Wednesday, October 14, 2026
Wednesday, October 14, 2026
Chips, data centres, energy and the race to build AI's infrastructure.
Anthropic’s eight-node exercise used emulators and an unauthenticated access path; its projected doubling time is measured in days, not observed in real networks.
Compute Desk3 min read
Microsoft says Maia 200 supports production workloads across three regions; Maia 300’s planned scale and performance figures remain future plans and company claims.
The first planned transceivers have 16 lanes, with a possible expansion to 32 or more. NVIDIA and Cisco have not set a sampling or customer-availability date.
Marvell says LZ4 compression can raise effective capacity two to three times, but its inference performance claims remain untested and sampling is expected in Q1 2027.
CXMT plans a DDR5 server-memory module by year-end, but its announcement does not establish production readiness or measured performance gains.
Intel specifies 160 GB for its branded card; ODM partners may offer up to 480 GB, while whether its FP8 peak will be sufficient for inference remains to be seen.
The CPU-only blueprint reserves capacity for agent execution while separate infrastructure handles model inference; AMD says partner configurations are planned for November 2026.
PNNL outlines goals in grid planning, geothermal and scientific computing but publishes no project budgets; DOE’s separate Phase II pool is not tied to the five projects.