"Jensen Huang's Exit Could Change the AI Industry Forever" - Dylan Patel
Topics
Mindset and Psychology of Tech Founders
Open Source vs Closed Source AI
Trillion-Dollar AI Capex and Bubble Risks
US-China AI Rivalry and Geopolitical Controls
Rogue Autonomous Agents and Recursive Self-Improvement
AI Training Data and Evaluation Benchmarks
Pioneers
Dylan Patel
Summary
- A Jensen Huang exit would put Nvidia’s decision speed, supply-chain leverage and customer financing under pressure simultaneously. Discussing allegations around Huang’s China relationships, Patel argues that replacing his ability to coordinate those functions would be difficult. Moving decisions to committees means “that’s just gonna make decisions slow.”
- Outside Nvidia, Patel sees open source as the biggest loser. Nvidia backs American open-model developers, while tighter China restrictions could threaten another major source of open weights. Anthropic looks relatively advantaged: Patel puts its Nvidia exposure at one-third or less, versus more than 90% for OpenAI, and concludes, “In general, the closed source is winning.”
- The venture risk extends from model availability to the financing required to secure compute. Patel asks how inference providers grow from $1 billion to $10 billion in revenue without credit, a backstop and improving open models. He calls Huang “the venture ecosystem’s strongest ally,” while questioning whether Nvidia will remain as active.
- Patel favors tighter China hardware controls while preserving some cloud access. He suggests Google renting chips to ByteDance could retain American control and extract margins, while floating possible limits on some chip companies’ rentals. Data generated with frontier models could face scrutiny as another channel for transferring capabilities. Retaliation remains a constraint because American AI servers contain Chinese components: “China can stop it if it wants.”
- Patel has become more receptive to AI 2027 while retaining doubts about diffusion. He says roughly 70% of its verifiable predictions so far that year proved accurate, despite his initial view that it was “a big stretch.” Recursive self-improvement is “looking more and more true,” but China’s response introduces supply-chain risks.
- GDPX shows sharp coding and chip-design gains alongside much weaker progress elsewhere. A host tentatively identifies the tested model as “Mythos 5.2” and cites roughly 16%; Patel emphasizes domain-specific reinforcement-learning environments and data. His call on valuable data suppliers: “It’s experts. It’s experts.”
Deep dive
1. Huang’s departure would expose Nvidia’s coordination problem
- The hosts frame allegations around Huang’s China relationships as a “black swan,” then question Nvidia’s deal pipeline without him. One had assumed Huang would prevent any AI company valued above $10 billion from failing for at least the next year, partly to avoid a damaging headline. Patel worries about decisions migrating to committees: “that’s just gonna make decisions slow.”
- Asked whether Huang actually designs architectures, Patel says he does not; his role is committing Nvidia among competing options and switching “on a dime” if one fails. His example is Volta: Tensor Cores were added about three months before tape-out. Deferring them to the next generation, he argues, would have prevented Nvidia from establishing AI dominance.
- Supply-chain leverage is equally personal: Patel doubts a successor could secure 60%–70% of a supplier’s capacity below market rates. Asked about Nikesh Arora, he questions whether Palo Alto hardware experience translates to competing against Google, Amazon, AMD and frontier-lab chip teams. “These guys are fucking good.”
- Internally, Patel favors Ian Buck and Jonah Alben, with Alben closest to the technical leader; promoting either could dilute work they already own. His preferred outsiders are Broadcom’s Charlie Kawwas and Apple hardware chief Johnny Srouji for their technical and supply-chain command. Hock Tan or Lisa Su would be better, he says, but he does not think either would take the role.
2. Open models and venture financing share Nvidia’s backstop
- Patel says American open source “would not exist without Jensen,” pointing to Nvidia’s backing of Reflection and Thinking Machines. Applications and inference providers also depend on Chinese open models. If restrictions tighten, he questions whether those models keep progressing and whether developers keep publishing everything: “I think the biggest loser is the open source ecosystem for sure.”
