Finn Puklowski

Inference Compute

Demand for fast, low-cost inference is attracting premium customers and long-term capacity commitments. The opportunity spans purpose-built silicon and neocloud providers, though several deployments remain years from launch.

• Cerebras (CBRS): Reported Ultrafast inference revenue reaches about $200M per megawatt annually at 750 tokens per second; capacity is reportedly sold out, with Jane Street among premium customers.
• Positron AI (POSI): Raised $875M at a $5B valuation and secured a multi-year General Compute cluster deployment, with hardware scheduled for Q1 2027.
• Nebius (NBIS): Shares were cited at $169.72, up 10.74%; an Anthropic S-1 is viewed as a potential neocloud-stock bottom signal.

Sovereign Chips

Governments are emerging as strategic buyers of AI processors, while the broader thesis is that inference silicon could become a persistent supply bottleneck. The key uncertainty is how quickly announced programs translate into scaled demand.

• Tenstorrent and Preferred Networks: Selected for a five-year Japanese government AI-processor project integrating photonics and electronics; sovereign and cybersecurity demand may broaden within 12 months.
• NVIDIA (NVDA): Reportedly plans to invest in d-Matrix, whose inference chips target Nvidia MGX rack integration; systems are expected in Q4 2027, but performance claims remain early-stage.