Who was in the room?
The CEO said a lab spending $100–130M a day on inference was in the audience. That's a club of about five companies. We don't know which one.
Identify an attendeeTHE CHRONOLOGY / A PUBLIC RECORD2025 — 2026
FOLLOW THE THREAD THROUGH TIME
Nothing happens
all at once.
A paper. A working protocol. A difficult test. A new direction.
Read the events in sequence—and see where the story is still open.
Dates belong to the event described. Source labels distinguish public records, team announcements, recorded statements and our interpretation. An announcement is not evidence of delivery.
How we know:On tapetheir own recorded wordsPublic recordGitHub, arXiv, HuggingFace, DNSOfficial postPearl's own announcementOur readclearly-labeled inference
2025 — MAY 2026
Before the noise.
The research arrives first. Then the people, the code and the first model checkpoints.
Apr2025
The paper that starts everythingPublic recordResearch
Pearl-GEMM / proof-of-useful-work lands on arXiv (2504.09971) — PoW on arbitrary matrix multiply A×B with vanishing overhead, miner chooses (A,B).
Apr2025
NVIDIA Inception bars crypto startupsPublic recordBusiness
As of April 2025, cryptocurrency companies are on Inception's published list of excluded business types. Remember this date.
Feb2026
First engineer on the benchPublic recordThe bench
Earliest dated bench member in public commit history (GPU kernel engineer). The first Pearl model checkpoint — Llama-3.1-8B-Instruct-pearl — is published the same month.
Mar2026
“Initial Pearl implementation”Public recordThe bench
The infra/DevOps engineer's very first commit message. The repo that becomes the chain starts here.
Apr2026
The bench doublesPublic recordThe bench
Protocol and mining-systems engineers join. Two more model checkpoints published (Gemma-4-31B, Llama-3.3-70B).
May2026
ZK and client engineers joinPublic recordThe bench
Cryptography/ZK and desktop-client hires. Six of fifteen bench members have dated starts; all six fall in Feb–May 2026.
MLSys2026
Hawkeye presented at MLSys 2026Public recordResearch
Reproducing GPU-Level Non-Determinism — Research-Track oral plus poster. Authors include Stanford's Dan Boneh and co-founder Ilan Komargodski. The paper never says “blockchain.”
JUNE — JULY 2026
Into the open.
The protocol evolves in public. Papers, releases and recorded talks begin to tell a larger story.
Jun 82026
Rafael Pass paper on PoUW economicsOfficial postResearch
Answers the “if miners already get paid for AI work, does security collapse?” objection.
Jun 112026
MoE checkpoint publishedPublic recordProtocol
Qwen3-30B-A3B-Instruct-2507-pearl — a real mixture-of-experts checkpoint — published to HuggingFace.
Jun 122026
The MoE hard fork shipsOfficial postProtocol
Network moves v1 (generic matmul) → v2 (MoE certificates), governed by PIP-2. Pearl can now mine mixture-of-experts models.
Jun 162026
Vision support addedPublic recordProtocol
Ongoing model work — real commits, not a launch dump.
Jun 182026
BLAKE3 GPU-kernel writeupOfficial postResearch
Engineering blog on their hashing kernel implementation.
Jun 282026
Benchmark results publishedPublic recordProtocol
Added to the model cards.
Jun 292026
“Money, rethought”On tapeResearch
First of three teaser clips cut from the keynote, released ahead of the full talk.
Jul 12026
“They were wrong”On tapeResearch
Omri on PoUW being repeatedly declared impossible — including Vitalik's “probably infeasible.”
Jul 22026
Pearl-GEMM explainedOn tapeResearch
The low-rank-noise trick, in his own words. Same day: congratulations to partner Together AI on an $800M Series C.
Jul 62026
Rafael Pass on the economicsOn tapeResearch
Full 34-minute talk posted — the security-collapse objection, answered on tape.
Jul 92026
The keynote dropsOn tapeResearch
“The Science Behind Pearl” — the full ~40-minute talk, and the source of every teaser clip. Also the tape where the CEO describes his own invite-only audience.
Jul 142026
Inference cloud + mining rewardsOfficial postBusiness
Discord announcement on the compute marketplace direction.
Jul 152026
NVIDIA Inception — and “Indeed”Official postBusiness
Pearl announces Inception membership. A crypto commentator notes Inception bars crypto and that Pearl is “functionally an AI inference company.” Pearl's verified account replies: “Indeed.” The reply supports the AI-facing description; the admission rationale is not established.
Jul 162026
Floating-point roadmap goes publicOfficial postProtocol
Komargodski: next-gen PoUW will support FP natively — FP8 first, FP4 to follow — described at the time as removing quantization. The later September explanation still includes randomized quantization; Hawkeye and deployment are separate milestones.
Jul 192026
Wallet v2.0.3Official postProtocol
Routine client release.
Jul 202026
The W7A7 quantization formatOfficial postProtocol
Komargodski: PoUW forced them to invent a quantization format that didn't exist — GPTQ + SmoothQuant fused, claimed to beat FP8 baselines. Paper promised.
Jul 212026
The roadmap, in their own wordsOfficial postProtocol
A Pearl Discord roadmap post splits the work in two. AI: inference optimisations, floating-point support, and “exploring distributed inference architectures.” Crypto: exploring L2 and stablecoin frameworks, and building relationships with qualified custodians to support institutional adoption.
