SpaceX just became an AI landlord

Follow the power cables. SpaceX’s $6.3B Reflection deal puts a third frontier lab on Nvidia GB300 chips at the Colossus 2 site near Memphis, with Reflection AI paying $150 million a month from July 1, 2026 through 2029. That sits below SpaceX’s Anthropic deal at $1.25B a month and its Google deal at $920M a month. Three labs are now renting from the same landlord. The compute layer is consolidating faster than the model layer.

Reflection builds open-weight models, and openness is becoming a security pitch too. OpenAI’s Patch the Planet ships the full GPT-5.5-Cyber, scoring 85.6% on the CyberGym benchmark versus 81.8% for standard GPT-5.5, with more than 30 open-source projects signed on, including cURL, Python, and Sigstore. Built with Trail of Bits across 19 projects, it surfaced a 23-year-old use-after-free flaw in OpenBSD’s kernel and five exploitable bugs in Chrome’s V8 engine. The defense race now runs through public code.

The talent layer is shifting the same way. Google’s twin AI defections send Noam Shazeer, who co-led Gemini and co-authored the Transformer paper, to OpenAI, while John Jumper, who led AlphaFold and shared the 2024 Nobel Prize in Chemistry, joins Anthropic. Alphabet shares fell about 5% to 6% on June 22, with reports tying the drop to AI talent-retention concerns. Two of the field’s most cited researchers left within days. The market is now pricing people, not just product.

Latest from PulseMark

GLM-5.2 API Tutorial: Build With Z.ai’s 1M-Token Model A working build guide for Z.ai’s open-weight GLM-5.2 — OpenAI-compatible calls, tool use, MCP, and a 1M-token window at $1.40 per million input tokens. Read the tutorial →