When Jeff Dean, the 27-year Google veteran who built the infrastructure that trained the world’s models, walks out the door to start his own lab, and Demis Hassabis gets moved into a chairmanship that smells like a golden parachute, the market can read the signal: the money is no longer on the science. It’s on the steel, the concrete, and the power lines.

Google’s stock dropped 4% on the news — roughly $200 billion of market value — the day the departures were reported. Speaking on the All-In podcast, Brad Gerstner of Altimeter Capital, alongside hosts David Friedberg, David Sacks, and Jason Calacanis, spent the next hour explaining why this is not a personnel story. It’s a capital-allocation story. And it explains everything from SpaceX’s first public earnings blowout to why a once-$11.7-billion software company just sold for roughly a tenth of that.

The boardroom math that pushed the scientists out

Friedberg, who was an early Google product lead, laid out the arithmetic that every board member at a big AI company is now doing. Google has committed to roughly $200 billion of AI capital expenditure this year. Under current US tax rules, accelerated depreciation means every dollar deployed into domestic compute infrastructure returns about 26 cents as a tax benefit — a government-subsidized, high-confidence investment at a time when compute demand is extreme and Google is among the best operators of infrastructure on earth.

Frontier model development, by contrast, costs tens of billions per training run with far less certain profits. Open-weights models from China and the open-source community are closing the capability gap every quarter, compressing the premium that once accrued to being first.

“I would frame it as capex is high alpha, low beta in data center infrastructure,” Friedberg said. “Model development theoretically could be high alpha but it’s very high beta — it’s a very risky way to deploy capital.”

Compute infrastructureFrontier model developmentCapital requiredTens of billions per gigawattTens of billions per frontier training runTax treatment~26% effective discount via accelerated depreciationNoneDemand profileExtreme, visible, contracted by frontier labsUncertain; open-weights compress pricingReturn profile“High alpha, low beta”“High alpha, very high beta”Google’s positionWorld-class operator, model-agnostic hostingNo longer necessary to the enterprise and consumer moat

Jeff Dean and the four senior researchers who left with him to found Discovery Loop can do this math. More importantly, they can read the capital-budget spreadsheet that made it. A scientist of Dean’s caliber, Friedberg noted, can walk down the road and “raise a couple billion dollars at a multi-billion-dollar pre-money with a PowerPoint deck because I’m the greatest in the world at doing this.” The talent is following the same signal as the capital: out of the integrated labs and into independent frontier plays, or into infrastructure.

Gerstner added the structural explanation. Every big cloud owner now has a channel conflict. Google Cloud wants every available GPU to rent to Anthropic, one of the two labs paying the highest spot prices for compute. DeepMind and the Gemini team want that same silicon to compete with Anthropic. Microsoft faces the identical tension renting to OpenAI while building its own models. xAI, under Elon Musk, rents its Colossus data-center capacity to Anthropic while simultaneously building the Grok model and absorbing the Cursor acquisition. Meta, Gerstner reported, is reportedly entering the infrastructure-as-a-service game.

Only two companies are conflict-free. “It looks like it’s being resolved in favor of being more of an infrastructure company,” Gerstner said of Google’s direction. The scientists who built the frontier era are leaving because the boardroom has already made its choice.

“And then there were two”: the frontier duopoly

If capital allocation explains who left, market structure explains where the money went. David Sacks framed the new reality with a line that became the episode’s backbone: “And then there were two.” A year ago, five companies were plausibly in the hunt for the leading model. Today, Sacks argued, the race has narrowed to an Anthropic-OpenAI duopoly, with xAI still chasing.

His proof is in the revenue. Anthropic started the year at $10 billion in annualized recurring revenue. It is now reportedly north of $80 billion, guided toward $100 billion by year-end, and Sacks said it is tracking to $110 billion to $120 billion. The guidance was widely dismissed as impossible when it was first issued. It no longer looks that way. OpenAI is likewise seeing acceleration.

The analogy Sacks reached for was Apple versus Android. Android has more users globally, but the monetization flows to the premium platform. In AI, token consumption is rising fastest on the open-weights side, but the revenue — the actual economics — is concentrating at the frontier. Gerstner captured the bifurcation in a single sentence: “Token consumption is going up for the open-source guys while share of economics is going up for the frontier labs.”

