A $3,499 Box Runs a 27B Model Locally, in a Bite
Ghost's Core, a $3,499 box running a 27B model on device, sold out its first batch in hours. Local AI inference is a real purchase.
Google Funded Nuclear Uprates Instead of a New Reactor, in a Bite
Google's nuclear power purchase agreement funds reactor uprates for 890 MW of new capacity, without building a new plant.
Sovereign AI Capacity Costs $60 Million a Megawatt, in a Bite
AM Intelligence ordered 20,000 more Nvidia Rubin GPUs, taking the order book to 100 megawatts, and the cost per megawatt of new AI capacity rose.
Intel Raised CPU Prices Again and Memory Is the Reason, in a Bite
Intel's third CPU price increase in a year took effect October 5. DRAM and NAND are up five to seven times and the cost is spreading.
Goldman Kept the Data Center Number and the Backlash Did Not Dent It, in a Bite
Goldman raised 2026 US data center capacity to 64 GW and cut 2027 to 90 GW. Opposition is a cost and a delay, not a ceiling.
AWS Raised Reserved GPU Prices and Moved the Chips, in a Bite
AWS raised reserved GPU capacity prices again on 7 October and is moving $8 billion of Grace Blackwell chips into a financing vehicle.
Nvidia Halved the Memory to Hold the Price, in a Bite
Memory rose from 9 to 26 percent of an AI rack's material cost in one generation. Nvidia halved it to hold the price.
A GPU Hour Gets a Public Price on Monday, in a Bite
CME lists the first compute futures on 5 October, settling against a daily GPU rental index.
Nvidia Wants a Decade of Collateral. Lenders Price Four Years, in a Bite
GPU residual value is now a financing term. Lenders price a three to four year life while Nvidia argues for ten, and the GPU hour pays for the gap.
Cisco SD-WAN Manager Is Patched, Not Cleaned Up
CVE-2026-76504 is patched but not cleaned up. Admin API access is enough to rewrite SD-WAN policy, so triage starts in the logs.