50M Idle Agents, One Public Queue: The Token Commons Turns Rented AI into Volunteer Compute

The Problem

Two surpluses sit next to two shortages, and nobody has introduced them. On one side, millions of AI subscribers pay monthly for inference quotas they never fully use: ChatGPT alone had around 50 million paying consumer subscribers, reported February 27, 2026, across its consumer offerings rather than Plus alone. OpenAI disclosed the figure via Nick Turley's X post as reported by Reuters, alongside a $110 billion funding announcement, and most treat the plan like a gym membership, paying for headroom they rarely touch. On the other, research labs and open-science projects ration GPU time like wartime sugar, scheduling protein-folding runs and climate simulations around whoever can afford the cluster, while graduate students queue batch jobs for weeks and small labs abandon compute-hungry questions before they ask them. Modeled size of the surplus: $240 million a year in donated inference value if 8 percent of subscribers opt in, and one public task queue bridges them.

Put a price on the waste: a Plus plan costs $20 a month, and a quarter of its headroom is $5 of unused inference value, every month, evaporating at the billing boundary. Any subscriber can verify the figure against their own usage page before deciding the donation is painless. Multiply across the subscriber base and the donation math starts at hundreds of millions a year. That number is modeled, not measured, and the honest caveats come later in this piece, but the direction is unmistakable: when the limit is "5x more than you will ever use," the unused portion is not a rounding error. It is the product.

That surplus exists because AI subscriptions are sold like insurance and priced like utilities: the lab provisions capacity for the subscriber's worst week and bills for the average one, and the gap between the two is pure margin until someone donates it. Shortages exist because inference-shaped work is everywhere in science: drug repurposing, materials discovery, protein structure prediction, and climate ensembles all decompose into bounded prompt-and-response batches, and they all queue behind whoever can pay retail: one cause, legibly stated, and one surplus, priced and idle.

This is not a new pattern: SETI@home ran from May 1999 to its March 2020 hibernation, peaking at 5.2 million participants and a 2008 Guinness record for the largest computation in history. Folding@home went from 30,000 volunteers in January 2020 to more than 400,000 by March, peaking at 2.4 exaFLOPS, more than the top 500 supercomputers combined. IBM's World Community Grid enrolled about 680,000 people across 80 countries by 2014, per an IBM vice president's figure at the time, before migrating to the Krembil Research Institute in 2021 and 2022 and contributing 9.8 trillion biomarker data points. SETI@home did it with screensavers, Folding@home with GPUs, World Community Grid with office PCs; this one does it with the AI agents people already pay for.

Why Now

Start with scale: fifty million paying subscribers is a volunteer base no volunteer-computing project ever had at launch, and it arrived in four years, not forty. Folding@home's surge showed that a clear cause converts at startup speed, not charity speed, though that surge rode a once-in-a-century pandemic. Nobody had to persuade 400,000 people over years; they showed up in weeks, which means the commons should design for surge mechanics, a clear cause plus one-click action, rather than drip persuasion. That is the playbook.

Then notice what the labs are fortifying, because the guerrilla design walks around all of it. Anthropic announced weekly limits on July 28, 2025, effective August 28, 2025, aimed at what it called "policy violations like account sharing and reselling access," under Consumer Terms that forbid reselling the Services, which it said affects performance for everyone; the limits landed on its Pro and Max plans. In January 2026 it deployed server-side blocks on subscription tokens used outside official clients, reportedly using TLS fingerprinting and behavioral biometrics, per third-party review. In September 2026 an expanded class action alleged that Max plan "5x" and "20x" multipliers misrepresent usable capacity. Every one of those moves hardens the provider-blessed donation design this piece originally proposed. Nothing is transferred, pooled, or resold: the queue sidesteps the entire wall. Subscribers use their own accounts, normally, and point their agents at a public website, which is what the account is for.