Jul 212026
An L2 is on the tableOfficial postProtocol
Same day: “the current structure of the blockchain is not a good fit for agentic payments, or even regular payments, so it makes sense to develop a faster L2.” Open questions posted to the community include whether to reuse Bitcoin-L2 architectures, whether it needs smart contracts, whether it can mint other tokens, whether Pearl needs a native stablecoin, and whether to issue an asset pegged to the value of a GPU.
Jul 212026
Inside a closed NVIDIA eventOfficial postBusiness
A photo posted to Discord puts CEO Omri Weinstein on stage with Prof. Freddy Gabbay — Hebrew University, formerly VP of Chip Design at Mellanox, the networking-silicon company NVIDIA acquired — discussing “the intelligence economy” at a closed NVIDIA event.
Jul 212026
On the circuitOfficial postBusiness
The same roadmap post lists where the team has been showing up: the RAISE AI summit in Paris (attended), AMD Advancing AI in San Francisco (that week), and GTC among others — plus “a strong network of advisors” producing meetings with senior AI-company executives. Their own caveat: they can't discuss what was said in those meetings.
Jul 212026
Stable diffusion as a 2-for-1 use caseOfficial postResearch
A brief note that they're working on stable diffusion, called “a great use case for the 2-for-1 model” — image generation as mineable work, not just LLM inference.
JULY — AUGUST 2026
The stress test.
An exploit interrupts the momentum. The response and post-mortem become part of the record.
Jul 242026
The MoE mining exploitOfficial postProtocol
Official statement: an exploit in the MoE mining mechanism let certain miners generate disproportionate hashrate. Mitigated by a soft fork that temporarily disables MoE mining while a permanent fix is built. A technical report was promised for the following week.
Jul 262026
The MoE post-mortemOfficial postProtocol
The promised report lands. Pearl's own account: the soft fork went in at noon on 23 July, hours after the vulnerability was identified that morning with community help. MoE proofs ran at 28% of newly mined blocks across the affected period and peaked at 73% on fork day; 4.8M coins were minted with the vulnerable construction, ~0.23% of total supply. Root cause: sharing the commitment and noising of the activation matrix across experts to cut overhead introduced a uniqueness condition — each token routed at most once per expert — that the verifier never enforced. Pearl reports no reorg, double-spend, theft, forged transactions or key compromise, and no change to the emission schedule; the stated harm was reward dilution for honest miners. MoE mining stays off pending a revised construction and external review, with a bug-bounty programme promised.
Aug 312026
Omri keynotes ESA 2026Public recordResearch
Omri Weinstein gives an invited keynote at the European Symposium on Algorithms (ALGO 2026, L'Aquila): “Proofs of Useful Work from Arbitrary Matrix Multiplication” — the only consensus/PoW talk among the invited keynotes.
SEPTEMBER 2026
The next proposition.
Floating-point work and verification enter the foreground. The dates matter; so do the qualifications.
Sep 142026
The FP8 floating-point scheme goes publicOfficial postProtocol
Pearl publishes its FP8 proof-of-useful-work scheme for NVIDIA Blackwell — whitepaper (“Pearl Floating Point Scheme Specification”) plus code (GitHub PR #311), with a claimed near-zero overhead over optimized vLLM. Released for community review before forking to mainnet — the Hawkeye/Badash reproducibility work shipping as product.
Sep 142026
Partnership with attestableOfficial postBusiness
Pearl announces a partnership with @attestable (a $20M-seed ZK-verification startup): combine attestable's zero-knowledge proofs with Pearl's proof-of-useful-work to prove, with cryptographic guarantees, that an entity is running the computation it claims — a “trust layer for AI,” framed for AI-pacing and compute treaties.
Sep 152026
The community AMAOn tapeBusiness
A one-hour AMA with the founders. Headlines: Pearl is being built as a store of value (“unapologetically”); FP4 is next (no date); FP8 is not yet forked to mainnet and no date was set; and the silent third co-founder Idan Sugarman appears in this archive’s September recording. Full transcript in the reading room.
Sep 172026
“Fastest FP8 MoE kernels for Blackwell”Official postResearch
Omri: Pearl-GEMM is ~5% faster than Quack (Tri Dao) and ~42% faster than FlashInfer on B200s. Self-reported; not independently verified.
THE UNFINISHED FILE / FOUR OPEN QUESTIONS
The story still
has missing pieces.
These questions remain open in the dated record. A source, a name, a public announcement could move the investigation forward.
Who are the two providers?
At the Sept AMA the team said it's “in close touch with two other major inference providers,” and knows of “several neolabs already mining Pearl” — none named.
Name a counterpartyWhat's the institutional-access product?
Idan teased an unannounced way to “democratize access… to institutional investors” — the on-tape match for the Anchorage & Coinbase RPC endpoints already on the board.
Tell us what you've seenWhich chip-maker collaboration?
A joint reveal was floated roughly a month out from the July roadmap. Still unannounced as of the Sept AMA.
Tell us what you've seenDATES THAT CHANGE THE READING
New to the file.
Not all new events.
Research edition / 22 September 2026
A research connection predates the announcement.
Bar-On, Komargodski and Weinstein co-author mining-incentives research.
Paper record ↗The ESA proceedings record.
The publication date of the invited-talk paper is separate from its presentation date.
Publisher record ↗Implementation work continues.
The reviewed FP8 PR history records further commits. BzMiner v100.31 reports Pearl mining improvements. Neither establishes mainnet FP8 activation or measured customer adoption.
FP8 commits ↗ Miner release notes ↗Product evidence gets more specific.
Pearl advertises an inference API and managed services. Together’s model page reports Serverless unavailability. Onset dates and successful access have not been established.
Pearl services ↗ Updated ledger ↗Following the thread · 7 of 12