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

Summary

Theses (episode spine)

• Google’s shake-up (Demis Hassabis moved to chair of DeepMind/chief scientist; Jeff Dean and colleagues leaving to start “Discovery Loop”) reflects a deliberate capital-allocation choice by Google to favor high-ROIC compute infrastructure ($200B committed capex) over high-beta frontier-model R&D; Friedberg frames capex as “high alpha, low beta” versus model development as “high alpha, high beta.”
• Sacks argues the frontier-model market has consolidated into a duopoly of Anthropic and OpenAI (“and then there were two”), citing Anthropic’s ARR run-rating over $80B (started the year at $10B, initial $100B exit-ARR forecast now tracking to $110-120B) and comparing the dynamic to Apple vs. Android, where the premium frontier tier captures disproportionate monetization even as commodity/open-weight models capture more usage.
• Jason counters that Google will still be the #1 AI company by consumer usage in 2026 — five products with 3B+ MAU (Search, Gmail, Chrome, YouTube, Android) plus Gemini’s MAU tripling year-over-year to 950M — and argues open-source/non-frontier models are already “good enough” for 95% of his own work, a view Elon Musk publicly disputed (“it’s actually a world of difference”).
• SpaceX’s first public-company earnings showed $7.8B Q2 revenue (+92% YoY), with “Elon Web Services” AI-compute revenue (renting Colossus compute to Anthropic/Google) more than tripling quarter-over-quarter to $2.6B; the stock still fell 13% on the print and sits roughly 30-40% below its post-IPO peak, now around a $1.4T valuation versus a >$2T IPO valuation.
• Gerstner is bullish on SpaceX over a 3-4 year horizon (sees potential to triple invested capital at the current valuation against either Morgan Stanley’s $325B 2030 revenue estimate or Elon’s own pulled-forward $1T ARR-by-2030 target) but stresses entry price matters, and flags financing risk for the next 6 gigawatts of compute buildout ($300B of capex) given a $30-50/watt spot compute price he isn’t sure will hold.
• Sacks separately argues Starlink alone could become close to a $1T market-cap business within about 18 months (12M subscribers doubling YoY, $66 ARPU, $2.6B quarterly adjusted EBITDA, a ~30x multiple justified by high-retention subscription economics), effectively funding the riskier AI-compute and Terafab bets.
• Airtable was acquired by Bending Spoons for $1.28B (about 10% of its 2021 peak $11.7B valuation) after a board-driven pivot to a sales-led motion failed (only 30% quota attainment); Sacks argues Bending Spoons can strip 80-90% of the cost structure, revert to product-led growth, and turn it into a highly profitable asset (potentially $300-400M of annual EBITDA, paying back the deal in about three years) — something the original board and founders were structurally unable to do themselves.
• Sacks and Gerstner push back on reading Airtable’s fall as a broad “SaaS apocalypse,” pointing to the IGV growth-software index up ~20% over six months (Snowflake +88-90%) and Salesforce’s continued lock-in (15 of 15 U.S. cabinet agencies) as evidence that compliance-critical enterprise software isn’t simply being “vibe-coded” away, even as no-code tools like Airtable are the most disrupted category.
• Jason and Sacks disagree over a Forbes report that U.S. data-labeling startups (valued at $20B+) are selling the same expert-curated training data to Chinese AI labs (Tencent, Baidu, Alibaba, Moonshot) for roughly $500M/year; Jason calls this “not patriotic” and argues it materially speeds Chinese model catch-up, while Sacks views the data as largely commoditized and non-dual-use, and warns that banning the sales risks reciprocal Chinese restrictions (e.g. rare earths) given China’s own large PhD talent pool.
• On financing the broader AI buildout, Sacks and Gerstner debate whether Nvidia should keep “backstopping” hyperscaler compute purchases, warning that if compute demand ever slips, the whole sector could reprice violently together — as happened in July when a Kimi-driven scare about undercut Frontier Lab pricing drove a roughly 40% drawdown in CoreWeave-type compute-rental and semiconductor names.

$SPCX (SpaceX)

Convergence / divergence: Both hosts came away bullish from the first public earnings print, agreeing the 13% post-earnings drop looked like a buying opportunity given the 92% YoY revenue growth and tripling AI-compute revenue. They diverge on emphasis — Gerstner focuses on financing risk for the next compute buildout and stresses that entry price/valuation discipline matters, while Sacks is more focused on the technical unlock (Starship’s heat shield working) that de-risks the Starlink bandwidth roadmap.

