Tracking the infrastructure behind foundation models, inference, and agents.
The reported evaluation raises questions about government data, accelerator availability, and what Malaysia's vendor-neutral policy means in practice.
Terminal-Bench-Science measures agents on research workflows, not textbook questions. A vendor-run result released days later shows how quickly its baseline can move.
DOE is funding a grid model designed to evaluate 1 billion scenarios a day. But queue delays arise as much from process, staffing, speculative requests, and construction as from solver time.
Four jurisdictions, four approaches: an unapproved Ohio settlement, a TVA board-set rate, a contested FERC proceeding in Wisconsin, and a Texas audit of its large-load queue.
Compute scales about 3x every two years; memory bandwidth scales under 2x. NVIDIA, AMD, Google, Intel, and OpenAI each answered that gap differently.
Brazil pairs a Huawei-iFlytek project with a supercomputer tender officials expect NVIDIA to win. The tender buys portability on paper; the silicon stays foreign.
Eight converted bitcoin miners hold about 5.7 GW of AI leases. Under 1 GW is live. The rest is contracted and awaits phased construction and energization.
NVIDIA has priced training across 1,000 km; Google trains synchronously across data centers. Missing: what one job pays to span mixed silicon and fabrics.
Google outbid a training-data startup for Spirit Airlines' internal records. One auction does not prove a data wall, but it shows how far labs will go.
Rental indices, secondary venues and syndicated credit already exist for GPU compute. What lenders don't have is a residual-value history to underwrite against.
As data-center developers add on-site generation to bridge utility delays, the grid queue is being joined by queues for turbines, air permits, gas infrastructure, storage, and EPC capacity.
