Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI
In a Nutshell
Google is hemorrhaging top AI talent—Demis Hassabis, Jeff Dean, and others—to startups because its $200B capex push favors renting out compute over building frontier models, creating internal channel conflict. SpaceX posted $7.8B Q2 revenue, tripled AI compute rentals to $2.6B, and guided to $100B ARR by year-end while expanding to 2GW of data-center capacity. Meanwhile, Airtable collapsed from an $11.7B peak to a $1.28B sale after its growth stalled, and U.S. data-labeling firms are feeding proprietary training data to Chinese labs, accelerating their catch-up.
These notes were generated by AI and may contain inaccuracies.
On Wednesday, Demis Hassabis moved to chair of DeepMind and chief scientist at Google. Reports describe this as Demis stepping down or being kicked upstairs, though Google framed it as a promotion and says he was stepping up. Axios reported that Google's Gemini 3.5 Pro is months behind, with some company sources telling Axios that it's in part due to low morale. Several top researchers, including Gemini's co-lead, have left the firm for competing AI labs.
Jeff Dean, plus three other AI superstars, are leaving Google to start a company called Discovery Loop. Dean is a legend who worked with Freeberg at Google. He was employee number 30, joined in 1999, and has worked there continuously for 27 years. Discovery Loop will be focused on deep scientific breakthroughs in AI. Google shares are down 4% on the news of Dean leaving, representing $200 billion in lost market cap.
Google has made a commitment to deploy $200 billion in capex this year in AI infrastructure data center buildout. Because of the capex and accelerated depreciation, making an investment in AI compute in the US right now is hugely tax advantaged. Due to the extreme demand for compute, there is a clear ROIC model where companies making this investment have significant demand for that compute infrastructure, are good at running it, and the capital can deliver massive profit returns with high confidence.
Building the most advanced frontier lab driven model takes tens of billions of dollars of capital. The question is whether companies can deliver profits from the model, especially in a world where open-source is becoming so good and open weights models are catching up quickly, with all frontier labs catching up to each other rapidly.
Early interviews with Jeff Dean from a couple years ago revealed that Google had a ChatGPT equivalent internally a year before ChatGPT came out from OpenAI. Google chose not to release it for fear of cannibalizing search. When Sergey stepped in, there was a revitalization effort, but as time has gone on and everyone has competed on models, it has become clear that it is very hard to get the same sort of return on capital invested in model development as it is in capital invested on compute infrastructure and being model agnostic.
Google has one of the greatest install enterprise bases in the world for compute, with the most enterprise customers and the most consumers. In both cases, they don't necessarily need to have the best model to make an incredible business. They can be model agnostic, working with Anthropic, OpenAI, SpaceX, and have a significant ownership stake in SpaceX and Anthropic. They can work with all the open weights models and host them all.
When top computer scientists like Demis, Jeff Dean, and others are inside at Google and the company is allocating capital not to their models but to infrastructure and data centers supporting the broad ecosystem of models, they may consider that given they can raise a couple billion dollars at a multi-billion dollar pre-money with a PowerPoint deck because they're the greatest in the world at doing this, that might be a better path.
Capex is high alpha low beta in data center infrastructure, while capital in model development theoretically could be high alpha but is very high beta and a very risky way to deploy capital. If the board and management are deploying more capital in computing infrastructure and less capital into model development, that explains the pattern of departures.
The same dynamic is occurring at Microsoft, with Morgan Stanley's report showing over a 30% return on invested capital in tokens as a service in the infrastructure business. Scientists who want to be involved in super intelligence, who want to cure cancer, who want to be on the frontier of these models are dealing with channel conflict at Google because Google Cloud wants all of the compute in order to rent it out to Anthropic, while those building the frontier models internally want that compute in order to compete with Anthropic.
SpaceX has channel conflict as they rent out their compute to Anthropic at the same time they're trying to build their own model with Grok and Cursor. Google and Microsoft have this channel conflict, while Meta is talking about getting into the infrastructure as a service game. At Anthropic and OpenAI, there is no channel conflict because they are not in the infrastructure business and only in the model business.