Frontier intelligenceCommodity / lagging intelligencePlayersAnthropic, OpenAI; xAI huntingOpen-weights models, Gemini, Kimi, Qwen, GLMTiming gapLeading edge6–12 months behindPricingPremium on tokens and modelsWeights ~free; monetize compute, inference, hostingMonetizationThe model layer itselfInfrastructure layer, consultingAnalogyAppleAndroidRevenue evidenceAnthropic: $10B to $80B+ ARR in one yearGoogle Cloud: +82% year over year

The hosts tested that thesis against the open-source wave unfolding in parallel. Jason Calacanis disclosed that for 95% of his work, the difference between non-frontier models and frontier models is “negligible already.” Elon Musk fired back on X that “it’s actually a world of difference,” and both Gerstner and Friedberg noted that for high-stakes specialized work — genomics, life-sciences modeling, video generation where Gemini is currently best-in-class — the frontier premium is real and growing. Enterprises, Friedberg’s synthesis predicted, will run a blend: cheap open-weights models for simple workflow tasks, premium models for specialized applications, with orchestration layers managing the mix.

Gerstner surfaced two non-consensus views from the week that complicate the open-source-equals-free narrative. Musk has been telling anyone who will listen that “we’re entering the singularity and the frontier models are way further ahead than people think.” And Jensen Huang, the Nvidia CEO, has been arguing that closed models are actually cheaper, once you factor in the training cost, expertise, fine-tuning, guardrails, and maintenance of running your own. Both claims, if true, imply that the two-tier structure has more staying power than the open-source bulls believe.

SpaceX’s earnings: the infrastructure thesis in one spreadsheet

If the Google departures are the theory, SpaceX’s first quarterly report as a public company is the evidence. The company reported Q2 revenue of $7.8 billion, up 92% year over year and 67% quarter over quarter. “Elon Web Services” — the nickname the hosts gave to the compute-rental business — more than tripled quarter over quarter to $2.6 billion. Capital expenditure was $18.4 billion in the quarter, six times year over year. Management guided to $100 billion in annualized revenue by the end of the year and pulled the $1 trillion ARR target forward from 2031 to 2030.

For scale: the company did $18 billion in revenue in all of 2025.

MetricQ2 2026ChangeRevenue$7.8 billion+92% YoY, +67% QoQCapex$18.4 billion6x YoY (~$75B annualized run rate)AI compute rental revenue$2.6 billionMore than tripled QoQStarlink subscribers12 million2x YoY, +20% QoQStarlink ARPU$66 per month—Market capitalization~$1.4 trillionDown ~30% from the June IPO above $2 trillion

Yet the stock fell 13% on the report. Gerstner’s explanation was unsentimental: the IPO priced ahead of the fundamentals, and what the market is doing is a normal post-IPO consolidation. Nearly every large-cap tech stock in history has drawn down roughly 50% peak-to-trough within six months of listing. At its peak, SpaceX traded at roughly 160 times revenue. It is now closer to 45 times as the revenue catches up. On the day of the IPO, Gerstner told CNBC: “I would want to own this company but I’m not sure today is the day I would buy the company.” His forward view, at the current valuation on a three-to-four-year horizon: he can see tripling the money. “The price of entry matters.”

Starlink: the cash machine that funds the science projects

Inside the segment results, the hosts identified a business that barely needs AI to justify its valuation. Starlink now has 12 million subscribers, doubled year over year and growing 20% quarter over quarter, at a $66 monthly average revenue per user. It generated $4.3 billion of revenue and $2.6 billion of adjusted EBIT in the quarter. Friedberg walked through the standalone math: at this trajectory, Starlink alone could reach roughly $40 billion of revenue and $30 billion of free cash flow within a year. At a 30 times multiple — standard for a high-renewal subscription business — it is a trillion-dollar market-cap company on its own within 18 months.

“The Starlink business alone could be a trillion dollar market cap within 2 years, within 18 months, let’s say,” Friedberg said. “That, I think, funds all of the rest of this as kind of science projects and upside.” The “rest of this” includes Starship, the AI data centers, and the terafab semiconductor project that Friedberg called potentially “the greatest semiconductor fabrication site on planet Earth” and a pathway off the US dependency on Taiwan and China.

The technical engine that makes Starlink’s growth possible is the V3 satellite and Starship’s deployment economics. The recent 13th Starship test flight — heat shield intact, booster floating in the ocean — validated the reusable architecture. A test deployment of 20 V3 satellites proved the deploy-and-connect sequence. The step-change in capacity is orders of magnitude.

Falcon 9 (today)Starship (next phase)Satellites per launch27 (V2)60 (V3)Added network capacity per launch2.6 terabits per second~60 TbpsBandwidth per satelliteBaseline10x the V2Capacity per launch vs. Falcon 91x~20x

With direct-to-cell service on the roadmap and every future Tesla carrying Starlink hardware, Sacks noted that an acquisition of T-Mobile is “easily within their range of purchases.” Musk, meanwhile, has claimed the network could eventually carry roughly half of all internet traffic.

