General Compute

General Compute runs an inference cloud for chips other than the usual GPUs. It buys the hardware itself, places racks in colocation data centres and takes care of cooling, networking, energy and software, so that customers can run models quickly without operating the machines.

What it does

Customers can take bare metal capacity or a contract with a tokens per second service level agreement. The site shows logos for Cerebras, SambaNova, Positron AI and d-Matrix, and details one system, the Cerebras CS-3: two systems to a 48U rack, 47 kW average and 55 kW rated power, water cooled, with 44 GB of on-wafer SRAM. It quotes more than 1,000 tokens per second per user on models with multi-trillion parameters.

The company also describes splitting each request across chip types, running GPUs for prefill and specialised chips for decode. Its site claims up to 16.2 times faster end-to-end performance in its open benchmarks and 2.8 times the throughput per rack at a given speed per user, and compares 110 tokens per second per user on a GPU cloud with 2,000 on General Compute.

Capacity and pricing

General Compute reports more than 30 MW of capacity in existing United States colocation sites, with deployments going live from the first quarter of 2027. Cerebras capacity is priced per contract, sized to term, volume and site; no other prices are published, and the site names no customers.

Key facts

Funding: $100 million Series A co-led by Blackbird and Abstract (October 2026), after a $400 million debt facility from Upper90, as the company states it Capacity: more than 30 MW in United States colocation Hardware: Cerebras, SambaNova, Positron AI, d-Matrix Products: bare metal inference capacity, tokens per second SLA