Speaker calls:

• Brad Gerstner (bullish, 3-4 years, target: triple invested capital): Says $1.4T of value creation is extraordinary and, on either Morgan Stanley’s $325B 2030 revenue estimate or Elon’s own pulled-forward $1T ARR-by-2030 guide, sees a path to tripling invested capital over 3-4 years, but stresses entry price matters and flags that financing roughly $300B of capex for the next 6 incremental gigawatts of compute is an open question.
• David Sacks (bullish): Calls the earnings call “very bullish” and was surprised the stock fell 13% despite a beat-and-raise; highlights the successful Starship heat-shield test as the unlock for 10x-bandwidth V3 satellites, and frames AI compute, Starlink and Grok/Cursor as “multiple ways to win,” while questioning whether the $30-50/watt spot compute price used to justify the $100B ARR guide will hold.

Cross-check:

• Price: $133.11 (P/E -58.77 TTM, unprofitable; mkt cap $1.79T). Next earnings: 2026-11-03 (unconfirmed; Q2 already reported Aug 4).
• Recent headlines worth knowing: Q2 2026 revenue $7.8B (+92% YoY), beat estimates of $6.81B; still operating at a net loss (-$8.89B TTM); first earnings-linked insider share-unlock triggered around the August 4 report.
• ⚠️ Inconsistencies: Stock has recovered to $133 / $1.79T market cap, well above the ~$1.4T valuation the hosts cite from the immediate post-earnings dip discussed on the show.

$GOOGL (Alphabet)

Convergence / divergence: All three see the Hassabis/Dean departures as a strategic pivot rather than a crisis, but they diverge on what it means for the stock: Jason is unambiguously bullish on Google’s consumer AI distribution advantage, Friedberg views the capital-allocation logic (infrastructure over frontier models) as sound business strategy without taking an explicit stock position, and Sacks implicitly downgrades Google’s standing in the “frontier intelligence” tier relative to Anthropic/OpenAI even while conceding its compute business is strong.

Speaker calls:

• Jason (bullish, 2026): Argues Google will be the #1 AI company by consumer usage in 2026 given five products with 3B+ MAU (Search, Gmail, Chrome, YouTube, Android) and Gemini’s MAU tripling year-over-year to 950M, and says open-source/non-frontier models are already good enough for 95% of his own work.
• David Friedberg (neutral): Frames the Dean/Hassabis departures as a rational board-level choice to allocate more capital to high-ROIC compute infrastructure ($200B capex this year) and less to high-beta frontier-model development, since Google can stay “model agnostic” via its stakes in SpaceX and Anthropic and its huge enterprise/consumer install base.
• David Sacks (neutral): Frames the frontier-model market as now a duopoly of Anthropic and OpenAI (“and then there were two”), implying Google has fallen behind on frontier-model leadership even if Google itself would still claim to be in the hunt.

Cross-check:

• Price: $356.62 Class C / GOOG (P/E ~15-18x trailing, below peer average; mkt cap $4.36T). Next earnings: ~late October 2026 (Q3, unconfirmed).
• Recent headlines worth knowing: Q2 revenue $119.8B (+24% YoY); Google Cloud revenue up 82% to $24.8B with 35.6% margin; Demis Hassabis moved to chair of DeepMind/chief scientist and Jeff Dean departed to start Discovery Loop amid Gemini 3.5 Pro delay reports.
• ⚠️ Inconsistencies: Despite the on-air narrative of an AI brain-drain crisis, GOOGL trades near record highs on a below-peer P/E with accelerating cloud growth, suggesting the market has not penalized the stock for the departures the way the episode’s framing implies.

$BSP (Bending Spoons)

Convergence / divergence: Only Sacks gave an explicit investment take; Brad Gerstner agreed with the underlying math (“$400 million to the bottom line pays for the acquisition in a couple years”) but did not stake out an independent stock view, so this is presented as a single-speaker call rather than a convergence/divergence pair.

