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

Source: Authorized transcript of the All-In Podcast. The episode states that it is August 6. Jason Calacanis appears to host, with David Friedberg, Brad Gerstner, and David Sacks participating; Chamath Palihapitiya is described as traveling and absent from the main discussion.

Source status: This document records what the speakers said. It does not independently verify their claims, financial figures, company developments, political assertions, predictions, or reported news.

Transcript quality: The transcript is usable but contains significant speech-to-text errors. Examples include names and terms rendered as “Hasabas,” “Gersner,” “Chimov/Chimap,” “Sotcha,” “GROC,” “Curser,” “Starling,” “Gwynchawall,” “MeriCorps,” “Micro One,” “Kimmy,” “Quinn,” and others. Speaker labels are also absent. Attribution below follows the conversational handoffs where reasonably clear; uncertain wording is marked [transcript unclear] rather than silently corrected.

The principal subjects are:

1. Google's AI leadership changes and the economics of frontier models versus AI infrastructure.
2. The emerging structure of the AI-model market and open-source competition.
3. SpaceX's reported public-company earnings, Starlink, AI compute, Starship, and future capital requirements.
4. The All-In Summit.
5. Airtable's reported acquisition by Bending Spoons and implications for SaaS.
6. U.S.-generated AI training data reportedly being sold to Chinese AI companies.

The hosts open by saying it is summer and August 6 and joke that it has been difficult to assemble a quorum. David Friedberg and Brad Gerstner are present. Chamath Palihapitiya is described as traveling, while David Sacks joins after initially being expected to arrive late.

The opening banter includes a photograph said to show Chamath visiting the progress of a data center, possibly in Colorado or Nevada [transcript unclear], jokes about an associated meme, and reference to Brad Gerstner appearing on CNBC.

Jason introduces reports of major changes in Google's AI organization.

He says:

• Demis Hassabis was moved to chair of DeepMind and chief scientist at Google.
• Some reports characterized the move as Hassabis stepping down or being “kicked upstairs.”
• Google characterized it as a promotion.
• A report attributed in the transcript to “Exeosis” [transcript unclear; likely publication/name error] said Gemini “3.0.5 Pro” [verify exact model name] was months behind schedule and that low morale was partly responsible.
• Several important researchers, including a Gemini co-lead, were said to have departed for competing AI laboratories.
• Jeff Dean and three other researchers were reported to be leaving Google to start a company called Discovery Loop.
• Dean was described as Google employee number 30, having joined in 1999 and remained for approximately 27 years.
• Discovery Loop was described as intending to pursue major scientific breakthroughs using AI.
• Google's shares were said to have fallen approximately 4%, representing roughly $200 billion of market capitalization, although Jason explicitly framed the connection to Dean's departure as a possible correlation rather than an established causal relationship.

David Friedberg rejects a simple choice between the departures representing failed scientists or entrepreneurial temptation. He proposes a third explanation: Google's management and board may be deciding that capital invested in AI infrastructure has a more attractive risk-adjusted return than capital invested in frontier-model development.

Friedberg says Google has committed roughly $200 billion of CAPEX during the year to AI infrastructure and data-center construction.

His reasoning is:

• U.S. accelerated-depreciation treatment makes domestic AI infrastructure unusually tax-advantaged.
• Demand for compute is extremely high.
• Google already knows how to operate large compute infrastructure efficiently.
• Therefore, money invested in compute can potentially generate large profits with comparatively high confidence.
• By contrast, creating the most advanced frontier model requires tens of billions of dollars without an equally predictable path to profit.
• Open-source and open-weight models are improving quickly.
• Frontier laboratories are also rapidly matching one another.
• As a result, the economic return from spending heavily to produce the single best model may be less compelling than operating infrastructure capable of supporting many models.

Friedberg says Google had something equivalent to ChatGPT internally roughly a year before OpenAI released ChatGPT.

He attributes Google's decision not to release it to fear that the product would cannibalize Google Search. He says Sergey [surname omitted in transcript] subsequently became involved in a wider revitalization of Google's AI effort.

Friedberg's larger argument is that although Google was early in frontier AI, the competitive economics subsequently changed.

Friedberg argues that Google's existing strategic position lets it profit even if it does not possess the world's best frontier model.

He says Google has:

• an enormous installed enterprise-computing customer base;
• enormous consumer distribution;
• the ability to host multiple model providers;
• relationships with Anthropic, OpenAI, SpaceX and open-weight-model developers;
• significant ownership interests in SpaceX and Anthropic, according to the source.

That permits Google to become model agnostic rather than forcing the company to win the frontier-model competition itself.

Friedberg argues that the same capital allocation that is attractive for Google's shareholders may be unattractive to leading researchers.

If scientists such as Hassabis or Dean want to pursue frontier intelligence but Google is directing incrementally more money toward data centers and ecosystem infrastructure, those scientists have another option: leave and raise billions of dollars from venture investors based largely on their reputations and a presentation.

He summarizes the risk profile as:

• Data-center infrastructure: potentially “high alpha, low beta.”
• Frontier-model development: potentially high alpha but also very high beta.

His conclusion is that, if he were allocating Google's capital, he would put more into compute infrastructure and comparatively less into frontier-model development.

Brad Gerstner says Friedberg's interpretation fits what is happening elsewhere, including Microsoft.

He cites Satya Nadella [rendered “Sotcha” in transcript], in turn citing Morgan Stanley, as saying Microsoft was seeing more than a 30% return on invested capital in a “tokens as a service” or infrastructure business.

Gerstner says this creates a powerful internal conflict at Google:

• Google Cloud wants compute available to rent to customers such as Anthropic.
• Google's internal frontier-model researchers want the same compute so they can compete against Anthropic.

That creates what Gerstner calls an inherent channel conflict.

He thinks the conflict appears to be resolving in favor of Google becoming more infrastructure-oriented.

Gerstner says the same structural issue exists at several companies:

• SpaceX, according to the discussion, rents compute to Anthropic while also pursuing its own model-related businesses involving “GROC” and Cursor [terminology/ownership relationships require verification].
• Microsoft has infrastructure incentives and, in Gerstner's view, is no longer obviously pushing the model frontier.
• Meta has discussed entering infrastructure-as-a-service.
• Anthropic and OpenAI differ because their primary business is the model layer rather than renting infrastructure.

Gerstner suggests this might ultimately remove some diversified technology companies from the very front of model development, particularly if their leading researchers leave.

