General Compute is an AI inference neocloud, founded in 2025 and headquartered in San Francisco, California. The company buys compute racks, installs them in colocation data centers across the United States, and manages the power, cooling, networking and software required to run them. It targets latency-sensitive inference workloads rather than training, serving customers such as frontier AI laboratories, AI application businesses, coding-agent developers, voice-application builders, autonomous-systems firms and asset-light cloud providers. Its colocation sites represent more than 30 MW of capacity, with deployments scheduled to go live from Q1 2027.
Its central technical idea is a disaggregated inference architecture. Each request is split between different hardware types: NVIDIA GPUs handle prompt prefill, while specialised processors from Cerebras, SambaNova, Positron AI and d-Matrix perform token decoding. The company reports throughput of up to 2,000 tokens per second per user and up to 16.2x faster end-to-end performance than a traditional GPU cloud. Engineering work spans the full stack - high-performance computing hardware, inference workload scheduling and orchestration, and optimisation of large language model token generation.
Customers can take capacity in two forms: dedicated bare-metal compute on the company's colocated racks, or an inference API with service-level commitments measured in tokens per second. General Compute publishes open benchmarks alongside its performance claims. It raised a $100m Series A together with a debt facility of up to $400m.
