Interview
Google’s AI Brain Drain, SpaceX's Huge Quarter, Airtable’s 90% Collapse, US Data Fuels China AI
- Google plans a $200 billion capital expenditure this year to build AI infrastructure data centers, leveraging accelerated depreciation tax advantages, while shifting capital allocation from model development to computing infrastructure to improve return on invested capital and resolve channel conflicts with tenant competitors.
- The frontier intelligence market is projected to evolve into a duopoly between Anthropic and OpenAI, with Anthropic expected to finish the year with $110 billion to $120 billion in ARR and OpenAI experiencing accelerated growth, while lagging models face commoditization.
- Google is projected to become the leading AI company by consumer usage by year-end due to AI integrations in Android, Search, Gmail, Chrome, and YouTube, where Gemini's monthly active users are tripling year over year, and consumers are expected to pay a $20 to $40 monthly premium for full-stack subscription services.
- A blended model approach will dominate enterprise adoption, utilizing cheap open-weight models for simple workflows and premium frontier models for high-stakes applications, despite open source token consumption rising and economic share shifting away from frontier labs due to downward pricing pressure.
- SpaceX is projected to reach $100 billion in ARR by year-end, up from $7.8 billion Q2 revenue, driven by AI revenue growing 67% quarter over quarter and web services tripling to $2.6 billion, with a pulled-forward target of $1 trillion ARR by 2030.
- SpaceX's compute capacity is expected to expand from 1.4 gigawatts to 2 gigawatts by year-end at spot prices of $30 to $50 per watt, with further growth targeted to 5 to 10 gigawatts next year, though securing memory, land, and power shells for 6 incremental gigawatts presents a significant risk.
- Starlink is forecast to double its subscriber base to a 24 million run rate within 18 months, potentially generating $30 billion in free cash flow and a $1 trillion market cap, while V3 Starship satellites could increase network capacity by 100 times to handle half of global internet traffic.
- The market faces a "wall of worry" regarding whether AI compute demand will sustain the current CapEx pace for 12 to 24 months, with risks that spot price drops could extend payback periods for infrastructure investments beyond one year.
- Late-stage venture capital is prioritizing capital preservation over 10x returns, illustrated by Bending Spoons acquiring Airtable for $480 million against $480 million in revenue, a fraction of the 2021 peak valuation of $11.7 billion, with a strategy to generate $400 million in EBITDA within three years by cutting 85% to 90% of costs.
- Non-frontier open source models are becoming negligible for the majority of work tasks, with expectations that no-code tools like Airtable and Retool will face disruption from AI agents, while large enterprise SaaS companies with compliance needs remain resilient against replacement.
- China is expected to continue narrowing the AI capability gap with the US through data acquisition and domestic efforts, suggesting that US data export bans may not provide a decisive advantage as Chinese models have already demonstrated significant quality.
- Investor focus in the SaaS sector is shifting from revenue growth multiples to free cash flow and profitability, signaling the end of the "SaaS Apocalypse" narrative as the market recognizes that regulated or infrastructure-heavy software is less vulnerable to AI disruption.
- The "channel conflict" between hyperscalers renting compute and competing with tenants is expected to be resolved by prioritizing infrastructure over internal model development, while investors view SpaceX's $1.4 trillion valuation as reasonable for a three to four-year horizon.
- The US AI race against China remains active, with the US currently leading but the pace of China's catch-up dependent on data and talent access, while the market is underpricing potential combined ARR of $10 to $20 billion by year-end from SpaceX's "Grok" and "Cursor" businesses.
- Open source model adoption is expected to continue growing, yet frontier models will retain a premium for specialized use cases with low margin of error, and compute spot prices of $30-$50 per watt are viewed as temporary due to current memory shortages and demand constraints.