Gerstner also says that, assuming a 26% corporate tax rate and immediate CAPEX depreciation, each dollar spent on eligible CAPEX effectively receives approximately a 26% tax benefit in the current year.

This claim should be checked against the exact applicable tax law and accounting treatment before publication.

Jason cites a Polymarket market concerning which company would possess the leading AI model on December 31.

The probabilities he gives are approximately:

• OpenAI — 32%
• Google — 20%
• Alibaba — 14%
• Moonshot — roughly 10%
• xAI — roughly 10%
• Meta — roughly 10%
• ByteDance — roughly 10%

He says Anthropic had won the previous version and therefore was not included [exact market rules should be verified].

David Sacks says his reaction to Google's personnel news was essentially “and then there were two.”

He argues that a year earlier roughly five companies were serious candidates to operate the world's leading frontier AI lab, whereas the market now appears to have consolidated around Anthropic and OpenAI.

He acknowledges:

• Elon Musk/xAI remains in pursuit.
• Google would almost certainly dispute the claim that it has left the frontier race.
• Google's infrastructure economics may nevertheless weaken its incentive to spend aggressively to win that race.

Sacks therefore describes frontier intelligence as a powerful duopoly, not as a commodity.

Sacks predicts a two-tier structure:

Tier 1: Frontier intelligence

• The most capable intelligence available.
• Providers can charge a premium for the model itself.
• He places Anthropic and OpenAI in this group.

Tier 2: Commodity or lagging intelligence

• Roughly six to twelve months behind the frontier, in his characterization.
• Still potentially a huge market.
• Model weights themselves become difficult to monetize.
• Revenue instead comes from compute, inference and services needed to implement the models.

His core distinction is that a provider at the frontier can monetize intelligence itself; a provider behind the frontier mostly monetizes infrastructure surrounding the intelligence.

Sacks says:

• Anthropic had exceeded $80 billion of ARR.
• It started the year at approximately $10 billion.
• It had projected a $100 billion exit ARR for the year.
• That target had initially been considered implausible by many observers.
• It now appeared, in his assessment, that Anthropic could reach the target months early.
• He speculates that exit ARR might therefore rise to $110 billion, $120 billion or more.
• OpenAI was also seeing acceleration.

These are source claims and require independent verification.

Sacks compares the AI market to Apple versus Android.

Android has greater global user volume, he says, while Apple captures disproportionate monetization because customers pay for a premium experience.

He believes frontier AI could work similarly: inexpensive or open models may win vast usage while genuine frontier intelligence captures premium economics.

Jason distinguishes having the best model from being the largest AI company.

He cites reported Google Cloud revenue growth of 82% year over year and predicts Google could become the number-one AI company by consumer usage because AI is being integrated into products already used by billions of people.

He says Google has five products with more than three billion monthly users:

• Android
• Search
• Gmail
• Chrome
• YouTube

He also says Google has 13 products with more than one billion users and that Gemini had more than “950 monthly active users” in Q2, triple the prior year. The transcript omits a unit after 950, so this is [transcript unclear] and must not be silently converted to millions.

Jason says that he now uses non-frontier models almost exclusively and that, for approximately 95% of his own work, they are good enough.

He had publicly argued that the difference between the open-source models he was using and frontier models had already become negligible for his work.

Elon Musk reportedly responded that there was actually a “world of difference.”

Jason nevertheless argues that many ordinary applications no longer need the frontier and that enterprise adoption of open-source models is partly slowed by implementation difficulty. He expects implementation to become progressively easier.

One sentence concerning copying video games, speed requirements and when frontier models are necessary is internally garbled and should be treated as [transcript unclear] rather than reconstructed.

Sacks agrees that commodity intelligence can be sufficient for many use cases but argues that “good enough” does not eliminate demand for the best.

He gives the analogy that a cheap Android phone would probably serve his practical needs, yet he still pays a premium for a device he uses continuously.

He applies this to businesses:

• A hedge fund operating in a highly competitive industry may refuse to risk using second-best intelligence.
• Any sector where marginal intelligence creates competitive advantage may willingly pay for frontier capability.
• Organizations searching for new AI use cases should also begin with the strongest model because they do not yet know where the economic value is.
• In that exploratory phase, the return from discovering a valuable use case can dwarf the additional token cost.
• Companies beginning AI deployment may prefer the convenience of a complete frontier stack.

His conclusion is that the token-price premium can be trivial relative to the possible business benefit.

Friedberg rejects the assumption that companies must choose either “the best model” or an open-source model for everything.

He expects model blending and routing.

His examples:

• Open-source/open-weight models for simple, routine workflows.
• More capable premium models for difficult or commercially important workflows.
• Highly specialized models for fields such as life sciences and genomics.
• Best-in-class video models for media generation.

For a media company producing AI video, he says he would use whichever model was demonstrably best for video—giving Gemini and Sora as examples—rather than applying one universal model choice.

Friedberg expects consumers to be more likely to pay roughly $20–$40 per month for a complete service such as ChatGPT, Gemini or Claude rather than manage their own open-weight model.

Enterprises, by contrast, will actively select different models for different jobs:

• cheap models for simple workflows;
• stronger general models for important tasks;
• specialized models for specialist applications.

This favors cloud platforms capable of offering the whole mix.

Friedberg says it is far too early to write Gemini off.

He argues Google has advantages in:

• video data;
• life-sciences data;
• years of prior life-sciences research;
• protein-related research;
• Isomorphic Labs, which Hassabis was still expected to run.

He predicts Google could be particularly strong in vertical models for video, life sciences and protein folding even if the generic frontier market develops differently.

This reinforces his argument for Google Cloud: an enterprise may prefer a cloud provider that can supply a mixture of models rather than being tied to one model company.

The discussion notes large token-price reductions from OpenAI and Claude/Anthropic and increasing model orchestration within enterprise software.

Brad Gerstner argues this competition is healthy.

He describes a market containing:

• Chinese open-source models;
• domestic U.S. open-source models;
• frontier laboratories;
• major technology companies.

He says this competition is creating downward price pressure and characterizes the overall situation as one in which “America's winning.”

Although Gerstner agrees that Anthropic and OpenAI have emerged as major pure-play frontier companies and appear to be winning greater share of customer spending, he thinks calling the market a permanent duopoly is premature because Amazon, Microsoft and Google remain enormously capable competitors.