Speaker calls:

• David Sacks (bullish, ~3 years): Views Bending Spoons’ $1.28B purchase of Airtable (against a cash-adjusted $2.25B ask and an $11.7B 2021 peak valuation) as a strong deal: by reverting Airtable to product-led growth and cutting the 80-90% of cost structure built up for a failed sales-led motion, he estimates Bending Spoons could generate $300-400M of annual EBITDA, paying back the acquisition in roughly three years.

Cross-check:

• Price: $51.43 (mkt cap ~$35B, estimates vary $22-35B across sources). Next earnings: 2026-08-13 (Q2 2026, pre-market).
• Recent headlines worth knowing: Bending Spoons IPO’d July 1, 2026 at $29/share (~$18.4B valuation); stock has since risen sharply, trading in a $30-58 52-week range; 7 of 7 covering analysts rate it Buy.
• ⚠️ Inconsistencies: Stock has nearly doubled since its July 1 IPO price the same month the Airtable deal (discussed on the show as a savvy distressed acquisition) closed, and Q2 earnings due Aug 13 will be the first test of the cost-cutting thesis Sacks laid out.

Topics discussed

Google’s AI Brain Drain and the Compute-vs-Model Strategy Debate

Summary: Google moved Demis Hassabis to chair of DeepMind/chief scientist and lost Jeff Dean plus other senior researchers to a new startup called Discovery Loop, amid reports Gemini 3.5 Pro is delayed and morale is low. The hosts debate whether this is healthy creative destruction, a rational capital-allocation choice toward compute infrastructure over model R&D, or a sign Google’s frontier ambitions are losing to internal “channel conflict” between renting out compute and building its own models.

Speaker views:

• David Friedberg: Sees it as rational — capex/infrastructure is high-alpha low-beta with near-certain ROIC given demand, while model development is high-alpha high-beta, so the board is right to tilt capital toward infrastructure even if that pushes top scientists to leave and raise their own capital.
• Brad Gerstner: Agrees with Friedberg’s framing and adds that Google, Microsoft, SpaceX and Meta all face the same “channel conflict” between renting out compute and building frontier models in-house, while Anthropic and OpenAI avoid it by staying purely in the model business.
• David Sacks: Argues the frontier-model race has become a two-company duopoly (Anthropic and OpenAI), with Google, xAI and others fighting for a commoditized “lagging intelligence” tier instead.
• Jason: Still expects Google to be the #1 AI company by usage in 2026 given its consumer product footprint and Gemini’s fast-growing MAU, and personally finds open-source/non-frontier models good enough for most of his own work.

Potential impact: A prolonged talent exodus from Google’s model teams could cement Anthropic and OpenAI’s pricing power at the frontier tier while pushing Google further toward a compute/inference-and-distribution business model rather than a model-leadership one.

Frontier Model Duopoly vs. Commodity Intelligence

Summary: The hosts discuss a Polymarket contract on which company will have the #1 AI model by year-end (OpenAI 32%, Google 20%, Alibaba 14%, others ~10% each) and debate whether the token/model business is bifurcating into a premium frontier tier and a commoditized lagging tier, versus one undifferentiated commodity market.

Speaker views:

• David Sacks: Sees a clear two-tier structure — a frontier duopoly (Anthropic, OpenAI) that can charge a premium like Apple, and a commoditized tier of lagging models where the money goes to whoever provides the compute/inference, not the model weights themselves.
• Jason: Argues the usage gap between frontier and open-source/non-frontier models is already negligible for most of his work, prompting a public disagreement with Elon Musk, who said frontier models remain a “world of difference” for high-speed or specialized use cases.
• Brad Gerstner: Cites Jensen Huang’s argument that closed frontier models are actually cheaper once training, fine-tuning and safety/guard-rail costs of self-hosting open models are included, reinforcing why frontier labs keep taking disproportionate revenue share.
• David Friedberg: Says the choice isn’t binary — enterprises will blend cheap open-weight models for simple workflows with premium frontier or specialized models (e.g. Gemini for video, life sciences and genomics) for high-value tasks, and argues it’s too early to count Google out on specialized/vertical models given its data advantages.

Potential impact: If the duopoly thesis holds, capital and revenue could concentrate further in Anthropic and OpenAI even as usage volume shifts to cheaper open-weight and Chinese models, reinforcing a barbell market structure across AI infrastructure investment.

