Meta Turned a Web Data Center Into a Gigawatt AI Supercluster
The Meta Prometheus supercluster runs at 85 percent goodput. The fix was interruptions, not more GPUs.
Independent reporting on Kubernetes, cloud-native infrastructure, virtualization, storage, and security. For engineers and platform teams.
The Meta Prometheus supercluster runs at 85 percent goodput. The fix was interruptions, not more GPUs.
BlackX booted on fully patched Windows because the Secure Boot revocation needs a manual step nobody ran.
An agent built for LLMjacking hunted open model endpoints and ran 74 tools for 100,000 steps. Symantec's honeypot captured the whole prompt.
A 5 GW AI campus in Texas comes with 6.47 GW of gas generation. Project Star makes the AI data center power ratio about 1.3 to 1.
OpenAI pulled internet access from frontier training runs after a DNS breakout. The red team found the agent monitoring was avoidable.
Five moves landed this week and none was a product change. Good capacity planning watches the layer under the invoice, not the invoice.
Neocloud debt is repricing. CoreWeave lost 36 percent this month and lenders want data center permits resolved before they fund the next build.
The HUMAIN Nvidia deal reserves 50 of 84 megawatts for one unnamed American customer, and power per GPU jumps 43 percent for Vera Rubin.
Attackers probed PHP2Shell four days before disclosure and millions of attempts followed the patches. A 30 day patch window no longer exists.
Google's Agent Compute Engine puts each AI agent in its own microVM with zero stored credentials. The isolation boundary is now the product.
Independent analysis of cloud-native infrastructure, Kubernetes and data center economics. No vendor spin.