All-In — Jensen: doomer math fails; open models carry apps

Opening

Doomer extinction math is unscientific — open models and stack depth are the load-bearing claims. Jensen Huang (NVIDIA CEO) sits with Chamath, Calacanis, Sacks, and Friedberg for a live All-In Summit session (~47m). Ground covered: Amodei-adjacent "doomer" scorekeeping and regulation vs root-cause engineering; open vs closed models and China open-source volume; a mid-session speakerphone call from President Trump on data-center permitting; NVIDIA capital allocation across supply chain and NeoClouds; open-stack ambitions (Nemotron, Alpamayo, biology); a brief take on Elon's Terafab. Fidelity is asr — figures and quotes below are machine-transcribed and marked accordingly. Watch

Key takes

Failed apocalypse timelines are Huang's wedge against extinction-probability rhetoric. Radiology job wipeout, "90% of code in 6–12 months," and "50% of entry jobs in 6–9 months" are named as wrong, so "10% extinction" is treated as unscientific rather than precautionary. Huang: predictions "proven to be wrong"; quantification is "made up" and "irresponsible"; whistleblowing (Coxin) is separated from the scientific claim. [book]

Safety vs American AI leadership is a false choice — slowdown is the wrong strategy. Incidents concentrate at frontier labs because they hold the most compute; prescription is root-cause engineering, voluntary pause only if a lab is out of control, and multiple independent evaluators. Huang: "Slowing down is definitely the wrong strategy"; auditors should be plural so "one company doesn't become… pilled." [book]

Open models are load-bearing for the application layer. ~$400bn of AI-native venture funding in ~6 months, with ~80% said to use open models; closed models are "bottled water" — both needed for sovereignty, privacy, and proprietary stacks. Huang cites the funding mix as why founders can build dreams that differ from frontier labs; China is the majority contributor to open-source volume today, but forks make downloaded models "yours." [book]

NVIDIA's posture is enablement depth, not lab capture. "Go up as far as we need to and as low as possible," with the platform claimed to have broadened from running mainly OpenAI ~18 months ago to Meta/Grok/Gemini and Anthropic scaling on-stack. Prefers helping "everybody succeed instead of taking a slice"; NeoClouds exist because hyperscaler annual planning is "always almost wrong" on land/power/shell. [book]

Narrow superintelligence is already here in Huang's definition. Self-driving at far-below-human accident rates and protein synthesis / virtual screening count; China native advanced lithography dated to ~2030; Terafab shrugged as feasible if anyone can do it. Huang: "I think we're already there" on AGI-as-human-parity in narrow domains. [book]

Trump's live call hardens a federal pro-build narrative against data-center permitting pushback. Brands opposition a "hoax"; data centers called "the oil of the next 20 25 years"; cites Google Finland as permitting failure. Trump: "$20 trillion" investment claim vs "<1 trillion" prior-term contrast — political arithmetic, not NVIDIA guidance. Huang after the call stresses jobs / re-industrialization / energy as the bridge he wanted.

Key math

~$400bn VC into AI-native companies in ~last 6 months; ~80% said to use open models (asr) — application-layer funding mix, not lab spend. [book]

China advanced lithography ~2030 for native-grown systems (asr) — supply-chain clock; "just around the corner" in Huang's decade framing. [book]

Platform diversification: ~18 months ago "the only thing we ran was OpenAI"; now Meta, Grok, Gemini, Anthropic scaling on NVIDIA (asr) — stack breadth, multi-lab not single-customer. [book]

Terafab ~100m sq ft (host framing); Huang endorses executability without his own size/capex figure (asr) — competitive fab narrative, not NVIDIA guidance.

Trump: ~$20tn investment "coming into the country" in one year vs "<1 trillion" over four years prior (asr) — political claim, unconfirmed here; permitting/sentiment, not a cash-flow line.

Radiology / code / entry-job counterexamples inverted or wrong on Huang's scorekeeping (asr) — used to discount extinction percentages. [book]

Quotes

"We have to take account for all of the stupid predictions that were made." — Jensen Huang [asr] [book]

"The way I think about closed models is kind of like bottled water." — Jensen Huang [asr] [book]

"Our strategy is go up as far as we need to and as low as possible." — Jensen Huang [asr] [book]

"Slowing down is definitely the wrong strategy." — Jensen Huang [asr] [book]

"I think we're already there." — Jensen Huang [asr] [book]

"I'm telling you, it's all a hoax. The data centers are great… it's the oil of the next 20 25 years." — Donald Trump [asr]

Variant perception

Priced in — NVIDIA as full-stack AI infra; open/closed coexistence; land/power/shell bottlenecks. Hyperscaler + NeoCloud constraints, China open-weight competition, and political sensitivity of US data-center buildout were already consensus.

What's new — prediction scorekeeping plus a concrete open-model funding claim. The falsifiable wedge is failed apocalypse timelines against Amodei-adjacent extinction framing, paired with ~$400bn / ~80% open-model VC. "Already there" superintelligence is defined narrowly (driving, protein work). Live presidential alignment with the pro-data-center narrative is a sentiment / permitting datapoint. Hugging Face treated by hosts as a consequential platform move (acquisition / aqua-hire framing in-room).

Bear case — confidence that every incident is fixable inside frontier labs is the single point of failure. If danger concentrates because of compute, open-model China volume undercuts "American stack" nationalism even as he argues exploitation > origination; the $400bn / 80% figures are CEO assertions without a cited dataset here; Trump's $20tn line is political rhetoric that can reverse with permitting reality; "already there" AGI dilutes the word without proving durable moats for NVIDIA customers.

Discount — Huang is talking demand, open-source strategy, and anti-slowdown policy. All [book]. All-In hosts are summit principals with AI/tech exposure; Sacks's White House AI role shapes the regulation segment. Trump is talking his industrial / energy agenda on speakerphone. Doomer-dismissal sells GPUs and permitting; treat extinction rebuttals and funding-mix claims as advocacy until corroborated.

Positioning

AI capex durability — STRENGTHENS. Huang rejects slowdown, cites AI-native VC + compute demand, NeoCloud agility on land/power/shell, and a presidential pro-data-center frame against permitting pushback — sustained build as base case.

Enterprise agent stall — NEUTRAL. Agentic / RSI discussion stays at lab and tooling altitude; no Fortune 500 production agent with $ or headcount is named.

Frameworks

Bottled water vs free water (2026-09-14, Huang on All-In). Closed frontier models are paid convenience; open models are the commodity layer required for sovereignty, privacy, and divergent founder products. [book]

Go up as far as needed, as low as possible (2026-09-14, Huang on All-In). Build enabling stack depth so labs and NeoClouds succeed on NVIDIA rather than NVIDIA taking a slice of their P&L. [book]

Plural auditors (2026-09-14, Huang on All-In). Third-party AI evaluators modeled on financial auditors — multiple firms so no single evaluator becomes captured. [book]