Anthropic Is Renting Its Future from Elon Musk

The AI industry's structural dependency on infrastructure it doesn't control is becoming impossible to ignore — and the bill is arriving all at once.

The $15 Billion Problem

SpaceX's IPO filing revealed that Anthropic is paying roughly $15 billion annually to access the Colossus data centers in Memphis — facilities controlled by Elon Musk, who is simultaneously running his own competing AI lab. That's not a vendor relationship; it's a structural vulnerability with a price tag attached. And it's not the only exposure: Anthropic is also reportedly in talks to license Microsoft's custom AI chips. The pattern isn't diversification — it's dependency in multiple directions. When training runs cost this much and you don't own your silicon, every competitor with a data center has leverage over you. Anthropic is building one of the most sophisticated AI systems in the world while renting the foundation from people who would prefer it didn't exist.

GitHub Is Losing the War It Was Supposed to Win

Cursor reportedly sees a clear opening as GitHub flounders, and if that's accurate, it's the most underreported story in enterprise software right now. GitHub Copilot was Microsoft's presumed moat in AI coding — the product closest to where developers actually work. But proximity to the workflow only matters if you're the best tool there. If Cursor is eroding that lead, Microsoft faces an uncomfortable reckoning: does owning the repo host still matter when the editor is where the real leverage lives? The assumption that platform incumbency translates automatically into AI tooling dominance is looking shakier by the month.

Waymo Gets Wet

Waymo has suspended service in both Atlanta and San Antonio after its vehicles repeatedly drove into flooded roads. This is worth sitting with: the cars couldn't reliably detect standing water. Flooding isn't an edge case in Atlanta; it's a seasonal certainty. Meanwhile, Wayve just announced a deal with Stellantis to ship its self-driving software in US vehicles by 2028 — a reminder that the space still runs on optimism even as operational reality bites. The gap between "navigates a sunny afternoon in San Francisco" and "handles a Tuesday thunderstorm in Atlanta" remains uncomfortably wide.

Systems Behaving Badly

Two quiet failure modes surfaced recently that deserve more scrutiny. Gemini randomly exposed its own system prompt to users — the confidential configuration instructions governing its behavior, leaked without any apparent trigger. Separately, scammers are exploiting a legitimate Microsoft internal email account to send phishing links that clear reputation-based filters because the sender address is genuinely Microsoft's. Neither is a catastrophic breach, but both illustrate the same underlying problem: complex systems with many interacting layers produce failure modes their builders didn't model. The confidence with which these platforms are deployed doesn't match the control actually exercised over them.

The Valuation Is the Product

Hark raised a $700 million Series A at a $6 billion valuation for an AI interface described publicly only as "universal" and "secretive." Brett Adcock, who previously founded Figure (robotics) and Archer (eVTOL aviation) — both still very much pre-revenue bets — is now running a third swing at a moonshot. A $6 billion valuation for something that can't yet be described is either visionary or a sign that AI-era FOMO has fully consumed the rational portion of VC decision-making. Possibly both.

Across all of it, the picture that emerges is an industry spending at a pace that assumes perpetual growth, building on infrastructure it doesn't own, deploying systems it can't fully predict. The math has to work eventually — the question is who's still holding the bag when it doesn't.