UK AI chip startup Fractile closed a C round of over $720 million at a $7 billion valuation, backed by the UK government's sovereign AI fund and the British Business Bank alongside Google Ventures and BlackRock. The company is building inference hardware that avoids trade-offs among throughput, model size and context length, and its analog in-memory compute chips are reportedly eyed by Anthropic for 2027 deployment.
AI Inference
AI 推理
Forward-pass computation serving user requests after training, now the dominant driver of AI compute demand.
Curated Intelligence
AWS CEO Matt Garman said in an interview that today's AI inference market is not even 1% of what it will eventually become, arguing there is a massive runway ahead. Coming from the head of the world's largest cloud provider, the remark points to sustained expansion of inference-side compute demand.
Qualcomm has introduced the Cloud AI 100 Ultra, a new data center AI inference accelerator card optimized for Generative AI and Large Language Models. Featuring 128GB of memory and up to 548GB/s bandwidth, it aims to support large-scale AI deployments by balancing high performance with power efficiency.
Korean AI chip startup FuriosaAI has raised $620 million in a Series D round, quadrupling its valuation to $2.9 billion. The company aims to challenge Nvidia's dominance in the data center market with its energy-efficient RNGD inference chips, which reduce power consumption by minimizing data transfer. Its products are already adopted by major clients like Samsung SDS and LG, with plans to co-develop next-gen accelerators with Broadcom.
Oracle announced an expanded cloud AI chip partnership with Nvidia, planning large-scale deployment of Vera Rubin servers. To address diverse AI workload needs, Oracle is diversifying its supply chain by becoming the first hyperscaler to offer Google TPUs, alongside inference chips from Cerebras and SambaNova, with services expected to roll out starting in 2027.
AMD CEO Lisa Su declared 2026 the 'true year of AI inference' at the OCP Summit, predicting that agentic workloads will drive a ~50% CAGR in the CPU market, reaching over $200 billion by 2030. She also confirmed that AMD's Helios production systems began shipping to major clients like OpenAI and Meta in September, with inference expected to account for 60% of the market share.
Microsoft announced its third-generation Maia 300 AI chip, designed as the primary inference engine for its cloud services. Optimized for long-context reasoning and agentic tasks, it claims a 40% increase in tokens per megawatt compared to leading GPU systems, with production rollout scheduled for 2027.
OpenAI revealed initial results for its custom AI inference chip, Jalapeño, claiming industry-leading speed and efficiency. The announcement highlights qualitative improvements in throughput and latency but lacks specific benchmark data or architectural details.