OpenAI and Anthropic Race to IPO, SpaceX's S-1 Exposes Frontier Economics, Microsoft Build Puts Agents in Windows: Sunday Digest, June 21, 2026

Both OpenAI and Anthropic are now actively pursuing public listings. OpenAI is preparing a confidential filing with a target of a possible September listing at approximately $730 billion. Anthropic is closing its latest private round at a $965 billion valuation and has its own filing in progress.

The scale of these valuations demands context. At $730 billion, OpenAI would be more valuable than all but a handful of companies globally — on day one, before public market investors have had any chance to evaluate its financial performance. Anthropic’s $965 billion private valuation is even higher, implying that the market assigns a premium to Anthropic’s safety-focused approach, its commercial momentum (the ~$30 billion run rate), or both.

The simultaneous filings create an unavoidable comparison that will force both companies to articulate, in public financial documents, exactly how they make money, how much they spend, and what they believe their competitive advantages are. For an industry that has operated largely behind the veil of private funding, this transparency will be clarifying — and potentially uncomfortable.

SpaceX’s IPO filing contained a remarkable disclosure: xAI, Elon Musk’s AI company, lost $2.4 billion in Q1 alone on $7.7 billion in capital expenditure. Buried in the same filing is the detail that xAI is renting 300 megawatts of power capacity to Anthropic for approximately $1.25 billion per month through 2029.

These numbers reveal the true economics of frontier AI in a way that no press release or conference talk ever has. A $2.4 billion quarterly loss means xAI is burning roughly $10 billion per year — and that is with $7.7 billion in capital expenditure that presumably includes infrastructure that will generate future revenue. The Anthropic power rental arrangement suggests that Anthropic alone is spending $15 billion per year on infrastructure from a single provider. The capital requirements of frontier AI are not just large; they are on a scale that only a handful of organizations on earth can sustain.

Microsoft Build 2026 placed agents at the core of Windows through sandboxed Execution Containers — isolated environments where AI agents can operate with access to system resources while maintaining security boundaries. The architecture represents a thoughtful approach to the fundamental tension in agent deployment: agents need access to be useful, but access creates risk.

However, the Build announcements were counterweighted by reports that production agent deployments at several enterprises are being rolled back due to PII exposure and hallucination issues. The gap between the promise of sandboxed agent containers and the reality of agents leaking sensitive data in production is a reminder that the infrastructure for safe agent deployment is still being built, even as the pressure to deploy accelerates.

Anthropic shipped its dual release of Claude Fable 5 and Mythos 5, adding two new models to an already crowded June release calendar. Foxconn deployed “MoMClaw” factory agents in its manufacturing operations, representing one of the most concrete examples of AI agents operating in physical-world industrial settings. NAVER and NVIDIAannounced a sovereign-AI infrastructure buildout in South Korea. And KAIST presented its VOTP robotics work at ICML 2026, advancing the state of preference learning for robotic systems.