Gerstner highlights two recent arguments he considers important:

1. Elon Musk's view: frontier models remain much farther ahead of open-source and Chinese alternatives than current commentary suggests.
2. Jensen Huang's reported view: closed models may actually be cheaper when the user includes all the cost of operating open alternatives—training, fine-tuning, expertise, maintenance, guardrails and safety.

Gerstner therefore thinks:

• open-source models may gain a large share of token consumption;
• frontier laboratories may continue gaining a large share of the economic profit.

He again compares this to Android versus iPhone: one can dominate usage while the other captures more profit.

The discussion then turns to what the speakers describe as SpaceX's first earnings report as a public company.

Jason states:

• Shares fell approximately 13% after the report.
• The stock was about 30% below its June public listing.
• SpaceX was then valued at approximately $1.4 trillion, after having traded above $2 trillion.
• Q2 revenue was approximately $7.8 billion, up 92% year over year.
• The AI/compute business, which Jason calls “Elon Web Services,” had more than tripled quarter over quarter to approximately $2.6 billion.
• The revenue was attributed primarily to compute rented from the Colossus infrastructure to Anthropic and Google.
• Cursor was not included because the relevant transaction reportedly had not yet closed.
• Quarterly CAPEX was approximately $18.4 billion, about six times the prior-year level.
• Jason translates that into a rough annual run rate near $75 billion.

All of these claims require verification before publication.

Brad Gerstner calls the creation of roughly $1.4 trillion in public-market value extraordinary and does not find the post-IPO decline surprising.

He notes that many major technology IPOs have experienced approximately 50% peak-to-trough declines during their first six months.

He nevertheless considers the reported quarter strong.

Gerstner cites:

• a target of $100 billion of ARR by year-end;
• a $1 trillion ARR target moved forward from 2031 to 2030;
• a Morgan Stanley 2030 revenue estimate of approximately $325 billion;
• approximately $18 billion of company revenue in the previous year.

He says that, if either the Morgan Stanley or Elon Musk forecasts prove broadly correct, the current market price does not reflect the eventual scale.

Gerstner says investors have to decide how to value two different AI activities:

Compute rental

• SpaceX reportedly rented a very large block of compute to Anthropic.
• Investors must ask whether SpaceX should rent its hardware or retain it for its own models.
• Very large planned compute deployments create financing and customer-offtake questions.
• Pure GPU-rental companies generally receive lower valuation multiples.

Frontier/model businesses

Gerstner calls this the potential “sleeper.”

He says:

• “GROC” [verify as transcript likely inaccurate] tripled tokens during July.
• Cursor was reportedly on a path from approximately $3 billion to $10 billion by year-end.
• He speculates that Cursor plus “GROC” could reach $10–$20 billion by year-end.
• If so, those businesses could deserve much higher valuation multiples than data-center rentals.

Gerstner identifies two particularly important indicators:

1. Year-end revenue for “GROC” and Cursor.
2. Starlink's success in displacing conventional mobile carriers.

He also notes that venture funds holding SpaceX shares had begun distributing stock to their investors, which could create selling pressure as long-time private holders obtain liquidity.

He says he personally intends to hold the shares he receives for his grandchildren.

Sacks calls the earnings call bullish and says he was surprised the stock declined.

He emphasizes Starship.

According to his account:

• a recent Starship flight succeeded;
• the heat shield worked;
• the vehicle was left floating in the ocean;
• the success should permit a faster test cadence;
• that in turn enables deployment of a new generation of Starlink satellite;
• the new satellite generation increases network bandwidth;
• the increased capacity is necessary for a much larger direct-to-cellular business.

He therefore connects Starship engineering success directly to Starlink's telecom economics rather than viewing it only as a space-launch project.

Sacks says SpaceX expected to increase compute from approximately 1.4 gigawatts to 2 gigawatts by year-end.

He reports Elon Musk as quoting a compute spot price of approximately $30–$50 per watt.

His arithmetic is:

• 1 gigawatt = 1 billion watts.
• $30 per watt = roughly $30 billion per gigawatt.
• $50 per watt = roughly $50 billion per gigawatt.
• Two gigawatts operating at $50 per watt would therefore correspond to approximately $100 billion.

He presents that calculation as one explanation for the company's reported $100 billion ARR optimism.

He stresses that the calculation would exclude Starlink, launch revenue and the model/software businesses.

Friedberg focuses on Starlink as an unusually powerful cash-generating business.

He cites reported segment information:

• Starlink/connectivity adjusted EBITDA: approximately $2.6 billion.
• Space segment: roughly negative $200 million, which he rounds toward break-even.
• AI segment: approximately +$1.1 billion.
• Starlink quarterly revenue: approximately $4.3 billion.
• Subscribers: approximately 12 million.
• Subscribers doubled year over year.
• Monthly ARPU: approximately $66.6 [transcript formatting unclear].
• Growth: approximately 20% quarter over quarter.

Friedberg distinguishes the apparent durability of Starlink from the current AI compute economics.

He says AI rental revenue may currently benefit from an unusually high shortage premium. Customers desperate for scarce compute may be paying SpaceX rates that will not persist indefinitely.

That creates uncertainty around the long-run profitability of the AI infrastructure segment.

Friedberg extrapolates that Starlink could approach a 24-million-subscriber run rate and, if consumer and enterprise products such as airline connectivity continue scaling, might reach approximately:

• $40 billion in annual revenue;
• possibly $30 billion in annual free cash flow.

He then applies a hypothetical 30× multiple, arguing the recurring, high-retention subscription economics could support it.

From that, he speculates that Starlink alone could eventually justify a $1 trillion valuation, possibly within roughly 18–24 months. This is explicitly a bullish valuation case, not an established fact.

Friedberg says Starlink has already damaged legacy satellite businesses such as HughesNet and Viasat and refers more broadly to pressure on traditional communications providers.

As a personal example, he says his Sonoma County ranch once used HughesNet for roughly $200 per month and that the service was poor. Starlink, in his view, destroyed the economics of that older satellite-internet market.

Friedberg says Starlink was then adding approximately two million consumer subscribers per quarter.

He predicts direct-to-handset service could push additions to four or five million per quarter.

He cites approximately 400 million mobile subscriptions in the United States as the size of the domestic opportunity.

His broader argument is that Starlink alone can generate enough cash to fund much of Musk's other high-risk work.

The speakers discuss Starlink as a potential funding engine for:

• Starship;
• AI compute infrastructure;
• model development;
• a project rendered as “TerraFab”/“tariffab” [transcript unclear].

