All-In Summit 2026 (LA, Sep 13-15) — what the headliners actually said
Four sessions from the All-In Summit are now out. A straight recap of the main
points — reporting, not recommendations.
JENSEN HUANG (Nvidia), Sep 14 — ft. a live on-stage phone call from President Trump ("the robots will not be taking over... the whole thing is a hoax")
* Dismissed AI-extinction odds as "made up," citing a run of failed doom predictions (radiology, "90% of code," "50% of entry jobs").
* Says AGI and even "superintelligence" are "already here" -- but only in narrow domains (self-driving at ~1/10th the human accident rate, protein work).
* Notable non-answer: asked directly about Nvidia being "the bank of AI" (the BlackRock/Goldman GPU-financing structures), he reframed to "supply chain"
and "downstream" and did not address the financing question. Gave no Rubin or performance-per-watt numbers.
* China reaches its own advanced chip lithography "by 2030"; most open-source AI contribution now comes from China; "$400bn of VC into AI-native startups
in 6 months, 80% on open models."
ELON MUSK & GWYNNE SHOTWELL (SpaceX/Tesla), Sep 15
* SpaceX is public: $75bn raised, ~$1.7T debut market cap. Compute rental called "a heck of a business," "tens of billions per quarter" of buildout --
but no capacity or price figures disclosed.
* Concrete AI-safety idea: major labs should run each other's security test harnesses, all logged to expose IP theft -- no new regulation, must be
China-acceptable. "Dario is right"; rates Anthropic above OpenAI on safety.
* Terafab (own chip fab) framed as "build it or fail to scale," a hedge against Taiwan; advanced packaging is the tightest bottleneck. SpaceX/Tesla merger
left pointedly open. Starship Flight 14 window Sep 22; Tesla reveal Oct 1.
SATYA NADELLA (Microsoft), Sep 15
* Backs "pacing the frontier" in substance but as engineering discipline, not doom -- "serves humanity first and is in human control," welcomes third-party
testers, concedes no growth.
* The plan is visible de-risking from OpenAI: "one of our largest customers...but we need more"; Microsoft building its own models "from the very bottom,
not distilling," and spreading silicon across Nvidia, its own Maia, OpenAI's chip and AMD. Copilot at "30M+" seats.
* "$80 billion" Azure capex, deliberately below Meta/Google; value, he argues, migrates off the commoditizing model layer to apps and middleware.
JD VANCE (VP), Sep 15
* H-1B: $100k fee plus a rule barring H-1B hiring after mass layoffs -- by executive action, already in litigation; aimed at low-wage IT outsourcing,
"actual genius" exempted.
* Energy is the master variable: US generation "flat since 2005" vs China ~3x; goal "electricity too cheap to meter." Iran "shooting at commercial shipping
even during peace talks" -- energy risk premium stays.
* Fiscal: deficit target only "closer to 3% of GDP" over time, needs a bipartisan deal he calls unlikely -- 30Y ~5.2%. No federal AI statute; rejects "one-world governance."
THE THROUGH-LINE
Every headliner is routing around the model layer: Nadella ("value moves to apps + middleware"), Huang (Nvidia's own open models), Musk (own models +
fab). And the safety-pact idea now has three backers (Altman, Musk, Nadella) -- converging on industry-run mutual testing, not government reporting.
Individual 2-3 page notes per session available on request. Source is podcast commentary (T3/T2 for principals on their own firms), figures as stated,
unaudited. Not investment research.
Laurent Chekroun
Graham Advisors Sarl, Pfaffikon SZ