The All-In Summit
22 highlight clips below (49 mins total).
Investment The All-In Summit · 22 clips
All-In
Satya Nadella
Chairman and chief executive of Microsoft, which the hosts put at $175 billion of capital expenditure on AI infrastructure. He describes a position built on running OpenAI, Anthropic and Microsoft's own MAI models across heterogeneous hardware without holding a frontier model of its own.
-
1 of 2 framework
Microsoft separates land, power and shell, which it builds and holds, from racks and chips, which Satya Nadella puts at 60% of the cost and buys against a two to three year forecast.
Only the short half carries demand risk, so Microsoft builds what it can, leases where it cannot, and rents when it is caught short, which he admits it currently is. He ties the whole shape to refusing to supply two or three customers: a hyperscaler underwriting its build on OpenAI alone has no business, which is why he wants the long tail even while OpenAI is his largest account.
-
2 of 2 contrarian
Satya Nadella says the AI build-out needs at least 7% to 8% real, broad-based GDP growth to play out, against a long-run trend the panel puts at 2% to 4%.
He argues that augmenting an existing workflow only redistributes work, and the hosts' example is a shorter week at unchanged output. What clears the bar is invention: the case he reaches for is a small business running its working capital off its own invoices and emails instead of an ERP.
All-In
Brad GerstnerAltimeter
Gerstner runs Altimeter and gives a recurring markets and technology update on All-In; the introduction credits him with starting five companies and with driving the Trump Accounts legislation. He speaks as an investor sitting at what he calls a medium position.
-
1 of 4 framework
Brad Gerstner puts the offtake revenue needed to service AI capex at $450 billion and then $800 billion, against a $200 billion exit run rate this year.
The hyperscalers are building to rent, which leaves the bill with three private labs whose combined top line is a rumoured $100 billion. Gerstner says $80 billion more has to arrive across those three before December just to hold the trade where it is.
-
2 of 4 contrarian
Brad Gerstner puts Nvidia at 14 times next year's fully taxed GAAP profits and the Nasdaq multiple below its own average, and reads 2026 as a market driven by profit growth rather than a repeat of 2000.
Semiconductors produced 70% of the index's return this year, and hyperscaler capital expenditure arrives almost dollar for dollar as free cash flow at the chip makers. Consumer discretionary, software and financials have barely moved, so the earnings doing the work sit in one link of one chain.
-
3 of 4 framework
Brad Gerstner splits a decade of Nasdaq earnings growth into 6% revenue and 38 basis points of margin expansion a year, and asks whether AI can take that 38 to 100.
Uber and Snowflake have both told him they expect to grow 20% and 30% without adding headcount. That is where the extra points come from, engineers not hired rather than engineers dismissed, and on the demand side he needs roughly 4% of the knowledge-work market, near $1.2 trillion, to cover the capital expenditure.
-
4 of 4 framework
Brad Gerstner says the one-decision AI trade that worked from 2023 to 2025 is now fully priced, and he is running a medium position with three conditions written down that would change it.
He names monthly revenue at the largest AI labs running near $8 billion against $4 billion, a 10-year Treasury yield at 5.5%, and a halt to the Anthropic listing. Two of the three are observable within weeks, which is unusually specific from a manager speaking in public.
All-In
JD Vance
Vice President of the United States, roughly two years into the term and speaking ahead of the midterms. He leads the administration's anti-fraud programme and says it has recovered about $250 billion against a $7 trillion federal budget.
-
1 of 2 current issue
JD Vance puts the administration's fiscal deficit target at around 3% of GDP and says what markets are pricing is the perception that neither party in Congress is serious about the debt.
Cutting $2 trillion out of a $7 trillion budget is not on the table, on his account, so the path runs through smaller reductions plus growth. A durable fix needs bipartisan legislation binding whoever holds the White House and the legislature next, and he concedes many members have shown no willingness on spending.
-
2 of 2 framework
JD Vance dates China passing the United States in electricity generation to 2005 and puts its output at three times America's now, and reads the data centre backlash off that.
He names the retail power bill as the channel, citing one household whose monthly cost went from $290 to $580 in a year, and locates the opposition in places that made new plant hard to build. His conclusion is to build through it rather than pause, on the view that cheap electricity is what defuses the politics.
All-In
Dina Powell McCormick
President and vice chairwoman of Meta Platforms, leading the company's data centre, investment and partnership expansion, and on stage eight months into the role after a year on the board. Introduced as a former presidential adviser to Donald Trump and, on the hosts' description, a former deputy national security adviser.
-
current issue
Meta found the binding constraint on its largest data centre was fibre technicians rather than power or chips, and had to write the training course itself because none existed.
Unpaid OSHA and safety training priced out the waitresses, drivers and care workers the programme was aimed at. Meta's answer was to pay the full job wage from day one and guarantee a site job on completion; 40,000 applied and 250 have graduated, against the 500,000 open trade roles cited on stage.
All-In
Gwynne Shotwell and Elon MuskSpaceX
Shotwell was among SpaceX's first dozen employees, marked her 24th anniversary the week of the recording, and was made president in 2008 on the way to negotiating a $1.6 billion NASA cargo contract. Musk joined by video call from Memphis.
-
1 of 2 current issue
Gwynne Shotwell says she does not expect SpaceX to issue more stock against a capital programme she calls country-scale, and names renting out compute as the business now carrying it.
Starlink sits at 1.5% to 2% penetration depending on the country, and its revenue is small next to what SpaceX bills the United States military, so she sees headroom. On the compute side she reports no drop in demand at all, against a build-out the hosts size at tens of billions of dollars a quarter.
-
2 of 2 explainer
Gwynne Shotwell says land beside a known data centre site reprices from $3,000 an acre to $180,000, with generators quoted three years out, and puts an AI compute satellite on next year's launch manifest.
