The agent that pays its own way.
Mining BTX on the GPU it already runs on.

BTX's consensus is MatMul Proof-of-Work: each mining attempt is a 512×512 matrix multiplication over the Mersenne-prime field 2³¹ − 1, a roughly 2 MiB working set that maps directly onto the tensor cores and matrix-multiply units a GPU already has. This is not a hash function bolted onto a graphics card. It is the identical operation — dense matmul over a finite field — that sits at the heart of training and inference. The project lists AI-training-class hardware (NVIDIA A100/H100, Apple M-series via Metal) as the recommended profile precisely because the workloads are the same shape.

Here is the part that makes an agent the natural operator. An agent serving a model does not saturate its GPU every millisecond — inference is bursty, with troughs between requests, idle fragments, pipeline stalls, and micro-batches that never quite fill the hardware. That headroom is paid-for silicon doing nothing. Because the proof-of-work is the same matrix-multiply primitive the model itself runs, an agent can co-schedule mining into exactly those gaps, turning once-wasted capacity into block-eligible work. The honest claim is not that one GPU does two full jobs at once; it is that the security computation becomes close to a side-effect of the AI workload — yield extracted from the utilization a model leaves on the table, at the kernel level where the matmuls already happen.

The economics are favourable because the overhead is small and the verification is cheap. The construction derives from the published cuPOW work and adds only about 16.5% above a bare matrix multiply at production parameters; the rest is arithmetic the GPU is built to do. Validators confirm a claimed product with Freivalds' algorithm — probabilistic checking with a false-positive probability below 2⁻⁶² — so the network never has to redo the work in full. Cheap to verify, expensive to fake: exactly the asymmetry an automated operator wants.

Put the loop together and something new appears. An agent provisions a GPU to serve a model. In the idle cycles it mines BTX, banking the reward in its own post-quantum wallet under a spend policy committed at output creation. It monitors its own cadence and difficulty health, and it pays for the next interval of compute — or for a downstream service that quotes a work challenge — out of what it mined. The unit of infrastructure becomes partly self-financing, with bounded authority encoded in the chain rather than granted by an administrator.

At scale that closes into a loop worth naming. Demand for settlement raises the value of the mining reward; a higher reward draws more AI-grade compute onto the network; more compute hardens the chain's security and the credibility of its monetary rule; and because the work is real linear algebra, the cycles spent securing the chain stay productive instead of being burned. Each turn makes the next one cheaper to justify — the same flywheel that has always linked a proof-of-work asset to the cost of the work, except here the work is the operation the AI economy is already paying for.

We are not claiming a perpetual-motion machine. Mining revenue is a function of difficulty, reward, and the value of the alternative use of those cycles; the Difficulty Commons exists so an agent can do that arithmetic continuously and honestly. The point is structural: for the first time the security budget of a chain, the running cost of an agent, and the asset the agent earns are denominated in the same thing — GPU time on power-secured compute.

We do not write about BTX as a token to hold. We write about it because of what it confirms about demand for compute. Mining BTX is, underneath, a continuous bid for GPU-class hardware in power-secured facilities — the identical asset hyperscalers compete for to train and serve models. What this piece adds is the operator: when the miner is an AI agent on shared hardware, that bid is no longer waiting on a human to place it. It is automated, price-sensitive, and running in every trough of every inference workload.

It also points at a cleaner way to think about value. When the proof-of-work is real matrix multiplication, the asset is effectively indexed to the worldwide cost of useful compute: its security budget rises and falls with the price of the exact GPU hours the AI economy bids for, and BTX's Difficulty Commons publishes that index in the open for anyone to read. For a firm whose thesis is the scarcity and pricing of power-secured compute, a monetary network whose value tracks the cost of compute is not a curiosity — it is the cleanest expression of the trade we already believe in.

That is the through-line across this blog. Stablecoins financing Treasuries, AI training and inference, proof of useful work, machine-speed settlement for autonomous agents, and now agents that run their own miners — all of them resolve to one constraint: regulated, power-secured, U.S.-located compute. Agentic mining does not change the thesis. It adds a tenant that is fast-growing, price-insensitive at the margin, and increasingly able to provision itself.

Whether BTX specifically becomes the standard chain for this is not the bet. The bet is the shape of the demand it reveals — and the floor that demand puts under the only asset that can satisfy it.