General Compute is turning its next infrastructure expansion toward wafer-scale inference, signing a multi-year agreement with Cerebras to deploy its AI hardware through the company’s cloud platform. The agreement will bring Cerebras systems into General Compute’s commercial offering beginning in the first quarter of 2027, giving customers access to a different compute architecture without purchasing or operating the underlying machines themselves. General Compute plans to finance the systems directly and sell the resulting capacity as inference, positioning hardware ownership as part of its cloud business rather than leaving that capital burden with customers. The company has not disclosed the contract’s value or the total deployment scale, leaving the size of the commitment open even as it describes the deal as its largest single hardware commitment to date.

The timing matters because General Compute is making the commitment after raising $400 million in debt financing in July. The Cerebras agreement represents the first hardware commitment of this scale since that financing, tying the company’s capital strategy directly to the expansion of its inference infrastructure. For an AI cloud provider, the move reflects a growing need to assemble heterogeneous compute rather than build an infrastructure stack around one accelerator architecture. General Compute is effectively betting that customers will pay for the performance and economics of a specific workload without wanting to own the specialized infrastructure required to deliver it.

Cerebras targets the economics of agentic AI

Cerebras CTO and co-founder Sean Lie framed the partnership around the growing computational demands of AI agents, where latency can accumulate across long sequences of model interactions. He said: “In AI, speed is productivity. An agent that takes hundreds of steps to finish a task is only as fast as its slowest step. Working with General Compute puts Cerebras speed in front of the developers building these agents, on a platform they already trust.” The argument goes beyond benchmark performance because an agentic workload can turn small delays at the model-serving layer into a much larger productivity penalty across an entire task. General Compute’s cloud model gives Cerebras a route into those workloads without requiring every developer or enterprise to make a direct infrastructure investment.

General Compute CEO Finn Puklowski described that infrastructure gap as the central reason for the deal. He said: “The chips that win inference are not going to come from one vendor, and most customers cannot put a wafer-scale system on their own balance sheet. That is the gap we exist to close. We buy the hardware, and our customers get Cerebras speed on a contract they can actually sign. Agentic coding is where that speed is worth the most right now, so that is where we are starting.” His comments point to a cloud strategy built around matching different processors to different stages of AI workloads rather than treating accelerator choice as a single-platform decision. The commercial proposition becomes less about selling access to a particular chip and more about packaging specialized compute into a service that customers can consume without taking on hardware ownership.

Nvidia, AMD and SambaNova expand the compute mix

General Compute is not replacing its existing accelerator fleet with Cerebras systems, and its infrastructure already spans multiple processor families. The company uses Nvidia GPUs for prefill workloads, with Puklowski saying the approach “gives a major step up in reducing the cost of delivering inference – meaning more intelligence per dollar.” It combines that Nvidia capacity with AMD and SambaNova hardware, creating a multi-vendor environment in which different processors can handle different phases of model execution. That architecture gives General Compute more room to optimize inference around workload characteristics rather than forcing every stage through the same accelerator.

SambaNova GN50 chips handle decoding calculations within the platform, while AMD MI300X graphics cards manage the remaining inference workload phases. Cerebras now enters that mix with wafer-scale systems aimed at workloads where speed can carry particular economic value, especially agentic coding. The architecture suggests that General Compute sees inference as a pipeline that can benefit from assigning individual stages to the hardware best suited to them. Meanwhile, the use of Nvidia for prefill indicates that the company intends to combine conventional GPU infrastructure with specialized accelerators rather than treating wafer-scale compute as a universal replacement.

Cerebras expands AI cloud footprint

The General Compute agreement marks Cerebras’ second major AI cloud deal of the week, reinforcing the company’s push to make wafer-scale compute available through infrastructure providers. Gimlet Cloud has announced plans to deploy 100 megawatts of Cerebras wafer-scale compute for inference workloads through its cloud platform, with the first data center under that agreement expected to come online later this year. That deployment carries a substantially different disclosed scale from the General Compute arrangement, where neither the capacity nor financial value has been released. Still, the two agreements point toward a broader route for Cerebras systems to reach customers through cloud operators rather than through direct ownership by individual enterprises.

For General Compute, the Cerebras deal represents a more strategic shift than a straightforward hardware purchase. The company is building an inference platform around the premise that no single accelerator can serve every workload efficiently, while its financing model absorbs the capital intensity that specialized systems can impose on customers. That approach could become increasingly relevant as agentic applications increase the number of inference steps required to complete software, research and enterprise tasks. The key question now moves from whether wafer-scale hardware can deliver speed to whether cloud providers can translate that speed into durable inference economics at commercial scale.