vLLM Hit 7.8x More Per-GPU Throughput on Nvidia’s Rubin
vLLM already runs on Vera Rubin NVL72 at 7.8x the per-GPU throughput of GB200, and memory bandwidth explains most of the jump.
Cloudflare Acquires Deno, and Deno Deploy Has Six Months Left
Cloudflare acquires Deno, and the runtime built as a hedge against lock-in now carries a six month migration clock.
The Biggest Data Center Subsidy Is the Sales Tax on the GPUs
A Senate report puts data center subsidies on the hardware, not the land. Staffing runs one worker per megawatt and GPUs are 39 percent of the spend.
Last Week’s NetScaler Patch Is Still Vulnerable as a SAML IdP
NetScaler CVE-2026-107406 is a SAML memory overflow. Last week's fix is still open when the box acts as an identity provider.
The Gas Peaker Just Lost to a Battery in Every Market Modeled
Battery storage is now cheaper than open-cycle gas turbines in all 43 markets Wood Mackenzie modeled, and the turbine waitlist runs to the early 2030s.
AI Bought the Memory First, and PC Buyers Got the Invoice
The AI memory shortage put DRAM and NAND at nearly 40 percent of a PC component bill. Shipments fell 20.1 percent and prices are not coming back.
The Cheapest Inference on the Market Is a Classifier, Not a Chat Model
Microsoft-Decision-1 lands in a crowded decision model market. The pitch is not accuracy, it is $0.042 per million input tokens with output free.
A Single Prompt Owned Every Agent in the AWS Region
A Zenity Labs chain turns one prompt into every agent in an AWS region. The AgentCore vulnerability reaches the instance metadata service.
AI Infrastructure Got Its Invoice This Week. Every Line Was a Constraint.
AI infrastructure costs moved on four fronts in one week. Texas billed the grid queue, memory repriced the rack, and a chip vendor financed its customer.
The Government Booked $2.4 Billion of AI Compute in Vendor Credits
The Genesis Mission drew $2.4 billion in AI compute credits from eleven vendors, led by Nvidia, AMD and OpenAI.