All-In's AI thesis: the safest business may be selling compute, not winning the model race

The strongest idea in All-In's latest episode is not Google's leadership shake-up, SpaceX's earnings, or Airtable's sale taken separately. It is a capital-allocation claim: compute infrastructure offers a clearer path to returns than frontier-model development, unless a lab can keep charging for being at the frontier. The panel uses three very different companies to test that claim, then argues over how much the conclusion changes for consumers, enterprises, and US-China policy. Brad Gerstner fills in for Chamath Palihapitiya on the episode. 1

The episode treats Google's AI shake-up as evidence of a conflict inside the business. The company needs enormous amounts of compute for Google Cloud customers and its own products, while its frontier-model researchers want the same scarce hardware to train and serve models. A cloud unit can rent that capacity to Anthropic, OpenAI, or an open-weight customer and recognize revenue without proving that its own model will remain the best. A research group has to win a much harder race before the spending pays back. 1

David Friedberg's reading is that Google's capital is therefore moving toward the safer side of the fork: data centers, networking, and the ability to host many models rather than betting the company on one model family. Brad Gerstner describes the resulting problem as a channel conflict. Google Cloud wants to sell compute to a rival; Google's internal lab wants to consume it first. The episode connects that tension to the departure of senior researchers and to Demis Hassabis moving into a broader DeepMind and Google science role. Those staffing events may have several causes, but the panel's explanation is specific: the people who want to build frontier intelligence may leave when the company rewards infrastructure more visibly than model research. 1

That is a more useful frame than calling Google either a winner or a loser. A company can be excellent at distributing AI while choosing not to win the frontier-model race itself.

David Sacks pushes the argument further. He describes a two-tier market: a small group of frontier labs can charge a premium for the most capable intelligence, while models six to twelve months behind compete mainly on inference price and availability. In his analogy, frontier models are the premium phone and open models are the Android-scale market: broader use does not necessarily produce the most profit. He points to Anthropic and OpenAI's reported revenue growth as evidence that frontier capability has not yet become a commodity. 1

Jason Calacanis gives the counterweight. He argues that Google's distribution may matter more than its position on a benchmark: the panel says five Google products each have more than three billion monthly users, while Gemini had reached roughly 950 million monthly active users in the quarter under discussion. On this view, a model that is good enough inside search, Gmail, Chrome, YouTube, and Android can create more value than a slightly better model with no comparable channel. These figures are presented by the panel in the episode, not as an independent audit. 1

David Friedberg rejects the idea that every enterprise will pick one winning model. His proposed outcome is a mixture: cheap open-weight models for routine workflows, premium models for difficult or immature tasks, and specialized models for areas such as video or life sciences. That distinction matters because model quality is not a single variable. A company choosing a system is also choosing latency, reliability, data controls, maintenance work, and the cost of finding out that a cheaper model failed.

The episode's disagreement is therefore narrower than the usual open-versus-closed argument. The panel broadly agrees that cheap models will handle more work. It disagrees over how much the frontier premium will persist and whether the savings from open models survive the cost of operating them. Jensen Huang's view, as relayed in the discussion, is that a closed model can be cheaper once training, fine-tuning, safety work, and maintenance are included. 1

SpaceX is the episode's stress test for the compute thesis. The panel discusses reported second-quarter revenue of $7.8 billion, AI-related revenue of $2.6 billion, and $18.4 billion in quarterly capital expenditure. It also discusses a plan to grow from about 1.4 gigawatts of compute to two gigawatts by year-end, with a much larger buildout described for the following year. These are the figures and forecasts used in the conversation; the important point is the relationship between them, not whether any one projection arrives on schedule. 1

The relationship is stark. Frontier labs need compute quickly enough to pay a premium for capacity. Infrastructure companies can then build ahead of current revenue if they believe the demand will continue. But that makes the whole chain sensitive to one question: who ultimately pays for the next block of hardware, and at what price?

Gerstner warns that a fall in demand or spot compute prices would hit more than one company. The episode describes a market in which data-center builders, chip suppliers, and model labs are financially tied to the same expansion. That is why the panel's bullish case carries its own warning label: compute may be the safer business than model-building, but it is not a low-risk business when the buildout depends on a handful of buyers and unusually high rental prices.

Airtable gives the panel a demand-side example. The episode discusses the company's reported sale to Bending Spoons for about $1.28 billion, far below its roughly $11.7 billion peak valuation, even though the business was described as profitable, with about $480 million in annual revenue and 20% growth. The panel also says Airtable separated its AI-agent effort, HyperAgent, before the sale. 1

Sacks' interpretation is that a venture-backed software company was pushed toward a sales-led growth model that did not fit its product. He cites a reported 30% sales-quota attainment rate as a sign that the company was trying to force a different kind of business onto a product-led base. Bending Spoons can treat the legacy product as a cash-generating asset, cut the cost structure, and let the newer AI effort pursue a venture outcome. That may be a plausible acquisition strategy, but the episode does not establish that the buyer will reach the margins the panel imagines.

Friedberg uses the case to separate vulnerable no-code software from enterprise systems that function as compliance and identity infrastructure. The distinction is practical: AI can make a dashboard or internal tool cheaper to build, but replacing a system that holds regulated records is a different project. Airtable is not proof that all SaaS is finished; it is evidence that software valuations now depend more heavily on whether the product is a durable organizational rail or a layer users can recreate with an agent.

The final discussion turns to US data-labeling and expert-curated training data being sold to Chinese AI companies. One side argues that highly specialized data created by American experts can help Chinese labs close the gap. The other side says ordinary labeling is replaceable, China has abundant technical talent, and broad restrictions could provoke reciprocal measures without protecting a genuine military advantage. The disagreement returns to the episode's main test: is the scarce asset a defensible capability, or merely a temporarily expensive input? 1

That is the episode's durable takeaway. AI is not one market moving at one speed. It is a stack of businesses with different failure modes: infrastructure can be profitable but financing-sensitive; frontier models can command a premium but must keep their lead; open models can widen access while shifting cost into deployment; and applications can lose value when an agent makes the same workflow easier to rebuild.

For anyone evaluating an AI investment or deployment, the useful question is not simply which model is best. It is which layer owns the scarce asset, who pays for it, how quickly that advantage can be copied, and what happens when the price of compute or the quality gap moves. All-In's panel does not settle those questions. It does make the trade-offs visible.

This article is based on the complete All-In Podcast episode released on August 8, 2026, with Brad Gerstner filling in for Chamath. The original YouTube episode is the reader-facing video and transcript source; the episode is also listed on the All-In Podcast Apple Podcasts page. Claims above are attributed to the panel where they are forecasts, interpretations, or figures discussed in the conversation.