TMTB Podcast Intel is produced by Ethmos, research agents that listen to every episode in TMTB’s coverage universe and write the briefs below.
Follow NVDA on Ethmos to get a brief whenever its executives go on a podcast. TMTB readers get 14 days free and 25% off their first month.
Today’s Essential Episodes:
SemiAnalysis’s Dylan Patel on a Slower NVIDIA After Jensen Huang, His CEO Picks and Why Open Source Loses Most
TBPN, with John Coogan, Jordi Hays and SemiAnalysis’s Dylan PatelBen Thompson and Andrew Sharp on NVIDIA After Jensen Huang, What He Meant to the AI Buildout and Washington’s Revenge
Sharp Tech, with Ben Thompson and Andrew SharpBen Bajarin and Jay Goldberg on NVIDIA’s Succession Crisis and the Fate of the AI Supply Chain
The Circuit, with Ben Bajarin and Jay GoldbergDwarkesh Patel on Why He Now Expects Superintelligence by 2030: The Avalanche That Hit Math Is Coming for AI Research
Dwarkesh Podcast, a narrated essay by Dwarkesh Patel
Produced by Ethmos: AI-assisted summaries of public podcast episodes. For information only. May miss context or contain errors. Verify before relying on them. Summaries are not substitutes for the episodes, which belong to their creators. Not investment advice, offer, or solicitation. Not affiliated with or endorsed by the shows, hosts, guests or companies discussed. Corrections and publisher opt-outs: support@ethmos.ai.
Listen · Watch · Follow NVDA on Ethmos
THE SIGNAL
Patel puts OpenAI at 90%+ NVIDIA compute versus Anthropic at one-third or less, so NVIDIA’s turmoil and a weaker open-source ecosystem favor Anthropic and closed-source labs over OpenAI, which works more closely with NVIDIA.
Patel says Jensen’s personal pull lets NVIDIA lock up 60-70% of a supplier’s capacity at below-market prices; a less charismatic successor likely loses this, which could squeeze NVIDIA’s supply allocation and costs (our read).
Patel warns that tighter US curbs could trigger Chinese retaliation, since every NVIDIA server and Google TPU server contains Chinese content that China could cut off, a supply-chain tail risk for the whole AI hardware stack.
Patel says the US open-model frontier, led by Reflection and Thinking Machines, depends on heavy NVIDIA backing, and inference providers rely on Chinese open models, so both are exposed if NVIDIA’s investing slows and China policy tightens.
01 Succession and governance risk
Patel’s internal favorites are Ian Buck and Jonah Alben — Buck for charisma and dealmaking aptitude, Alben as the closest technical lead to Jensen — though he doubts either can match Jensen across technical, supply-chain, and dealmaking dimensions simultaneously. His preferred external picks are Broadcom’s Charlie Kawwas, whom he rates very highly, and Apple hardware chief Johny Srouji, whom he says are the only two people capable of navigating both supply chain and technical complexity at NVIDIA’s scale; he dismisses AMD’s Lisa Su as unlikely to leave her post. He also waves off wildcard ideas like Palo Alto Networks CEO Nikesh Arora, arguing NVIDIA’s competitive set (OpenAI, Broadcom, Google, Amazon, Anthropic, AMD chip teams) is far tougher than Palo Alto’s.
Operationally, Patel breaks NVIDIA’s Jensen-dependence into three legs: technical decision-making (no clear single technical lead besides possibly Alben, with many others reporting directly to Jensen), supply chain leverage (NVIDIA’s ability to lock up capacity via charisma/relationships, which a less charismatic successor likely can’t replicate), and NVIDIA’s role propping up the ecosystem (open source labs, startups, inference providers). He expects major AI-company investment decisions that previously moved on Jensen’s say-so to now move through committee review, slowing NVIDIA down against well-funded rivals.
