A proof-of-work network based on the Bitcoin codebase, only five months old, uses 320,000 GPUs to price computing power through real market transactions.
Written by: @Decentralisedco
Compiled by: AididiaoJP, Foresight News
The Bloomberg terminal displays GPU price indices from Silicon Data, Ornn, and Compute Desk. All three track H100 rental prices, yet they cannot agree on what an H100 is actually worth.
Most GPU compute changes hands through private bilateral contracts at discounted rates, so these indices cannot effectively price it.
But a crypto network running 320,000 GPUs is already pricing GPU compute through real market transactions.
Why can no one price GPUs?
When you pay $3 to $10 per hour for an H100, what are you actually renting?
On Vast.ai, that $3 buys you a GPU card in someone's spare server rack, connected over ordinary Ethernet, with no guarantee of machine uptime. On the other side, the same H100 costs $10 per hour, but it sits in Oracle Cloud, a purpose-built facility where GPUs are interconnected via 400-gigabit-per-second InfiniBand and monitored around the clock by a full team of engineers.
The price is more than three times higher because Oracle is selling you an entire computing environment, not just access to a chip. That difference alone creates a 6.7x spread for the same piece of silicon across different providers.
The spread exists because the GPU itself is not the product. The GPU is merely one raw material within a larger service. A cluster of 512 GPUs connected via NVLink and InfiniBand in one building can run distributed training workloads, whereas 512 individual GPUs scattered across five low-cost data centers cannot efficiently accomplish the same task even with an identical number of chips. These are entirely different products that happen to share a name, just as a studio apartment in Queens and a penthouse on the Upper East Side are both technically called "a New York apartment."
The debt markets have already figured this out. CoreWeave has repeatedly borrowed against the same class of H100 hardware. When that debt is backed by a long-term take-or-pay agreement with Meta, the interest rate is about 5.9%. But when the company borrows against the same hardware without a contract wrapper, it pays 9.75%. The 400 basis points of rate spread is entirely the customer contract layered on top of the hardware. This reflects what the bond market believes it is actually underwriting, not the GPU itself.
The entire logic of a futures contract depends on being able to grade the underlying product. You inspect it, classify it, and the gap between the best grade and the worst grade must be narrow enough for the pricing mechanism to absorb the variance. Gold, for example, deviates less than 1% from London Good Delivery bars, which is why gold futures have worked well since the 1970s.
WTI crude oil allows roughly 5% deviation from the benchmark grade, with the rest absorbed by a differentials table. Even live cattle—arguably the strangest thing people have tried to stuff into a futures contract, since every animal is biologically unique—keeps quality bands within 15% to 20%. But when Silicon Data benchmarked 3,500 H100S GPUs across 11 different facilities, performance varied by 38%.
And that is only chip performance, before accounting for networking, uptime, cooling, and everything else that separates a low-cost rack from an enterprise-grade cluster. Once the full service configuration is factored in, the spread balloons to 200% to 700%. No grading table on Earth can absorb a seven-to-one quality range. Believing a futures contract can cleanly price this kind of thing is genuinely delusional.
So what happens when you try to build a delivery mechanism on top of something that cannot be graded? You get textbook adverse selection. If a seller owes you 1,000 H100 hours under a futures contract, they can deliver time on a high-end cluster or time on a $3-per-hour scattered low-cost spot instance. Both technically satisfy the contract. And they will deliver the cheap one every time, because the difference goes into their pocket and you have no recourse.
George Akerlof described this phenomenon in his 1970 paper on the market for lemons, concluding that when buyers cannot verify quality before delivery, the worst version of a product systematically drives out every better version until the entire market hollows itself out from within.
CME clearly knows this, which is why its GPU futures contract planned for October 5, 2026, avoids physical delivery entirely and instead settles in cash against a Silicon Data survey index. Except that index collects prices cloud providers publish on public rate cards, while the prices large buyers actually negotiate are typically 40% to 50% lower. The settlement mechanism is an estimate based on list prices, and most large buyers in the market clearly are not paying those prices.
You cannot fix this with a better survey. The information needed to accurately price a specific GPU hour in a specific facility exists only in the heads of the people operating that hardware, and they have every reason to keep it private. Moreover, compute is a flow commodity. A GPU hour that no one buys in that instant is gone forever. Any pricing signal for GPU compute must operate in real time, or the number is already stale by the time you publish it.
The only way to capture this knowledge is through real transactions at real scale, where counterparties have real money on the line. Which brings us to a solution that likewise evolved out of electricity: a network descended from Bitcoin, with 320,000 GPUs currently pointed at it.
What if you could compute the price?
Bitcoin miners run SHA-256 hashing. This computation exists for one purpose only: to prove that a miner burned a certain amount of electricity. Once that proof is recorded in a block, the hash—or the work done—is worthless. It cannot be reused, repurposed, or sold to anyone outside the blockchain. The work done to mine a Bitcoin block is not fungible.
The problem is that thousands of miners do this work for every block, but the network accepts only one block from one miner. The rest of the work is discarded. Every SHA-256 hash ever mined but not accepted by the network represents a GPU cycle that could have done something commercially valuable, and in 2026, "commercially valuable GPU work" means only one thing: AI compute.