Thanks to the law passed on capex depreciation, assuming a 26% corporate tax rate, every dollar deployed in capex provides 26% off because companies get to write it off in the year of deployment.
Polymarket shows companies that will have the number one AI model by the end of this year on December 31st, with OpenAI at 32%, Google at 20%, Alibaba at 14%, and Moonshot, XAI, Meta, ByteDance all at about 10%.
The market for frontier intelligence has become a duopoly between Anthropic and OpenAI. Elon is still on the hunt, and Google would say they're still on the hunt, but they may have contradictory incentives because they can do quite well just with their compute. This represents a very powerful duopoly that is not being commoditized.
The market is evolving into a two-tier structure with a market for frontier intelligence and a market for commodity or lagging intelligence that is 6 to 12 months behind. For the commodity models, companies cannot charge anything for the weights but can charge for the compute, inference, and consulting services. If companies are not at the frontier, they cannot charge for the model layer itself, but if they are at the frontier, they can charge a premium.
Anthropic is now over $80 billion of ARR, having started the year at $10 billion. They had forecast $100 billion as exit ARR for the year, and most people said that would be impossible to achieve. Now it looks like they're going to do it with a couple of months to spare, with estimates going up to $110-120 billion or higher for end of year ARR. OpenAI is seeing acceleration as well.
The frontier model duopoly can charge a premium similar to how Apple competes against Android. Android is open source and has more users globally, but all the monetization goes to Apple because people are willing to pay for the premium experience. In a similar way, people are willing to pay a premium for true frontier intelligence if it's really at the leading edge.
Google Cloud posted 82% year-over-year revenue growth. Google has five products with over three billion monthly users each: Android, Search, Gmail, Chrome, and YouTube. These are becoming AI-first products. Google has 13 products total with over a billion monthly users, and Gemini had over 950 million monthly active users in Q2, tripling year-over-year.
For 95% of jobs being done, non-frontier models are good enough. The difference between open source models and frontier models is negligible for many use cases. However, for sophisticated use cases, frontier models remain significantly ahead. Companies set up frontier models because it's too hard to implement open source right now, but it will get easier and easier to implement.
For businesses in highly competitive industries like hedge funds, they don't want to take the chance that they're not getting the best intelligence to power their models. There are many industries where competitive dynamics drive companies to pay for the best intelligence. When use cases are immature and companies are searching for opportunity to use AI, they want to use the best because the return on finding those use cases will be much greater than the small premium being paid at the token level.
There is a blend happening where companies use open-source open weights for a vast majority of simple workflow applications, but when it comes to specialized applications requiring high quality model proficiency like life sciences and genomics modeling, they go for the premium model. For media companies doing AI rendering of video, they use Gemini's model that does video, or Sora, or whatever the best model is for that particular application.
Consumers are not going to be using some openweight model because they can pay $20-40 a month and get ChatGPT or Gemini or Claude and be very happy. For enterprise, companies will be very active in selecting a blend of models, using very cheap openweight models for simple workflow applications and individual employees spinning up an app, then more complex models for really key workflow tasks and specialized models.
It is way too early to count Gemini out on building incredible specialized models. They have the best video data and the best life sciences data, having been working on this for far longer than Anthropic or OpenAI in the life sciences side. Demis is still going to be running Isomorphic Labs. For specialized models in video, life sciences, and protein folding, Gemini will shine.
There is downward pressure on pricing as OpenAI and Claude have done massive price cuts for tokens. They are reacting and not taking it sitting down. The orchestration between these models is being built into a lot of harnesses inside of enterprises.
There is massive competition with Chinese open source, domestic open source, and frontier national labs. There is healthy downward pressure on pricing. It is hard to call it a duopoly when only a few years into this and there are giants like Amazon, Microsoft, and Google. While Anthropic and OpenAI have emerged as the pure plays gaining share of wallet, there are non-consensus views that the Chinese have not caught up as much as people think, and that closed models are actually cheaper when accounting for training costs, expertise to fine-tune and maintain, guardrails, and safety considerations.