The semiconductor-fabrication idea is framed as strategically important because it could reduce U.S. dependence on Taiwan and China for semiconductors.

One speaker describes the envisioned facility, if successfully executed, as potentially becoming the world's greatest semiconductor fabrication site.

That is a prediction and expression of enthusiasm, not a verified project outcome.

Gerstner contrasts Musk with executives who might simply preserve Starlink as a highly profitable cash-generating company.

He says many managers would avoid additional risks once they possessed an asset potentially worth one or two trillion dollars.

Musk instead continues investing in:

• fabrication;
• data centers;
• frontier models;
• other large technical projects.

Gerstner calls these investments highly risky but also important and describes Musk's willingness to reinvest as “heroic.”

He says he wishes more public-company CEOs accepted comparable levels of frontier risk.

Another speaker adds that companies are increasingly recognizing that share buybacks and dividends are not necessarily more important than investing in the future, while noting that markets have punished companies such as DoorDash and Google for unusually large CAPEX plans.

Several counterarguments are acknowledged.

One risk is whether current demand for tokens and compute can continue indefinitely.

Potential downward pressures include:

• local/on-premise computing;
• powerful desktop hardware;
• open-source models becoming smaller and more efficient.

The speaker ultimately rejects the bearish conclusion, saying he does not know whether there is an upper bound to demand for “on-demand intelligence.”

Another risk is that conventional Starlink is less compelling in dense locations already served by good fiber or cable connections.

The response offered is that future integration with vehicles and direct-to-device technology may dramatically expand Starlink beyond rural fixed broadband.

The discussion speculates that:

• future Teslas could contain Starlink hardware after a corporate combination [transaction/status requires verification];
• vehicles and robotaxis could provide connectivity to nearby phones;
• next-generation Starlink could connect directly to handsets with a suitable line of sight.

These are forward-looking claims.

The discussion uses the familiar analogy that private venture markets are a voting mechanism while public markets become a weighing mechanism.

One example cited is Tesla previously trading at roughly 160× price-to-sales and later moving closer to 45× as revenue grew relative to valuation.

The argument is that rapid revenue growth can eventually resolve apparently extreme early valuation multiples.

Gerstner considers SpaceX's public-market correction healthy and predictable.

At approximately $1.4 trillion, he says a three- or four-year investor might plausibly envision tripling their money if Musk or Morgan Stanley's forecasts prove accurate.

However, he emphasizes that even an extraordinary business can be a poor investment at the wrong price.

He recalls saying on CNBC on the day of the SpaceX IPO:

“I would want to own this company, but I'm not sure today's the day I would buy the company.”

His point is that investors who bought above $2 trillion should have anticipated a possible correction.

Gerstner says he has heard discussion of an Anthropic IPO later in the year at approximately $1.5–$2 trillion.

Using the earlier estimate that Anthropic might exceed $100 billion of run-rate revenue, he says this would represent roughly 10–15× revenue.

He argues this is not necessarily excessive for a company that had grown approximately tenfold and was rumored to have reached profitability in Q2.

Both the IPO possibility and profitability claim are presented as reports/rumours and require verification.

The speakers argue that Musk possesses an advantage not captured by comparisons focused only on AI software.

His companies have repeatedly demonstrated the ability to build physical infrastructure quickly, including factories and rockets.

The discussion cites Tesla gigafactories and SpaceX's hardware experience.

One speaker paraphrases Musk's argument that data centers are comparatively simple because building a data center is “not rocket science.”

The thesis is that AI software ultimately requires physical hardware, power and facilities, giving an operator skilled at constructing large industrial sites a strategic advantage.

Sacks gives a more detailed technical explanation.

According to the transcript:

• Starlink's existing network uses V2 satellites.
• The new V3 satellite has approximately 10× the bandwidth of V2.
• Falcon 9 carries roughly 27 satellites per launch.
• Such a launch adds approximately 2.6 “terabytes per second” of total network capacity [unit should be verified; transcript may be wrong].
• Starship can deploy roughly 60 V3 satellites per launch.
• That would add approximately 60 terabits per second of capacity per launch.
• He characterizes this as more than 20× the capacity added per launch.

Sacks says his understanding was that a recent Starship test deployed approximately 20 V3 satellites as a test.

He says:

• connections were established with the test satellites;
• cameras were installed;
• the satellites were not intended to remain permanently and later burned up.

He identifies the next major milestone as launching approximately 60 V3 satellites, placing them in the correct orbit, connecting them successfully and adding their bandwidth to the operating Starlink network.

Sacks argues that successful V3 deployment could eventually raise Starlink network capacity by 10× or even 100×.

That would enable more ambitious direct-to-cellular applications.

The conversation also references comments from Gwynne Shotwell concerning ground stations [exact statement unclear], speculates that SpaceX might someday be capable of acquiring T-Mobile, and attributes to Musk a suggestion that Starlink might eventually handle roughly half of global internet traffic.

These statements should all be independently checked.

Sacks says Musk indicated:

• roughly 2 gigawatts of compute by year-end;
• between 5 and 10 gigawatts the following year;
• closer to 10 than 5.

Sacks then chooses 8 gigawatts merely as an illustrative assumption, explicitly saying he is making the number up within Musk's stated range.

That would mean approximately 6 incremental gigawatts.

Sacks asks whether today's reported compute price of roughly $50 per watt can survive such rapid expansion.

He says he interpreted Musk as expecting continued high prices because the market was constrained by memory supply.

The numbers he recalls are:

• memory production might rise around 20% the next year;
• demand might rise more than 200%.

His point is that whichever component is most constrained at a given moment sets the pace of AI infrastructure expansion.

Using an estimate of $50 billion of CAPEX per gigawatt, Sacks calculates that six additional gigawatts could require roughly $300 billion.

He asks how that could be financed without heavily diluting existing shareholders.

Possible mechanisms raised include:

• borrowing;
• an equity offering;
• NVIDIA providing financing or backstopping purchases.

He notes the company had suggested compute investments could currently pay back in a year or less, but says that economics depends heavily on current spot prices.

Gerstner agrees that approximately $50 billion per gigawatt is a reasonable minimum construction estimate in the context being discussed.

He says the financing alternatives are essentially:

• debt;
• dilutive equity;
• vendor or NVIDIA-supported financing.

He warns that NVIDIA shareholders are unlikely to accept unlimited balance-sheet exposure because compute prices could fall and extend the investment payback period.