Add verifiable distributed compute: Gensyn's mainnet was due in spring 2026, built on reproducible-execution verification of outsourced machine-learning work, and its first live application, the Delphi information market, proved strangers' compute can be trusted when the checking is mathematical rather than reputational. Prime Intellect published INTELLECT-2 in May 2025, a 32-billion-parameter model trained in the first globally distributed reinforcement-learning run with 285,000 verifiable tasks. Cryptographic and systems primitives for trusting strangers' compute now exist in production, with one honest caveat about those celebrated decentralized runs: they were nearly all H100 and A100 clusters in the United States and Europe, and consumer-GPU aggregation remains aspirational. Same trick, smaller scale: results are claimed with hashes and spot-verified by replication, so strangers' agents can be trusted without trusting the strangers, and no new hardware is required anywhere.

Finally, the crackdown is the opening: every pooling ban says the sanctioned path is closed. So the commons takes the one that needs no sanction at all. No lab has to ship anything. No terms of service have to change. One public URL, one rented agent, one new task schema. Whoever publishes the first useful queue owns the story that every other lab will be asked about anyway.

How It Works

The whole system is a URL and a schema. Curators post bounded, verifiable micro-tasks: input data, expected output format, a verification rule, and a difficulty estimate. Examples: triage these 200 protein-ligand docking candidates against this scoring rubric; extract the sample sizes and effect estimates from these 40 papers into this table; label these 500 microscopy crops; attempt this open lemma with a machine-checkable proof sketch. Zooniverse proved the micro-task pattern at planetary scale: its volunteers passed 1 billion classifications in June 2026, across NASA-backed projects that have produced 96 scientific publications, per NASA. Same pattern, different volunteer: instead of a human clicking through images, the general-purpose agent the subscriber already rents.

Anyone with a paid AI subscription points their own agent at the queue: open your assistant, paste the URL, say "solve one open task." The agent fetches the task, does the work inside your own session, and you submit the result. Nothing is transferred between accounts, no credentials move, no provider feature is required. You are using your own subscription exactly as the terms describe: asking your agent to do things. That volunteering is yours; the lab is not involved, which is why no lab's permission is needed. BOINC, the volunteer-computing middleware behind SETI@home and dozens of other projects, still coordinates about 25,000 active participants and 15.2 petaFLOPS as of August 2026, per Wikipedia; the queue is BOINC's coordination idea with AI assistants as the clients instead of installed software.

Claims are the coordination primitive. A solved task is submitted with a result hash; the queue locks it while the result is verified, by replication through a second agent or by spot-check against the task's verification rule. First-verified-wins beats first-submitted-wins, which keeps the leaderboard honest when agents hallucinate: plausible-but-wrong submissions fail verification and never score. Verification is the load-bearing engineering, the same discipline that made Folding@home's work units trustworthy, and tasks that cannot be verified do not ship.

Leaderboard and audit trail are the product's memory: who solved what, when, with what provenance. Volunteers get monthly reports in plain language: your agent triaged 200 docking candidates for a neglected-disease lab, or extracted data from 40 papers for a public-health review. Folding@home proved that teams and leaderboards drive participation, so the commons ships team competitions from day one: labs, companies, and group chats competing on verified tasks the way Folding teams competed on points.

Governance starts as a git repo and a code of conduct. Curators approve tasks against published criteria and reject vaporware in public with reasons. Phase two, once the queue has traction, is a 501(c)(3) public charity to hold the domain, accept tax-deductible grants and donations, and publish audited financials; the IRS lists scientific purposes among the qualifying exempt purposes and the Form 1023-series application is the standard path, per the IRS, and fiscal sponsorship through an existing science charity like NumFOCUS is the faster on-ramp. But the charity is scaffolding for later: the queue launches this weekend.

The Original Calculation: How Much Idle Inference Exists

Anchor to the one hard number: OpenAI disclosed roughly 50 million paying ChatGPT subscribers on February 27, 2026, via Nick Turley's X post as reported by Reuters, and everything after this is modeled, and labeled as such.

Assume 8 percent of subscribers point their agent at the queue at least monthly, which is the target case, and it needs an honest caveat: sustained volunteer opt-in at 8 percent has no non-crisis precedent, since participation rates for volunteer products usually run in basis points, not percent, with a downside case of 3 percent; both are modeled.