SpaceX had their first earnings report as a public company with shares dropping 13% on concerns about surging AI capex. The stock is down 30% since going public in June and is now trading at a $1.4 trillion valuation, having gone public above $2 trillion. Q2 results showed $7.8 billion in revenue, up 92% year-over-year and 67% quarter-over-quarter. AI revenue more than tripled quarter-over-quarter to $2.6 billion from renting out compute specifically to Anthropic and Google from the Colossus collection of servers. Capex was up to $18.4 billion in the quarter, representing 6x year-over-year growth for a run rate of about $75 billion.
SpaceX guided to $100 billion in ARR by the end of the year and pulled forward the $1 trillion target in ARR by a year from 2031 to 2030. Morgan Stanley's 2030 revenue estimate is $325 billion. The company did $18 billion in revenue last year. The market has questions about the rental business and whether companies will be able to find people with capital to offtake the compute at 10 to 20 gigawatt scale, and about the frontier model business where Grok tripled tokens in July, not including Cursor which was already on a path to go from $3 billion to $10 billion by the end of the year.
Cursor was already on a path to go from 3 billion to 10 billion by the end of the year. Cursor plus Grok could be at 10 to 20 billion by the end of the year. That would be an extraordinarily valuable asset going to trade at a much higher multiple than the data center business. Starlink is positioned to run the table on mobile.
Normal consolidation is occurring with funds across Silicon Valley distributing their shares. The stock has traded down, which is unsurprising. The two most important things to watch are how Grok and Cursor revenues end the year, and the traction they get on continuing to replace traditional mobile carriers with Starlink. Distribution started recently, creating downward pressure from people wanting to cash out after being in it for a long time.
The Starship test flight was successful with the spacecraft floating in the ocean and the heat shield working. This enables more flights of Starship at an accelerated rate, paving the way for the V3 satellite which enables much more bandwidth for the Starlink network and powers the direct-to-cell play. The telecom aspect appears very on track and the team is bullish about that direction.
SpaceX expects to go from 1.4 gigawatts of compute to about 2 gigawatts by the end of the year. The spot price for compute is in the $30 to $50 per watt range. A gigawatt is a billion watts, so 30 to 50 billion per gigawatt. They are at the high end of that range. When Elon says they will end the year at 100 billion of ARR, all you have to believe is that they are at 2 gigawatts of compute running for $50 a watt to hit that target. This does not include Starlink, the launch business, or the Grok Cursor piece.
There are multiple ways to win with the stock. Starlink is an unbelievable juggernaut cash machine. On the connectivity side, Starlink generated $2.6 billion in adjusted EBITDA in the quarter, which approximates operating cash flow. Space was negative 200 million, essentially break even, and AI was plus 1.1 billion.
There is a question mark about whether the pricing they are getting on compute rental today is temporary and at a premium because of the lack of compute available in the market. On the connectivity piece, Starlink generated 4.3 billion in the quarter and 2.6 billion in adjusted EBIT. They have 12 million subscribers, which doubled year-over-year. The average revenue per user is $66 per month, and they grew 20% quarter over quarter. Extrapolating this out, they are close to a 24 million subscriber run rate on this multiple.
Assuming enterprise stuff like airlines and other verticals scale as the consumer business has, Starlink alone could be generating on the order of $40 billion of revenue topline with a huge amount of that flowing to free cash flow. That could be $30 billion free cash flow within the year. That alone provides the cash flow to fund much of what Elon is doing. At a 30x multiple on that business, given these subscription businesses have very high renewal rates and very low capex, the Starlink business alone could be a trillion dollar market cap within 2 years.
Starlink has decimated traditional satellite providers. AT&T, Verizon, HughesNet, and Viasat have all been impacted. The market got decimated by Starlink. The company is adding 2 million subscribers on the consumer side a quarter, with potential to go to 4 to 5 million a quarter. There are 400 million mobile subscribers just in the United States.
Every single Tesla sold is going to have Starlink in it when they get that merger done. This means Wi-Fi networks connecting any phone to any Tesla, including all those robotaxis out there. Phones will be able to connect directly with the next generation of Starlink if they have clear line of sight, connecting to any Tesla on the road.