Gerstner says a more traditional calculation had recently been:

• spend about $50 billion;
• earn approximately $10–$15 billion annually;
• recover the investment over four or five years;
• a sixth year of useful economics pushes returns well above 20%.

Current shortage pricing, however, can imply a payback period close to one year.

He does not expect a one-year payback to represent normal long-term economics.

Gerstner says compute scarcity is so severe—and frontier labs believe the strategic stakes are so high—that they may pay three, four or even five times ordinary market pricing to secure very large blocks of compute.

He presents the Anthropic/SpaceX deal as an example.

He believes Anthropic and OpenAI would buy substantially more capacity immediately if they could obtain it.

The reported $30–$50-per-watt price is presented as a reflection of that strategic scarcity.

Gerstner says relatively few customers can afford compute at this scale.

He identifies Anthropic, OpenAI and NVIDIA as responsible for much of the large-scale offtake commitments discussed in the market.

He says hyperscalers are therefore building infrastructure partly to serve this narrow group of exceptionally large AI buyers.

Gerstner believes Musk may be uniquely capable of constructing six additional gigawatts on the proposed timetable.

However, even if SpaceX can build the facilities, it must still secure:

• memory;
• chips;
• land;
• power;
• completed building shells and associated infrastructure.

He presents component availability rather than construction management alone as the primary constraint.

Gerstner says Anthropic and OpenAI together began the year with roughly five gigawatts of total compute.

Adding six gigawatts would therefore exceed their combined starting capacity.

However, he argues such expansion is plausible if:

• Anthropic is growing roughly 10×;
• OpenAI is growing roughly 4× or more.

He believes demand for the next 12–24 months is real.

Gerstner nevertheless describes significant market anxiety.

He says a July scare involving “Kimmy” [likely model/company name incorrectly transcribed] made investors fear that cheaper models could undermine frontier-lab revenue.

The chain of risk is:

1. Cheaper/open models pressure frontier-model revenue.
2. Frontier labs therefore need less expensive compute.
3. Demand for enormous data-center construction falls.
4. Highly leveraged compute suppliers and semiconductor companies reprice sharply.

He attributes roughly a 40% decline in CoreWeave-type companies and semiconductor/AI-related shares to this fear.

The group discusses a hypothetical next 10 gigawatts costing roughly $500 billion.

Potential financing structures include:

• secondary/equity offerings;
• NVIDIA holding exposure on its own balance sheet;
• special-purpose vehicles kept outside operating-company balance sheets.

Gerstner sees it as positive that investors are already discussing these risks because markets can adjust in real time if expansion slows.

Although he is bullish on near-term demand, Gerstner doubts this construction and financing pace can continue for more than roughly another two years.

He invokes Bill Gurley's concerns about extensive seller financing and what Gurley reportedly calls “circular revenues.”

The concern is that vendors and financiers effectively help customers construct capacity ahead of the customers' own revenues.

That leverage magnifies downside if demand eventually slips.

Gerstner says he does not presently see demand slippage during roughly the next 12–18 months, while explicitly acknowledging “unknown unknowns.”

He notes that credit spreads on these deals remain wide, showing that debt markets continue to price meaningful risk.

The show promotes the fifth annual All-In Summit, scheduled in the transcript for September 13–15 in Los Angeles, again using Universal Studios.

Guests mentioned include:

• Jensen Huang, NVIDIA;
• Satya Nadella, Microsoft;
• Jared Isaacman, identified with NASA;
• Brad Gerstner;
• Bill Gurley;
• Gwynne Shotwell, SpaceX;
• Jake Paul;
• Nick Shirley;
• possibly Martin Shkreli.

The hosts also mention:

• casino night;
• a concert with the performer not yet announced;
• attendees from more than 60 countries;
• a focus on learning, meeting/networking and memorable experiences.

Brad notes that the timing would place the event within approximately 60 days of a midterm election and possibly within about 30 days of an Anthropic IPO, assuming the latter occurred as speculated.

Jason introduces Airtable as part of the broader “SaaSpocalypse” discussion.

According to the transcript:

• Airtable had roughly $480 million of annual revenue.
• Revenue was growing around 20% annually.
• The company had nearly $1 billion of cash.
• The headline sale price was approximately $1.28 billion.
• Including cash, the transaction was described as approximately $2.25 billion.
• Airtable's peak private valuation had been approximately $11.7 billion in 2021.
• The acquisition therefore represented roughly 10% of that peak headline valuation.
• Bending Spoons is described as Milan-based.
• Its portfolio was said to include AOL's legacy business, Evernote, Eventbrite [rendered “Eventbride”], Vimeo and Meetup.
• Bending Spoons itself was said to have gone public the previous month, with shares rising approximately 15% on the Airtable announcement.

These transaction and company-status details require verification.

Sacks says Airtable spun out an AI-agent business called Hyperagent into a separate independent company before the acquisition.

His interpretation is that the company's founders and strongest talent wanted to focus on the new AI opportunity rather than operate the mature Airtable product.

He describes:

• the AI company as the venture-style bet, where enormous new value might still be created;
• the mature Airtable business as a private-equity-style asset, where the opportunity comes from efficiency and cash generation.

Bending Spoons, in this interpretation, is buying the second type of opportunity.

Sacks cites a reported data point that only approximately 30% of Airtable's sales team was achieving quota.

He treats that figure as highly revealing.

His interpretation—not a reported fact—is:

1. Airtable originally succeeded through product-led growth (PLG).
2. Organic growth remained around 20%.
3. Investors who had financed the business near an $11 billion valuation needed a much larger venture outcome.
4. The board therefore pressured management to add a traditional enterprise-sales operation.
5. This motion was unnatural for the product.
6. Hundreds of salespeople were hired.
7. Only about 30% achieved quota.
8. The expensive sales organization therefore failed to materially accelerate growth.

Sacks describes the sales effort as trying to “push on a string.”

Sacks believes Bending Spoons can eliminate the costly sales-led structure and return Airtable to its original product-led model.

He speculates that an acquirer could:

• eliminate approximately 85–90% of the cost structure;
• retain much of the organic 20% growth;
• operate Airtable as a highly profitable mature product.

The speakers speculate about margins as high as 80–90% and annual EBITDA of approximately $300–$400 million.

At that level, the purchase price could theoretically be repaid in roughly two or three years.

These are analytical estimates from the speakers, not reported transaction guidance.

