AWS Gave the Cheap Lustre Tier an SSD Ceiling. Only New File Systems Get It.
FSx for Lustre metadata IOPS on Intelligent-Tiering rises from 12,000 to 192,000. Existing file systems keep the old ceiling.
A Converted Aircraft Engine Beats the Grid by Sixty Months
Some AI campuses run on rebuilt jet engines because a grid connection takes 84 months. Time to power is the scarce asset, not megawatts.
SambaNova Is Selling Rack Capacity, and the Proof Is a Number Nobody Printed
SambaNova and Supermicro plan racks that split disaggregated inference between GPUs and SN50 chips. The 10,000 chip plan has no performance numbers.
A $6.5 Billion AI Lease Missed Its Closing. The Condition Was Financing, Not Demand
Fermi's 222 MW TensorWave lease carries $6.5 billion over 15 years. It still is not effective, because data center lease financing has to close first.
A Gigawatt of Demand Response Meets the Workload That Can’t Stop
OpenAI promised Georgia Power 1,000 MW of AI data center demand response. Inference is the load that cannot pause, and it is 87 percent of the fleet.
Anthropic’s Revenue Growth Halved. Its Compute Bill Grew Faster Anyway
Anthropic compute commitments reached $895.4 billion in the amended S-1, while quarterly revenue growth fell to 47 percent.
Marvell Paid $115 Billion for the Wire, Because Copper Stops at the Rack
Marvell is buying Lumentum for $115 billion. AI optical interconnect just became the line that decides how large a cluster can get.
NVIDIA Shipped Vera Rubin With 13 Percent Less Bandwidth Than It Advertised
NVIDIA cut Vera Rubin bandwidth 13 percent on its own spec page and never announced it. The HGX page still prints the old per GPU figure.
The Token Fell 77 Percent. The Rented Macs Found Out First.
AI token prices fell 77 percent in five months. Darkbloom told 800 Mac owners to expect no recovery, and the fixed GPU leases priced on that unit are next.
Microsoft Cut $5 Billion From Capex. The Reason Was Buildings, Not Demand.
Hyperscaler capex lands between $803B and $828B this year. The first cut came from late data center capacity, and that reprices 2027 GPU capacity.