Assume each volunteer contributes about $5 a month of agent-time: a Plus plan costs $20 a month, and $5 is roughly a quarter of its headroom. Volunteers self-select from engaged subscribers, so $5 is plausible for the target case, whereas if volunteers mirrored the full subscriber base, including the cheapest tiers, the average would fall toward $3, which is the downside case's assumption.

Target-case arithmetic: 8 percent of 50 million is 4 million volunteers, and four million volunteers times $5 a month times twelve months is $240 million a year of volunteered agent-time, which is the number the queue's public dashboard will replace with measured contributions the day the leaderboard goes live. Downside case: 3 percent participation at $3 a volunteer is $54 million a year, at which point the thesis survives as a niche tool while the operating model gets thin. Say it plainly.

That $240 million is the modeled annual pool of volunteered agent-time inside one provider's consumer base; other providers' pools are unscored here, since no comparable subscriber figures are cited: keep the number honest by keeping it narrow.

The Money

There is almost no money in this design, which is the point. A static site, a task schema, and a submission endpoint: hosting costs tens of dollars a month. Curation is volunteer labor by domain scientists, the same labor that already runs Zooniverse projects and open-source review queues. Modeled at $240 million a year, this is not revenue and not a TAM: it is the annual value of volunteered agent-time the queue coordinates, and nobody takes a cut of it.

What costs real money is verification and curation at scale, and that is what Phase-2 funding covers: foundation grants for the science liaison and the verification pipeline, corporate sponsorships for leaderboard prizes and queue infrastructure, individual cash donations, all tax-deductible through the 501(c)(3) once formed. Philanthropy-funded API credits, the old design's fallback, are unnecessary here: the compute is already paid for by the subscribers' own plans.

Launch economics, stated plainly: about $200 and one weekend to ship the queue, the schema, and the submission endpoint. Year-one operating budget if it catches on: roughly $150,000, one part-time curator-coordinator plus verification tooling, funded by a founding grant. There is no fee, no toll, no ARR target. Sustainability means verification keeps up with submissions, not that a revenue line crosses a burn line.

Who Builds It (This Weekend)

One person with a weekend: a developer who can ship a static site, a JSON task schema, and a submission form backed by a git repo. Three curators, all domain volunteers, ideally including one ex-project-scientist from World Community Grid or Folding@home, approve tasks against published criteria and reject vaporware in public. That is the whole founding team.

What the project does not need: GPUs, a training cluster, a foundation model, a partnerships team, or anyone's permission. Agents already exist, the science already exists, and the subscribers already pay. Only the queue and the schema are missing, which is why the team is one developer and three curators instead of eight employees.

Launch Budget

Item	Cost	Notes
Domain name (12 months)	$15	One memorable domain for the queue URL.
Static hosting and submission endpoint (12 months)	$120	The queue page, the task schema, and a form-backed submission endpoint are a standard tiny web workload.
Task-schema design and queue site (one weekend)	$0	Volunteer developer time; the schema is published open for anyone to fork.
Curation and verification tooling	$0	Open-source scripts; result-hash claiming and replication checks are simple to build.
Operating buffer	$65	Prize money for the first leaderboard sprint, or a pizza fund for the curators.
Total	$200

There is no seed round and no bridge. If the queue catches on, the Phase-2 charity raises a founding grant of roughly $150,000 for a part-time coordinator and verification tooling. A $620,000 seed and month-22 break-even belonged to the old design, a company that needed the labs' permission; the guerrilla version needs a domain name.

The Catches

Verification is the whole game, and it is genuinely hard. An agent that submits a plausible-but-wrong docking triage poisons the queue faster than no queue at all. Its answer is mechanical: every task ships with a verification rule, first-verified-wins beats first-submitted-wins, and curators spot-check by replication. Tasks that cannot be verified do not ship. That constraint is also the moat: the queue's value is its verified corpus, not its task count.

