A cryptocurrency network running 320,000 GPUs is already pricing GPU computing power using real market transactions.

Written by: @Decentralisedco

Translated by: AididiaoJP, Foresight News

The Bloomberg terminal displays GPU price indices from Silicon Data, Ornn, and Compute Desk. All three track the rental prices of H100s, but they cannot agree on how much one H100 is worth.

Most GPU computing power circulates through private bilateral contracts at discounted rates, making these indices ineffective for pricing.

But a cryptocurrency network running 320,000 GPUs is already pricing GPU computing power using real market transactions.

Why can't anyone price GPUs?

When you pay $3 to $10 per hour for one H100, what are you actually renting?

On Vast.ai, that $3 gets you a GPU card from someone’s spare server rack, connected via regular Ethernet, with no guarantee of machine uptime. On the other hand, the same H100 costs $10 per hour, but it’s located in Oracle Cloud, a purpose-built facility where GPUs are interconnected via InfiniBand at 400 gigabits per second, and a whole team of engineers monitors it around the clock.

The price is more than three times higher because Oracle sells you a complete computing environment, not just access to silicon. This single difference causes the price disparity of the same chip between different providers to reach 6.7 times.

Price disparity exists because the GPU itself is not the product. GPUs are just a raw material within a larger service. A cluster of 512 GPUs connected by NVLink and InfiniBand in one building can run distributed training tasks, whereas 512 individual GPUs scattered across five low-cost data centers, even with the same number of chips, cannot operate efficiently. These are entirely different products, merely sharing a name, much like a studio apartment in Queens and a penthouse in the Upper East Side, both technically called "a New York apartment."

The debt market has understood this. CoreWeave has repeatedly borrowed against the same class of H100 hardware. When this debt is backed by a long-term, non-negotiable agreement with Meta, the interest rate is about 5.9%. But when companies borrow with the same hardware and no contractual packaging, they must pay 9.75%. The 400 basis point spread is entirely attributed to the customer contracts above the hardware. This reflects what the bond market believes it is actually insuring, rather than GPUs themselves.

The entire logic of futures contracts depends on being able to grade the underlying product. You check it, classify it, and the gap between the best and worst grades must be narrow enough for the pricing mechanism to handle variance. For example, the deviation of gold from London Good Delivery gold bars is less than 1%, which is why gold futures have performed well since the 1970s.

WTI crude allows a relative benchmark grade deviation of about 5%, with the rest absorbed by the contango and backwardation. Even live cattle—probably the strangest thing people try to squeeze into futures contracts—are kept within a quality band of 15% to 20%. However, when Silicon Data benchmarks 3,500 H100S GPUs from 11 different facilities, the performance difference reaches 38%.

And that’s just chip performance, not accounting for network, uptime, cooling, and everything that distinguishes low-cost racks from enterprise-grade clusters. Once full-service configurations are taken into account, the price differences can swell to 200% to 700%. There is no grading table on earth that can absorb a seven-to-one quality range. To think futures contracts can cleanly price such things is indeed a delusion.

George Akerlof described this phenomenon in his 1970 paper on lemon markets, concluding that when buyers cannot verify quality before delivery, the worst versions of the product systematically drive out every better version until the entire market is hollowed out from within.

The CME clearly knows this, which is why their GPU futures contract scheduled to launch on October 5, 2026, completely avoids physical delivery and instead settles in cash against the Silicon Data survey index. However, that index collects prices published by cloud providers on public rate cards, while large buyers typically negotiate prices that are 40% to 50% lower. The settlement mechanism is based on the listed estimated prices, whereas most large buyers in the market clearly do not pay that price.

You cannot fix this with better surveys. The information needed to accurately price specific GPU hours in a particular facility only exists in the minds of the people operating that hardware, and they have good reason to keep it private. Moreover, computing power is a time-sensitive commodity. The GPU hours that no one buys in that instant simply disappear. Any pricing signal for GPU computing power must operate in real-time; otherwise, the numbers are outdated by the time you publish them.

The only way to gain this knowledge is through real transactions at true scale, with counterparties betting real money. This brings us to a solution that has also evolved from electricity: a network rooted in Bitcoin, currently with 320,000 GPUs pointing at it.

What if you could figure out the price?

Bitcoin miners run the SHA-256 hash. This computation exists for one purpose: to prove that miners have burned a certain amount of electricity. Once that proof is recorded in the block, the hash or the work done has no value. It cannot be reused, repurposed, or sold to anyone outside the blockchain. The work done to mine one Bitcoin block is not interchangeable.

The problem is that thousands of miners do this work for each block, but the network only accepts one block from one miner. The rest of the work gets discarded. Each SHA-256 hash that was mined but not accepted by the network represents a GPU cycle that could have been used for something commercially valuable, and in 2026, "commercially valuable GPU work" only means one thing: AI computing power.

