Castle Labs Puts Bitcoin (BTC) Miners' AI Revenue Share at 22%

From 2% to 22%

Artificial intelligence and high-performance computing produced 22% of revenue at the eleven publicly listed Bitcoin (BTC) mining companies covered in an analysis Castle Labs released on Tuesday, October 13, up from 2% in the first quarter of 2025. The figure is a revenue mix for those eleven firms alone, not a read on the mining industry as a whole, a limit the research firm writes into the report itself, and the sample is the same eleven companies at both endpoints of the comparison. It describes a shift in what these companies sell rather than a one-off gain: electricity and data center space once dedicated to hashing are being routed toward AI computation instead. Revenue mix is the operative phrase here, since the percentages describe proportions of each firm's sales, not totals, and hashing income still exists alongside the newer streams. High-performance computing, in this setting, means selling capacity for workloads such as AI inference, where matrix operations dominate the compute profile.

The assets behind the number explain the direction. Listed Bitcoin (BTC) miners have spent years securing large power allocations and erecting data center buildings, and Castle Labs reads that power access and infrastructure as a new income source in the AI computing market, where operators need floor space, cooling and grid interconnection as much as they need machines. Data centers standing next to power facilities are the physical shape of that advantage, because proximity to generation decides how much capacity can be served at all. Five quarters of data stand behind the 22% line, so the reallocation is gradual rather than a sudden pivot, and the baseline quarter predates most of the AI data center buildout. The Bitcoin (BTC) price still sets what each rig earns from hashing, while AI and HPC income follows its own demand cycle.

Pearl Replaces Hashing With Matrix Math

Pearl, the protocol the second half of the analysis examines, is a layer-1 blockchain forked from the Bitcoin (BTC) codebase and launched in April 2026. Its consensus mechanism drops the SHA-256 hashing that secures proof-of-work networks and substitutes matrix multiplication, the operation type behind AI inference. Results are verified with zero-knowledge proofs, a layer that lets the network confirm a computation ran correctly, and miners are paid in PRL for supplying the work. For a miner, the design converts hash capacity into an income stream paid in a new token rather than in Bitcoin, which is where the revenue-mix discussion and the protocol discussion meet.

Emission follows its own curve. PRL's maximum supply is 2.1 billion tokens, and issuance declines linearly instead of stepping down through the periodic halving written into Bitcoin's schedule. At the time of the analysis, roughly 1.01 million PRL reached miners each day, and about 122 million tokens were issued across the first six days after launch. Castle Labs attaches two checks to the design. Its June research found that about 8,000 machines equipped with inference-supporting hardware were running no AI-related code, a direct test of whether such fleets end up serving real workloads. Efficiency is the second: the network's simplified INT8 format currently carries about 10-15% of overhead, which weakens the incentive for AI operators to choose a miner's compute resources. Pearl-GEMM, the implementation running on testnet, aims to process the standard format with a performance loss of 0-3%, a target Castle Labs keeps separate from any measure of actual customer demand or service utilization. The firm's own closing line draws the same boundary: a computation style that resembles AI operations does not, by itself, prove that real AI services or customer demand exist.

What the 22% Does Not Yet Prove

The 22% share is evidence about power assets and revenue composition, and the same report's caveats mark what it is not evidence about. A mix drawn from eleven listed companies does not extend to the wider mining economy, and a compute design that resembles AI work does not create AI customers. The idle-rig count from June and the 10-15% INT8 overhead stand between infrastructure and utilization, and the compute-for-token model long present across the altcoin market now sits inside a fork of Bitcoin's code, territory Bitcoin maximalism has historically opposed. The figures to watch are whether Pearl-GEMM's 0-3% target survives mainnet and whether inference workloads arrive at all; until then, the line between revenue earned and demand proven stays where Castle Labs drew it.