OpScale’s Breakthrough Slashes GPU Waste by 36%
OpScale is a new operator-level autoscaling framework that cuts GPU count by 36% and power by 28% for LLM inference without breaching latency targets.
Coverage of the chips that train and run AI, including GPUs, custom accelerators, supply constraints and the companies competing to build them.
OpScale is a new operator-level autoscaling framework that cuts GPU count by 36% and power by 28% for LLM inference without breaching latency targets.
AMD has acquired Toronto-based Taalas, a startup building chips that run only one AI model.
QumulusAI signed a Blackwell GPU capacity deal with an agentic hedge fund, a first-of-kind deal showing AI trading firms now compete for compute like model.
Nvidia CEO Jensen Huang says memory is now AI’s biggest constraint, a diagnosis that could deepen the company’s grip on the global AI hardware market.
AMD Helios has won OpenAI, Microsoft, Meta and Oracle as buyers, giving AMD a stronger foothold against Nvidia in large-scale AI racks.
Samsung AI chip demand helped Samsung post a record $62 billion quarterly operating profit, underscoring surging AI data center spending.
The U.S. government awarded GlobalFoundries $300M to advance silicon photonics, the technology that could break AI’s data-transfer bottleneck at the chip level.
Safe Superintelligence gains a $5 billion Nvidia investment, giving the frontier lab preferential GPU access and reshaping AI competition.
Nvidia has made a substantial investment in Safe Superintelligence and will supply the lab with next-gen Vera Rubin AI systems, Bloomberg reported July 27.
AI ETFs have grown to $65 billion in assets, per J.P. Morgan data, as institutional investors make the sector one of the fastest-growing thematic categories in.