Zettascale Computing

AI models have largely exhausted the gains available from reading more text, and the field is shifting toward learning from experience through agents, simulators, and recursive self-improvement. This kind of workload is dominated by moving data rather than doing arithmetic, and existing chips like GPUs (built for graphics) and TPUs (built for tensor math) were not designed for it. Datacenters are capped by power, so energy spent moving data instead of computing is a direct ceiling on how much discovery can happen per watt. Companies pursuing frontier AI and agentic workloads need hardware that keeps data close to computation rather than generic accelerators adapted after the fact.

Zettascale is building XPU Grasshopper, an AI chip described as reconfigurable ('polymorphic') silicon that changes with the workload and keeps data close to reduce the energy spent moving it. It currently runs end-to-end inference live on an FPGA (VU47P), with a serving API coming soon, and the company says it is heading toward dedicated silicon. Developers can bring existing frameworks (PyTorch, tinygrad, JAX) and target the XPU with a single import, and the company also exposes libxpu, a planned open-source C ABI that gives direct control over buffers and data movement for custom kernel development. The company is also building Monolith, a cluster of XPUs intended to behave as a single larger chip for long, stateful workloads like agents, experience generation, and training, currently in early access rollout to pre-order ticket holders.

The company argues language-based scaling has hit diminishing returns and that models now improve through experience and recursive self-improvement, which is far more compute- and data-movement-intensive, making chip-level efficiency the new bottleneck.