Gerstner’s broader point was about risk appetite. Most public-company CEOs with a business like Starlink in their portfolio would stop taking risk — buybacks, dividends, safe bets. Instead, Musk is plowing the cash into Starship, AI compute, and the terafab. “I don’t know who else I give it to to do what he’s doing with Starship and with AI compute and the terafab,” Friedberg said. The irony the hosts highlighted: the market punishes capex — DoorDash for investing too much, Google for its AI buildout — which is precisely why most public companies won’t do what SpaceX is doing.

The $300 billion question

That ambition carries a price. Sacks sharpened the two most important questions about SpaceX’s compute buildout. First, is the $30 to $50 per watt spot price durable? Musk’s answer, relayed on the episode: the market is memory-constrained, with memory production growing perhaps 20% next year while demand is up more than 200%. The bottleneck persists, so prices may rise.

Second, how do you finance the next tranche? SpaceX is going from roughly 2 gigawatts of compute at year-end to a guided 5 to 10 gigawatts in 2027 (“closer to 10,” per management). That is an incremental 6 gigawatts at roughly $50 billion per gigawatt — about $300 billion in one year of capex.

MetricValueCompute today1.4 GW, heading to ~2 GW by end of 20262027 guidance5–10 GW, “closer to 10”Spot price$30–$50 per wattBuild cost~$50 billion per gigawattARR implied by 2 GW at $50 per watt~$100 billionIncremental capex for ~6 GW~$300 billionPayback at spotUnder one yearPayback under traditional pricing4–5 years at $10–$15 per watt per yearMemory market dynamicsSupply: +20% next year; demand: +200%+

Gerstner laid out the three options starkly: “In order to finance that, it seems to me you either have to go into the market and borrow the money or you have to do a dilutive equity raise, neither of which they want to do. Or you get Nvidia to backstop it.” He confirmed that Musk has said Nvidia backstopping “is happening more of,” but he flagged the fragility of the whole structure. The payback math is extraordinary at current spot prices — under a year — but it rests entirely on frontier labs continuing to pay three to five times market pricing because they are desperate for at-scale compute. If the July open-source scare (Chinese models undercutting frontier revenues) proves to be an early warning rather than a false alarm, the offtake collapses.

Gerstner invoked Bill Gurley’s warning about the circular financing that defined the zero-interest-rate era: “I can’t believe that we’re all just taking in stride this level of seller financing. He would call it circular revenues.” The same warning now applies to the AI infrastructure boom, where Nvidia and the hyperscalers are, in effect, financing the demand for their own products. Credit spreads on these deals remain wide — fear is already priced into the debt market.

The July pullback Gerstner referenced was not small. CoreWeave and semiconductor stocks fell roughly 40% on fears that Chinese open-weights models (Kimi, Qwen, and GLM were the names cited) would undercut frontier lab revenues, breaking the chain that justifies $50 per watt. Gerstner’s timeline for the infrastructure cycle is two more years, give or take. “The pace cannot be sustained at current levels for more than roughly two more years,” he predicted, but the counterweights are real. Jensen Huang told Gerstner directly: “Nobody comes close. Microsoft doesn’t come close. Google doesn’t come close in terms of standing it up in that time frame.” The physical constraint is not demand, in Huang’s view — it is the ability to build, and only one company has that capability at scale.

Jason added the demand-side conviction: “There’s no upper bound for on-demand intelligence.” Even as on-device and open-weights models improve, token consumption keeps rising. The open question the episode could not resolve: whether the incremental demand comes from premium frontier revenue, which is durable and high-margin, or from commodity usage, which is elastic but thin.

The SaaS cautionary tale: Airtable’s 90% collapse

The same two-tier logic that is reshaping AI is producing casualties in the application layer. The week’s most instructive deal was Bending Spoons’ acquisition of Airtable, once valued at $11.7 billion in 2021, for a headline price of $1.28 billion — roughly a tenth of peak value. Airtable had roughly $480 million to $500 million in run-rate revenue, 20% growth, profitability, and a billion dollars of cash on the balance sheet. Sacks called the postmortem a case study in structural misalignment.

Airtable was a product-led-growth company growing healthily. Its board, composed of investors who paid peak-2021 valuations, pressured management to bolt on a sales-led motion. The result: sales representatives hitting 30% quota attainment, “pushing on a string,” as Sacks described it. Hundreds of reps burning cash for no return. Bending Spoons, the Milan-based acquirer of challenged assets — AOL’s legacy business, Evernote, Eventbrite, Vimeo — went public last month and can do what Airtable’s own owners could not: cut 85% to 90% of the cost structure, return to product-led growth, keep most of the revenue, and generate $300 million to $400 million of EBITDA that pays for the acquisition in two to three years.