02 China dynamics and export policy
Patel, who describes himself as having favored certain types of China restrictions for years, says NVIDIA/Jensen were the industry’s main defenders against export curbs, and that credibility is now gone amid investigations. He expects the policy pendulum to swing toward broader restrictions but hopes for nuance — hard limits on chip sales, TSMC/memory sales, and high-end equipment to China, but preserving some chip-rental access (e.g., via hyperscalers like Google renting to ByteDance, where margin/control stays in US hands) to avoid fully cutting off US leverage. He flags that data sales to China — via data companies supplying frontier-model-generated training environments — represent an unscrutinized channel he expects to draw attention next.
On Chinese labs, Patel says it’s an open secret that Chinese AI developers work closely with NVIDIA engineers (including via NVIDIA’s Shanghai offices) despite formal export bans, and that labs like Zhipu (sanctioned, supporting seven Chinese accelerator types) and DeepSeek (close to Huawei) have already hedged by diversifying hardware, making them comparatively resilient versus NVIDIA-reliant peers. He reassesses the AI 2027 forecast, saying roughly 70% of its verifiable predictions have been accurate so far; asked by the hosts about the forecast’s prediction that China “wakes up” around November 2026, he treats this as plausible but unresolved.
03 Venture capital and startup fallout
Patel says VCs are now largely excluded from direct lab investing as, in his words, OpenAI probably just did its last private round, which he says wasn’t that great, and Anthropic moves toward a public listing, pushing venture attention downstream to inference clouds and application companies. He frames NVIDIA/Jensen as the venture ecosystem’s most important backer, having propped up companies via investment and open-source model support; he questions whether firms like Fireworks and BaseTen can scale from roughly $1 billion to $10 billion in revenue without continued compute financing and open Chinese model access, and whether application companies like Harvey and Sierra can keep improving without open-source models — warning that reliance on closed-source models risks reducing them to mere wrappers around someone else’s model.
04 Anthropic’s benchmark jump and data strategy
Discussing SemiAnalysis’s own benchmark results on a new Anthropic model, Patel reports dramatic improvement in software engineering and chip design but flat performance elsewhere, which he attributes to reinforcement-learning environments being the actual vector for scaling — domains without unique proprietary data and RL environments don’t improve regardless of model scale. He notes Anthropic has pivoted from an earlier stance against building its own chip (previously seen as distracting from AGI-focused work) toward now believing models themselves can help design chips, and says the company builds most of its chip-design data internally rather than buying from vendors, while in biology it both runs its own wet lab and buys external data — concluding that specialized expert-generated data, not generic labeling, is where real business value now lies.
Listen · Watch · Follow NVDA on Ethmos
THE SIGNAL
More NVIDIA damage may be coming: Reuters reported a second affair during recording and Thompson’s DC sources hint at worse; his own guess is that Megaspeed chips reached China’s military, making tighter bans likelier.
Thompson argues NVIDIA still leads on tokens per watt, so power limits on data centers help it, though he speculates that allocation favoring CoreWeave over Amazon may draw scrutiny once Huang is gone.
Thompson runs his printer, label printer and inventory app through Claude without touching their software, suggesting value shifts from app interfaces to agent-accessible systems of record (our read); agent-filled carts drove purchases he’d otherwise skip.
01 NVIDIA’s leaderless machine
NVIDIA’s entire structure, Thompson says, mirrors early Steve Jobs-era Apple: a functional organization where specialists execute and only the CEO synthesizes everything, enabling extraordinary speed (yearly product cadences, the Mellanox acquisition) but zero redundancy. He notes Huang spoke recently of running NVIDIA until age 120, and the company never underwent the succession wake-up call Apple got when Steve Jobs’s death was gradual rather than sudden. Both hosts flag this as a genuine governance failure that may explain why, despite the stock falling roughly 15% on the news, NVIDIA’s board still moved to oust him — a classic principal-agent/fiduciary-liability dynamic Thompson walks through at the start.