The speakers ask why Airtable's existing board and management could not have carried out the same restructuring themselves.

Sacks's answer is structural and psychological.

Founders:

• are loyal to employees;
• generally do not want to eliminate 80–90% of the organization they built;
• tend to prefer building the next growth business rather than operating a mature asset for cash.

Venture-capital board members similarly seek venture-scale upside rather than private-equity-style efficiency.

Sacks also says the cap table created poor incentives because later investors held liquidation preferences and had invested at very high valuations.

A major restructuring might generate a profitable company without creating enough equity value to produce the venture outcome expected by those investors.

He therefore characterizes the incentives as broken for the incumbent organization but potentially attractive for a new owner.

Gerstner broadens the Airtable lesson.

He says revenue multiples can collapse very quickly.

A high multiple can make sense while a company is growing above 50%, but price-to-revenue remains merely a rough Silicon Valley heuristic.

He says observers calling Airtable a roughly two-times-revenue sale may miss the cash-flow economics.

On a “look-through” basis, he estimates the deal may have been closer to 30× free cash flow, depending on assumptions.

Gerstner is more cautious than Sacks about the restructuring opportunity.

He argues that, if producing $400 million of EBITDA were simple, Airtable's board probably would already have done it.

When a growth company slows:

• employee morale deteriorates;
• customers begin leaving;
• product momentum weakens;
• each problem reinforces the others.

He says managing such a decline can become extremely difficult.

Gerstner speculates that Airtable's core product may have begun losing momentum as product advances slowed.

He suspects customer churn increased.

He also argues that non-AI software companies increasingly struggle to retain strong sales and product-development talent because many high-performing workers want to move into AI.

Sacks sees a particular advantage for Bending Spoons.

The acquirer already maintains a centralized operating team capable of managing many mature software properties.

Historically, buyers could not remove most of a software company's staff because too much institutional knowledge lived in the people who understood the codebase.

Sacks argues AI changes that.

An AI system can inspect and learn a codebase and reconstruct much of that historical knowledge.

Therefore:

• fewer original engineers may be needed;
• mature software may be maintainable with much smaller teams;
• the economics of acquiring and harvesting older software businesses may improve.

Sacks summarizes the idea as maintenance mode becoming much easier with AI.

Jason asks Friedberg whether today's huge AI valuations resemble the zero-interest-rate SaaS boom, when extremely high software valuations were widely accepted.

Friedberg says the two periods are structurally different.

During the SaaS boom, investors often reasoned that one dollar of recurring revenue could receive a 20× valuation and therefore rewarded revenue acquisition very aggressively.

Today's largest AI capital commitments, in his view, are different because they primarily fund:

• physical AI infrastructure;
• compute;
• model training.

He therefore does not view the AI CAPEX buildout as simply another version of the ZIRP SaaS valuation trade.

Sacks identifies two reasons SaaS companies became overvalued:

1. Artificially low interest rates created a broad speculative asset bubble.
2. Investors treated SaaS recurring revenue like a guaranteed, perpetually growing annuity.

An example was 120% net-dollar retention.

Investors effectively assumed that a company with that retention could automatically grow roughly 20% every year indefinitely, even before acquiring new customers.

AI disruption demonstrated that these revenues are not permanent annuities.

Products can be displaced, churn can rise and the presumed growth floor can disappear.

Sacks strongly cautions against using Airtable as evidence that every SaaS company is doomed.

He says Airtable was unusually vulnerable because the product was always difficult to categorize and explain.

At its peak, supporters described it as:

• a new Excel;
• a new Google Sheets;
• perhaps even a new Microsoft Office;
• effectively a “spreadsheet for words.”

But Airtable never achieved spreadsheet-like ubiquity.

It retained dedicated fans or a “cult following” without becoming universally understood.

Customers still often had to ask what Airtable was for and when they should use it.

Sacks says Airtable belonged broadly to the no-code category, which he believes is among the areas of SaaS most heavily disrupted by AI.

Earlier no-code products such as Airtable and Retool removed the need to write traditional source code, but users still had to learn a new tool.

In that sense, the products functioned somewhat like alternative programming languages.

AI systems change the interface again.

With products such as Claude Code, Lovable and Perplexity tools, the user can increasingly describe the desired result in ordinary language.

That can eliminate the learning curve attached to traditional no-code platforms.

Sacks therefore says no-code may be the most disrupted segment even though all SaaS faces some AI pressure.

Sacks does not believe enterprises will casually replace fundamental systems such as:

• CRM;
• ERP;
• HR systems;
• compliance-sensitive software

with software that has simply been “vibe coded.”

The reason is not necessarily superior underlying code. It is trust, integration, regulatory approval, accumulated data and operational dependence.

Jason gives the opposite example from his own firm.

His team reportedly built an internal portfolio-management system in approximately one month.

He says buying an equivalent system would have required:

• around $250,000 of off-the-shelf software;
• approximately $1 million of integration work;
• two to three years of implementation.

The internally built tool reportedly gives the team broad insight into its portfolio, competitors and founders.

His example supports the argument that bespoke internal tools are increasingly vulnerable to AI-enabled replacement even if core enterprise systems remain sticky.

Sacks references an investor called “Leopold” [identity should be verified] who had reportedly bet on a “SaaSpocalypse.”

According to Sacks, that investor was:

• heavily long semiconductor/chip stocks that corrected;
• short Adobe and other SaaS companies that moved against the short positions.

He uses the example to argue that treating all SaaS as doomed is financially dangerous.

A speaker quotes or paraphrases a post arguing that enterprises do not buy Microsoft because Microsoft writes the world's best code.

They buy Microsoft because Microsoft products have become the rails on which the organization operates.

Examples given:

• Active Directory holds employee identities.
• Excel contains important corporate and board numbers.
• Teams hosts compliance-recorded communications.
• Azure reportedly holds FedRAMP High authorization and Department of Defense Impact Level 5 clearance.

A defense contractor therefore cannot casually replace Azure with a cheaper improvised alternative.

The broader economic argument is that spending tens of millions of dollars to replace a trusted system costing roughly $1 million per year can make no sense even if a new AI-generated alternative looks cheaper on a software-license basis.

The discussion cites Marc Benioff as having just tweeted that 15 out of 15 cabinet agencies use Salesforce.

Sacks says the U.S. government is not going to replace Salesforce with something “vibe coded.”

