Can AI inference secure a blockchain?

The Bitcoin network used an estimated 189 TWh of electricity in 2025, more electricity than entire countries such as Argentina and Poland.

So what did all that electricity buy? Simply, security for a shared ledger.

At the same time, AI's demand for electricity is surging. Electricity use by data centers built for AI rose 50% in 2025 alone, and is expected to more than triple, to around 465 TWh, by 2030. That's roughly two and a half times what Bitcoin uses today.

The two industries are increasingly competing for the same power. Some Bitcoin miners, including Core Scientific, IREN, and Hut 8, are already using their electricity and sites to serve AI customers.

But what if companies didn't have to choose?

What if a single computation could help secure a blockchain and serve an AI inference request at the same time?

That is the promise of proof of useful work.

First, a quick refresher on Bitcoin mining

You likely know the basics. Miners gather transactions into a candidate block, then hash its data over and over, changing a small piece each time, until someone finds a result below the network's target. Other nodes check the winning block and build on it, and the winner collects newly issued BTC plus the block's fees.

Only one miner wins each block, but the losing tries aren't wasted: all that work is what makes Bitcoin's history so hard to rewrite. To change an old transaction, an attacker would need to:

1. Redo the work for the altered block and the blocks after it.
2. Catch up with honest miners as they keep adding new blocks.
3. Produce a valid competing chain with more accumulated work.

With less computing power than the honest network, the attacker's odds of catching up shrink with every new block.

It's worth pausing on why a hash puzzle does this job so well, because anything that replaces it has to do the same job. A good proof of work is:

• Hard to produce, cheap to check. Finding a winning hash takes sextillions of tries. Confirming it takes one.
• Tied to the block. Each try commits to the block's contents and the block before it, so no one can do the work in advance or reuse it.
• A fair lottery. Every try has the same small chance of winning, and no try makes the next one more likely. A miner's odds match its share of the network's computing power.
• Free of shortcuts. There's no clever way to find a winning hash faster than trying.
• Adjustable. The network can make the puzzle harder or easier to keep blocks arriving on schedule.

The hash puzzle gives Bitcoin all five almost for free.

In contrast, most versions of useful work so far don't check all of these boxes, and that's where the difficulty lies.

What proof of useful work could change

The hash puzzle has one big and fairly obvious downside. All that hashing is completely useless outside of Bitcoin. The electricity buys Bitcoin's security, but nothing else.

Now imagine the computation miners spend competing for blocks also produced something people wanted, like the answer to an AI inference request. The same computation could serve a customer and give the miner a chance to add a block.

What might that look like?

1. A customer requests an AI inference, such as an answer from a language model.
2. A GPU runs the model, performing eligible calculations along the way.
3. Those same calculations double as mining attempts, taking the place of Bitcoin's hash puzzle.
4. The customer receives an answer. If one of the mining attempts wins, the miner can submit a block and earn a reward.

This wouldn't be an upgrade to Bitcoin. Bitcoin's mining machines are built for one job, SHA-256 hashing, and can't run AI models. Proof of useful work means a different chain, mined with GPUs.

It also isn't a new idea. Primecoin, launched in 2013, had miners search for long chains of prime numbers. It worked as proof of work, but the results had little use outside math research. Gridcoin rewarded volunteers for scientific computing, but ended up securing its chain with proof of stake and paying for the science separately.

Between them, they show a similar pattern. Work that's genuinely useful struggles to pass the proof-of-work checklist from earlier: be a fair lottery, tie to a specific block, stay cheap to check. Work that passes it tends not to be very useful.

What's different now is AI inference, and it might break that pattern. It's enormous, in growing demand, and built mostly on matrix multiplication: multiplying large grids of numbers as a model processes a request.

So the big question is: could matrix multiplication behave like a hash puzzle?

One proposed approach works roughly like this:

1. Mix random noise, generated from the block being mined, into the grids.
2. Multiply them, as the model would anyway.
3. Hash the intermediate results. If the hash is below the target, the miner wins, just like a Bitcoin try.
4. Strip the noise back out. It has a simple structure, so this is cheap, and it leaves the customer's real answer.

Komargodski and Weinstein, who proposed this approach, show that the extra cost becomes negligible for large matrices.

The biggest benefit of this approach: even when a miner doesn't win a block, its computation still went toward an answer someone wanted.

Sounds great, so what's the catch?

If the math holds up, a GPU really could serve a customer and mine at the same time. But that's only the first of three questions a network has to answer, and each one is harder than the last.

1. Can the same work do both jobs?

The approach above suggests the math can work. Checking it is harder.

