NVIDIA: OpenAI, Future of Compute, and the American Dream | BG2 w/ Bill Gurley and Brad Gerstner
Key insights
Companies
- OpenAI - Central subject; Nvidia's new deal makes it a preferred compute partner and invests up to $100B as OpenAI self-builds AI infrastructure for the first time
- Microsoft / Azure - Existing multi-year buildout partner for OpenAI workloads, continuing alongside the new self-build arrangement
- Oracle (OCI) - Building out 5-7 gigawatts of contracted OpenAI/SoftBank capacity
- CoreWeave - Third leg of existing OpenAI compute buildout; also an Nvidia investment
- SoftBank - Partner in the OCI/Stargate buildout for OpenAI
- Nvidia - Huang's own company; frames it as an AI infrastructure company, not just a chipmaker, building six to seven new chips per year on an annual cadence
- xAI - Cited as an Nvidia equity investment and as Elon Musk's Colossus 1/2 cluster buildout example of extreme-speed data center construction
- Meta - Zuckerberg's comment that Meta could overspend by $10B and still consider it worth the existential risk; also an example of the CPU-to-GPU recommender-engine transition
- Google / Alphabet - TPU program cited as the one credible ASIC challenger, built on 'foresight' starting TPU1 before the market existed; now on TPU7
- Intel - New Nvidia partnership (NVFusion) fusing Intel's enterprise CPU ecosystem with Nvidia's accelerated-computing stack, despite Intel historically trying to put Nvidia out of business
- Alibaba - Eddie Wu comments cited on 10x data center power growth by end of decade and token generation doubling every few months
- Huawei - Cited as the beneficiary of the US China chip export ban, using monopoly profits in China to fund a three-year plan to catch Nvidia
- Amazon - Referenced via Trainium as one of several ASIC efforts competing at the chip-component rather than full-system level
- Databricks / Snowflake / Oracle SQL - Cited as still CPU-bound structured-data processing workloads Nvidia plans to move to accelerated AI data processing
- LSI Logic - Huang's prior employer, credited with inventing the ASIC business model; used to explain why ASIC vendors extract 50-60 points of margin until scale justifies customer-owned tooling
Techniques and frameworks
- Three scaling laws (pre-training, post-training, inference/test-time reasoning) - Huang's framework for why compute demand keeps compounding: pre-training scaling, post-training (reinforcement learning practice), and inference-time 'thinking' before answering
- Extreme codesign - Optimizing model, algorithm, system, and chip simultaneously rather than just making one chip faster; Huang's explanation for Blackwell's 30x jump over Hopper despite the end of Moore's law
- Architectural chips vs. ASICs vs. customer-owned tooling (COT) - Huang's three-tier taxonomy of chip businesses; argues most AI ASIC projects will stay stuck as components of a larger system rather than reaching Apple-iPhone-scale COT economics
- Dynamo - Nvidia's open-sourced disaggregated AI workload orchestration system, built for the coming disaggregated AI factory (separating prefill, decode, memory/KV-cache processing)
- Invest America - Program passed in the 'big beautiful bill' giving every child born in the US from 2026 onward a $1,000 investment account at birth, seeded by companies including Nvidia
Summary
Jensen Huang joins Brad Gerstner and Clark Tang about a year after their last conversation, opening with the newly announced OpenAI-Nvidia partnership: Nvidia will invest up to $100B in OpenAI and become a preferred compute partner as OpenAI self-builds its own AI infrastructure for the first time, layered on top of existing OpenAI buildouts through Microsoft Azure, Oracle Cloud, and CoreWeave. Huang frames the equity investment and the compute contracts as two separate decisions: he believes OpenAI is likely to become the next multi-trillion-dollar hyperscale company, making the equity stake an attractive opportunistic bet independent of whatever chips OpenAI ultimately buys. Much of the early conversation works through why Nvidia's own compute demand keeps compounding: Huang describes three now-simultaneous scaling laws (pre-training, post-training reinforcement learning, and inference-time "thinking") that together are driving inference compute up by what he calls a billion times versus the old one-shot approach, and he walks through a GDP-based model implying the addressable AI infrastructure market could grow from roughly $400B today to as much as $5 trillion a year.
A recurring thread is Huang's defense against skeptics on two fronts: Wall Street's flat 2027-2030 growth consensus, and financial-media concerns about a compute "glut" or "circular revenue" roundtripping between Nvidia, OpenAI, and other partners. On the glut question, Huang argues the industry is still years away from having fully converted general-purpose computing to accelerated computing, so oversupply risk stays low until that conversion completes. On circular revenue, he distinguishes OpenAI's compute buildout (funded by OpenAI's own growing revenue, equity raises, and debt) from Nvidia's separate equity investment, arguing there is no obligation for OpenAI to keep buying Nvidia chips if a competitor's architecture proves better. He extends this into a broader total-cost-of-ownership argument: because Nvidia's performance-per-watt advantage is so large (roughly 30x Hopper-to-Blackwell), a rival chip would have to be given away for free before it beat the revenue a customer could generate from a fixed, power-constrained gigawatt of Nvidia hardware.
The conversation turns to competitive moat and system design. Huang lays out Nvidia's shift to an annual release cadence (Hopper, Blackwell, Rubin, Rubin Ultra, Feynman) enabled by what he calls "extreme codesign," simultaneously optimizing chip, networking, and software across six to seven chips a year, something he says would be impossible without using AI internally to help design Nvidia's own products. He offers a three-tier taxonomy of chip businesses (architectural platforms, ASICs, and customer-owned tooling reserved for Apple-iPhone-scale volume) to argue that most current AI ASIC efforts, Google's TPU program aside, will remain stuck building one component of a much larger, fast-changing AI factory system rather than displacing Nvidia's full-stack platform.
The second half shifts to geopolitics and policy. Huang argues that cutting Nvidia out of China (from roughly 95% market share previously) amounted to unilateral disarmament, letting Huawei build monopoly profits it is now using to fund a stated three-year plan to catch up; he frames continued competition in the Chinese market as good for both US economic interests and for keeping Chinese engineers inside the US-aligned technology ecosystem. Both hosts and Huang treat the broader US-China AI contest as an existential, Manhattan Project-scale race funded by private companies rather than government, citing a sharp reported drop (from roughly 90% to 10-15%) in the share of top Chinese AI PhDs choosing to stay and work in the US as a leading warning indicator worth watching closely. The episode closes on immigration and the "American dream": Huang gives qualified support to the $100,000 H-1B visa fee as "a good start" against abuse while cautioning it should not be the final answer, and both speakers discuss the newly passed "Invest America" program, which gives every US-born child a $1,000 investment account at birth, as a way to spread AI-era wealth broadly. Huang closes by rejecting mass AI unemployment fears, arguing that because AI has made Nvidia itself more productive and led it to hire more people rather than fewer, intelligence is not a zero-sum game.
Notable Quotes
"I think that OpenAI is likely going to be the next multi-trillion dollar hyperscale company." - Jensen Huang
"Nobody needs atomic bombs. Everybody needs AI." - Jensen Huang
"Even if they gave it to you for free, your opportunity cost is so insanely high." - Jensen Huang
"Fundamental to the American dream is the right to rise." - Brad Gerstner
"Just get on it while it's going kind of slowly, and go exponential along the way." - Jensen Huang