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

Google committed to deploy 200 billion dollars in CAPEX this year on AI infrastructure, with accelerated depreciation rules providing an effective 26 percent discount at a 26 percent corporate tax rate. Freeberg drew a distinction between data center infrastructure capital, which he called high alpha and low beta, and model development capital, which he called high alpha but very high beta, arguing the former is the more reliable bet. Google Cloud posted 82 percent year over year revenue growth, described as unprecedented in cloud provider history, and Gemini reached over 950 million monthly active users in Q2, tripling year over year. Google now has 13 products with over 1 billion users each, with Gemini newly joining that list.

Despite those metrics, Google faces a serious internal talent problem. Jeff Dean, employee number 30 who spent 27 years at the company, is leaving with three other AI researchers to start Discovery Loop, focused on deep scientific breakthroughs. Google shares fell roughly 4 percent on the news, implying about 200 billion dollars in lost market cap. Gemini's co-lead and several other top researchers have also departed for competing labs, and Axios cited company sources saying Gemini 3.5 Pro is months behind schedule partly due to low morale. Saks argued these departures have effectively reduced the frontier model market to a duopoly of Anthropic and OpenAI, with everyone else six to twelve months behind and unable to charge for the model layer itself. Anthropic's ARR was cited at 10 billion dollars at the start of the year and is now over 80 billion, well ahead of the company's own 100 billion exit ARR target. Freeberg countered that Google does not necessarily need the best model given its enterprise install base, its ownership stakes in both Anthropic and SpaceX, and GCP's ability to host open weights and proprietary models alike. Gerstner noted a structural tension: Google Cloud wants compute to rent to Anthropic while internal model teams want that same compute to compete with Anthropic, a conflict Anthropic and OpenAI do not face.

SpaceX reported Q2 revenue of 7.8 billion dollars, up 92 percent year over year. Elon Web Services revenue more than tripled quarter over quarter to 2.6 billion dollars, driven by renting Colossus compute to Anthropic and Google. Starlink generated 4.3 billion dollars in revenue and 2.6 billion dollars in adjusted EBITDA, with 12 million subscribers doubled year over year, an ARPU of 66 dollars per month, and 20 percent quarter over quarter subscriber growth. SpaceX CapEx was 18.4 billion dollars in the quarter, up six times year over year, implying an annual run rate of roughly 75 billion dollars. Musk guided to 100 billion dollars in ARR by year end and pulled forward the 1 trillion dollar ARR target from 2031 to 2030, a figure Morgan Stanley's 325 billion dollar 2030 estimate falls well short of. SpaceX shares rose 13 percent after the report but remain down roughly 30 percent since the June IPO, settling at a 1.4 trillion dollar valuation. Speakers attributed Musk's edge to a core competency in standing up physical infrastructure faster than competitors, with Jensen Huang specifically stating no one at Microsoft or Google comes close to Musk in that capability. On the Starlink capacity trajectory, a Starship launch deploying 60 V3 satellites adds 60 terabits per second of bandwidth, more than 20 times the capacity per launch of a Falcon 9 carrying V2 satellites, and speakers predicted Starlink could eventually handle roughly half of all internet traffic.

Airtable was acquired by Milan-based Bending Spoons for 1.28 billion dollars, approximately 10 percent of its 2021 peak valuation of 11.7 billion dollars. Including nearly one billion dollars in cash, the effective transaction value was roughly 2.25 billion dollars. Airtable had approximately 480 million dollars in annual revenue growing at 20 percent. Saks argued that only 30 percent of the sales team was making quota and that the board had pressured founders to layer a traditional sales-led motion onto a product-led growth business, which failed. He predicted Bending Spoons could eliminate 80 to 90 percent of costs, retain most of the revenue growth, and generate EBITDA margins of 80 to 90 percent, potentially paying back the acquisition in roughly three years. Gerstner pushed back, arguing that if those cuts were straightforward the existing board would have already made them. Airtable was characterized as part of the no-code category, described as the most disrupted segment of SaaS right now because AI tools perform the same function without requiring users to learn a new paradigm. Speakers cautioned against extrapolating this outcome to core SaaS systems like CRM, ERP, and HR platforms, where compliance requirements make replacement by AI-generated alternatives unlikely.

Forbes reported that US data labeling startups Surge AI and Mercor, both valued above 20 billion dollars, sell training data to OpenAI and Anthropic while also selling to top Chinese AI labs including Tencent, ByteDance, Alibaba, and Moonshot. The top six Chinese AI labs are reportedly spending 500 million dollars per year on PhD-written content and reinforcement learning pipelines from these vendors. Saks argued data labeling is largely a commodity, that China has no shortage of domestic labor, and that banning sales could prompt retaliation such as rare earth export restrictions.