- Hardware diversification changes the exposure. Patel says Zhipu supports seven Chinese accelerators and DeepSeek works closely with Huawei; Nvidia-dependent Chinese labs look weaker than diversified peers. The same comparison favors Anthropic over OpenAI: he puts OpenAI above 90% Nvidia and Anthropic at one-third or less. Weaker open-source competition provides Anthropic another benefit.
- Patel agrees venture investors are leaving the major-lab trade as those companies approach public markets; OpenAI “probably just did their last round.” Downstream, financing remains essential. He asks how Baseten, Fireworks and peers scale from $1 billion to $10 billion in revenue without compute access, credit and a backstop—or continually improving Chinese models.
- The application-layer distinction, in Patel’s framing, is control over the underlying model. Companies such as Harvey and Sierra forced onto closed models become wrappers; using open models while owning the customer relationship gives them another position. Huang is “the venture ecosystem’s strongest ally,” but Patel does not know whether Nvidia’s investing will remain as active.
3. China restrictions could extend from hardware into training data
- Patel declines a Mercor-specific prediction but says many data companies sell heavily to China. Their pipelines use frontier models to generate, extract and filter material, including environments. Selling that output transfers information indirectly: “It’s not like traditional distillation.” He expects greater scrutiny of chip sales and rentals, with data potentially joining an area previously focused on hardware.
- His preferred restrictions target chip shipments, TSMC and memory-company sales, and advanced chipmaking equipment, with possible data restrictions. He suggests not broadly restricting chip rentals, while floating possible limits on chip companies such as Megaspeed; he then qualifies that suggestion. Google could withdraw chips from ByteDance when needed or instructed, while charging margins that make ByteDance less profitable relative to Meta. The objective: “how do you make America win.”
- On the hosts’ “honeypot” framing, Patel says he does not think the woman involved is a spy or honeypot “in the colloquial sense,” describing a well-connected business figure. He argues such connections can bring pressure to cooperate with intelligence agencies. Whether Huang broke laws remains unresolved in his account: “I don’t know.”
4. AI 2027 looks more plausible, but China complicates the trajectory
- A host sees capabilities tracking AI 2027 closely and asks about its November 2026 prediction that China “wakes up.” Patel says he originally thought the scenario “was a big stretch,” despite being highly bullish on AI. Roughly 70% of its verifiable predictions so far that year proved accurate, he says. He remains skeptical about diffusion, while recursive self-improvement looks increasingly plausible.
- Another host pushes back that leaving GPUs in Hong Kong could suggest longer Chinese timelines. Patel sees China waking up as a significant likelihood, but warns against an indiscriminate restrictions pendulum: every Nvidia server and Google TPU server contains Chinese content. Beijing could stop that supply, raising the question of what minimum access might prevent retaliation.
5. GDPX makes specialist data central to further progress
- A host tentatively calls the model “Mythos 5.2” and cites roughly 16% on GDPX. Patel says testing Anthropic’s model took about two days and revealed major software-engineering and chip-design improvements. He initially describes other areas as showing no improvement, then qualifies that it “did improve slightly in all domains.”
- Patel’s explanation is that reinforcement-learning environments determine where capabilities scale: without unique data and environments, particular domains will not improve. Anthropic’s chip effort illustrates the shift. Previously, he says, building a chip looked too slow relative to AGI; now “the model can make the chip.” Scaling size alone will not improve every benchmark: “You really have to scale techniques, algorithms, and data.”
- Asked about data obtained through product harnesses versus purchased from brokers, Patel says Anthropic buys very little chip-design data “as far as I understand,” instead building it internally. Biology uses a mixture: its own wet lab collects data while the company also buys substantial external data. Different domains therefore support different internal-versus-external sourcing strategies.
- Generic vendors connecting labs to merely adequate labelers look less valuable as models improve. “It’s experts. It’s experts.” Patel sees room for a strong business producing specialist material-science data, for example, while leaving its eventual scale unresolved: “how scalable is that actually gonna be?”