Airtable metricValueRun-rate revenue~$480–$500 millionGrowth~20% per yearCash on balance sheet~$1 billionHeadline sale price$1.28 billion (~$2.25 billion including cash)Peak valuation (2021)$11.7 billionSales-rep quota attainment30%Pro forma EBITDA potential$300–$400 million per yearPayback for buyer2–3 years

Notably, Airtable’s AI-agent business, called Hyper Agent, was spun out as an independent company before the sale. The talent went to a venture play. The legacy product went to a private-equity play. Sacks added a structural reason this could not have happened internally: venture boards and founders are incapable of shifting into the ruthless cost-cutting mode that private equity demands. Liquidation preferences, emotional attachment to what was built, and the lingering hope for a venture outcome all block it. And AI now makes maintenance mode vastly cheaper — an AI can learn a codebase instantly, so the institutional memory that used to require retaining expensive engineers is no longer a constraint.

Gerstner pushed back on the apocalyptic “SaaS is dead” narrative. The IGV software ETF is up 20% in six months. Snowflake is up 88% in six months. Databricks and ClickHouse are strong. Some companies, he argued, will make the jump to AI-first products — he named Figma as a likely survivor. Some applications — spreadsheets, dashboards, no-code tools — are being absorbed directly into the AI layer. And some, like Microsoft, are uncancellable not because they write the best code but because they are the rail everything else runs on: Active Directory holds identities, Excel holds board numbers, Teams holds compliance-recorded conversations, and Azure holds FedRAMP and Department of Defense Impact Level 5 clearances you cannot casually swap out. Gerstner’s honest close on Airtable, however, was that “if this is a failure, this is a pretty good failure for Silicon Valley” — late-stage investors got their money back, early investors made multiples.

American training data, Chinese AI

The episode’s final subject widened the lens from company-level economics to geopolitics. A Forbes investigation — “These American startups are making China’s AI smarter” — reported that US data-labeling startups, each valued above $20 billion, sell the same expert-created training datasets to OpenAI, Anthropic, and federal agencies that they sell to top Chinese labs including Tencent, Baidu, Alibaba, and Moonshot. The top six Chinese AI labs reportedly spend $500 million a year buying what Forbes called the “secret sauce”: PhD-written, double-verified technical content in code, biology, and science, plus reinforcement-learning and knowledge pipelines.

Jason Calacanis, who disclosed personal investments in a couple of these companies, took the hardest line: “I don’t think it’s very patriotic to be giving them an advantage. I wouldn’t do it.” In his telling, this training data is a large part of why Chinese open-weights models — Kimi, Qwen, GLM — have caught up so fast, alongside distillation from Western frontier models.

Sacks pushed back with a targeted-export-controls argument. If the goal is a blanket prohibition on business with China, that is a policy choice, but the historical rule is to restrict technology with genuine dual-use or military applications. Is PhD-written training data in that category? His answer was nuanced: data is largely a commodity, data labeling is trivially replicable in a country that graduates more math and science students than the rest of the world combined, and a restriction would create friction and invite retaliation — rare earths, for example — without decisively altering the race. The model to emulate, he argued, is the first Trump administration’s restriction on EUV lithography exports in 2019 — a targeted, high-impact control, not a broad embargo on data.

Gerstner’s read was political and contingent. The reason these data flows pass muster in Washington today, he argued, is that the US is winning the race, so the administration has no incentive to impose costs before a September bilateral meeting with China. But he flagged the tripwire: “If the president asks his advisers, one of these days, are we winning against China? And all of a sudden he gets a response, no, we’re no longer winning, they’ve caught up, they’ve passed us — then these things will get a lot more scrutiny.” Gerstner spent 60 days testing Kimi, Qwen, and GLM 5.2. His verdict: they are very good. Jason added the domestic corollary: the US is graduating world-class talent and simultaneously “kicking them out of the country” through immigration policy, compounding the data-export problem from the other side.

The unresolved tension that ran through the entire discussion, from Google to SpaceX to the Senate hearing that hasn’t happened yet, is whether American leadership holds. If it does, the permissive posture toward data flows and open-source competition that characterized 2025–2026 continues. If it doesn’t — if the president’s advisers one day deliver the answer Gerstner imagined — the gates close. And they close fast. The near-term items to watch are concrete: Anthropic’s potential IPO later this year at a rumored $1.5 trillion to $2 trillion valuation, the close of the Cursor acquisition and whether the combined Grok-plus-Cursor business reaches the $10 billion to $20 billion ARR Gerstner predicted by year-end, the first Starship flight carrying a full complement of 60 V3 satellites, xAI’s progress from 2 to somewhere between 5 and 10 gigawatts, and the September US-China bilateral meeting. In a world where intelligence itself is rapidly commoditizing, the most durable moat is the physical ability to stand up infrastructure faster than anyone else — and the most valuable pricing power is being one of the two labs still ahead of the curve.