In this file 6 sections
PEARLIFIED / THE OPEN FILEOPEN SPECULATION / 05
PEARLIFIED / INVESTIGATION 05THE NAMES ARE STILL MISSING

THE EVIDENCE ROOM

Someone’s on
the other end
of the thread.

Unnamed providers. Familiar connections.
Follow the record as far as it goes.
Then decide what you believe.

Choose your investigation

THREE DOORS INTO THE FILE

Where does
your instinct
take you?

01 / THE COUNTERPARTY

Who serves
the next answer?

The provider roadmap, Fireworks, and the documented Cursor connection.

Follow the provider lead ↗
02 / THE PAYMENT

How does PRL
reach compute?

A stated ambition. A scaling question. A research trail through Simplex.

Trace the payment story ↗
03 / THE PEOPLE

What came
before Pearl?

Papers, coauthors and implementations. Connections with a history.

Follow the research trail ↗

Unnamed partners. Familiar names. Interesting coincidences. The clues are theirs; the red string is ours. Follow a lead, inspect the evidence, and decide how far you’d take it.

THE EVIDENCE ATLAS / FOLLOW A CONNECTION

Pull a thread.
See what holds.

People, papers and engineering artifacts. Choose a trail, then open a card to see exactly what connects it.

Source review: 22 September 2026 · research connections are not commercial agreements

THREE WAYS INTO THE FILE

Follow a connection. Inspect the leap.

Select a trail, then open a connection to see its evidence and limits. Every trail and connection has its own link.

FOLLOW THE THREAD

Compute payments

The ambition, the scaling question, the possible design.

Link to this trail ↗ · Choose another trail ↑
  1. 01 / Team statementPRL → inferenceInspect connection +

    The May AMA describes spending PRL on inference. This establishes the stated ambition, not a live checkout.

    Hear the compute flywheel ↗Link to this connection ↗
  2. 02 / Team statementThroughput → research beyond L1Inspect connection +

    The September answer discusses scaling approaches and references Pass’s consensus work. It does not choose a payment architecture.

    Read the full scaling answer ↗Link to this connection ↗
  3. 03 / Documented researchChan + Pass → SimplexInspect connection +

    Coauthorship connects Pass to the consensus research. The Simplex project lists implementations and research descendants; these are not Pearl payment integrations.

    Inspect the research map ↗Link to this connection ↗
  4. 04 / Our interpretationConsensus research → a possible payment systemInspect connection +

    A separate system could use consensus to agree on transfers. Funding, billing, settlement and exits still need a concrete design.

    Explore the four-step journey ↗Link to this connection ↗

FOLLOW THE THREAD

Inference providers

Follow the unnamed counterparties without filling in the names for them.

Link to this trail ↗ · Choose another trail ↑
  1. 01 / Team statementUnnamed partners → an open investigationInspect connection +

    The earlier recording describes negotiations without naming the counterparties.

    Hear the unnamed-partner clue ↗Link to this connection ↗
  2. 02 / Our interpretationServing economics → Fireworks as a candidateInspect connection +

    An inference-serving business could fit Pearl’s integration ambition. Several companies could fit the same description; fit does not establish talks.

    Read the candidate investigation ↗Link to this connection ↗
  3. 03 / Documented relationshipFireworks → CursorInspect connection +

    Fireworks’ own site carries Cursor statements about training and serving. This supports a relationship between those two companies, not a Pearl agreement with either.

    Inspect the company’s account ↗Link to this connection ↗
  4. 04 / Open questionCandidate → named deployment?Inspect connection +

    The decisive next connection would come from an operator naming its Pearl integration and documenting a usable service.

    Compare the historical case ↗Link to this connection ↗

FOLLOW THE THREAD

Research connections

People, papers and implementations—each with its own kind of relationship.

Link to this trail ↗ · Choose another trail ↑
  1. 01 / Documented researchBenjamin Chan + Rafael Pass → SimplexInspect connection +

    The paper identifies both authors. Coauthorship is a precise research connection.

    Open the Simplex paper ↗Link to this connection ↗
  2. 02 / Documented implementationSimplex → Commonware → TempoInspect connection +

    The existing research file follows a consensus implementation into payment-focused infrastructure. This is a comparison for Pearl, not an announced collaboration.

    Follow the implementation trail ↗Link to this connection ↗
  3. 03 / Research lineageSimplex → modified Alpenglow / VotorInspect connection +

    The Simplex project describes Votor as a heavily modified descendant. Design lineage does not mean Solana runs unmodified Simplex or establish activation status.

