TechFlow Logo
Login/ Sign up
ETH Gas
Gwei
Fear
gas
Pearl Runs AI Inference on Bitcoin Computing Power: The Story of Doubling Miners’ Revenue Has Only Just Begun

Pearl Runs AI Inference on Bitcoin Computing Power: The Story of Doubling Miners’ Revenue Has Only Just Begun

2026.10.13
Share
Pearl is making compute useful. Considering the growth and eventual widespread adoption of AI, its target market is highly attractive.

Author: Castle Labs

Compiled by: TechFlow

TechFlow Summary: The insatiable demand for compute from AI is pushing Bitcoin miners to an awkward crossroads: their power and data centers are becoming increasingly valuable, but the hash rates generated by mining itself are useless. Pearl attempts to merge these two functions, allowing miners to run AI inference while maintaining the network. This narrative has already driven the token price up threefold, but whether actual adoption can catch up is what will truly determine if it’s “the missing piece of the AI puzzle.”

Artificial intelligence is a critical technology. As its scale of usage continues to expand, its demand for compute throughput is rising sharply.

To deliver this throughput, AI companies need GPUs and energy. But even well-funded firms may struggle to secure sufficient resources due to limited GPU supply and scarce electricity.

A common solution is to lease infrastructure.

In recent years, Bitcoin miners have built out industrial-grade mining infrastructure and locked in the most critical resource: power access. With growing AI compute demand, this advantage is being re-priced by the market. Although these machines run ASICs and cannot be directly used for AI computation, miners can rent out their facilities and power to AI clients.

This gives miners the opportunity to establish a more sustainable revenue stream by putting their power to work on AI-related computations. Mining companies' revenue structures have already reflected this shift: among 11 publicly listed mining firms, the proportion of total revenue from AI and HPC has risen from 2% in Q1 2025 to 22% in Q2 2026.

As a result, Bitcoin hash rate may face greater competition from AI inference workloads.

An alternative way to observe the growing relationship between AI and miner infrastructure is to look at how BTC is mined. Under the Proof-of-Work (PoW) mechanism, miners deploy massive computing power to secure the Bitcoin network and validate transactions. This process requires continuously computing cryptographic hashes, which hold no value once computed. Mining is only economically viable when the value of the mined BTC exceeds the computational and infrastructure costs.

So why not have both?

If computation can run AI inference while continuing to secure the network and mint assets, such an approach is known as “Proof-of-Useful Work (PoUW).” The Pearl network, which we will discuss in detail next, focuses squarely on the intersection of AI adoption and blockchain.

What is the Pearl Network?

Pearl is an L1 blockchain that launched in April this year. Forked from Bitcoin's codebase, it replaces the SHA-256 hashing algorithm with matrix multiplication (matmul), which is the type of operation executed by GPUs during AI inference.

This is exactly what Pearl is doing: transforming Bitcoin's Proof-of-Work into compute useful for AI (PoUW).

To integrate AI inference into mining, Pearl has designed a system that replaces Bitcoin's standard SHA-256 hash verification with matrix multiplication verification.

Matrix multiplication is how AI models are trained and inference is run, and it demands significant compute power.

Matrix multiplication shares many characteristics with SHA-256:

Rapidly solving it requires heavy computation

The algorithmic outcome is not based on subjective judgment

Answers cannot be forged or guessed (you cannot predict in advance which function will be used for the next block)

Correspondingly, these operations must still be “expensive enough to execute, but objective enough that answers can be verified cheaply by others.” This is precisely why Pearl utilizes zero-knowledge (ZK) proofs, making verification inexpensive.

To tap into AI-related compute demand, Pearl partners with aggregate compute networks and inference service providers like Together AI. This allows enterprises to utilize the compute output from miners, who earn PRL tokens in return. Meanwhile, Together AI also operates its own miner infrastructure on the network and uses earned PRL rewards to offer discounts to clients.

Ultimately, Pearl aims to replace the calculations BTC miners perform to mine BTC with the more useful matrix multiplication routinely used by AI, prioritizing AI while preserving the core blockchain economic model.

However, the "useful" aspect of PoUW warrants further discussion. One criticism of Pearl is that the protocol cannot prove these matrix multiplications actually come from paying customers or useful tasks, since miners can input synthetic matrices and still claim PRL rewards.

In fact, a June paper found that results from mainstream mining software at the time consisted of “verifiable matrix multiplications, but with no external inference output.” Brian Breslow also highlighted this point in his investment thesis for PRL.

This is the crux of the matter: avoiding subjective judgments on the "usefulness" of the inference. Bittensor previously implemented a similar system:

Scarcity via the native token

Miners provide AI-related resources

Validators assess output quality

Rewards flow to where the network deems valuable

Yet this design introduces subjective judgment and greater discretion. There have been precedents of collusion because “validators are incentivized to collude on voting for usefulness.”