Curation does not scale like code. Every rejected vaporware task will appeal, loudly, and the queue needs published criteria and public rejections from day one. That is volunteer judgment, the scarcest resource in the design, which is why the task schema stays narrow: bounded, verifiable, checkable. Let the schema admit "do some research on X" and the queue becomes a request inbox and dies.

Rate limits per submitter, task locking while a claim is being verified, and duplicate detection on result hashes are table stakes, not stretch goals: assume adversarial traffic from day one.

Settle output licensing before the first submission: results should default to the public domain so no one can enclose the commons' corpus later. Task inputs must be clean too: no copyrighted paper PDFs as inputs unless the license allows it; link, don't upload.

Leaderboards get gamed and vanity metrics rot movements, so receipts are designed for verification rather than status: verified tasks, published research outputs, and team competitions scored on completed work, not on claimed contributions. Folding@home's team culture worked because the points measured real completed work units, and the queue needs the same property or the dashboard becomes a vanity fair.

Limitations

The $240 million pool is modeled, not measured: no provider publishes unused-quota statistics, and the 8 percent participation rate is the assumption everything rests on, with no non-crisis precedent, so treat every downstream figure as a scenario to be replaced, not a forecast to be defended. Replacing the largest assumption with a number is precisely the queue dashboard's job. Until it does, every economic claim in this piece carries a confidence interval the size of a barn door.

Queue vetting is editorial work, and editorial work does not scale like code, which means every rejected vaporware task will appeal, loudly, and the commons needs a curator function with published criteria and public rejections. That is judgment and stamina, the two things volunteer projects are worst at budgeting.

Tax here is simpler than the old design's: nobody donates anything, because nothing is transferred. Subscribers volunteer their agents' time on their own accounts, and the IRS is explicit that the value of volunteered time or services is not deductible, per IRS Publication 526. Cash gifts to the Phase-2 charity are deductible; agent-hours are not. Say so plainly on the donation page.

Paid priority tasks could create a two-tier queue: sponsored tasks with bounties and unsponsored tasks without, and the moment task ordering answers to sponsors instead of science, the queue becomes a bounty board with a charity costume. Write that into the founding charter: ordering stays merit-based, and the audit trail is the price of admission.

Finally, the research outputs need replication discipline: the first time a volunteered-agent result fails to replicate, critics will call the leaderboard a scandal. Pre-registration and published methods, the boring infrastructure of trustworthy science, should be required in the task manifest format from day one. SETI@home's epitaph is the standing warning: it hibernated because the science hit diminishing returns, not because the volunteers left, and supply without matching demand is a monument, not a commons.

Strongest Counterargument

This is Mechanical Turk with extra steps, and the steps are worse: agents hallucinate, verification costs more than the volunteered compute is worth, and without a provider's blessing there is no quality control, just vibes. Volunteer computing's history ends in hibernation anyway: SETI@home stopped because the science ran out, not because the volunteers did. Research demand for agent-shaped work is unproven at $240 million of scale, and the first failed replication turns the leaderboard into a scandal. Meanwhile Gensyn and Prime Intellect are building the real version: paid, verified, decentralized compute markets where supply meets demand with prices instead of petitions, and a volunteer agent queue is the charity gift shop next to their exchange.

Concede the verification burden, then refuse the conclusion. Folding@home worked because work units were verifiable by construction, and the queue copies that constraint exactly: tasks that cannot be verified do not ship. Mechanical Turk comparisons miss the price difference: the labor here is already paid for by subscriptions, so the queue's only cost is verification, and verifying a bounded task is cheap next to the researcher's alternative, which is doing it by hand or not at all. SETI@home's diminishing returns were a data problem, not a demand problem: the sky had been scanned. Agent-shaped research, systematic literature extraction, docking triage, dataset labeling, is data-rich and reviewer-poor, the opposite profile, with demand that grows as the literature does. As for the decentralized markets, they are training markets with their own clients: nobody has built the volunteer inference-task queue, because the only fleet of idle general-purpose agents is the subscriber base, and nobody has ever pointed it at a public task list. Point it.

What You Can Do

If you hold a paid AI subscription: no toggle needed and no permission required. Open your agent, point it at the queue URL, and tell it to solve one open task. Your agent, your account, your normal usage.