Few CEOs or founders would take the Starlink business, which is such an exceptional trillion dollar business going to two trillion, and take on these other risks. They would not do Terafab, would not try to build out the data center, would not try to build their own model. These are highly risky but highly important investments. It is heroic and important to have this level of unbridled enthusiasm for innovation on the frontier.
Other CEOs and management teams are starting to realize that buying back shares and giving dividends is not as important as betting on the future. DoorDash got taken to the woodshed because they are investing too much in capex. Google got smacked with their capex spend. The arguments that will turn out to be wrong but are valid to discuss are whether demand for tokens and compute will keep up, or if on-device and desktops and open source models getting smaller and better will mute demand at some point.
Starlink is for people in rural neighborhoods. If you have Verizon fiber to your building or Spectrum, you are not putting a Starlink on your building. The piece that will be explained in the next year or two is every single Tesla sold will have Starlink in it. This creates Wi-Fi networks connecting any phone to any Tesla, including robotaxis. Phones will be able to connect directly with the next generation of Starlink with clear line of sight.
There has been speculation about valuation. At liquidity, private companies and venture capital operate as a voting mechanism, and when it goes public it becomes a weighing mechanism. The hand ringing peaked in the last quarter with a 160 times price to sales ratio for Tesla. Taking their 2 or 3 trillion market cap and putting it against a smaller revenue number, now it is down to a 45 times price to sales ratio as revenue increases.
The SpaceX IPO was extraordinary. The consolidation is perfectly predictable. Now there is a company at 1.4 trillion that, taking a three or four year view, could triple in value at a very reasonable valuation. The bet has always been whether Elon is the greatest innovator and a great allocator of capital. Entry price matters. When you get carried away on day one of an IPO and buy over two trillion, you have to know that consolidation will happen.
Anthropic IPO is rumored later this year with some saying 1.5 or 2 trillion. The company is run rating over 100 billion maybe by the end of the year. That is 10 to 15 times revenue, which is not that much for a company that just grew 10x and is rumored to be profitable in Q2. The market consolidation in July saw some semi stocks down 30 percent.
Elon is exceptionally good at building factory specific physical sites, which is a core advantage in a world where everyone is competing for data centers and fabs. The software layer needs hardware in the physical world to deliver software services. No one is better than Elon at actually doing that. Looking at how gigafactories have been stood up around the world demonstrates this core competency.
Putting up data centers is nothing compared to the difficulty of putting up a rocket. Creating data centers is not rocket science. They take some of that hardware expertise from SpaceX and put it into data centers, which is why they have been able to stand up more data centers or bigger data centers faster than all competitors.
SpaceX has developed a new V3 satellite that has 10x the bandwidth of its V2 satellite. Currently the Starlink network is powered by V2 satellites deployed on Falcon 9 rockets, launching about 27 satellites per launch and adding about 2.6 terabits per second of total network capacity. Starship deploys 60 of these V3 satellites per launch, adding 60 terabits per second of total network capacity per launch, which is over 20 times more capacity per launch.
The last Starship test proved the heat shield worked and they deployed 20 V3 satellites as a test, making connection with those satellites. The reason the Starship was visible was because they put HD cameras on them. Those satellites burned up as it was just a test. The next big milestone will be when they launch Starship with 60 of these V3 satellites, put them in the correct orbit, make connection, and add the bandwidth to the network.
Playing this out to its logical conclusion, the bandwidth available to the Starlink network goes up 10x or eventually 100x times, enabling direct-to-cellular capabilities. The Starlink network could eventually handle roughly half of internet traffic. This thing could get so much bigger than just 12 million subscribers.
Elon mentioned they will be at 2 gigawatts by the end of the year and 5 to 10 gigawatts next year, closer to 10 than 5, so approximately 8 gigawatts. That is an add of 6 gigawatts from 2 to 8. The questions are how to know the spot price will stay where it is at $50 per watt, and how to finance the expansion if going from 2 to 8 gigawatts.