In contrast, Pearl's AI inference involves no subjective judgment.

This is crucial for adoption: the entire value proposition rests on the premise that "submitting two matrices for AI inference" is simple enough to incur virtually no overhead (the received PRL value must exceed integration friction, risk, throughput loss, and overhead).

Miners still must cover the full GPU cost to earn PRL. As difficulty rises, only miners with the highest-end GPUs will remain profitable, as they possess the strongest mining capability.

If enough "useful miners" join the network, others will no longer profit and will have no incentive to continue mining.

Avoiding factors that could introduce subjectivity helps prevent miner collusion issues seen in other networks like Bittensor.

But there is a caveat: currently, due to the network using simplified numerical formats (INT8), the overhead sits at around 10-15%. Only when this overhead drops sufficiently low will useful miners become the most efficient block producers. Efforts are already underway to address this: Pearl-GEMM has launched on testnet, reducing performance loss to just 0-3% when processing standard formats.

PRL: Tokenomics and Future Value Accrual

The PRL token has a hard cap of 2.1 billion coins. Unlike Bitcoin's halving mechanism, PRL emissions follow a linear decay curve, with approximately 50% of the total supply expected to be released by year four.

As of this writing, PRL's market cap stands at $475 million, with 339 million PRL mined (16.2% of the circulating supply). Approximately 305 million additional PRL are expected to enter circulation over the next 12 months, nearly doubling the circulating supply.

Currently, PRL is listed on two exchanges (BigONE and SafeTrade), with daily trading volume around $6 million and bid depth of approximately $22,000 within a 2% price range.

Despite this, PRL's repricing appears to have outpaced actual utilization. Since September, PRL has surged more than threefold. It gained the most traction after Brian Breslow of No Limit Holdings published an investment thesis.

Furthermore, PRL is currently undergoing significant dilution. Pearl lacks a halving mechanism, distributing roughly 1.01 million PRL to miners daily, valued at approximately $1.4 million at the time of writing. More critically, PRL supply issuance deviated from plan: in the first 6 days of lowest mining difficulty, roughly 122 million were issued (representing 36% of the current circulating supply), and 20,000 blocks were mined within 30 hours of launch. Block times have since stabilized near the 194-second target. This early first-mover advantage caused unnecessary dilution, rewarding miners in the first week and exacerbating supply concentration.

The token itself carries no inherent yield. Miners earn PRL, provide AI inference services, and generate stablecoin revenue, but PRL lacks a burn mechanism and supply will not contract.

Nonetheless, the team is actively addressing current challenges, including performance overhead, synthetic mining, and token value accrual:

Pearl-GEMM Upgrade: The currently running Pearl version only processes simplified numerical formats (INT8), which most compatible AI models do not utilize, causing model speed to drop by 10-15%. At this level of overhead, AI companies have little incentive to procure compute from miners. However, a solution is live on testnet: Pearl-GEMM handles standard formats with only a 0-3% performance penalty. This upgrade is critical, as AI operators would not tolerate sacrificing 10-15% of output.

Real AI Work Must Replace Synthetic Mining: Not all mining constitutes useful AI inference. A June study examining roughly 8,000 machines found they ran on inference-capable hardware without executing any AI-related code. Pearl's vision of a decentralized compute market could solve this through partnerships with AI enterprises: users pay and get compensated in PRL to run inference on the network. This includes reserving GPU capacity, fine-tuning models, or settling compute contracts without relying on traditional cloud solutions.

AI Usage Should Create Buy Pressure and Potential Burn Mechanisms: Currently, AI inference usage generates sell pressure. US cloud provider Together AI launched a model in May 2026 offering a 25% discount on access, funded by their PRL mining activity to meet compute demand. The team is considering pricing AI services in PRL or burning a portion of PRL with each purchase to move away from USD-denominated pricing and drive demand for the asset.

Conclusion

Pearl is making compute useful. Considering AI's growth trajectory and eventual widespread adoption, its target market is highly attractive.

Building a novel network inherently comes with challenges. For Pearl, these include high overhead, synthetic mining, and token value accrual mechanisms—all of which the team is systematically tackling.

Critical to note, Pearl currently prices model services in USD, meaning minimal value flows back to the token. Mandating PRL usage will be fundamental to driving token value and improving demand capture. Much of the minted PRL serves as incentives, which miners may ultimately dump for profit.

Additionally, PRL is slated to become the utility token for the decentralized marketplace envisioned by the team, making traffic and volume forecasts difficult. At this stage, direct comparisons with BTC's price metrics (e.g., Pearl reaching 1% of BTC's market cap) amount more to speculation than established fact.

Nevertheless, we believe this represents a compelling use case for BTC miners already selling AI inference, capable of capturing a significant share of miner hashrate, particularly given the future growth potential of the AI inference market.