Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI

Summary

• Google is shifting capital toward infrastructure over frontier models, and that may be pulling core model talent away: Demis Hassabis moved to chair of DeepMind and chief scientist at Google, while Jeff Dean plus 3 stars left to found Discovery Loop. Shares fell 4%—which Jason loosely correlated with roughly $200B of lost market cap. Friedberg’s framing: with $200B of CapEx this year and accelerated depreciation handing back “26% off” every dollar, data-center capital is “high alpha, low beta” while model development is “high alpha… very high beta.” Scientists can instead raise “a couple billion at a multibillion-dollar pre-money with a PowerPoint deck.”
• Sacks’s market-structure call: “and then there were two.” He calls frontier intelligence a duopoly — Anthropic (from $10B to over $80B ARR this year, potentially $110-120B at exit) and an accelerating OpenAI — that can charge a premium like Apple vs Android, while models 6-12 months behind “can’t charge anything for the weights,” only for compute, inference, and consulting. He acknowledges Elon and Google still say they are in the hunt.
• JCal’s counter — the difference between open-source and frontier models is “negligible already” for 95% of his work — drew Elon’s public rebuttal: “It’s actually a world of difference.” Brad sides with Elon and cites Jensen’s claim that closed models can be cheaper all-in, explaining why frontier labs “continue to run away with it on the revenue side.”
• SpaceX’s first public quarter: $7.8B revenue, up 92% YoY, xAI Web Services more than tripling QoQ to $2.6B, CapEx $18.4B (6× YoY), stock down 13% post-print and ~30% since the June IPO to a $1.4T valuation. Elon guided $100B ARR by year-end and pulled the $1T ARR target forward to 2030; Brad calls Grok+Cursor “the sleeper” at a possible $10-20B by year-end, deserving a far higher multiple than GPU rental.
• Starlink alone could support a trillion-dollar valuation within 18 months, per Friedberg: $4.3B quarterly revenue, $2.6B adjusted EBITDA, 12M subs doubling YoY at $66 ARPU — extrapolating to ~$40B revenue and ~$30B FCF on a 30× multiple, funding “all of the rest of this as kind of science projects.” Starship’s V3 satellites add 60 Tbps per launch versus Falcon 9’s 2.6 — over 20× capacity per launch — supporting direct-to-cell.
• The financing question is the real risk: Sacks’s illustrative 2→8 gigawatt case next year—Elon said 5-10, closer to 10 than 5—at $50B/GW implies ~$300B of CapEx. Payback claims of under a year depend on a $30-50-per-watt spot price; Brad says frontier labs are willing to pay 3-5× market pricing for scarce, at-scale compute. Gurley’s warning hangs over it all: “I can’t believe we’re all just taking in stride this level of seller financing.” If demand slips, “everything will trade down together… it becomes much more violent.”
• Airtable sold for $1.28B — ~10% of its $11.7B peak — to Bending Spoons after spinning out its AI agent business Hyperagent. Sacks’s tell: only 30% of sales staff made quota, meaning the board bolted a sales-led motion onto a PLG company; the buyer could cut 85-90% of costs, run it at possibly 80-90% margins, and AI makes maintenance mode much easier because “the AI can go in and reconstitute that historical knowledge.” Brad’s caveat: don’t extrapolate to all SaaS — IGV is up 20% in 6 months, Snowflake 88%.
• US data-labeling firms Surge AI and Mercor, both valued over $20B, are reportedly selling PhD-built training data to Chinese labs spending $500M a year. Sacks resists a ban unless the data is proprietary and dual-use — “targeted strategic controls make sense… make sure that this one actually meets that bar” — while Brad warns the story “will muddy the waters” and would face much more scrutiny if advisors told the president “we’re no longer winning.”

