Mining hardware is the clearest case in computing of the trade between general-purpose silicon and chips built to do one thing. It is also the case that taught the industry why that trade can be a trap, and the same question now shows up in a place with much better economics: zero-knowledge proving.
Why Bitcoin has no GPU alternative
SHA-256 is a fixed function. There is no algorithmic cleverness left to find, no memory hierarchy to exploit, so the only remaining lever is raw efficiency — how many hashes per joule a chip can produce. That is precisely what an application-specific integrated circuit is for, and the result is a hardware arms race that general-purpose silicon lost decisively years ago.
The consequence is stranded capital. A Bitcoin ASIC is worthless for anything except mining Bitcoin, so the operators who buy them are making a leveraged bet on the price of the asset and the cost of power at one specific site. That produces the industry's characteristic behaviour: fleets migrate to wherever electricity is cheapest and most stranded — excess hydro, flared methane, curtailed renewables — and the hardware's resale value collapses as newer generations arrive.
What happened to GPU mining
Every other proof-of-work chain could be mined with GPUs, and Ethereum was the largest of them. GPUs were attractive because they hold resale value and can be repurposed — a rented card is a cost, not a sunk one.
Then Ethereum moved to proof-of-stake in September 2022, and the GPU-mining market's largest tenant disappeared overnight. The hardware did not disappear; it was absorbed by other proof-of-work chains and, decisively, by AI workloads, where the same cards that hashed efficiently turned out to be exactly the accelerators the training and inference markets needed. That redeployment is the whole difference between the two hardware markets, and it is why no GPU miner faced the stranded-asset problem an ASIC miner faces.
Where acceleration pays now
The mining comparison gets the attention, but the interesting applications are elsewhere.
Zero-knowledge proving. Proof generation is compute-bound and the workload is predictable, which is the same profile that made ASIC mining viable. GPU provers already sit in most production proving pipelines, and FPGA and ASIC implementations beat general CPUs by a wide margin. This is the clearest case where specialisation is winning today.
FHE and MPC. Homomorphic encryption and secure multi-party computation are similarly compute-bound and similarly predictable, so the same economics apply.
Consensus is deliberately not accelerated. Proof-of-stake validation is designed to be cheap, which is why the Merge removed a hardware arms race rather than escalating it. That was a design decision, not a technical limitation.
Why this does not repeat mining's outcome
Three differences matter.
Hardware retains optionality. An FPGA can be reconfigured, and a GPU keeps resale value across a wide range of workloads. The capital is not stranded when the target changes.
Demand tracks activity, not emissions. Mining demand is a function of a fixed block reward and a token price. Proving demand is a function of transaction volume and how many applications choose proving — a broader and less reflexive base.
No perpetual hardware premium. Because the hardware is replaceable and general, whoever buys the first chip does not capture a durable advantage. That is worse for a hardware speculator and better for everyone building on top.
The practical read
For a protocol, the question is whether proving cost per transaction falls fast enough that ZK verification becomes the default rather than an opt-in. That is a hardware-roadmap question as much as a software one, and it is the main constraint on how much of the stack can be proven rather than trusted.
For an operator, the question is whether the hardware holds value when the workload changes. Mining ASIC buyers learned that the answer was no, expensively. Anyone buying proving hardware today should ask the same question and check that the answer is different.
References
Next, read Proof of Computation 101. More in Infrastructure & Scaling.