Pearl is a Layer 1 blockchain attempting to make GPU compute commercially valuable. It inherits the same proof-of-work consensus and difficulty adjustment that have secured Bitcoin since 2009. But instead of SHA-256 hashing, Pearl miners perform matrix multiplication on GPUs.
Why is matrix multiplication valuable? Because it is the foundation of every answer you get when you ask ChatGPT a question. The same applies when Nano Banana renders an image from a text prompt, when companies fine-tune language models on proprietary data, or when autonomous vehicles process camera feeds in real time.
Every large language model and every diffusion model that has consumed Silicon Valley's attention and hundreds of billions of dollars in capital over the past three years runs matrix multiplication at its computational core. Pearl gives miners additional incentives to perform that same operation as proof of work.
What about security? That is the primary job of a hash function. You cannot simply replace a hash function with matrix multiplication.
SHA-256 has no shortcuts. You cannot produce an output without doing the full computation. The existence of the output means the necessary work has been done. Matrix multiplication is different. You are essentially multiplying and adding strings of numbers. Patterns can therefore let miners compute the answer with very little effort, undermining the effectiveness of proof of work.
Pearl uses a mechanism called NoisyGEMM to keep miners honest. It works in three steps:
Miners commit to the input matrices before beginning any work. Once submitted, the inputs cannot be changed.
The protocol then adds random noise to the committed matrices. This eliminates any symmetry (shortcut) the miner might have planned. And because miners do not know what the noise will be, they must perform the work later.
Miners run the full matrix multiplication on the noised version, block by block, on the GPU.
Each output block is hashed and checked against the network's difficulty target.
If a block's hash falls below the target, the miner wins the block reward.But introducing noise prevents miners from getting paid without actually doing work. Yet it is precisely this noise that also means we cannot get the answer to the original workload. Pearl's solution is to give the noise a known structure. This lets it strip the noise after the miner completes the multiplication, at a cleanup cost that is small and cheap.
When a GPU mines Pearl, the network measures how many matrix multiplications that machine actually completed under the then-existing power, cooling, and network conditions within the protocol's target 194-second block interval.
An H100 in a well-cooled enterprise facility may complete more work per block than the same chip sitting in someone's garage on consumer-grade Ethernet with poor airflow. Pearl automatically captures this difference, which no price index averaged by model can do.
Now scale this to 320,000 GPUs across dozens of geographic regions, and the total compute becomes the largest continuous real-world AI compute benchmark ever, updated every 194 seconds, without anyone needing to run a survey or publish a rate card.
Compute also tracks configuration. Each of these 320,000 GPUs can earn rental income on Vast.ai, CoreWeave, Lambda, or dozens of other cloud providers competing for AI customers. These operators choose to point their machines at Pearl.
When real demand for GPU compute spikes somewhere in the world, operators pull machines off Pearl to serve paying customers, and compute falls.
When demand softens and rental income declines, those GPUs flow back, and compute rises. The network's difficulty adjustment responds to these migrations in the same way Bitcoin's difficulty has responded to changes in mining economics since 2009.
The difference is that Bitcoin miners weigh the electricity cost they will incur against the revenue they can earn from selling the bitcoin they mine. During the August 2023 Texas heat wave, Riot Platforms shut down most of its miners and collected $31.7 million in power credits, more than three times the value of the 333 bitcoin it mined that month. When enough miners make that choice, hashrate falls, and difficulty readjusts every 2,016 blocks to ensure blocks are produced roughly every 10 minutes.
This flow of GPUs between Pearl and the rental market produces a continuously updated equilibrium price for GPU compute. High and rising compute means computers are cheap on the open market, because operators find mining Pearl more profitable than renting them out.
Falling compute means someone somewhere has started paying a high enough price to pull machines away. And this price signal comes from revealed preference, from what thousands of GPU operators choose to do with their own hardware and their own money. It does not come from what cloud providers' marketing teams decide to print on a rate card, which is all Bloomberg indices and CME settlement mechanisms have ever had access to.
Together AI is the first to commercialize on this basis. On May 15, 2026, it announced an inference endpoint running on the Pearl mining network, offering Gemma-4-31B-it-Pearl at a price more than 25% below standard cloud rates. The economics work because the GPUs serving inference through the Together endpoint are simultaneously mining Pearl, so mining revenue offsets part of operating costs, and Together can price below the broader market without subsidizing the difference out of its own profits.
Most of the Pearl network does not yet serve this level of demand from paying inference customers, and closing that gap is clearly the challenge ahead. But pricing does not depend on it. It depends on the opportunity cost being real, and that opportunity cost exists for each of these 320,000 machines, as well as for several others in the inference market.
The GPU compute market is a market worth hundreds of billions of dollars a year, yet it has no reliable price. Bloomberg indices try to track it but cannot give an accurate direction in any given week. CME plans to launch GPU futures on October 5, 2026, cash-settled against list prices that large buyers demonstrably do not pay.
But a proof-of-work network based on the Bitcoin codebase, only five months old, with 320,000 GPUs continuously deciding whether to mine or rent out capacity on the open market, and the aggregation of those decisions, is producing the closest thing to an accurate price that the GPU compute industry has ever had.