In Bitcoin, verifying a block takes one hash.

In the basic version of this approach, other nodes check a winning try by redoing the multiplication, which means they need the inputs too. And all of it has to happen without slowing down the customer's answer.

Different projects are attempting to tackle this in different ways.

2. Can anyone tell whether the work was wanted?

Say you're a miner. Serving real customers means finding them, handling their requests, and keeping them happy. Or you could skip all that: send yourself prompts, or just multiply random grids of numbers.

The calculations would satisfy the mining rules either way, and you'd have the same chance of winning a block.

So checking the math is not the same as checking that someone wanted it. Researchers have started calling this the usefulness gap.

Also, charging a fee doesn't close the gap, because you can pay the fee on your own request.

If the fee can't be recovered, faking demand at least costs something, but it still doesn't prove that anyone besides you wanted the work.

3. How much of the work is actually useful?

Even if a network can do useful work, the real test is how much of it does. Of all the calculations miners perform, what share serves customers who pay independently? A chain could support useful inference while most of its miners run something else.

Some of that is unavoidable. Customer requests arrive in bursts, but a blockchain has to keep producing blocks at steady intervals, so during quiet periods miners need something else to compute.

So the share of useful work should be measured and disclosed, not assumed.

But this is where the second question comes back: a network can only count customer work if it can tell customer work apart from everything else.

If it can't, any number it reports is a claim, not a measurement.

Additional challenges

Even with good answers to all three, some practical problems remain:

• Security. Bitcoin's mining machines can't do anything else, so miners only earn back what they spent on them if the chain stays healthy. GPUs, on the other hand, can be rented by the hour from a huge market. Smaller GPU-mined chains, including Ethereum Classic and Bitcoin Gold, have already suffered 51% attacks carried out with rented computing power. A chain mined on general-purpose hardware has to ask what it would cost to rent a majority, and whether the revenue from useful work makes an attack cheaper.
• Privacy. Prompts, model weights, and other inputs may be confidential. The network needs to verify mining proofs without exposing that data.
• Subsidies and unit economics. Block rewards might help lower inference prices, but the GPUs, electricity, and extra protocol work still cost money. Can customer payments cover those costs as rewards shrink, or does the service depend on token rewards forever?

Conclusion

The hard part isn't getting GPUs to mine. It's proving the mining was useful.

There's no guarantee proof of useful work will succeed at scale, but several projects are already trying. In future posts, I'll look at how their approaches hold up against these three questions.

Sources

Opening

1. Bitcoin used about 189 TWh in 2025. Cambridge Bitcoin Electricity Consumption Index — Cambridge Centre for Alternative Finance. Best-guess estimates; public data through March 29, 2026.
2. More than Argentina or Poland. Electricity Data Explorer — Ember, 2025 electricity demand by country.
3. AI data center electricity up 50% in 2025, around 465 TWh by 2030. Key Questions on Energy and AI — International Energy Agency, April 2026.
4. Core Scientific moving into AI infrastructure. Core Scientific second-quarter 2026 results — July 28, 2026.
5. IREN moving into AI infrastructure. IREN delivers Horizon 1 to Microsoft — August 13, 2026.
6. Hut 8 moving into AI infrastructure. Hut 8 River Bend lease with Fluidstack — December 17, 2025.

Bitcoin mining

1. Sextillions of hashes per block. Mempool mining dashboard — network difficulty, checked September 29, 2026.
2. An attacker with less computing power keeps falling behind. Bitcoin whitepaper — Satoshi Nakamoto, 2008, sections 4–6.

Proof of useful work

1. Primecoin's prime-chain proof of work. Challenges of Proof-of-Useful-Work (PoUW) — Felix Hoffmann, arXiv:2209.03865, September 2022.
2. Gridcoin's move to proof of stake. Gridcoin — Wikipedia. Forked from proof of work to proof of stake in October 2014.
3. Matrix multiplication as proof of work, and how nodes check it. Proofs of Useful Work from Arbitrary Matrix Multiplication — Komargodski and Weinstein, arXiv:2504.09971.
4. The usefulness gap. The Usefulness Gap in Proof-of-Useful-Work — Abhinaba Basu, arXiv:2606.04819, June 2026.

Practical challenges

1. Ethereum Classic attacked with rented hash power. ETC Labs CEO: Rented hash power is "huge" problem for proof of work — Decrypt, August 2020. Hash power rented from NiceHash.
2. Bitcoin Gold attacked with rented hash power. Bitcoin Gold 51% attack — MIT Digital Currency Initiative, February 2020. January 2020 reorgs, estimated at about $1,200 per attack in rented hash power.