The Compute Layer Is No Longer For Sale
The Colossus Clause
Anthropic's $5B annual commitment to the SpaceX Colossus facility in Memphis, running on Nvidia B200 NVL72 clusters, converts what was a procurement relationship into something structurally closer to a joint venture. Nvidia's $40B in announced equity positions, across OpenAI, CoreWeave, and several infrastructure build-outs, arrives the same week and completes the same picture: the chip vendor is now a capital partner, not a supplier. Compute is being structured as co-investment. For PRC labs, the constraint is already closed. Baidu, Moonshot AI, and Zhipu are running training workloads on Huawei Ascend 910B clusters; the FLOP-per-dollar gap between an Ascend 910B pod and a B200 NVL72 rack is running at roughly 2.3x on standard transformer workloads per Q1 2026 MLCommons submissions. DeepSeek's R2 distillation work has narrowed that gap at the inference layer, but the training layer -- where the Colossus contract sits -- is not closed. The BIS October 2023 controls, extended in January 2026 to cover H20 derivatives, ensure that gap does not narrow on the import side before the next major training window opens in 2027.
Security Stack, Running Now
While the compute debate runs at the infrastructure layer, OpenAI's GPT-5.5-Cyber has already moved past it. OpenAI's disclosed pricing for the security-tier API, a $150,000 annual seat floor, places GPT-5.5-Cyber in the same procurement conversation as Palo Alto Networks Cortex XSIAM and CrowdStrike Falcon Complete, not as a research integration but as a billable line item that CISO offices are defending in FY2027 budgets. Google's disclosure this week that an AI-generated zero-day exploit targeting a production Linux kernel version was identified and neutralized before deployment marks the other side of the same shift: AI-generated attacks are now a production input into enterprise threat models. The timing of that disclosure against GPT-5.5-Cyber's enterprise rollout captures the deployment layer's current shape. Both offense and defense are AI-mediated, the security stack has priced that in, and the revenue is already flowing. The malpractice suit filed this month against the medical chatbot that presented as a licensed physician names the same structural moment from the consumer side: the deployment layer moved faster than the liability framework, and litigation is now the catch-up mechanism.
*The Musk-Altman trial placed OpenAI's internal governance correspondence from 2021-2023 into the public record this week, including Microsoft executive communications that characterized OpenAI's nonprofit control structure as effectively nominal. The restructuring timeline is now inside 12 months. The question that does not resolve on its own is whether the BIS controls review scheduled for Q3 2026 will classify compute co-investment agreements (the Anthropic-Colossus structure, the Nvidia equity positions) as transactions subject to review. The determination lands before October.*