China’s Largest AI Campus Is One-Fourteenth of a Tennessee Building, in a Bite
Epoch AI sized Huawei's Horinger data center from orbit at 81,000 H100-equivalents, against 1.1 million at Colossus 2 in Tennessee.
The KV Cache Speedup Is Real. The Fabric Is the Ceiling, in a Bite
A shared KV cache tier reports a 56x speedup. The IBM Redbook behind it says client-side networking, not storage, is the ceiling.
A Xen Guest Can Reach Another Guest’s Memory, in a Bite
A guest on Xen can read and write memory belonging to other guests, Dom0 included. No mitigation exists, only patches.
An Aircraft Engine Beats the Grid by Five Years, in a Bite
Converted jet engines power an AI campus in 24 months where the grid needs 84. Time to power is the scarce asset, not megawatts.
Europe Got a Later VMware Deadline, in a Bite
Broadcom split its VMware end dates by region. Europe keeps vSphere Standard 8 until October 2028, everywhere else ends in 2027.
The Cheap VMware Edition Came Back With a 128 Core Ceiling, in a Bite
vSphere Express is 128 cores, one pack per customer, no expansion and no published price. The cheap VMware edition has a ceiling.
A Gigawatt of Grid Relief Needs Workloads That Can Stop, in a Bite
OpenAI promised Georgia Power 1,000 MW of demand response. Ember's workload mix puts the pausable share of that campus nearer 3 to 7 percent.
Anthropic Put $895 Billion on the Books, in a Bite
Anthropic compute commitments hit $895.4 billion in the amended S-1, and only about a quarter of the total can be avoided.
The Wire Costs $115 Billion, in a Bite
Marvell pays $115 billion for Lumentum because copper stops at the rack. The optical layer now decides how big an AI cluster can get.
Eight vLLM Advisories Closed at Once, in a Bite
Eight vLLM advisories close in 0.32.0. The worst vLLM vulnerability skips the insecure serialization flag entirely.