Deep dive

1. Google’s AI leadership shakeup

• The news: Demis Hassabis moved to chair of DeepMind and chief scientist — Google framed it as a promotion, reports called it being “kicked upstairs” — while Jeff Dean, employee #30 with 27 continuous years, left with 3 other AI superstars to found Discovery Loop, focused on deep scientific breakthroughs in AI. Shares fell 4%, roughly $200B of market cap, though Jason presented that as a correlation rather than a proven causal loss.
• The morale backdrop, per the Axios quote read on air: “Google’s Gemini 3.5 Pro is months behind, with some company sources telling Axios that it’s in part due to low morale,” with several top researchers including Gemini’s co-lead already gone to competing labs.
• Friedberg’s insider detail worth keeping: Google had a ChatGPT equivalent internally a year before ChatGPT and chose not to release it for fear of cannibalizing Search. He said Sergey Brin then stepped in and there was a revitalization.

2. Why capital is shifting toward data centers

• Friedberg’s capital-allocation logic: Google committed $200B of CapEx this year, and thanks to accelerated depreciation at a 26% corporate rate, “every dollar you deploy in CapEx… you’re basically getting 26% off.” Compute is an obvious ROIC play with massive, high-confidence demand; frontier models cost tens of billions with open-weights catching up fast. His summary: “CapEx is high alpha, low beta… model development could be high alpha, but it’s very high beta.”
• The consequence for talent: if you’re Demis or Jeff Dean and capital flows toward infrastructure instead of your models, “given the fact that I can go down the road and visit Brad Gerstner… and raise a couple billion dollars at a multibillion-dollar pre-money with a PowerPoint deck because I’m the greatest in the world at doing this, that might be a better path.”
• Brad’s extension — channel conflict in multiple companies: Google Cloud wants the compute to rent to Anthropic while internal teams want it to compete with Anthropic; he described a similar conflict at Microsoft and SpaceX, which rents compute to Anthropic while building Grok. Brad said Anthropic and OpenAI have no such conflict because they say, “We’re not in the infrastructure business. We’re only in the model business.”

3. Duopoly or commodity? The frontier-model fight

• Sacks’s reaction to the Google news: “And then there were two.” Five companies were in the hunt a year ago; he now describes frontier intelligence as an Anthropic–OpenAI duopoly with a two-tier structure — a premium frontier market and a commodity tier 6-12 months behind where “you can’t charge anything for the weights,” only compute, inference, and consulting. He noted that Elon remains in the hunt and Google would say it does too. Anthropic went from $10B to over $80B ARR this year; its $100B year-end forecast, which many considered impossible, now looks achievable, with estimates rising to $110-120B or higher. His analogy: Android has more users, “but all the monetization goes to Apple.”
• JCal’s dissent, and the exchange that framed the episode: he tweeted that “the difference between the open-source models I’m using and frontier models is negligible already” for his use cases — and Elon replied, “It’s actually a world of difference.” JCal holds his ground for 95% of his work and expects Google to lead in consumer AI usage: 5 products over 3B monthly users each, Gemini at 950M MAUs in Q2, tripling YoY.
• Sacks’s rebuttal: for hedge funds and competitive industries, or immature use cases where “the return on finding those use cases is going to be so much greater than the small premium you’re paying at the token level,” you buy frontier. Friedberg says enterprises will blend: cheap open-weights for simple workflows, premium models for genomics or video, and “it is way too early to count Gemini out” on specialized models where Google has strong video and life-sciences data, with Demis still running Isomorphic Labs.

4. Price cuts, Jensen’s math, and “America’s winning”

• OpenAI and Claude both cut token prices hard; Brad’s read: “America’s winning. This is exactly what you want” — Chinese open source, domestic open source, and frontier labs all pressing prices down. He resists the duopoly label a few years in with Amazon, Microsoft, and Google still giants, but concedes the pure plays are gaining share of wallet.
• His two non-consensus points: Elon’s “we’re entering the singularity, and the frontier models are way further ahead than people think,” plus Jensen’s claim that closed models can be cheaper when users avoid building, training, fine-tuning, guarding, and maintaining them — “token consumption is going up for the open-source guys, while share of economics is going up for the frontier labs.”

