750 Tokens Per Second: Why Cerebras Wins the Inference Speed Race
Cerebras and OpenAI launched Ultrafast, running GPT-5.6 Sol at 750 tokens per second — 14x the standard rate. The secret is on-chip memory, not raw compute.
Read ArticleCategory: Infrastructure
Articles about Infrastructure
IBM's $240M Open Inference Bet Is the Inflection Point Enterprise AI Needed
IBM and Together AI just committed $240M to proving open-source inference can undercut the hyperscalers. Here's why this changes the math for every team running AI in production.
Read Article
Mojo 1.0 Is Live: The Serious CUDA Alternative That Just Became Production-Ready
Mojo 1.0 shipped today with a stability contract, an Apache 2.0 standard library, and MLIR-backed benchmarks that match CUDA on memory-bound workloads. Here's what it means for AI infrastructure.
Read Article
Ten Trillion Parameters, Thirty Thousand GPUs, and No H100s in Sight
ByteDance has started the largest AI training run ever attempted — on domestic Chinese chips. Here's what the infrastructure reality looks like, and why the hardware story matters more than the parameter count.
Read Article
SCTPhantom: The 18-Year-Old Linux Kernel Bug That Escapes Your Containers
A use-after-free lurking in Linux's SCTP stack since 2008 just became a full container-escape chain. Here's the operational breakdown and what you need to do right now.
Read Article
Terafab: Tesla and SpaceX's $16.8B Bet That the Chip Supply Chain Is Broken
Tesla and SpaceX just committed $16.8 billion to build a chip factory in Texas. The real story isn't the money—it's what vertical integration at this scale reveals about who actually controls AI compute.
Read Article
The Model IS the Chip: AMD's Taalas Acquisition and the End of the Memory Wall
AMD acquired Taalas, which etches AI model weights directly into silicon. At 16,960 tokens/sec on Llama 3.1 8B with no HBM--this isn't optimization, it's a paradigm shift for inference.
Read Article
GPUBreach: The Rowhammer Attack That Broke Through IOMMU
University of Toronto researchers showed at Black Hat 2026 that GDDR6 Rowhammer attacks can reach a host CPU root shell by bypassing IOMMU — the control the industry assumed was sufficient.
Read Article
The 10-Hour CUDA Clone: How AI Coding Agents Are Eating Nvidia's Software Moat
A one-year-old startup built CUDA-equivalent chip software in 10 hours using AI coding agents. The implications for Nvidia's moat—and for every team running AI infrastructure—are significant.
Read Article