TPU (Tensor Processing Unit)

TPU (Tensor Processing Unit) is an application-specific integrated circuit (ASIC) developed by Google to accelerate machine-learning workloads. Rather than being a general-purpose processor, a TPU is purpose-built for the dense matrix multiplications that dominate neural-network training and inference, trading flexibility for throughput and power efficiency on that narrow task. It is the same design philosophy that produced the Bitcoin mining ASIC: pick one mathematical operation the world needs at enormous scale, then burn the silicon to do only that, orders of magnitude more efficiently than any general-purpose chip.

Systolic array architecture

The heart of a TPU is a systolic array: a grid of multiply-accumulate (MAC) units wired directly to their neighbours so data flows through the array in lock-step, like blood pumping through a heart — which is where the name comes from. Google's first-generation TPU used a 256 x 256 array of 65,536 arithmetic units, letting it perform tens of trillions of multiply-add operations per second without constantly shuttling intermediate values back to memory. That dataflow design is the source of the efficiency win: in a conventional processor, moving data between memory and compute costs far more energy than the arithmetic itself, and the systolic array simply keeps the data moving through the compute fabric instead. Later TPU generations added high-bandwidth memory, floating-point support for training, and pod-scale interconnects linking thousands of chips into one machine.

The ASIC parallel miners already understand

A Bitcoin miner has lived this story once already. CPU mining gave way to GPUs, GPUs to FPGAs, and FPGAs to SHA-256 ASICs that do nothing else but do it thousands of times more efficiently. AI compute is partway through the same curve: GPUs still dominate because models keep changing shape, but wherever the workload stabilizes, specialized silicon like the TPU wins on joules per operation — the AI equivalent of a miner's joules-per-terahash discipline. The difference is who holds the hardware. Anyone can buy a mining ASIC and point it at any pool; TPUs live almost exclusively inside Google's data centres and are rented by the hour.

Why it matters for sovereignty

That rental model makes the TPU the clearest example of centralized compute: powerful, but permissioned, metered, and revocable. Your workload runs on someone else's silicon, under someone else's terms of service, with your data in transit to their building. For builders who want AI under their own roof, the TPU is the benchmark to route around rather than the tool to adopt. The encouraging part is that the same systolic-array idea scales down: it reappears inside the NPU in a laptop, in edge accelerator modules, and in the tensor units of consumer GPUs — hardware you can own outright and run offline for inference on local models.

Google has also shipped the idea at the edge: small accelerators in its Coral line brought TPU-style inference to single-board computers at a few watts, and that pattern — a specialized matrix engine you can actually buy and own — is the fragment of the TPU story most useful to a self-hoster. The rule of thumb transfers cleanly from mining: measure hardware by operations per joule and by who controls it. A rented exahash and a rented exaflop share the same weakness, and an owned accelerator, however modest, shares the same strength.

D-Central's interest in the TPU is exactly this convergence: hash-intensive and inference-intensive compute are both ASIC stories, both energy stories, and both sovereignty stories. Understanding how Google's chip wins on efficiency clarifies what to look for in self-hosted alternatives — memory bandwidth, operations per watt, and a workload stable enough to specialize for. See how these threads meet in distributed compute and the broader move toward sovereign AI, where the goal is the TPU's efficiency without the TPU's landlord.