5. SpaceX’s first public quarter: beat, raise, and a 13% drop

• The numbers: $7.8B revenue, up 92% YoY; xAI Web Services more than tripled QoQ to $2.6B (the Cursor acquisition had not closed); CapEx $18.4B in the quarter, 6× YoY (~$75B run rate). Jason attributed the 13% post-print stock drop to AI-CapEx concern; it was down ~30% since the June IPO above $2T to a $1.4T valuation. Brad called the decline normal: “within 6 months of the IPO, almost all these tech stocks are down 50% peak to trough.”
• Guidance Brad called extraordinary: $100B ARR by year-end, and the $1T ARR target pulled forward from 2031 to 2030 — against Morgan Stanley’s 2030 estimate of $325B, for a company that did $18B last year. His sleeper: Grok tripled tokens in July, and Grok plus Cursor “could be at $10 billion to $20 billion by the end of the year” — an asset trading at a much higher multiple than the data-center business.
• Sacks’s back-of-envelope on the guide: 1.4→~2 gigawatts by year-end at a spot price of $30-50 per watt ($30-50B per gigawatt, and he thought they were at the high end) — “all you have to believe is that they’re at 2 gigawatts running for $50 a watt” to hit $100B ARR, before Starlink, launch, or Grok-Cursor. On entry price, Brad’s discipline: at $2T on IPO day “I said I would want to own this company, but I’m not sure today’s the day I would buy” — versus Anthropic’s rumored IPO at $1.5-2T, which at $100B+ run-rate is “10 to 15 times revenue… not that much for a company that just grew 10× and is rumored to be profitable in Q2.”

6. Starlink: the cash machine that funds the science projects

• Segment math from the quarter: connectivity did $4.3B revenue and $2.6B adjusted EBITDA, 12M subscribers doubled YoY at $66 ARPU and grew 20% QoQ; Space was roughly breakeven, AI +$1.1B, with a question mark on whether rental pricing is a temporary scarcity premium. Friedberg’s extrapolation: ~24M-sub run rate, ~$40B top line, possibly $30B FCF within the year — at a 30× multiple, “the Starlink business alone could be a trillion-dollar market cap within 18 months,” funding everything else “as kind of science projects and upside.” Evidence: HughesNet and Viasat were described as decimated by Starlink.
• The Starship dependency, spelled out on air: Falcon 9 deploys 27 V2 satellites per launch (~2.6 Tbps of capacity); Starship deploys 60 V3 satellites at 10× bandwidth — 60 Tbps per launch, over 20× more capacity. Sacks said his understanding was that the last test also deployed 20 V3s, connected to them, and that the test satellites later burned up; the next milestone is 60 in the correct orbit. He said Elon had mentioned Starlink potentially handling “roughly half of internet traffic”; Friedberg mused they “might buy T-Mobile.”
• Brad’s meta-point: most CEOs would milk Starlink and skip Terafab, data centers, and models — “Elon refuses just to take the safe bet… It is heroic and important.” JCal adds a conditional Tesla angle: if the contemplated merger happens, he expects future Teslas to have Starlink built in, allowing phones to connect through Teslas and, with line of sight, to next-generation Starlink.

7. The $300 billion question: who finances the buildout?

• Sacks’s framing — the sharpest exchange of the episode: Elon said 2→5-10 GW next year, “closer to 10 than 5.” Sacks used 8 GW as an illustrative case, meaning ~6 incremental gigawatts at $50B per gigawatt = ~$300B of CapEx. The two risks are whether the $50/watt spot price holds—Elon thinks it rises because memory production may grow ~20% next year against 200%+ demand growth—and how to finance the buildout non-dilutively: “do you think NVIDIA gives them that financing?”
• Brad’s answer: borrow, dilute, or NVIDIA backstops it — but NVIDIA shareholders won’t backstop unlimited amounts because if spot goes against the business, payback could move from the current implied one year back toward the historical 4-5 years. Today’s 3-5× premium pricing exists because frontier labs “recognize they’re on the verge of some massive breakthroughs” — and the vast majority of offtake commitments are from Anthropic, OpenAI, and NVIDIA.
• The fragility, on the record: the July pullback happened because “Kimi scared people into thinking they’re going to undercut the frontier labs’ revenues… who the hell is going to pay for all this compute?” — a 40% trade-down in “the CoreWeaves of the world” that Brad described alongside a hoped-for “Leopold bottom.” Gurley’s line, invoked by Brad: “I can’t believe that we’re all just taking in stride this level of seller financing.” His own hedge: “Famous last words, I don’t see it today over the next 12 to 18 months” — but demand slippage would be “much more violent” given the leverage, and credit spreads on these deals remain wide.

