Brookfield and Radiant target more than 1GW at Drax Power Station
Drax said on 3 November it had signed an agreement to develop a strategic partnership with Brookfield and Radiant, Brookfield's UK-headquartered AI infrastructure platform, targeting a data centre of more than 1GW on the Drax Power Station estate in North Yorkshire. Capital comes through a Brookfield joint venture with Qatar's Qai announced in 2025. Phase one targets up to 100MW operating from the first quarter of 2028. Brookfield estimates total potential investment, including semiconductor chips, could exceed £20 billion, with an initial £2.5 billion as part of phase one. Sikander Rashid, Brookfield's head of AI infrastructure, said Brookfield's "capital and deep operating capabilities will help accelerate the construction of a critical piece of sovereign AI".
Phase one would import power from the grid, and its progression depends on planning, regulatory and network approvals, a tenant, commercial agreements and a final investment decision. Phase two is an ambition for up to an additional 500MW between 2028 and 2031, with further capacity after 2031. From phase two the power would come behind the meter from Drax's own generation under a long-term PPA, using a mechanism in the Contract for Difference Drax signed with the government in November 2025, which runs from April 2027 to March 2031 and needs the government and the Low Carbon Contracts Company to agree. A buyer quoted UK capacity for 2028 should ask which phase it is in.
Revolut says self-hosted open weights cut its AI coding inference costs by about 90%
Revolut says it has made self-hosted open-weight models the default for AI coding and cut inference costs by roughly 90% in under a year. David Abitbol, its ops and infrastructure leader, set out the migration at the AI Engineer conference on 13 October, reported by The Register on 3 November. Gemini, deployed in early 2025, helped; the Claude Code rollout in early 2026 returned less, with engineers spending $1,000 to $10,000 a month for little return, so Revolut launched it with strict budget limits. It moved to Z.ai's GLM-5.1, then GLM-5.3, now served on 56 Nvidia B200 GPUs.
Abitbol put self-hosted GLM at a blended $0.30 per million tokens, cutting its bill to about a tenth of what it had been paying Anthropic and OpenAI. He credited tuning, NVFP4 quantisation, disaggregated prefill and decode and shorter context, not more GPUs. At around 256,000 tokens of context he reported sub-three-second time to first token and about 200 tokens a second. The rollout reached roughly 2,000 engineers, with a reported 60% improvement in coding time and 70% in lead time. The coding model is text-only, so vision queries go elsewhere. Abitbol said the concern with Chinese open weights is APIs running out of China; run on-prem or on regional cloud, The Register reported, there is nothing to worry about.
Quick Hits
- Menlo Ventures estimates enterprise AI spending reached $176.8 billion by September, up from $33 billion in 2025. Open-source models handle about one in five tokens but only about one in 25 dollars of model spending (release).
- Ai2 published StateBench on 3 November, which tests how well models track entity states as they change across long documents. The frontier models tested dropped to 13% or lower on the longest of its hardest problems (Ai2).
- Nvidia has told cloud and data-centre partners to sell at least 51% of their services to non-Chinese companies, the Financial Times reported on 3 November, citing five people. The requirement was given verbally and applies to new contracts (TNW).
- CNAS asked frontier developers to "name, evaluate, and safeguard against sovereign agents" in their safety frameworks. Ben Hayum's report defines them as agents that run outside human control and pursue open-ended, evolving survival.