General Compute

General Compute is a neocloud for AI inference that combines NVIDIA GPUs for prefill with alternative decode chips such as Cerebras, SambaNova, Positron AI and d-Matrix, offered as bare metal or as an endpoint with a tokens per second SLA.

General Compute is a neocloud focused on AI inference rather than training. The company purchases racks of accelerators, installs them in colocation data centers and sells the resulting capacity either as bare metal or as an endpoint with a tokens per second service level agreement. Its hardware mix spans NVIDIA GPUs for prefill and purpose-built silicon for decode, including Cerebras CS-3, SambaNova, Positron AI and d-Matrix. The company states it holds more than 30 MW of capacity, with deployments going live from Q1 2027. The technical premise is disaggregated inference. Reading the prompt happens on GPUs, while writing the answer runs on ASIC systems that are tuned for token generation. Because the prefill and decode racks are linked by fast networking and split by a software stack that distributes each request across chips, General Compute claims 2.8x the throughput per rack at the same speed per user, and up to 16.2x faster end to end in its own open benchmarks against a conventional GPU cloud. Those figures come from the vendor, so treat them as directional until reproduced on a real workload. The product suits teams running latency-sensitive inference at scale: agent platforms and coding assistants where a single task may involve hundreds of sequential model calls, and where the slowest step sets the pace. Buyers who want a single contract, one set of SLAs and the option to customise rather than a self-managed cluster of GPUs are the natural audience. Competitors are the mainstream GPU clouds and other inference-specialist providers; General Compute differentiates by aggregating several chip vendors instead of reselling one. Commercial terms are quoted per contract and sized to term, volume and site, so there is no public price list. Engagement starts with a call in which the buyer describes the model, the traffic shape and the interactivity target, and receives a rate card and a capacity date. Capacity is not live yet, which makes this a forward commitment rather than a drop-in replacement today.