The news that Cerebras Systems has built its commercial traction around two names like OpenAI and AWS is more than a marketing win. It’s a snapshot of an AI accelerator sector where the game is played on a planetary scale and power balances can solidify or crumble around individual contracts.

Out-of-scale hardware that won over the giants

Cerebras is known for having shattered conventions with its Wafer-Scale Engine (WSE), a chip as large as an entire silicon wafer. The latest generation, the WSE-2, packs 850,000 cores and 40 gigabytes of on-die memory, with a bandwidth of 20 petabytes per second. On paper, these specs make the systems especially suited for training ever-larger models, eliminating the chip-to-chip communication bottlenecks that plague multi-GPU architectures.

For OpenAI, which constantly pushes the limits of Large Language Models, having access to hardware with a unified memory footprint and such high internal bandwidth can mean reduced training times and simpler model partitioning. AWS, for its part, has integrated Cerebras systems into its cloud offering, allowing customers to tap into these capabilities without managing the physical infrastructure.

When anchor customers become a double-edged sword

Having a handful of ultra-high-profile customers is not uncommon in the enterprise hardware world. But when those customers represent nearly all revenue, concentration risk becomes very real. Dependence on one or two large buyers exposes the company to sudden swings: a change in internal strategy, aggressive contract renegotiations, or the customers’ own development of alternative solutions could have devastating impacts.

OpenAI, for instance, is already working with Microsoft on large-scale GPU architectures, while AWS continues to invest heavily in its own Trainium and Inferentia chips. It’s not hard to imagine a scenario where interest in Cerebras cools, should internal roadmaps converge on more integrated or cost-effective hardware.

What’s the on-premise angle: sovereignty, TCO, and lock-in

For organizations evaluating on-premise deployments, the Cerebras story highlights familiar themes. On one hand, hardware solutions radically different from mainstream GPUs promise efficiency gains and a potential reduction in Total Cost of Ownership for specific workloads. On the other, the niche nature of these systems raises questions about support continuity, software portability, and the risk of vendor lock-in that would become untenable if the manufacturer ran into financial trouble.

Data sovereignty also hinges on the stability of the hardware supply chain. Choosing a supplier with an overly concentrated customer base means implicitly accepting a higher risk profile. In regulated environments—where GDPR compliance or data residency are critical—such uncertainty can weigh more than a favorable benchmark.

Beyond the headlines: a litmus test for the AI market

Cerebras’ strategy reflects a phase in which the AI accelerator ecosystem is still in flux. On one side, Nvidia dominates with GPUs and the CUDA ecosystem; on the other, new architectures seek their place through flagship customers and cloud partnerships. The fundamental question for those following deployment decisions isn’t just which chip delivers the best performance, but which manufacturer has the staying power to support workloads for the entire lifecycle.

In this landscape, Cerebras’ commercial concentration isn’t necessarily a death sentence, but a signal to watch closely. For anyone assessing on-premise deployments of specialized accelerators, the vendor’s commercial solidity becomes as critical as technical specs. AI-RADAR will continue to track these shifting balances, providing analytical tools to guide informed choices.