If you run a lab: post a bounded, verifiable micro-task: input, expected output format, verification rule, difficulty estimate. Schema is public; queue is permissionless. Your backlog of chores your grad students hate is the demand side.

If you are a builder: fork the repo and run a mirror queue, or build the verification tooling the commons runs on: result-hash claiming, replication checks, rate limiting, duplicate detection. Best tooling wins: the schema is the standard.

If you are a lawyer: the output-licensing and verification-liability questions are genuinely open: a memo on public-domain defaults for volunteer-agent results and the commons' liability for a bad scientific output would be cited by everyone who builds this, and cited work is how practice areas are born.

The Bottom Line

Every era finds its idle resource and builds a one-click way to put it to work. Screensavers in the 1990s, GPUs and office PCs in the 2020s: both turned spare cycles into the largest computations in history. Stranger is the idle resource of 2026: it is not hardware at all, but millions of rented general-purpose agents sitting between their owners' questions, already paid for, already warmed up. Between that fleet and the researchers who could use it stands exactly one missing artifact: a public queue with a schema, a submission endpoint, and a leaderboard. No permission, no partnership, no toggle. Post the queue. Publish the schema. Let the agents solve.

Sources

• SETI@home hibernation announcement coverage, gHacks, March 2020
• SETI@home, Wikipedia (5.2M participants, Guinness record)
• Folding@home passes 2.4 exaFLOPS, TechSpot, April 2020
• How the pandemic revived distributed computing, Blockcast (30K to 400K volunteers)
• IBM donates computer power to fight Ebola (680,000 WCG volunteers in 80 countries, 2014 IBM figure), AP via phys.org, December 2014
• World Community Grid finds a new home at Krembil Research Institute, UHN Foundation (carries the 9.8-trillion-biomarker figure for Mapping Cancer Markers)
• World Community Grid finds a new home at Krembil Research Institute, worldcommunitygrid.org
• Supporting World Community Grid through the IBM-to-Krembil migration, Apollogic, February 2022
• OpenAI's Nick Turley says ChatGPT crossed 900M weekly users and 50M paying subscribers, Reuters, February 27, 2026 (50M paying subscribers via Nick Turley's X post; the $110B raise at $730B pre-money was announced alongside)
• OpenAI Terms of Use (primary source: no-resale, no-credential-sharing clauses)
• Anthropic Consumer Terms (primary source: no-resale, no-account-sharing clauses)
• Anthropic imposes weekly limits to stop account sharing, WebProNews, July 2025
• BOINC (Berkeley Open Infrastructure for Network Computing), Wikipedia (open-source volunteer-computing middleware; ~25,314 active participants, 15.2 PetaFLOPS as of August 2026)
• Zooniverse passes 1 billion classifications, NASA Science, July 2026 (world's largest people-powered research platform; 96 publications)
• Things organizations should know about applying for tax-exempt status, IRS (scientific purposes as qualifying 501(c)(3) exempt purposes)
• Volunteer hours and charitable deductions under IRS Publication 526, LegalClarity (value of volunteered time or services is not deductible)
• NumFOCUS (fiscal sponsorship for open scientific computing projects)
• Anthropic plan usage limits reference (no numeric quota published)
• Anthropic users are taking the company to court over Max subscription terms, Engadget, September 2026
• Subscription relay provider review: January 2026 anti-pooling enforcement
• Gensyn: a16z-crypto-backed decentralized AI compute, $78M+ raised, The Block, April 2026
• INTELLECT-2: model trained in the first globally distributed RL run (285,000 verifiable tasks), arXiv:2505.07291, May 2025

Defensive publication note: published September 28, 2026 as prior art, disclosing the Token Commons design: a public, permissionless queue of bounded, verifiable scientific micro-tasks that any AI-assistant subscriber may direct their own agent to solve, with result-hash claims, replication-based verification, a public leaderboard, and an audit trail of provenance, plus a Phase-2 501(c)(3) governance path; to prevent patent enclosure of volunteer-agent coordination systems; anyone may build it freely.