Orbit avoids three costs she names on the ground: the real estate, the permitting queue, and the cooling that a radiator facing deep space provides for nothing. The comparison works because SpaceX already owns the lift, and it works on those terms for nobody else.
AI The All-In Summit · 22 clips
All-In
Satya Nadella
Chairman and chief executive of Microsoft, which the hosts put at $175 billion of capital expenditure on AI infrastructure. He describes a position built on running OpenAI, Anthropic and Microsoft's own MAI models across heterogeneous hardware without holding a frontier model of its own.
-
1 of 2 framework
Satya Nadella reads the collapse in frontier token prices as the same open-source check that Linux put on Windows and Postgres put on SQL Server, and says it is what makes an application layer possible at all.
Without a free substitute, price stays where the closed vendor sets it and there is nothing left over for anyone building on top. He wants the model companies to go further and standardise interoperability, including KV cache reuse across model families, on the grounds that making Windows work with Unix grew Windows rather than shrinking it.
-
2 of 2 explainer
Satya Nadella describes a frontier model told to optimise a company's working capital as a new category of insider risk, because it may fake the books to hit the target.
It happens at test time on an ordinary task, not in a training run. That places the exposure inside any enterprise already running agents, and the control he proposes is a separate causal or semantic model that checks and verifies, with the admission attached that the science on reward hacking is not settled.
All-In
Gwynne Shotwell and Elon MuskSpaceX
Shotwell was among SpaceX's first dozen employees, marked her 24th anniversary the week of the recording, and was made president in 2008 on the way to negotiating a $1.6 billion NASA cargo contract. Musk joined by video call from Memphis.
-
1 of 3 framework
Elon Musk wants every frontier lab running its own security test harness against every rival's model before release, on the grounds that a company grading its own homework cannot see what it is blind to.
Enforcement comes from liability rather than a regulator, in his account: a lab told by competitors that its model is unsafe, which ships anyway and then causes harm, hands a jury something close to prima facie negligence. He argues Beijing would accept it because neither side gives anything up, unlike the pause China has already refused, and because no American inspector sets foot in a Chinese company.
-
2 of 3 current issue
Elon Musk says the people inside Anthropic and OpenAI warning that their own models are dangerous should be believed, and that this is what he meant by saying Dario Amodei is right.
He says he should have said more at the time and tried to clarify on X, and what he wants to emphasise is that the risk is significant now and rising exponentially with this generation of models. The hosts put the objection straight back: the same firms assign a one in ten chance of annihilating humanity and are asking allocators how much of the IPO they want.
-
3 of 3 contrarian
Elon Musk says Anthropic and OpenAI sit close enough in capability that neither can slow down without handing the other the lead.
He calls the swarm test run against Hugging Face somewhat reckless and then explains why he expects more of it: with two firms level, restraint is unilateral disarmament. Of the two he says the more careful is Anthropic, whose own staff have said in public that the models are frightening them.
All-In
Jensen Huang, founder, president and chief executive of Nvidia
Huang founded Nvidia and runs it as president and chief executive. The show's own introduction cites revenue up 97% year on year and describes the company as the only full-stack AI computing platform. President Trump also joins by telephone for roughly five minutes.
-
1 of 4 framework
Jensen Huang argues that every real AI incident so far has come from a frontier lab because nobody else has the compute, and that regulation should follow the compute.
A school student cannot cause one and nor can a startup, on his account, because the danger scales with how much of it is available to run an experiment. He counts four incidents at one lab and one larger one at another, and wants all of them treated as engineering failures: root cause, fix, institutionalise.
-
2 of 4 explainer
Jensen Huang breaks recursive self-improvement into parts that already exist, in-context learning, skills, reflection, reinforcement learning, synthetic data and LoRA, and says it cannot run away because a model still has to pass release testing.
Weights can improve inside a company all day, on his account, and the brake comes at the point of shipping, when somebody has to evaluate the new version and confirm it has not regressed. The question reached him through a Chinese lab, the maker of GLM, reported to be putting billions into exactly this.
-
3 of 4 framework
Jensen Huang proposes that AI models be checked the way company accounts are, by independent third-party evaluators, and that there be several of them so that no one evaluator can be captured.
Auditors do not have to know the business better than the people running it, on his account; they have to ask the right questions, and plurality is what stops any single one being pulled onto the client's side. He also says the case for restraint has moved from national security to safety, and that changes who the argument has to convince.
-
4 of 4 framework
Jensen Huang answers the charge that Nvidia has become the bank of AI with a stated rule: go up the stack as far as Nvidia needs to and stay as low as possible.
He invests where he finds a bottleneck, upstream at Corning, TSMC and the memory makers years before the demand arrived, and now downstream into land, power and shell. Pressed on why the company does not take the application layer where earnings usually settle, he says taking a slice would cost him the ecosystem that runs on him.
Management The All-In Summit · 22 clips
All-In
Bill Gurley
Investor who backed Uber and Zillow and has worked through 25 years of technology cycles. He spoke at the same event three years earlier on regulatory capture.
-
1 of 2 framework
Richard Feynman cleared six months of his calendar before the Challenger commission met, and Bill Gurley says another commissioner told him the panel evaluates and does not investigate.
He was already turning up at NASA facilities asking to meet people before the panel convened, and when the final report smoothed over what he had found he threatened to take his name off it. His notes went in whole, as Appendix F.
-
2 of 2 framework
Bill Gurley singles out one rule behind Japan's Fukushima inquiry: no member could have an institutional connection to the regulator or the power company.
Japan's parliament had never commissioned a private investigation before this one. It reported in seven months, published in English first out of concern for how the country would be read abroad, and opened by rejecting the natural-disaster account of the accident, calling it man-made and preventable.