8. Airtable at 90% off: a case study in venture-to-PE handoff

• The deal: Jason described Airtable as profitable, with $480M revenue growing 20% and ~$1B cash — sold for $1.28B ($2.25B including cash), ~10% of its $11.7B 2021 peak, to Milan’s Bending Spoons (Evernote, Vimeo, Meetup), whose newly public shares jumped 15%. Airtable spun its AI agent business Hyperagent into a separate company first — Sacks’s read: the talent keeps the venture play, sells the PE play.
• Sacks’s forensic detail: only 30% of the sales team made quota — a board chasing a venture outcome bolted a sales-led motion onto a PLG company and got “hundreds of sales reps pushing on a string.” Bending Spoons could “do what Elon did at Twitter: eliminate 85% or 90% of the cost structure,” keep most of the 20% growth, maybe generate $300-400M EBITDA — and AI makes maintenance easier because “you don’t need the historical knowledge anymore. The AI can go in and reconstitute that.” Brad’s pushback: “I don’t think it’s easy to get it to $400 million in EBITDA… once they slow down, morale goes to hell, churn spikes, it feeds on itself” — on a look-through basis this sold for maybe 30× free cash flow, not 2× revenue.
• The don’t-extrapolate case: Sacks argues no-code is “the most impacted, the most disrupted area of SaaS” — “what is Claude Code really good at? That’s the ultimate no-code tool” — but quotes the compliance moat: “Nobody buys Microsoft because Microsoft writes the best code… Azure holds a FedRAMP High authorization,” and Benioff tweeted 15 of 15 cabinet agencies run Salesforce. Notably, Leopold’s blowup wasn’t just long chips — he was short Adobe and SaaS, and those trades moved against him too. Brad’s tape: IGV up 20% in 6 months, Snowflake up 88%.
• The cap-table coda: “this is one of those cases where the liquidation preference actually mattered” (Sacks). Brad said late-stage investors from the $2B/$5B/$11B rounds (Altimeter passed on all 3) appeared to get their money back, while early-stage investors made money. Sacks separately explained that clean terms generally mean a 1× liquidation preference, rather than a participating preferred double dip: “if this is a failure, this is a pretty good failure for Silicon Valley” (Brad).

9. Selling America’s “secret sauce” to Chinese labs

• The Forbes investigation: US data-labeling startups Surge AI and Mercor (both valued over $20B) reportedly sell PhD-written content, reinforcement-learning material, and knowledge pipelines to OpenAI, Anthropic, and federal agencies — and the same datasets to Tencent, ByteDance, Alibaba, and Moonshot, with China’s top 6 labs spending $500M a year. JCal discloses investments in the space and notes Micro1’s founder chose not to sell to China.
• Sacks’s test — worth keeping: is the data proprietary, is it dual-use, and does it have a military application? Data labeling is a commodity China can replicate with its own labor; a ban could invite reciprocal action (“maybe rare earths”) while “we still, at this moment in time, do have some dependencies.” He cited the first Trump administration’s EUV lithography export restriction—“I think” it was in 2019—as an example of a targeted control that packed a punch: “targeted strategic controls make sense. I would just make sure that this one actually meets that bar.”
• JCal’s dissent: this isn’t merely labeling — it is Western experts packaging “all the knowledge of the West” for LLMs, “a big part of why they’re catching up… I don’t think it’s very patriotic to be giving them an advantage.” Friedberg is unsure the material is truly secret sauce or that China cannot recreate it; he noted that China graduates “more math and science graduates every year than the rest of the world combined.”
• Brad’s political forecast: the story “will muddy the waters” alongside distillation and chip exports, but currently passes muster, in his view, because “we’re still leading the race” — the day the president hears “we’re no longer winning,” the issue would receive much more scrutiny. Timing note: Brad mentioned a September bilateral meeting with the president.