Anthropic Just Reset AI Expectations
The AI Daily Brief covers a transformative week for Anthropic, including the hiring of OpenAI co-founder Andrej Karpathy, the company's first-ever profitable quarter at a $44B annualized run rate, a massive $45B compute deal with SpaceX's Colossus, and NVIDIA's record-breaking quarterly earnings. These developments, combined with OpenAI's imminent IPO filing, signal a potential acceleration toward recursive self-improvement in AI research.
Summary
The episode opens with OpenAI news: the company has engaged investment bankers and expects to file IPO paperwork confidentially as soon as Friday, targeting a September IPO. This potentially leapfrogs Anthropic, which had targeted October but is still assembling a final private round. The host notes that compute constraints have made public market access more critical than anticipated, and frames the race as a fascinating game-theory scenario among three companies each valued at $1-2 trillion.
On the policy front, the White House held briefings with major AI lab CEOs about a new executive order, reportedly involving a voluntary framework for model disclosure and testing. The labs are pushing for a 14-day pre-release review window versus the government's proposed 90 days. The NSA would handle model testing, and the framework would take roughly 60 days to formalize, meaning it won't affect releases until late summer.
The central story of the episode is Anthropic's landmark week. Renowned AI researcher and OpenAI co-founder Andrej Karpathy announced he was joining Anthropic's pre-training team, where he will lead efforts to use Claude to accelerate pre-training research itself — a form of recursive AI research. Observers noted a pattern of top OpenAI talent (Jan Leike, John Schulman, now Karpathy) migrating to Anthropic, and many interpreted the move as a signal that recursive self-improvement (RSI) may be approaching, where AI systems meaningfully accelerate their own development.
Anthropicʼs financials then delivered another shock: the company is forecasting $10.9B in Q2 revenue at a $44B annualized run rate, along with a $559M operating profit — the first profitable quarter ever for any foundation model lab. This comes years ahead of prior forecasts (Anthropic had projected profitability in 2029). Analysts noted that compute scarcity is actually constraining even greater revenue growth, with some estimating Anthropic could reach $100B+ annually with unconstrained compute access.
NVIDIA's Wednesday earnings reinforced the AI infrastructure boom: revenue hit $81.6B (beating $78.9B estimates), data center revenue grew 92% year-over-year, and Blackwell chips are now firing on all cylinders. Jensen Huang declared AI factories represent 'the largest infrastructure expansion in human history.' Despite exceptional results, NVIDIA stock fell 3% after hours, reflecting investor uncertainty about how to value a $5T+ company.
Finally, Anthropic's compute partnership with SpaceX deepened significantly. SpaceX's IPO filing revealed Anthropic agreed to pay $45B over three years (~$15B/year) for Colossus capacity, instantly making Anthropic SpaceX's largest revenue source and overtaking Starlink's $11B annual revenue. Elon Musk signaled SpaceX is actively pursuing additional AI-compute-as-a-service deals. The host and various analysts framed this as potentially devastating for OpenAI, as a profitable Anthropic can fund compute directly without diluting equity, while Musk cannot afford to walk away from such a transformative revenue contract.
Key Insights
- Karpathy's role at Anthropic is specifically to use Claude to accelerate pre-training research itself, representing a form of recursive AI development where the model helps improve its own successor — a dynamic the host and analysts argue signals RSI may be approaching.
- Anthropic's profitability is partly a byproduct of compute scarcity rather than pure demand fulfillment; analysts argue that an unconstrained Anthropic could generate $100B+ in annual revenue, meaning current financials dramatically understate the company's true revenue potential.
- The $45B Anthropic-SpaceX Colossus deal makes Anthropic SpaceX's single largest revenue source, surpassing Starlink, and creates a mutual dependency: Anthropic needs the compute, and SpaceX as a public company cannot afford the revenue volatility of losing the contract.
- The host argues that the current AI infrastructure buildout differs fundamentally from the 1990s telecom bubble because leading-edge wafers and power infrastructure face structural supply constraints that cannot be resolved in months, meaning demand will outpace supply for the foreseeable future rather than triggering an overbuild crash.
- NVIDIA disclosed for the first time that hyperscalers represent only 46% of data center revenue, and that the company has effectively ceded the Chinese market to Huawei — yet still posted 92% year-over-year data center growth, suggesting the non-China, non-hyperscaler AI compute market is enormous and still expanding rapidly.
Topics
Full transcript available for MurmurCast members
Sign Up to Access