The market is memory constrained right now. There might be a 20 percent increase in memory production next year but demand is going up 200 percent plus. The market is constrained by whatever the bottleneck is at that time. Right now the bottleneck is memory. A gigawatt power data center is 50 billion of capex, so 6 incremental gigawatts of compute would be 300 billion of capex next year.
To finance 300 billion, you either have to go into the market and borrow the money or do a dilutive equity raise, neither of which they want to do. Or get Nvidia to backstop it, which they have indicated they will do more of. The problem is Nvidia shareholders do not want them backstopping unlimited because the fear is that spot price at some point may go against you and the payback period changes.
Nobody thinks the payback period is going to be one year even though the spot price is suggesting that today. A few months ago, people thought you would get payback over four years. You spend 50 billion, earn 10 to 15 per year, get payback over four to five years, and then hopefully get the sixth year, which takes returns well above 20 percent.
The shortage is so acute and the willingness to pay from frontier labs is so high because they recognize they are on the verge of massive breakthroughs that they are willing to pay three, four, 5x market pricing to get at-scale compute. That is what happened with the Anthropic deal with SpaceX. Anthropic would buy a lot more of that today if they could, same with OpenAI.
The vast majority of offtake commitments are coming from Anthropic, OpenAI, and Nvidia. When you hear about hyperscalers building compute out, they are building it out to sell to the people just mentioned. It was not the Chinese open source companies that were buying SpaceX's excess compute or building the 10 gigawatt plant in Ohio.
If Elon builds 6 gigawatts next year, only Elon can actually stand up that much in that timeframe. Jensen said nobody comes close, Microsoft does not come close, Google does not come close in terms of standing it up in that timeframe. The offtake is there, but the question is whether he can get the memory, chips, land, power, and shell all in time.
Anthropic and OpenAI combined had a starting total compute of about 5 gigawatts this year. Elon is talking about incrementally adding more than they had as combined companies. That is not that much of an increase when Anthropic is growing 10x year-over-year and OpenAI is maybe at 4x or higher now. The demand exists in the world today and will exist for the next 12 to 24 months.
There is a wall of worry in the market. The reason for the pullback in July is that people got scared into thinking frontier labs revenues would be undercut, and if they undercut frontier labs revenues, who is going to pay for all this compute. That is why there was a 40 percent trade down in companies like CoreWeave and all the semiconductor stocks and semiconductor related AI stocks.
The fact that this next 10 gigawatts is going to cost 500 billion dollars and questions about where that comes from, whether from secondary offering, Nvidia putting it on their books, or creating SPVs, means the market has been educated on it. People will be able to change in real time if it does not come to pass or if it slows down, which cannot keep up at this pace for more than another two years.
Bill Gurley reminds that it is hard to believe everyone is just taking in stride this level of seller financing, which he would call circular revenues. The market has gotten comfortable with this. If there is a scare about demand, the whole sector trades down together because that is the leverage being pumped into the system, effectively backstopping people's ability to build ahead of their revenue.
The fifth annual All-In Summit is happening in September at Universal Studios. Confirmed speakers include Jensen Huang, founder and CEO of Nvidia, Satya Nadella, CEO of Microsoft, Jared Isaacman from NASA, Brad Gerstner, Bill Gurley, SpaceX's Gwen Shotwell, Jake Paul, and Nick Shirley. Applications are open at allinsummit.com or theallinsummit.com.
Airtable was acquired for $1.28 billion by Bending Spoons, an Italian Milan-based company, representing about 10% of its peak valuation of $11.7 billion in 2021. The company had $480 million in annual revenue, growing 20% a year, and nearly a billion dollars in cash. Bending Spoons went public last month and saw shares drop 15% on the Airtable news.
Airtable spun out its AI agent business called Hyper Agent into a separate independent company prior to the acquisition. The founders and talent chose to focus on the AI venture play while selling the legacy product as a private equity play to Bending Spoons.
Only 30% of Airtable's sales team was making quota, indicating the company had a successful product-led growth motion generating 20% annual growth, but the board pressured founders to add a traditional sales-led motion to accelerate growth. This resulted in hundreds of sales reps achieving only 30% attainment.
