amazon servers. He cites NVIDIA’s largest PCB supplier being Chinese and Washington’s repeated unwillingness to absorb supply-chain pain: “We clearly don’t have the stomach to disentangle the supply chain.”
- His preferred equilibrium is the least Chinese compute compatible with avoiding a supply-chain rupture, under maximum American oversight. U.S.-controlled hyperscalers could rent GPUs abroad, keep data outside China and cut access during escalation: “You can turn it all off.” Closing that route would increase incentives to smuggle and build domestic alternatives, while removing Beijing’s reason to preserve the existing arrangement.
- Schneider’s pushback: Chinese leaders may welcome RASA as leverage over firms reluctant to abandon NVIDIA, and China’s “escalatory dominance” should not be assumed. He cites the absence of comparable retaliation after controls in 2022, 2023 and 2024. Patel replies that those rules contained loopholes and later trade fights ended in compromise because Washington would not stomach the pain; he says the pair are “generally pretty close.”
3. China’s rental bill can subsidize America’s AI ecosystem
- Patel puts the economics per megawatt at roughly $7 million of NVIDIA profit plus another $7 million–$10 million for the cloud provider at current prices. Requiring Chinese companies to rent rather than buy preserves that margin stack for America. In his ByteDance–Meta example, Meta could potentially have one-quarter of ByteDance’s compute cost because the American competitor avoids some of those markups.
- Schneider asks whether the better use of Malaysia-bound gigawatts is simply giving OpenAI, Anthropic and DeepMind more capacity to improve models and earn money. His counterexample to commercial accommodation is TikTok spending billions on Instagram ads: receiving a rival’s money does not settle whether letting that rival compete is strategically wise.
- Patel explicitly casts his response as devil’s advocacy: American application companies using open models generate “well north of $10 billion” in aggregate ARR. He says DeepSeek earns less than U.S. serving companies such as Baseten and Fireworks. “America’s making so much money off of the Chinese models,” allowing American businesses to accrue revenue and invest more.
- Schneider adds the engineering relationships: China-based NVIDIA and AMD teams bring up GPUs and work on low-level hardware and software, while Chinese-born researchers contribute at U.S. frontier labs. Escalation could sever relationships that are difficult to restore: “I think it’s a one-way street.” Schneider also says he learns from people in China. The panel discusses whether the U.S. should accept more restrictions in order to widen its lead.
4. MATCH targets the equipment China needs to scale
- Schneider describes MATCH as extending restrictions to advanced DUV lithography alongside EUV, which he says has not been able to get into China. He wants Japanese, Dutch and Korean suppliers prevented from filling gaps left by American equipment controls. Here the panel converges; Patel’s closing instruction is blunt: “Pass the fucking MATCH Act.” The shared objective is constraining China’s ability to manufacture advanced chips at scale.
- Patel’s analogy is a jigsaw puzzle: “You remove one piece, and they’ll engineer that piece. If you blow up the board, it’s harder.” His prescription is to block lower-value production inputs and force China to purchase increasingly expensive outputs: equipment gives way to chips, chip ownership to rentals, and rentals to purchased AI. That raises China’s cost disadvantage in a scale-driven competition.
- Schneider says substitution is happening regardless of RASA. In February, GLM supported seven domestic accelerators: Huawei, Moore Threads, Hygon, Cambricon, Kunlun, MetaX and Enflame. DeepSeek published a megakernel with Huawei Ascend for V4; Meituan claims LongCat was trained entirely on Ascend. His comparison is pointed: “People don’t even train on AMD,” while Chinese developers support a broad domestic stack.
- Schneider judges the model-capability gap wider but the chip gap “as close as it’s ever been,” comparing Ascend with Hopper and claiming Huawei’s Kirin phone chips are as good as or better than Apple’s. He returns to volume: making 1,000 chips is different from making 100,000. Useful economic deployment requires megawatts of advanced hardware; equipment denial constrains that scaling.
5. A few billion dollars cannot guarantee an open-model champion
- Schneider asks why an American player has not spent a few billion dollars to beat Kimi or DeepSeek and open-source the result. Patel’s answer: “It’s not just money.” Talent is concentrated at Anthropic, OpenAI, DeepMind and potentially Meta, with the remainder scattered across a hundred startups. He rates Moonshot “probably just as good as the U.S. labs” skill-wise, with deficits in compute and deployment; Schneider adds data.
- Schneider points out that NVIDIA already tried: its spending on Nemotron did not produce a model better than the Chinese open-source alternatives. “You need compute, people, and data.” Patel’s conclusion is that engineering and research quality at Chinese labs is extremely high; spending more does not automatically reproduce that capability.
6. NVIDIA’s moat can endure while its execution slows
- Schneider rejects the leap from Huang’s exit to NVIDIA going to zero. Near-term orders require execution against an established roadmap for Rubin and Rubin Ultra; supply-chain control, sales and support, neocloud ties, lab investments and balance-sheet backing remain. “It’s still NVIDIA’s to lose.” He sees a “very high chance” the company is still doing fine two, three or four years out.
- His one-year test is retention. NVIDIA has had few high-profile executive departures; he singles out Rene Haas’s move to Arm. Now venture capitalists can offer billions for new chip companies or labs. A year of departures would make him “really concerned”; a team still moving together under Colette Kress or an emerging technical leader such as Ian Buck would make him more confident.
- Patel’s pushback is immediate execution risk: Rubin manufacturing problems and yield ramps could take longer to resolve as decisions move through committees. SemiAnalysis itself has needed to email Huang to get open-source software issues fixed. “Things get solved more slowly. Things get solved by committees. Things don’t get escalated up.” Slower adaptation to DeepSeek and Moonshot software, delayed Rubin Ultra changes, or weaker supplier allocations could each erode shipment volumes and performance.
- His best example is Volta: Tensor Cores arrived three months before tape-out because Huang demanded a change engineers said required six months. Multiply such interventions across a hundred decisions, Patel argues, and the next architecture, Feynman, could arrive later and less aggressively. In 2029, NVIDIA might lack both the best chips and a majority of the supply chain. His categorical conclusion: “Hardware progresses more slowly without Jensen at the helm of NVIDIA, period.”
7. Succession leaves an unresolved fight over autonomy
- Schneider stresses that 2026 has more capable competitors than 2022 and asks whether departing executives’ new bets could benefit the Western ecosystem. Nanos invokes Apple after Steve Jobs to show that losing a founder need not destroy shareholder returns. Patel favors Charlie Kawwas’s technical-business balance; Schneider says he favors Colette Kress. Patel warns of possible “civil war” among finance, technology and sales, with finance now in charge.
- Schneider says he cannot fathom Kress being unaware of alleged chip diversion and predicts congressional investigations. Patel acknowledges, “We don’t know how deep this web goes.” Nanos counters that immediately appointing a fully complicit insider would be “really strange” after intelligence briefings to the board. The exchange offers conflicting inferences about executive knowledge; it does not establish Kress’s involvement.
- Schneider compares the intervention with Jack Ma disappearing and argues they are watching NVIDIA’s “nationalization.” Patel and Nanos reject that characterization outright, framing it as export-law enforcement. Patel says Huang could sell H20 or H800 within the rules; Schneider says lawful advocacy remained available. The disagreement survives the closing banter: “There’s so much between ‘don’t break the law’ and ‘nationalized.’”
Verification Notes
The transcript uses anonymous “Speaker 0” and “Speaker 1” labels in several stretches. Attributions in those stretches are softened where the speaker is not explicit.