He therefore concludes that SaaS businesses with:

• deep infrastructure;
• regulatory entrenchment;
• strong founders;
• passionate users

can transition successfully into AI-first products.

He mentions personally buying Figma and places Figma in that more resilient category.

Gerstner notes that the IGV software index was approximately:

• 20% higher over the previous six months;
• also roughly 20% higher over five years.

He says Snowflake was approximately 88–90% higher over six months.

He also cites Databricks and ClickHouse as software/data companies performing extremely well.

His conclusion is that software experienced a panic and a major rotation, but strong companies recovered or continued performing.

The weakest no-code and application-software companies may still conclude that their best option is to sell, but it is inaccurate to treat the entire software category as one trade.

Gerstner says his firm passed on Airtable's final three funding rounds, which he recalls at roughly:

• $2 billion;
• $5 billion;
• $11 billion valuations.

He says the late-stage investors in those rounds nevertheless appear to have received their original capital back, while early investors made substantial returns.

He therefore describes Airtable as a remarkably good outcome if it is to be classified as a Silicon Valley “failure.”

The original late-stage investment logic, he says, was that the company already had sufficient product quality, revenue and team strength that an investor might plausibly recover principal even if the high-upside case failed. He notes that similar reasoning could apply to some large AI investments today.

The speakers discuss whether Airtable's liquidation preferences protected later investors.

The exact transaction terms are uncertain in the transcript.

Sacks explains the standard 1× liquidation preference:

• preferred investors receive their invested capital back before common shareholders begin receiving proceeds;
• once the preference is satisfied, remaining upside is shared according to the security structure.

He distinguishes this from a participating preferred security, which allows investors both to take their preference and then participate again—a “double dip.”

Sacks says his firm does not use participating preferred structures and favors “clean terms.”

His rationale is that investors should not lose money while other shareholders make money from the same sale, but neither should the structure be unnecessarily punitive to founders.

Jason introduces a Forbes investigation described as “These American startups are making China's AI smarter.”

According to his description of the report:

• U.S. data/training companies sell high-value training material to American frontier laboratories.
• Some sell the same or similar material to major Chinese AI companies.
• This may help Chinese laboratories reduce the capability gap with American models.

Two companies are rendered in the transcript as “Sergei” and “MeriCorps” [company names unclear and must be checked].

Jason says both were valued at more than $20 billion.

Customers were said to include:

• OpenAI;
• Anthropic;
• federal agencies;
• Tencent;
• ByteDance;
• Alibaba;
• Moonshot;
• other major Chinese laboratories.

Jason says the six largest Chinese AI laboratories were collectively spending approximately $500 million per year on this type of data.

He characterizes the material as including:

• PhD-written content;
• reinforcement-learning material;
• specialist knowledge;
• other high-quality expert-generated training content.

He calls this material part of the “secret sauce” behind improving AI models.

The Forbes findings, companies, valuations and spending figures all require independent verification.

Jason says he has investments in several companies operating in this area, including one rendered as “Micro One” [transcript unclear].

He says the founder of that company chose not to sell training material to China.

He discloses this while arguing that training data has become increasingly important because easily accessible public web data is being exhausted.

He also references a previous discussion about books having their spines removed so they could be scanned for AI training.

Sacks cautions against immediately prohibiting such sales.

He asks whether the goal is:

• to begin a full economic war with China;
• to prevent American companies doing business there generally;
• or specifically to restrict strategically sensitive technology.

He says U.S. policy has historically focused export restrictions on dual-use technology, particularly technology with possible military applications.

He initially views ordinary data and data labeling as comparatively commoditized.

Sacks argues China has enough skilled labor to perform its own data labeling and enough scientists to generate sophisticated training material.

Therefore, restricting U.S. companies may create friction without necessarily denying Chinese labs access to equivalent data.

He says the relevant questions should be:

• Is the data genuinely proprietary?
• Is it strategically irreplaceable?
• Does it have dual-use value?
• Does it have a military application?

If the answer is no, he questions whether an export ban would materially change the AI race.

Sacks warns that banning American companies from selling services or data to China could lead China to impose reciprocal restrictions, potentially involving resources such as rare earths.

He says he wants the United States to be as independent and sovereign as possible, with minimal dependencies, but acknowledges that the two countries remain economically interdependent.

Restrictions therefore have costs as well as possible strategic benefits.

Sacks is not opposed to technology restrictions in principle.

He cites restrictions from the first Trump administration on exporting EUV lithography machines to China, which he places around 2019, as an example of a strategically important control.

His distinction is:

• controls that materially deny a difficult-to-replace strategic capability can make sense;
• controls that merely inconvenience China while provoking retaliation may not.

He therefore says this particular data issue must first meet a meaningful strategic threshold.

Gerstner broadly agrees with maintaining maximum competition while the United States remains ahead.

He says:

• U.S. frontier laboratories are winning;
• U.S. open-source efforts are also strong;
• regulation remains comparatively limited;
• bilateral U.S.-China relations were improving on several fronts.

The transcript then says “She's coming here in September in a bilateral meeting to meet with the president” [transcript unclear: person referenced is not reliably identifiable from the text].

Gerstner predicts policymakers will be irritated by evidence that U.S. companies may be helping Chinese laboratories catch up.

He links the issue with other concerns including:

• model distillation;
• exported chips;
• other transfers that reduce the difficulty of Chinese AI development.

Even if none individually justifies an immediate policy reversal, they create cumulative political pressure.

Gerstner says the reason he does not expect an immediate major policy change is simple: the United States is still winning.

However, if the president asks advisers six months later whether America is still ahead and receives the answer that China has caught up or passed it, all of these activities will receive far more scrutiny.

His argument is therefore conditional: current tolerance depends partly on continued U.S. technological leadership.

Jason takes a harder line.

He says he has personally used Chinese models rendered as “Kimmy,” “Quinn,” and “GLM 5-2” [transcript unclear] during the previous 60 days and found them extremely capable.

He says he does not consider it patriotic for American companies to provide Chinese developers an avoidable advantage.

He then explains why he regards these datasets as different from ordinary commodity labeling.

According to Jason:

1. AI users submit highly technical queries.
2. Some responses receive negative feedback because the answers contain mistakes.
3. Those difficult examples can involve code, biology, science and other specialist domains.
4. Data companies hire PhDs and other experts.
5. Those experts supply current specialist knowledge and corrected answers.
6. The answers are verified and sometimes double-verified.
7. The corrected expert material is then used to improve models.

