For more than a decade, training and running artificial intelligence models has revolved around an almost undisputed protagonist: the NVIDIA GPU. However, a company called Cerebras Systems has decided to rethink the problem from the ground up with an idea as ambitious as it is counterintuitive: instead of manufacturing many small chips and connecting them together, why not turn an entire silicon wafer into a single giant processor? The result is the Wafer-Scale Engine, and it represents one of the most radical architectural bets the AI computing sector has ever seen.
NVIDIA’s traditional approach: many GPUs working as a team
NVIDIA’s model starts from a classic premise of the semiconductor industry. From each silicon wafer, dozens of individual chips (the GPUs) are cut out, each with its own package, its high-bandwidth memory (HBM) and its communication interfaces. A modern GPU such as the H100 or the B200 from the Blackwell family concentrates enormous power in a single package, but training large models requires connecting hundreds or thousands of them.
That interconnection is precisely where NVIDIA has built much of its advantage. Technologies such as NVLink, NVSwitch and InfiniBand allow many GPUs to communicate at high speed, while the CUDA software ecosystem acts as the glue that binds everything together and makes it easy for developers to take advantage of that hardware. The challenge inherent to this design is that, when data has to travel from one chip to another through cables and switches, latency is introduced and energy is consumed. At the scale of thousands of GPUs, that device-to-device communication becomes one of the main bottlenecks.
Cerebras’s proposal: a single wafer-sized chip
Cerebras turns the approach on its head. Instead of slicing up the wafer, it keeps it whole and turns it into a single, colossal processor. According to the company itself, its third generation, the WSE-3, measures about 46,225 mm² and houses around 4 trillion transistors along with 900,000 cores optimized for AI, delivering 125 petaflops of compute. To put those figures in context, Cerebras claims that its chip has “19x more transistors and 28x more compute than the NVIDIA B200”.
The great advantage of keeping everything on a single piece of silicon is that communication between cores happens within the chip itself, at speeds far higher than any external wiring. All the memory is integrated into the wafer and very close to the compute units, which boosts bandwidth and eliminates much of the latency associated with moving data between separate devices. In practice, a problem that in the GPU world requires coordinating a great many cards can, in the Cerebras world, be solved with the simplicity of a single device.
The manufacturing yield problem
The obvious objection to manufacturing a wafer-sized chip is production yield. In any semiconductor manufacturing process, defects appear, and in such an enormous chip the probability that imperfections exist is practically 100%. Cerebras’s solution does not consist of avoiding defects, but of designing to live with them: it incorporates redundant compute cores, redundant communication routes and a “fail-in-place” architecture capable of isolating the defective areas and rerouting the work around them.
What does all this translate into?
Where the difference is most noticeable is in inference, that is, in running already-trained models. Cerebras maintains that its cloud platform can be up to 15 times faster than GPUs in this type of task, something especially relevant for interactive applications, conversational assistants, agent-based workflows and reasoning systems that chain together many calls to the model. When each response arrives in a fraction of the usual time, it not only improves the user experience, but also opens the door to running more reasoning steps within the same latency budget.
It is worth noting that the comparison figures come from the manufacturer itself and depend heavily on the model, the workload and the specific scenario. In addition, NVIDIA maintains an advantage that is hard to match in ecosystem maturity, availability, tooling and developer community thanks to CUDA and years of massive adoption.
Two philosophies for the same goal
At their core, Cerebras and NVIDIA pursue the same thing (more computing power for AI), but they start from opposing philosophies. NVIDIA scales “outward”, adding many powerful chips and perfecting the network that connects them, relying on a dominant software ecosystem. Cerebras scales “inward”, concentrating an enormous amount of compute and memory into a single piece of silicon to minimize communication between devices.
In practice, this divergence translates into different use cases. NVIDIA’s approach fits particularly well where flexibility, software portability and the ability to deploy very varied workloads on the same platform matter, from large-scale training to general-purpose inference. Cerebras’s approach, on the other hand, shines when the goal is to minimize latency and squeeze the maximum inference speed out of a specific model, avoiding the cost of coordinating and synchronizing many devices.
There is also a factor that often stays in the background but carries more and more weight: operational complexity and energy consumption. Concentrating compute on a single piece of silicon reduces the number of interconnects, cabling and nodes that have to be powered, cooled and kept in sync. For certain organizations, that deployment simplicity can be just as decisive as raw performance figures.
It is not clear that one approach has to eliminate the other. The most likely outcome is that they coexist, each excelling in the scenarios that best suit its design. But the mere existence of a competitor capable of questioning the dominant paradigm is good news for the whole sector: innovation in AI hardware is far from having stalled, and the competition between these two visions promises to accelerate the pace of progress even further. The question is no longer whether GPUs will remain the standard, but how much room alternative architectures like Cerebras’s full wafer will manage to carve out as demand for AI compute keeps growing.