    Read the lineage description ↗Link to this connection ↗
  4. 04 / Documented researchBar-On + Ilan + Omri → a shared paperInspect connection +

    The paper’s author list is the evidence for coauthorship. It does not establish employment or a commercial agreement.

    Inspect the author list ↗Link to this connection ↗

Drawn from the existing research file, reviewed September 24, 2026. Labels describe what each source supports; the trails do not announce new partnerships.

01 Pass + Chan co-authored → Simplex influenced → implementations

02 Coauthorship precedes → the later collaboration

03 Kernel engineering connects to → the agent-integration ambition

PERSONRafael PassInspect connection +

Co-authored Simplex with Benjamin Y. Chan. The September AMA also places his research in Pearl’s scaling discussion.

Original author record ↗Read the research history ↗
PAPER / 2023Simplex ConsensusInspect connection +

A documented line from academic research to consensus implementations. Its assumptions remain part of the story.

Protocol research site ↗Hear the AMA connection ↗
IMPLEMENTATIONCommonwareInspect connection +

Commonware describes its consensus module as inspired by Simplex and documents differences from the original protocol.

Implementation documentation ↗Follow the full trail ↗
PROJECTTempoInspect connection +

Tempo names Simplex via Commonware in its payment-focused architecture. Compare that with Pearl’s PRL-for-inference ambition and its L1/L2 discussion; no Pearl integration is established.

Project architecture ↗Read the context ↗
COAUTHORSHIP / 2025Bar-On · Ilan · OmriInspect connection +

A May 2025 paper puts Yogev Bar-On, Ilan Komargodski and Omri Weinstein together on an author list.

Inspect the author list ↗Before the partnership ↗
COMPANY RESEARCHAttestableInspect connection +

The later collaboration has a research prehistory. The technical questions concern what a proof establishes and how a complete system would use it.

Read the proposed approach ↗Hear the September explanation ↗
ENGINEERINGThe kernel workshopInspect connection +

Pearl describes agent-assisted grouped-GEMM optimization. Its reported gains apply to the tested kernel shapes.

Engineering account ↗Explore the workshop ↗
PROPOSALFloating-point miningInspect connection +

The reviewed FP8 implementation is a proposal. Follow code review and activation as separate milestones.

Inspect PR #311 ↗Compare release milestones ↗
ON TAPE / SEPTEMBERAgents in the stackInspect connection +

The AMA describes a future integration scenario. This links a recorded ambition to the engineering work without treating the ambition as deployed.

Read the passage ↗Hear the scenario ↗

THE SPECULATION DESK / THREE OPEN FILES

The next chapter
has missing names.

Three stories worth following. The clues are sourced. The possibilities are our read.

Editorial briefing · updated 24 September 2026 · no inside information

01 / THE LEAD INVESTIGATION · OPEN THEORY

Two names on the board.
One documented connection.

Fireworks and Cursor both appear on our candidate board. Fireworks’ own site connects them through training and serving infrastructure. The open question is whether either trail leads to Pearl.

MAY / THE EARLIER CALL

The names stay off tape.

The earlier AMA discusses unnamed partners. It gives the investigation a starting point, not a list of confirmed counterparties.

▶ Hear the negotiations · 29:05 ↗
SEPTEMBER / THE NEXT STEP

More providers. More capacity.

The September roadmap connects expansion with FP8 and Blackwell. The question becomes which operator will describe an integration in public.

Hear the roadmap · 8:48–10:27 ↗

READ THE CONNECTION PRECISELY

FireworksCursor

Documented relationship

Pearl?Either company

A Pearl connection remains an open theory.

ONE CANDIDATE TO EXAMINE / FIREWORKS AI

A business fit, with an interesting adjacent connection.

Fireworks describes serverless, dedicated and reserved inference services. Its own site also carries Cursor statements about training and serving infrastructure. That places two names from our historical board in a documented relationship with each other.

Our interpretation: a serving operator that already supports other model builders is an interesting place to investigate Pearl’s integration thesis. The Fireworks–Cursor relationship does not establish a Pearl relationship with either company.

Inspect Fireworks’ own account ↗

Source checked 24 September 2026. Relationship context; no newly dated Pearl announcement.

The exciting possibility

An operator integrates Pearl into an existing serving stack, making it available through a service customers already know.

The alternative explanation

The clues describe a broad market. Another provider could fit them equally well; conversations may remain experiments or never become a service.