Jason argues that this process effectively collects high-quality knowledge from elite Western experts and packages it for American frontier models.

If the identical packages are then sold to Chinese companies, the Chinese models can improve far more quickly than if those companies had to assemble the expert network and correction pipeline themselves.

He compares this acceleration effect with model distillation.

He therefore believes expert datasets may be an important reason Chinese models are catching up.

Sacks challenges the assumption that China cannot reproduce these datasets.

He argues that China graduates more mathematics and science students each year than the rest of the world combined [source claim requiring verification] and therefore does not lack intelligent technical people.

Jason responds that reproducing the datasets at the same level and scale would require hiring many of the best scientists and experts in the West.

Sacks says that if there truly is proprietary “secret sauce,” he would not want the United States to sell it to China.

However, he remains unconvinced that the material described automatically meets that standard.

His final test remains:

• Is it genuinely proprietary and difficult to reproduce?
• Is it dual use?
• Does it have military importance?

Without those elements, he doubts a restriction would seriously disadvantage China.

The speakers briefly mention the separate problem of technically skilled foreign graduates being trained in the United States and then forced to leave, but Sacks says that is a different policy issue and should not be conflated with the data-export discussion.

Exact quote:

“CAPEX is high alpha, low beta in data center infrastructure, that capital. And model development, theoretically could be high alpha, but it's very high beta.”

Context: Friedberg's explanation for why Google might direct incremental capital toward AI infrastructure instead of frontier-model development.

Topic: AI capital allocation; Google strategy.

Exact quote:

“My reaction was, and then there were two.”

Context: Sacks reacting to Google's reported AI personnel changes and arguing the serious frontier-model race had narrowed to Anthropic and OpenAI.

Topic: Frontier AI market structure.

Exact quote:

“I think that the market for frontier intelligence has become a duopoly.”

Context: Sacks's view that Anthropic and OpenAI had separated from the rest of the market.

Topic: AI competition.

Exact quote:

“If you're not at the frontier, you can't charge for the model layer itself.”

Context: Sacks arguing that lagging/open models can monetize compute and services but not intelligence itself.

Topic: AI economics.

Exact quote:

“A lot of people are just like, give me the best. I'm willing to pay a premium for the best.”

Context: Sacks explaining why businesses may continue paying for frontier models despite cheaper alternatives.

Topic: Frontier-model pricing.

Exact quote:

“I think the idea that there's kind of a model that you pick for everything, I think is the false assumption.”

Context: Friedberg arguing enterprises will use different models for different workloads.

Topic: Enterprise AI architecture.

Exact quote:

“Token consumption is going up for the open source guys, while share of economics is going up for the frontier labs.”

Context: Gerstner's synthesis of the Android/iPhone-like AI market.

Topic: Open-source versus frontier AI.

Exact quote:

“One platform makes the profit. One gets the majority of users, at least globally, and usage.”

Context: Gerstner comparing AI competition with Android and iPhone.

Topic: AI monetization.

Exact quote:

“Now it's all about execution.”

Context: After arguing SpaceX's post-IPO consolidation was normal.

Topic: SpaceX investment thesis.

Exact quote:

“I think Starlink's just an unbelievable juggernaut cash machine.”

Context: Friedberg introducing his Starlink valuation and cash-flow argument.

Topic: Starlink economics.

Exact quote:

“Elon refuses just to take the safe bet.”

Context: Gerstner praising Musk for reinvesting Starlink cash into risky infrastructure and technology projects.

Topic: Capital allocation; innovation.

Exact quote:

“I don't know there's an upper bound for on-demand intelligence.”

Context: Discussion of whether open-source models and local hardware could eventually reduce demand for centralized AI compute.

Topic: AI compute demand.

Exact quote:

“I would want to own this company, but I'm not sure today's the day I would buy the company.”

Context: Gerstner recalling his CNBC comment on the day of SpaceX's reported IPO.

Topic: SpaceX valuation; entry price.

Exact quote:

“Entry price matters.”

Context: Gerstner's warning that an exceptional company can still be overpriced.

Topic: Investing discipline.

Exact quote:

“Creating data centers is not rocket science.”

Context: Argument that SpaceX's rocket-building expertise gives it a construction advantage in AI data centers.

Topic: AI infrastructure.

Exact quote:

“You're effectively backstopping people's ability to build ahead of their revenue.”

Context: Gerstner explaining why vendor/seller financing can magnify a downturn if AI demand weakens.

Topic: AI infrastructure financing.

Exact quote:

“They've got hundreds and hundreds of sales reps here trying to push on a string.”

Context: Sacks arguing Airtable's enterprise sales effort was poorly matched to its product-led roots.

Topic: Airtable; SaaS go-to-market.

Exact quote:

“I think maintenance mode becomes way easier with AI because you don't need the historical knowledge anymore.”

Context: Sacks explaining why an acquirer such as Bending Spoons might operate mature software with fewer people.

Topic: AI and software operations.

Exact quote:

“It was basically a spreadsheet for words.”

Context: Sacks recalling the way Airtable was described during its peak hype.

Topic: Airtable product positioning.

Exact quote:

“No code has to be the most impacted, the most disrupted area of SaaS right now.”

Context: Sacks arguing conversational AI removes even the learning curve associated with traditional no-code applications.

Topic: AI disruption of SaaS.

Exact quote:

“I don't think this is going to give us a decisive advantage in the AI race.”

Context: Sacks questioning whether restricting expert-generated training data would meaningfully slow Chinese AI.

Topic: U.S.-China AI policy.

Exact quote:

“I think targeted strategic controls make sense. I would just make sure that this one actually meets that bar.”

Context: Sacks distinguishing important technology controls from restrictions that primarily create friction.

Topic: AI export controls.

Exact quote:

“The reason I don't think it will cause us to change our stance with respect to China is because we're winning.”

Context: Gerstner arguing that expert-data sales receive less scrutiny while U.S. AI maintains a lead.

Topic: U.S.-China AI competition.

Exact quote:

“I don't think it's very patriotic to be giving them an advantage.”

Context: Jason arguing against American companies supplying high-quality expert datasets to Chinese AI laboratories.

Topic: AI training data; China.

Exact quote:

“If there's something truly proprietary here, I don't want us to sell our secret sauce to China.”

Context: Sacks acknowledging that his position would change if the datasets contain a genuinely hard-to-reproduce strategic capability.

Topic: AI export controls; proprietary data.