Pearl is a Layer 1 blockchain that attempts to give GPU computing power commercial value. It inherits the same proof-of-work consensus and difficulty adjustment that has secured Bitcoin since 2009. But the miners in Pearl are not running SHA-256 hashes; they are performing matrix multiplication on GPUs.

Why is matrix multiplication valuable? Because it underpins all the answers you get when you ask ChatGPT. The same is true when Nano Banana renders images from text prompts, when companies fine-tune language models using proprietary data, or when self-driving cars process camera footage 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 has its computational core running on matrix multiplication. Pearl incentivizes miners to perform the same calculations as proof of work.

What about security? That’s the main job of the hash function. You cannot directly substitute matrix multiplication for the hash function.

SHA-256 has no shortcuts. Without performing the complete computation, you cannot produce output. The existence of output means the necessary work has been done. Matrix multiplication, on the other hand, is different. You are essentially multiplying and adding strings of numbers. Therefore, a pattern can allow miners to compute answers with minimal effort, undermining the effectiveness of proof of work.

Pearl employs a mechanism called NoisyGEMM to ensure miners remain honest. This is done in three steps:

Before starting any work, miners commit to the input matrix. Once submitted, the inputs cannot be changed.

The protocol then adds random noise to the committed matrix. This eliminates any symmetry (shortcuts) the miner might plan. And because the miner does not know what the noise will be, they must perform the work later.

Miners run a full matrix multiplication on the noised version, processing in chunks on the GPU.

Each output chunk is hashed and checked against the network's difficulty target.

If a chunk's hash falls below the target, the miner earns the block reward.

But introducing noise prevents miners from earning rewards without actually doing work. However, this noise also means we do not get answers to the original workload. Pearl's solution is to give the noise a known structure. This allows it to strip the noise away after miners complete the multiplication, incurring minimal and inexpensive cleanup costs.

When a GPU mines Pearl, the network measures how many times that machine has successfully completed matrix multiplication under the existing power, cooling, and network conditions within the 194-second block interval of the protocol.

An H100 located in a well-cooled enterprise facility may complete more work per block than the same chip placed in someone's garage, connected via consumer-grade Ethernet with subpar airflow. Pearl automatically captures this difference, which any price index averaging by model cannot do.

Now scale this to 320,000 GPUs across dozens of geographical regions, and total computing power becomes the largest continuous benchmark for real-world AI computing power ever, updated every 194 seconds, without anyone needing to conduct surveys or publish rate cards.

Computing power is also tracked by configuration. Every GPU among the 320,000 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 the true demand for GPU computing power surges somewhere in the world, operators pull machines off Pearl to service paying customers, decreasing computing power.

When demand softens and rental income declines, these GPUs flow back, increasing computing power. The network’s difficulty adjustment reacts to these migrations in a manner similar to how Bitcoin's difficulty has responded to changes in mining economics since 2009.

The difference is that Bitcoin miners weigh the electricity costs they will incur against the profits they can earn from selling mined Bitcoin. During the Texas heatwave in August 2023, Riot Platforms shut down most of its mining rigs and received $31.7 million in energy credits, more than three times the value of the 333 Bitcoins mined that month. When enough miners make this choice, computing power decreases, and difficulty is readjusted every 2016 blocks to ensure blocks are produced approximately every 10 minutes.

This flow of GPUs between Pearl and the rental market creates a balanced price for GPU computing power that continues to update. High and rising computing power means computers are cheap on the open market because operators find that mining Pearl is more profitable than renting.

A decrease in computing power means someone is beginning to pay a sufficiently high price to pull machines away. And that price signal is derived from revealed preference, from thousands of GPU operators choosing what to do with their own hardware and their own money. It does not come from the cloud provider marketing teams deciding what to print on rate cards, which is all the Bloomberg index and CME settlement mechanisms have ever had access to.

Together AI is the first to commercialize based on this. On May 15, 2026, it announced a reasoning endpoint running on the Pearl mining network, offering Gemma-4-31B-it-Pearl at over 25% below standard cloud rates. The economics hold up because the GPUs providing reasoning through the Together endpoint are simultaneously mining Pearl, so the mining income offsets part of the operating costs, allowing Together to price below the broader market without having to subsidize the difference from their own profits.

Most of the demand for this level of service from paid reasoning customers has not yet been met on the Pearl network, and bridging this gap is clearly a challenge ahead. But pricing does not depend on it. It relies on the opportunity cost being real, and that opportunity cost exists for every one of these 320,000 machines and for several others in the reasoning market.

The GPU computing power market is a multi-billion dollar annual scale market with no reliable prices. The Bloomberg index attempts to track it, but it cannot provide accurate direction in any given week. The CME plans to launch GPU futures on October 5, 2026, settling in cash against the prices that large buyers clearly 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 lease capacity in the open market, is generating something that is the closest we've ever come to an accurate price for the GPU computing power industry.