Ryzed.hl

AI spends hundreds of billions on GPUs that mostly multiply matrices. Bitcoin spends its entire security budget on hashes nobody can reuse. @prlnet makes them the same computation, and could become the Bitcoin of the AI age. Finally longed pearl-2:native in the recent (smol) dip via @entropyIO on Hyperliquid. Here's why ↓ Proof of useful work has been an open problem since Bitcoin launched. A key challenge is that in a permissionless chain, the miner picks its own inputs. Say the work is "multiply two matrices." A cheater just submits all-zero matrices, does no real compute, and farms tickets for free. So earlier designs mostly had two bad options: let someone else choose the inputs (not permissionless, and miners can't bring their own AI jobs), or pick some niche task nobody actually needs (e.g. Primecoin hunting prime chains). The breakthrough? Komargodski and Weinstein (Pearl's founders) published a PoUW mechanism for arbitrary matrix multiplication with 1+o(1) overhead in April 2025: - Miner commits to A and B, the chain seeds low-rank noise into both - The GPU runs its normal kernel on the noised operands - Every output tile doubles as a lottery ticket, hashed against the difficulty target - A winning tile gets a Plonky2 zkSNARK that nodes verify cheaply - The noise is low-rank, so the miner strips it and keeps A·B at FP8-level accuracy One forward pass, two outputs: the AI result and a shot at the block. Why can't miners cheat? The noise is seeded by the chain itself, so nothing can be precomputed, and a jackpot policy rejects "easy" crafted inputs. The only way to get a ticket is to actually run the full multiplication. And the team keeps pushing it further. The September whitepaper: - Moves mining to FP8, the low-precision format modern inference runs on across Nvidia's Hopper and Blackwell GPUs - Extends it to mixture-of-experts models, the architecture behind most large open models today - Uses Hawkeye, which replays the GPU's exact math on a normal CPU so anyone can check a proof bit for bit (co-authored by Dan Boneh, accepted at MLSys 2026) "But if the work is useful, can't an attacker sell the output and attack for free?" That's the classic objection, and Rafael Pass (Cornell Tech cryptographer, known for the first formal security proofs of Bitcoin's consensus) took it head on. His model shows that once the market adjusts, a majority attack still costs the full block reward, same as Bitcoin. What changes is where the reward ends up: instead of paying for wasted hashes, it subsidises inference. Same security, but cheaper tokens. Everything else is plain Bitcoin: UTXO, longest chain, Taproot, a 2.1B hard cap and a smooth emission curve instead of halvings. Plus a post-quantum path, since the mining is hash-based and XMSS signatures are already in consensus. So where's the demand? Most people still point to the Together AI endpoint (Together is one of the biggest inference providers for open models), which was pulled from serverless on the 27th of August. The more interesting signal tho imo is that Pearl now runs its own inference platform: 7 open models behind an OpenAI-compatible API, served on GPU capacity across the Pearl network. And it's cheaper. On all four models it shares with Together's serverless lineup (input / output per 1M tokens): - DeepSeek V4 Pro: $1.10 / $3.30 on Pearl vs $1.32 / $3.96 on Together - GLM-5.3: $1.10 / $3.50 vs $1.40 / $4.40 - GLM-5.3 Flash: $0.07 / $0.24 vs $0.15 / $0.50 - DeepSeek V4.1 Flash: $0.14 / $0.42 vs $0.30 / $1.20 Still worth keeping in mind tho that this is launch pricing, and Pearl doesn't say how much of the discount PRL is funding. But it's the direction the Pass model predicts: block rewards flowing back to users as cheaper inference. Ok, cool but how and where does this thing trade? At $1.70 and a ~$594M mcap, PRL is ~19% off its $2.10 ATH from the 14th of October rn. Spot is basically two small CEXs, BigONE and SafeTrade, with ~$5M of combined daily volume. The books are paper thin: per CoinGecko it takes ~$4K of buys to push BigONE up 2%, and ~$9K on SafeTrade. That's why the perp matters. @entropyIO listed PRL as a HIP-3 market on Hyperliquid on the 14th of Oct, and it's filling up fast: - ~$3.05M OI, already 61% of Entropy's $5M OI cap for the market - ~$2.05M daily volume, around 40% of all spot volume - Funding averaged ~0.03-0.04% an hour over the last 3 days, longs paying - Lighter runs a smaller PRL perp too, ~$0.56M OI There is a catch tho: Until a properly liquid venue lists PRL, io:PRL prices 90% off SafeTrade, so the perp inherits the spot book's fragility. Why am I long? Bc the hard part is (mostly) done. The math is public, the verification work is accepted at MLSys, the chain is live, and Pearl already sells inference cheaper than Together AI, one of the largest providers of open-model APIs, while awareness of what they're actually do still seems low across CT / my timeline. What's missing tho is proof that paid demand actually runs through the mining. Until Pearl publishes the share of mining that serves paid workloads and unless there is real adoption, PRL will trade like a mining coin and has to absorb ~1M new PRL a day (float +87% over the next year), and two thin spot books. The trade is definitely not without risk. That above gap is also the discount I see rn tho, and it's why I bought into this (still smol compared to recent run) dip. Bitcoin turned electricity into security. And if Pearl succeeds at turning GPU cycles into security and inference at the same time, it could indeed be smth like the "Bitcoin of the AI age", and with that, nothing short of huge. NFA ofc tho. Always DYOR. Once you have and want to trade pearl-2:native too, join me on Entropy (my ref link): entropy.io/?r=logix