OpenAI shipped Jalapeño, a custom ASIC built with Broadcom and TSMC in nine months, claiming 50% lower inference cost per token than Nvidia GPUs, with late-2026 deployment planned. This breaks Nvidia's inference hardware monopoly and signals hyperscalers are moving to proprietary silicon to cut per-token costs—a critical margin driver as inference pricing collapses.

Groq closed a $650M round to expand its global AI inference cloud as GPU rental pricing plummeted from $6.11/hour (B200) on May 30 to $4.22 by June 21. Inference-optimized competitors are consolidating capital while Nvidia's inference margins compress—builders should expect further price wars and custom-silicon competition.

Nvidia and AWS expanded tools to reduce operational complexity and latency in large-scale AI inference while lowering per-token costs. This signals the vendor partnership is racing to lock in hyperscalers before custom silicon (Jalapeño, Dragonfly, Cerebras) fragments the market further.

Anthropic is scaling AI compute capacity in Australia and Japan to meet demand, but copyright laws and power access remain blockers. Geographic data center fragmentation due to power constraints is forcing vendors to localize inference infrastructure—a shift that complicates global model serving.