General Compute Cloud launched into general availability on May 22, positioning itself as the first neocloud built around inference-optimized ASICs rather than general-purpose GPUs. The company runs SambaNova SN40 and SN50 dataflow silicon, claiming the fastest independently benchmarked speeds on the MiniMax M2.7 model family.
The more consequential feature is the onboarding. According to the company’s announcement, General Compute Cloud is designed so that coding agents, not human operators, complete the entire signup flow. When invoked through OpenClaw, OpenCode, or any compatible client, an agent can create an account, verify the workspace, claim the $200 launch credit, and return a scoped API key back to the developer’s environment. The entire process runs through natural language. No dashboard, no forms, no manual provisioning.
“Our goal is simple: we want General Compute to be the fastest inference provider on the market, and we want to ship the fastest inference API any developer or AI agent can call,” said Jason Goodison, CTO and co-founder of General Compute, in the announcement.
Agents as Cloud Customers
The idea of treating agents as first-class infrastructure users, rather than tools operated by humans who manage the infrastructure, has been gaining traction in 2026. Agentuity, another infrastructure startup, published a conceptual framework in February arguing that bolt-on AI services on existing cloud platforms are insufficient: infrastructure needs to be rebuilt from the ground up with agents as the primary persona. Google made a similar move at Next ‘26 in April, transforming GKE into what it called an orchestration engine for agent-native workloads.
General Compute’s implementation is narrower than either vision but more concrete. The company is not rebuilding cloud orchestration. It is solving a specific friction point: agentic workloads issue dozens or hundreds of model calls per task, and the onboarding flow for most cloud providers still assumes a human clicking through a dashboard. By removing that bottleneck, General Compute lets agents provision their own inference infrastructure as part of a broader task chain.
The Limits
General Compute is an early-stage company with a narrow product surface. The announcement covers inference on open and frontier models, not training, fine-tuning, or general compute workloads. The independently benchmarked speed claims are not linked to a specific benchmark report in the announcement. The $200 launch credit offer runs through May 27, 2026.
The agent-native onboarding pattern itself raises questions about access control and billing authorization that the announcement does not address in detail. When an agent can sign up for cloud services autonomously, the authorization chain between the human who owns the agent, the agent that provisions the account, and the billing entity needs to be explicit. For teams already managing agent permissions through OpenClaw or similar frameworks, this adds another surface to govern.