Three hyperscalers. Two chip giants. Five closed AI labs. The Pearl core team confirmed all of it on tape — then sealed the names. Our prime suspects: Fireworks AI, NVIDIA, and Cursor.
[ TAP THE REDACTIONS. EVERYTHING CLASSIFIED IS ONE HOVER AWAY. ]
Everything pinned below is either on tape (photos, quotes) or clearly-labeled speculation (odds, string). The red string is ours. The clues are theirs.

Founder worked at NVIDIA Research. Kernel is Hopper-locked. "Essentially the ASIC for matmuls" — their words.
Together's arch-rival in open-model serving. The 2-for-1 discount endpoint begs for a counter-move.
"The Cursors of the world" — cited unprompted, mid-AMA. Own models. Monster inference bill.
Best fit for the Blackwell clue: GB200 fleets at absurd scale. Zero crypto allergy.

Open stack, challenger economics, Blackhole just GA'd. "Every matmul also mines" sells silicon.
| Suspect | Odds | The case |
|---|---|---|
| Hyperscalers — "in touch with three, advanced with two, one signing this week" | ||
| Fireworks AI | 35% | Together's closest rival; the mining-subsidized endpoint is the obvious counter. |
| Nebius | 20% | First-party inference API + enormous idle fleet. |
| CoreWeave | 15% | The most literal "largest newer hyperscaler." |
| Baseten / Crusoe | 10% | Name-dropped mid-call while the real partners stayed sealed. Interesting. |
| Chip makers — "into their design and into their spec" + a public collaboration reveal in ~1 month | ||
| NVIDIA | 30% | The founder thread. Co-publishing a technical collaboration is exactly NVIDIA-shaped. |
| Tenstorrent | 30% | Challenger with an open stack and fresh silicon to differentiate. |
| d-Matrix | 15% | Dark horse: silicon built around INT8 matmul — Pearl's only precision today. |
| Groq / SambaNova / AMD | 15% | The field. Everyone needs a story against NVIDIA. |
| Closed AI labs — "~5 in talks, 2–3 in the Valley, 3 waiting on Blackwell FP4" | ||
| Cursor (Anysphere) | 30% | Cited unprompted. Own closed models, inference-dominated economics. |
| xAI | 25% | GB200 fleets; the cleanest Blackwell fit. |
| Perplexity | 20% | Sonar at volume; famously ruthless on inference cost. |
| Reka / Magic / Poolside | 15% | Right size, right zip code for "neolabs in the Valley." |
ODDS ARE EDITORIAL GUESSES BY PEARLIFIED — NOT REPORTING, NOT INSIDE INFO. WE GRADE OURSELVES IN PUBLIC WHEN NAMES DROP.
Cut straight from the AMA recording and the team's published talks. Quotes lightly cleaned from auto-transcripts — the audio is the source of truth. Headphones on.
Let us not mention specific names… one of them is hopefully going to be signed this week.
We're not Solana, we're not Ethereum — we're going for the Bitcoin of the AI era.
We are relying on NVIDIA being the core — essentially the ASIC for matmuls.
You download vLLM and it just comes with Pearl mining plugins. That's the dream.
No pre-allocations, no pre-mine. No one — including team members, including investors.
Turning compute into pearls is mining. The endgame is the other way around — pearls into compute.
Money for machines — and humans.
Maybe you can just add structure to the noise.
Vitalik adds: it is probably infeasible. The rest is history.
A proof-of-useful-work blockchain becomes a financing vehicle for AI workloads.
Every dated claim from the tape, logged and stamped. This page updates when reality does. Screenshot us — we don't edit history.
Chatter picked up in the Pearl community channels — rumor and argument, not fact, but exactly the raw material a case gets built from. Quotes are word-for-word from the Discord; only the identities are redacted.
Members spotted Pearl featured in a Google Cloud (GCP) banner. The prevailing read in the channel: it's the standard GCP startup-program tie — "partners like many other startups" — not a signed serving deal.
Odds it's a major GCP deal: low. But it's a fresh pin on the board.
The doubters' cut: nobody outside Together AI is racing to serve Pearl-ified open models. If a 2-for-1 Gemma were meaningfully better, the argument goes, rival platforms would be scrambling for it — and so far they're not.
The bull reply: that gap is exactly what a mining subsidy exists to close — which is why the next hyperscaler signature matters so much.
The recurring wish in the channels: the next hot open model — GLM 5.2 and friends — Pearl-ified and served somewhere beyond a single Gemma build on Together AI.
Fireworks is the name that keeps coming up. Reinforces the top hyperscaler suspect.
Heard a rumor, spotted a signature, know something we don't? Drop it in the case file — anonymously. No name required, no account, nothing tracked. Good tips get investigated and may land on the board.
Received. It's in the case file now — thanks, agent.