Jensen Huang on BG2: the capex math that beats the bubble case
Key points
Key takeaways from the BG2 Pod interview with Nvidia founder and CEO Jensen Huang (October 2026), his third conversation with Brad Gerstner and Clark Tang, recorded days after the White House accord on superintelligence and Huang receiving the National Medal of Science:
The capex rebuttal. Huang splits Nvidia's business into roughly three equal parts — classical accelerated computing, frontier AI labs and startups/neoclouds renting compute — and argues the market is putting "way too much pressure" on two labs. OpenAI and Anthropic "don't have to be anywhere near a trillion dollars in revenues to be successful"; the classical third alone grows 30–40% a year with no frontier lab involved.
2027's real bottleneck. With street estimates of roughly 50GW of announced chips against about 35GW of land, power and finished "shell", Huang says power and buildings — not silicon — will be the binding constraint, and scarce watts push buyers towards the architecture with the most tokens per watt. Nvidia's revenue per gigawatt has roughly doubled from about $18bn to over $40bn as racks gained networking, CPUs, storage and security.
Why reasoning eats compute. Thinking models explore, roll out and prune options, which takes on the order of a billion times more inference compute than simple generation; the past year's unlock was the "harness" giving models working memory, knowledge graphs and tool use (SPICE, Verilog, CAD). Huang expects the cost of intelligence to fall the way computing costs did under Moore's law.
Moore's law is over. "If you want twice as many transistors now, you have to pay twice as much" — Huang sees computing spend heading towards $1.5tn a year by the end of the decade, infrastructure build-out of $3–4tn a year by 2030 (closer to $4tn), and says annual spending of $10tn would not surprise him.
The sight-unseen moat. Customers will place Vera Rubin Ultra orders before the hardware can be benchmarked, and Huang argues only Nvidia can be trusted at that scale: "sight unseen, who can you place a $50 billion order on?" — "specialization is the enemy of fungibility", so general-purpose GPUs stay re-rentable and easier to finance as model architectures shift.
Safety as engineering. Alignment is "going to be a forever problem", so safety — like airbags and brakes — has to be built in: zero trust, minimum access to tools and data, and out-of-band monitoring that can escalate or shut a system down. Nvidia's OpenShell effort spans more than 100 companies, and Huang argues building responsibly means moving fast, not perfecting first.
New revenue streams. Enterprise software licensing — a few thousand dollars per GPU per year across millions of shipped GPUs — and revenue sharing with clouds are becoming "a very big part" of the business; physical AI alone brought in $10bn in the first half of 2026, after $5bn and then $10bn in the two prior years.
Demand beyond two labs. New orders arrive weekly — a recent 2GW deal in Australia is worth about $80bn — and Elon Musk's xAI has moved from Nvidia's enterprise segment to its cloud segment, with 10GW planned by the end of 2027. Nvidia and SpaceX engineers are designing a new server architecture for orbital data centres, where only the highest-value, most rentable compute justifies launch costs.
Read more: YouTube · BG2 episode recap