The clue that would move this forward

A named provider explains its Pearl integration, links a usable service and states which workloads participate. A partner’s own account would tell us more than another inference from the board.

WHERE THIS LEAVES THE INVESTIGATION

What we know
Fireworks documents a relationship with Cursor. Neither that connection nor the unnamed AMA discussions establish a Pearl partnership.
What remains open
Which provider will name a Pearl integration and explain what customers can actually use? Fireworks remains a candidate; the historical odds are unchanged.
FOLLOW THIS LEADHear the provider roadmap before weighing the names →

02 / ACCESS

What opens the next door?

The recorded clue. At the September AMA, the team said it was working on institutional access but was not ready to reveal details.

▶ Hear the tease · 56:48 ↗

The exciting possibility. A concrete access product opens Pearl to an audience that cannot use its current routes. The institutional audience is named; the product remains the mystery.

Explore the alternative explanation and next clue

The open gap. Institutional access could take several forms, and preparation may precede availability by a long time. The remarks do not choose a custodian or exchange; hostnames cannot settle that question.

Watch for. A named service, published availability and an attributable announcement from the parties involved.

Read the source passage ↗

03 / THE DEVELOPER STACK

Could agents carry the integration?

The recorded clue. The AMA describes agents integrating Pearl into an inference stack. Pearl’s engineering article separately describes agent-assisted kernel optimization.

▶ Hear the scenario · 45:30 ↗

The exciting possibility. A repeatable integration workflow lets another team add Pearl without rebuilding its serving stack. The next revealing exhibit would be that team documenting the process itself.

Explore the alternative explanation and next clue

The open gap. Agent-assisted kernel work could remain an internal engineering advantage without becoming a reusable integration product. The named lab in the AMA is a hypothetical example.

Watch for. A released tool, reproducible integration steps or a third-party team documenting its own deployment.

Read the engineering story ↗

THE QUESTION BEHIND THE BOARD

Who’s on the other
end of the thread?

Start with a lead. Separate the recorded clue from our interpretation. Then decide what evidence you would need next.

Historical case file · September 2026 · no new confirmations implied

CASE 01 / HARDWAREOPEN THEORY

Who might turn useful work into a silicon advantage?

Names on our radar: NVIDIA · Tenstorrent

01 / THE RECORDED CLUE

Pearl’s existing file connects a founder’s NVIDIA research background, GPU-focused kernel work, and statements about chip makers.

02 / WHY WE SUSPECT IT

Hardware fit and professional history make this a plausible place to look. They do not identify an unnamed counterparty.

03 / What doesn’t fit?

Research collaboration, program membership, and a commercial agreement are different relationships. Other chip makers could fit the same broad clues.

04 / What would change our mind?

A joint announcement naming the relationship and its scope would strengthen the case. A different named counterparty would change the ranking.

Find the clues on the wall ↓
CASE 02 / INFERENCEOPEN THEORY

Who follows the first inference provider?

Names on our radar: Fireworks AI

01 / THE RECORDED CLUE

The September 15 AMA describes contact with two other major inference providers without naming them.

02 / WHY WE SUSPECT IT

Fireworks is an editorial candidate because the original board compares its serving business with Together’s. That is a competitive-economics argument.

03 / What doesn’t fit?

The AMA does not identify Fireworks. Other providers could fit; a good business rationale does not demonstrate a negotiation.

04 / What would change our mind?

A named announcement, provider documentation, or an attributable statement about an integration would move this beyond a fit-based theory.

Find the clues on the wall ↓
CASE 03 / CLOSED LABSOPEN THEORY

Which labs are behind the unnamed adoption?

Names on our radar: Cursor · xAI

01 / THE RECORDED CLUE

The archived case includes the example “the Cursors of the world” and statements about labs waiting for Blackwell kernels.

02 / WHY WE SUSPECT IT

Cursor is a named example; xAI is a hardware-fit hypothesis in the original board. These are two different paths to a theory.

03 / What doesn’t fit?

An illustrative name is not a customer disclosure. Hardware requirements are shared by many organizations and cannot identify a lab on their own.

04 / What would change our mind?

A lab naming its use of Pearl, or a verifiable technical integration, would strengthen the case. A denial or incompatible implementation would weaken it.

Find the clues on the wall ↓
CASE 04 / INSTITUTIONAL RAILSOPEN THEORY

What do the two hostnames actually tell us?

