Pearl Draws Attention at Token2049 as the 'AI Version of Bitcoin'

Article by Xiao Bing

Walking around Token2049, you'll notice a name appearing with an unusually high frequency: Pearl.

Not a new L2, not an AI agent platform—it’s a mining coin. Whether Chinese KOLs or overseas professionals and KOLs, everyone is being repeatedly exposed to this name. What’s its secret?

Bitcoin miners consume electricity to compute hashes, which are discarded after calculation and serve no purpose other than block creation. Pearl says: Since miners are going to use electricity anyway, why not have them compute something useful?

This is the core logic of Pearl's consensus mechanism, Proof-of-Useful-Work.

Pearl's miners do not compute hashes; they perform matrix multiplications. Matrix multiplication is the fundamental operation underlying large language model inference, meaning that while mining, miners are theoretically helping AI run inference.

The publicly known founder, Omri Weinstein, holds a Ph.D. in Mathematics from Princeton University and is a professor of computational complexity theory at Columbia University. The mining mechanism is called NoisyGEMM: miners perform matrix multiplication using GPUs, generate commitment values with BLAKE3, and verify results using Plonky2 zero-knowledge proofs. The protocol is forked from Bitcoin’s btcd, with a maximum supply of 2.1 billion coins, no halving events, and a gradually decreasing emission rate.

Claiming a completely fair launch with no pre-mine, no founder allocation, and no VC round is extremely rare for new projects in 2026.

On October 7, No Limit Holdings sponsored the first Pearl Connect event, where bullish supporters of Pearl primarily cited three key reasons.

The first layer is "AI Bitcoin."

PRL is a sovereign-free digital currency backed by computational power and energy efficiency. The absence of cash flow is actually an advantage, as Bitcoin also lacked it in its early stages. Early-stage inflation is an inevitable phase for all PoW coins—surviving it leads to monetary premium. By this logic, the current FDV of under $3 billion is undervalued, with long-term potential resting on AI-native value storage, computational power settlement, and even Agent-based payments.

The second layer is the theory of energy migration.

Bitcoin mining companies are subleasing their facilities to AI data centers, and publicly traded mining firms reduced their hash rate by approximately 75 EH/s in the first half of the year. Bitcoin can only compete with AI for electricity, while Pearl "expands alongside AI adoption." When the entire industry's power and GPU resources are shifting toward AI, a PoW chain that moves in the same direction as AI naturally occupies a superior ecological position.

The third layer is the “2-for-1” unit economics: the same GPU computation generates both inference revenue and PRL mining rewards. The mining income acts as a cashback on inference services, driving inference prices lower. Lower prices stimulate greater demand, which in turn requires more computing power, and more computing power produces even more PRL.

Stack these three layers together, and the bullish conclusion is: Pearl is not just another mining coin—it’s the Bitcoin of the hashpower economy.

Pearl's narrative is that "miners' computational power is performing useful AI inference," but the on-chain consensus protocol cannot actually distinguish whether miners are performing genuine AI inference tasks or simply generating random matrices.

Pearl's own GitHub clearly states: mining is a "byproduct of matrix multiplication," with the underlying mechanism called cuPOW, which proves "computational difficulty," not "computational utility." The consensus layer only sees that "you performed a matrix multiplication," but it cannot and does not know whether that multiplication was used to run Gemma-4 inference for Together AI or whether the miner simply generated a random set of numbers and computed them for no purpose.

This issue is technically called a "job-binding failure": the consensus mechanism cannot bind the on-chain proof of work to the off-chain real reasoning task.

Worse still, computing random matrices is cheaper than performing actual inference tasks. Real inference requires loading model weights into GPU memory, incurring significant I/O overhead, while random matrices require no loading at all—they can be computed directly. This creates a natural economic incentive for “garbage mining.”

Academic papers have already provided empirical evidence. An analysis of 8,012 mining nodes found that no inference code exists in mainstream mining software. Researchers submitted 44 shares generated by random matrices, and all were accepted by the network. This means that, at least under the current protocol version, it is possible to mine normally and receive rewards without performing any useful work.

The team’s response was candid: core developers acknowledged that “there is no on-chain metric to measure the proportion of useful mining,” and the founder admitted that “allowing miners to choose their own matrix inevitably permits spam mining.” Their argument is that, in future protocol upgrades, when the cost of solo mining far exceeds the cost of performing inference as a byproduct, the economic incentive for spam mining will approach zero.

On May 15, Pearl announced a partnership with the AI inference platform Together AI to run Gemma-4 model inference using Pearl's network compute power, at prices more than 25% below market rates.

Bullish investors see this as the commercialization of "useful work," while bearish investors question whether Together AI is performing inference on its own servers and subsidizing prices with PRL tokens, without actually utilizing miners' GPU compute power for Together AI's inference tasks.

In other words, this is more of a marketing subsidy than a hashpower transaction—AI customers pay in USD, not PRL.

The Pearl mainnet launched on April 27, with approximately 450 million PRL in circulation, representing about 16% of the total supply and a fully diluted valuation of $2.8 billion.

Approximately $1.35 to $1.4 million in new token supply is added daily, with daily trading volumes ranging between $2 million and $6 million. Currently, the token is listed on only a few small exchanges and has an OTC market.

Pearl has tapped into two of the strongest narratives of 2026: AI and fair launch.

In an era when VC coins are under fire and AI is at its peak, "AI mining tokens without VC backing" represents a textbook narrative positioning—their popularity at Token2049 is no accident.

But there is a gap between the narrative and the mechanism. Pearl speaks of "useful proof of work," but the protocol itself cannot verify that the work is useful.

Therefore, Pearl can also be understood as an option—you're betting on the future possibility that GPU mining and AI inference can truly merge. If protocol upgrades resolve the issue of spam mining, and if AI inference demand genuinely begins settling transactions in PRL—not just using PRL as a subsidy—this option has substantial and highly attractive upside potential.