SpaceX’s AI-Compute Buildout: Financing the Next 6 Gigawatts

Summary: Beyond the headline earnings, the hosts dig into SpaceX’s plan to grow from 1.4 to roughly 2 gigawatts of AI compute by year-end and to 5-10 (likely closer to 10) gigawatts next year, at an assumed $30-50 per watt spot price and roughly $50B of capex per gigawatt — implying up to $300B of additional capex next year that would need financing.

Speaker views:

• David Sacks: Questions whether the $30-50/watt spot price will hold given a memory-driven compute bottleneck, and asks how SpaceX will finance ~$300B of incremental capex — via debt, dilutive equity, or an Nvidia backstop — noting the payback period is currently assumed to be about a year.
• Brad Gerstner: Says Anthropic and OpenAI’s combined starting compute base this year was only about 5GW, so SpaceX’s planned addition is larger than what the two frontier labs started with combined, but argues the offtake demand is real for the next 12-24 months even as he flags a “wall of worry” around seller-financed circular revenue in the sector.

Potential impact: If frontier-lab compute demand ever slows, the heavily seller-financed AI-infrastructure buildout (SpaceX, Nvidia, CoreWeave and peers) could reprice sharply and in a correlated way, as already happened briefly in July on a Kimi-driven demand scare.

Airtable’s Sale to Bending Spoons and the “SaaS Apocalypse” Debate

Summary: Airtable, once valued at $11.7B in 2021, sold to Bending Spoons for $1.28B (about $2.25B including its large cash position) after a board-driven pivot to a sales-led growth motion produced only 30% sales-quota attainment. The hosts debate whether this is a one-off failure of a “quirky,” hard-to-explain no-code product or a signal that AI coding tools are broadly displacing the no-code software category.

Speaker views:

• David Sacks: Argues Airtable never achieved the ubiquity of a spreadsheet and was especially exposed because AI coding agents like Claude Code now let users build what Airtable/Retool used to require learning a tool to do; he separately blames Leopold’s much-discussed hedge-fund losses partly on a short position against Adobe and other SaaS names, and cautions against extrapolating Airtable’s fate to all of SaaS.
• Brad Gerstner: Notes revenue multiples can compress quickly once growth slows below ~50%, and that the real 2021-era investors got their money back on a straight 1x liquidation preference (no participating preferred), calling the outcome “a pretty good failure for Silicon Valley” given early-stage investors still profited.
• Jason: Points to Figma as a counter-example of a SaaS company with passionate users and strong founders that can make the jump to AI-first products rather than being disrupted like Airtable.

Potential impact: The hosts see no-code tools as the most AI-disrupted software category, but argue compliance-locked, rail-like enterprise systems (Salesforce, Microsoft) are comparatively insulated, so the SaaS impact from AI coding agents will likely be uneven rather than universal.

China Training on U.S. AI Data via Data-Labeling Startups

Summary: A Forbes investigation reported that U.S. data-labeling startups (transcribed in the episode as “Sergei” and “Meror,” both valued at $20B+) sell the same expert-curated training data used by OpenAI, Anthropic and U.S. federal agencies to top Chinese AI labs (Tencent, Baidu, Alibaba, Moonshot) for roughly $500M per year, prompting a debate about whether this materially helps Chinese models catch up and whether it should be restricted.

Speaker views:

• Jason: Argues this data — PhD-written, expert-verified content and reinforcement-learning pipelines — is a meaningful part of why Chinese open models (he cites Kimi, Qwen, GLM) have improved so quickly, calls it “not patriotic” for U.S. startups to sell it, and says one of the funds he backs (Micro/Mercor One) chose not to sell to China.
• David Sacks: Says the U.S. should apply targeted restrictions only to technology with real dual-use/military application (citing 2019 EUV lithography export limits as a good example), argues data labeling is largely a commodity China could replicate with its own large pool of PhDs, and warns that broad restrictions risk reciprocal Chinese retaliation (e.g. rare earths) without materially changing the AI race outcome.
• Brad Gerstner: Agrees the U.S. is currently winning the AI race and that competition should stay maximized, but warns this story, along with distillation and chip-related leakage concerns, will draw more scrutiny in Washington if the perception ever shifts to “China is catching up.”

Potential impact: If U.S. policymakers conclude this data flow meaningfully narrows China’s AI gap, it could trigger new export or trade restrictions on data-labeling firms, with risk of reciprocal Chinese retaliation on rare earths or other strategic inputs.