Names on our radar: Anchorage · Coinbase

01 / THE RECORDED CLUE

The existing file records two company-named RPC hostnames on Pearl’s domain and an AMA reference to institutional access.

02 / WHY WE SUSPECT IT

The naming pattern suggests infrastructure worth investigating. Connecting it to custody or exchange access is Pearlified’s interpretation.

03 / What doesn’t fit?

A hostname does not prove a signed agreement, custody support, or a listing. Testing, preparation, and internal naming remain alternative explanations.

04 / What would change our mind?

Confirmation from the named company, supported-asset documentation, or a joint announcement would resolve more than DNS naming can.

Find the clues on the wall ↓
The odds are part of the game. The evidence does the work.

The percentages below are preserved historical editorial guesses. They are not calibrated probabilities, and the categories are not a single probability distribution.

OUR SPECULATION Subjective editorial odds · historical picksBrowse every pick ↓
OPEN CASES / PARTNER WATCHconnect the dots.

FILE 01 / HARDWARE

Pearl founder on stage beneath Vitalik's 'probably infeasible' slide
the talk: standing under Vitalik's "probably infeasible"… bold bold bold
Suspect · chip maker
NVIDIA
OUR ODDS 30%

Founder worked at NVIDIA Research. The early kernel work focused on Hopper. "Essentially the ASIC for matmuls" — their words.

"…we're relying on NVIDIA being the ASIC for matmuls."+ co-founder = ex-NVIDIA Research →
Read the public NVIDIA record ↗

The missing piece: Research ties and program membership do not name the still-unannounced commercial collaboration.

FILE 02 / INFERENCE

Suspect · hyperscaler
Fireworks AI
OUR ODDS 35%

Together's arch-rival in open-model serving. The 2-for-1 discount endpoint begs for a counter-move.

"One of them is hopefully going to be signed this week."Together's arch-rival, forced to copy? →
Pearl-GEMM protocol slide
the protocol slide. PoW from ANY matmul = anyone's GPU counts
Suspect · chip maker
Tenstorrent
OUR ODDS 30%

Open stack, challenger economics, Blackhole just GA'd. "Every matmul also mines" sells silicon.

Read the calls and transcripts ↗

The missing piece: The early call did not name Fireworks or Tenstorrent as the counterparties. These remain our picks.

FILE 03 / CLOSED LABS

Suspect · closed lab
Cursor
OUR ODDS 30%

“The Cursors of the world” — an example in the earlier AMA around 22:04. An illustrative name, not a customer disclosure.

"Three of them are waiting for Blackwell kernels."who runs FP4 on giant B200 fleets?? →
Suspect · closed lab
xAI
OUR ODDS 25%

Best fit for the Blackwell clue: GB200 fleets at absurd scale. Zero crypto allergy.

Read the September AMA ↗

The missing piece: A named example or hardware fit does not establish a signed deal. We are waiting for a named counterparty.

FILE 04 / INSTITUTIONAL RAILS

Two names.
One intriguing trail.

anchorage.rpc.pearlresearch.ai
coinbase.rpc.pearlresearch.ai
two partner-named hostnames, and only two, out of 89 probed →
Explore the infrastructure record ↗

The missing piece: A hostname establishes a naming clue. It does not confirm custody support, a listing, or a commercial agreement.

Public-record clue · custody
Anchorage Digital
OUR ODDS 85%

One of two dedicated, partner-named RPC endpoints on Pearl's domain: anchorage.rpc.pearlresearch.ai. The name is an intriguing integration clue, not proof of an agreement. The institutional rail.

Public-record clue · exchange
Coinbase
OUR ODDS 82%

The other one: coinbase.rpc.pearlresearch.ai — found on a wider re-probe, resolving cleanly to Pearl's own Google Cloud infra. Same pattern, same logic, bigger name: the largest US crypto exchange, publicly traded (NASDAQ: COIN). The retail rail.

Not two leads. One pattern. Two hostnames, built at the same time, on the same infrastructure, named for two companies that between them cover both ends of how money actually gets into a token: Anchorage is how institutions (funds, family offices, eventually an ETF-style product) legally hold an asset; Coinbase is how everyone else buys and trades one. Our read is that the pair could point toward broader access. Test infrastructure or preparatory naming could also explain it; the hostnames alone do not settle intent. Anchorage's endpoint carries its own dedicated TLS certificate (a real provisioning step); Coinbase's rides the domain's wildcard cert, reading as newer or less far along. Speculation, clearly labeled — but built on two hostnames Pearl themselves chose to publish, not guesswork. Sept 15 AMA update: co-founder Idan teased an unannounced push to "democratize access… to institutional investors" — the on-tape echo of exactly these rails. He wouldn't reveal it yet, but he said it's coming.