Bending Spoons can potentially eliminate 85-90% of the cost structure, return to product-led growth roots, and generate 300-400 million in EBITDA annually while growing 10-20%. At 80% EBITDA margin, the acquisition could pay for itself in a couple of years.
Multiples of revenue can compress very quickly. During the zero interest rate period, companies growing greater than 50% received high valuations, but this heuristic doesn't work as well when growth slows. Airtable's core product appears to have started fizzling as product advances slowed, with elevated customer churn.
The structural problem is that it's difficult for VCs on boards and founders to shift into private equity mode because they have loyalty to teams and don't want to eliminate 80-90% of cost structure. The cap table structure is problematic with liquidation preferences from investors who bought at the $11 billion valuation.
The AI capex buildout and model training differs fundamentally from SAS investments. During ZERP, capital flowed into SAS companies seeking 20x multiples on revenue, treating them like guaranteed annuities with 120% net dollar retention. AI disruption means these companies are no longer annuities, and elevated churn is occurring.
Airtable was always a quirky product positioned as a new Excel or Google Sheets - a spreadsheet for words rather than numbers. It never achieved the promised ubiquity and had a cult following but never reached broad acceptance. Most people still don't know what Airtable is.
Claude Code, Perplexity, and similar AI tools are now doing what Airtable did. The advantage of AI is that users can simply tell Claude what they want to create without learning alternative programming languages or interfaces. No-code tools required learning specific systems, whereas AI eliminates this learning curve.
Not all SAS is equal in this disruption. Microsoft is purchased not because it writes the best code, but because it's the infrastructure everything runs on - Active Directory for employee identities, Excel for board numbers, Teams for compliance, Azure for FedRAMP authorizations. Fifteen out of 15 cabinet agencies run on Salesforce.
IGV, the high growth software stock index ETF, is up 20% in the last six months and 20% over five years. Snowflake is up 88% in the last six months. Late-stage investors in Airtable at $2 billion, $5 billion, and $11 billion valuations got their money back, while early-stage investors made substantial returns.
Forbes published an investigation revealing that US data labeling startups Scale AI and Mercer are selling training data to Chinese labs including Tencent, Baidu, Alibaba, and Moonshot. The top six AI labs in China are spending $500 million annually on data including PhD-written content, reinforcement learning, and knowledge pipelines.
The founder of Mercer made the decision not to sell to China. Data labeling is largely a commodity, and China has abundant labor to perform this work. The concern is whether this represents technology transfer with dual-use military applications, though data labeling itself doesn't have obvious military applications.
The Trump administration previously limited export of EUV lithography machines to China in 2019, representing a targeted strategic control. The question is whether data sales meet the threshold for similar restrictions, given potential reciprocal actions from China such as rare earth export limitations.
The US is currently winning the AI race with frontier labs leading and open source winning, with limited regulations. However, if US leadership determines they're no longer winning against China, data sales and distillation issues will receive greater scrutiny.
The concern is that American companies are sending proprietary data sets to Chinese companies like Kimmy, Quen, and GLM 52. These data sets are created by American PhDs who receive queries with errors and provide thumbs down feedback on highly technical content including code, biology, and science. This process involves experts putting in the latest content and double verifying it, which improves LLM results.
The argument is that this data leakage helps Chinese models catch up to American ones. The same data packages created for Western LLMs are being resold to Chinese companies, allowing them to improve at the same rate.
China graduates more math and science graduates every year than the rest of the world combined. The concern isn't whether China has smart people, but rather that Western expertise is being packaged and sold to them.
The discussion notes that America is graduating top talent and then kicking them out of the country, which compounds the issue.
The process of data being leaked to Chinese companies is described as very similar to distillation. The packages of verified, expert-curated data are being sent overseas, allowing Chinese models to improve rapidly.
The principle is that truly proprietary data with dual-use or military applications should not be sold to China. However, there is debate about whether the specific data sets in question rise to that level of sensitivity.
The episode concludes with closing remarks and promotion of the All-In podcast store, including 1,000 signature Chamath autographed white sweaters with proceeds going to charity.
Keep All-In Podcast in your library
Save the videos and channels worth coming back to, and find them again in one place.
