General Compute has partnered with Positron AI to split AI inference workloads between GPUs and memory-first chips, the two companies said, a move aimed at allocating computing resources more efficiently as demand for AI processing strains existing infrastructure.
The arrangement pairs General Compute's platform with Positron AI's memory-first chip design, divvying inference tasks across two distinct hardware types rather than routing everything through GPUs. Inference — the process of running a trained model to generate answers — is the step most AI services perform constantly, and it has become a major cost center for operators.
How the work gets divided
Under the partnership, inference jobs are split between GPUs and Positron AI's memory-first processors. The companies position the split as a way to optimize resource allocation, matching each part of an inference workload to the hardware that handles it best.
GPUs remain the default engine for much of AI computing, prized for parallel throughput. Memory-first chips, by contrast, are built around moving data efficiently, which matters for inference tasks that are often limited by how fast memory can feed the processor. Putting the two to work side by side is the core of the companies' pitch.
Neither company disclosed the financial terms of the partnership, the specific hardware configurations involved, or a timeline for deployment.
A challenge to GPU-only infrastructure
The collaboration could challenge GPU-dominant AI infrastructure models, where operators typically buy large fleets of graphics processors and run both training and inference on them. If inference can be shifted partly onto memory-first silicon, operators may be able to spread workloads across a more varied hardware mix instead of competing for the same scarce GPUs.
That matters because inference demand has grown alongside the rollout of AI products, and GPU supply has been a persistent constraint. General Compute and Positron AI are betting that a hybrid approach can ease that pressure and improve efficiency at the same time.
What the companies are saying
The two firms framed the partnership around resource allocation and efficiency rather than raw performance claims. General Compute brings its platform to the table; Positron AI brings memory-first chip technology. The companies did not release benchmarks comparing the split approach with GPU-only inference.
Their announcement lands as AI infrastructure providers look for ways to stretch available computing capacity without simply buying more of the same hardware. Splitting inference by workload type is one route; the partnership tests whether it works in practice.
What hasn't been settled
Key details remain undisclosed. There's no public timeline for when the split inference setup will be available to customers, no indication of which customers will use it first, and no performance data to evaluate the claims. How the two companies will divide revenue and responsibility also wasn't disclosed.
For now, the partnership is a signal about direction: AI infrastructure may become less monolithic, with different chips handling different jobs. Whether that translates into meaningful efficiency gains will depend on results the companies have yet to publish.