THE FULL LINEUP

Every name on our radar.

Historical editorial picks. Read the clue, then weigh the leap.

SuspectOddsThe case
Hyperscalers  — "in touch with three, advanced with two, one signing this week" · Sept AMA: "two other major inference providers" in touch + "several neolabs already mining Pearl"
Fireworks AI35%Together's closest rival; the mining-subsidized endpoint is the obvious counter.
Nebius20%Inference-provider fit; fleet utilization not established here.
CoreWeave15%The most literal "largest newer hyperscaler."
Baseten / Crusoe10%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
NVIDIA30%The founder thread. Co-publishing a technical collaboration is exactly NVIDIA-shaped.
Tenstorrent30%Challenger with an open stack and fresh silicon to differentiate.
d-Matrix15%Dark horse: silicon built around INT8 matmul — the early integer scheme.
Groq / SambaNova / AMD15%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.
xAI25%GB200 fleets; the cleanest Blackwell fit.
Perplexity20%Sonar at volume; famously ruthless on inference cost.
Reka / Magic / Poolside15%Right size, right zip code for "neolabs in the Valley."
Custody, exchange & institutional rails  — DNS clues; commercial relationships remain speculative
Anchorage Digital85%One of two partner-named RPC endpoints on the whole domain. The hostname is the clue; the commercial relationship is our inference.
Coinbase82%The other partner-named endpoint, found on a wider 89-name re-probe. Largest US exchange, publicly traded.

ODDS ARE EDITORIAL GUESSES BY PEARLIFIED — NOT REPORTING, NOT INSIDE INFO. WE GRADE OURSELVES IN PUBLIC WHEN NAMES DROP.

New leads — from the Sept 15 AMA

EXHIBIT B.2 · ON TAPE

The reveal cluster answered a few of these on the record — and opened sharper ones. Every quote below is from the Sept 15 community AMA; the reads and odds are ours.

Now on tape

Store of value

CONFIRMED ON TAPE

The "SoV" theory, said plainly. Omri: "We're going for store of value… we're doing this unapologetically." The Bitcoin-of-AI thesis is the design goal, not a side effect.

Hear the excerpt Audio + transcript ↗
The bull case · proposed

Agents make it a one-liner

OUR READ HIGH

Omri: "An OpenAI engineer can just… tell their CUDA agent to just 'Pearlify' their own optimized vLLM. It's very, very close to that scenario." Their bet: once agents do all inference optimization, integrating Pearl is a drop-in with "no fine-tuning, no preprocessing." An integration ambition, not disclosed OpenAI adoption.

Hear the excerpt Audio + transcript ↗
Teased · unannounced

Institutional access

STATUS WATCH

Idan: "We're working hard on… democratizing access… to institutional investors… not ready yet to reveal it." The on-tape echo of the Anchorage + Coinbase RPC rails already pinned above.

Hear the excerpt Audio + transcript ↗
Adoption · unnamed

Who's already mining

OPEN LEAD

Omri: "in close touch with two other major inference providers" and "we know of several neolabs mining Pearl already, without formal announcements yet." Names withheld — prime red-string.

Capability · not a gap

Privacy is a switch

OUR READ NOTABLE

Why public, not private like Zcash? Omri: Pearl is already "one of the largest zero-knowledge networks in the world" (every block is ZK-proven), so adding shielded transactions later is "quite easy." A choice, not a limitation.

Hear the excerpt Audio + transcript ↗
Person of interest · who's noticing

Emad Mostaque

SPECULATION LOW

The Stability AI founder (now building Intelligent Internet, a rival PoUW-for-AI project) congratulated Pearl on the FP8 drop — "Congrats! Not easy at all." His bio line: "Consistent inference is possible" — Pearl's exact reproducibility thesis. Peer recognition; no known business tie.

QUOTES FROM THE SEP 15 AMA (LIGHTLY CLEANED). READS & ODDS ARE PEARLIFIED'S OWN, CLEARLY LABELED.

THE QUESTION WE’RE LEAVING OPEN

Which unnamed partner will leave the next public clue?

These theories remain open. A named announcement or attributable integration would let us revisit the picks with stronger evidence.

Explore the whole file