OpenAI Needs a Compute Token Economy Before Anthropic Builds One — A Strategic Proposal

Community
, , ,
1

I’m part of the Codex for Open Source program and have been building on the OpenAI API for my security research work. I don’t usually post strategy essays on forums, but the last 60 days have made me genuinely uneasy about where things are heading, and I think the developer community should be talking about it.

Anthropic just closed at $965B — above OpenAI’s $852B. Mythos/Glasswing found 23K vulnerabilities in a month across critical open-source infrastructure, with AWS, Apple, Google, Microsoft, and NVIDIA all inside the consortium. Meanwhile, OpenAI lost three senior execs in a single day (Weil, Peebles, Narayanan), shut down Sora after burning through $15M/day in compute for $2.1M in lifetime revenue, killed the Disney deal, dissolved OpenAI for Science, and restructured the Microsoft partnership to drop exclusivity — with models appearing on AWS Bedrock literally the next day. Only 2 of the original 11 co-founders remain.

I’m not saying the sky is falling. But I am saying that OpenAI’s current IPO pitch — “we’re a SaaS company with great models” — is the same story every competitor is telling. It’s not a moat. And I think there’s a structural play that could be one, if OpenAI moves before someone else does. It’s weirder than it sounds. Bear with me.

The Idea: Turn Unused Compute Into a Tradeable Asset

Every ChatGPT Plus/Pro/Enterprise subscriber pays for a compute allocation. Not all of it gets used. Right now, that unused compute is breakage — profitable for OpenAI, invisible to the user, and a ticking clock on churn (because eventually, the user notices they’re paying $200/month and using $40 worth).

What if unused compute automatically converted into a tradeable token?

User subscribes → gets X compute units → uses 70% on ChatGPT, API calls, Codex → remaining 30% mints into tokens in their account. Tokens can be used later, traded to other users who need more compute, or spent in an adjacent ecosystem.

This isn’t a loyalty program. The token has a market price bounded by real economics: its floor is the marginal cost of compute, its ceiling is the retail API price. It can’t decouple from reality because every token is redeemable for actual inference. And critically — no one gets free credits. A high-usage subscriber who wants more compute buys tokens at market price. OpenAI clips a fee on every transaction.

Why this changes the business model:

Subscription revenue becomes pre-paid infrastructure financing. Users are funding GPU buildout before compute is consumed — like Starbucks gift card float ($1.6B unredeemed at any given time), except the “gift card” is a productive asset rather than a depreciating one. Churn drops because users with token balances don’t cancel — they’re not “wasting money,” they’re “accumulating an asset.” And OpenAI gains a new metric for Wall Street that no other AI company can offer: token velocity as a real-time proxy for actual compute demand.

The difference between breakage margin and marketplace margin is the difference between linear and exponential. Breakage earns once per subscriber per month, scales 1:1 with subscriber count. Marketplace margin compounds with network effects — transaction volume grows with the square of participants, and OpenAI takes a cut on every exchange. This is why exchanges trade at fundamentally different valuation multiples than SaaS companies.

Add an Entertainment Layer and You Lock Users 24/7

OpenAI’s models are already excellent at creative writing, roleplay, and interactive storytelling — anyone who’s used GPT-5.5 for tabletop RPG sessions knows this. But the company abandoned consumer entertainment (Sora) because video generation was compute-intensive for the revenue it generated.

Text-based AI entertainment is the opposite: low compute, high engagement, extreme stickiness.

Imagine a partnership with a major gaming platform — Steam, with its 130M+ monthly active users — where AI-powered campaigns, character roleplay, and interactive fiction run on OpenAI models. Users spend subscription compute on work during the day, entertainment at night. Same subscription, same compute pool, full 24-hour GPU utilization instead of business-hours-only spikes. That directly improves unit economics.

And here’s the lock-in that no benchmark can break: after 300 hours of building a campaign with an AI that remembers every NPC, every plot twist, every inside joke — you’re not switching to Anthropic because Claude scored 2 points higher on MMLU. Your digital life is in this ecosystem. The character memories, the social graph, the campaign history — none of that transfers.

The Mining Bridge: PoUW as Settlement Layer

Pearl (PRL) just launched a Proof-of-Useful-Work blockchain where mining = matrix multiplication = AI inference. Together AI is already running a discounted Gemma endpoint subsidized by PRL mining emissions. The tech is real, but the structure is backward — Pearl has miners producing tokens that no one needs, because there’s no demand-side ecosystem.

Flip the direction. If OpenAI partnered with (or licensed the technology from) a PoUW protocol, unused subscription compute is the mining — idle GPUs produce tokens as a byproduct of being available for inference. The token’s value is anchored by the two platforms’ actual economic activity, not by miner speculation. And OpenAI never touches financial operations directly — a partner handles the exchange, keeping OpenAI outside SEC jurisdiction for token issuance.

This is how a three-party structure works: OpenAI provides models and compute, a gaming platform provides users and distribution, a token protocol provides the financial settlement layer. Each party does only what it’s already good at. No one needs to build capabilities outside their core. The alliance holds because defection is value-destructive for every member — leaving collapses the token price, which damages everyone’s balance sheet.

What This Means for the IPO

Right now, OpenAI’s valuation comp is SaaS — 10-20x revenue, same bucket as Salesforce or ServiceNow. The proposed structure puts OpenAI in a different bucket entirely: compute marketplace + financial infrastructure + consumer platform. That’s three separate investor constituencies, each with their own (higher) valuation framework.

More importantly, it creates a moat that isn’t “our model is smarter.” Model intelligence is a depreciating asset — whatever lead you have gets closed in 3-6 months. But a liquid compute marketplace with network effects, an entertainment ecosystem with migration costs, and a token economy with financial lock-in? That’s a defensible position that gets stronger over time instead of eroding.

What I’d Like to Hear From Other Developers

I know this sounds like a lot. But the pieces already exist separately — Pearl has the PoUW math, Steam has the users, OpenAI has the models. The question isn’t whether this structure will emerge. It’s whether OpenAI will be at the center of it or on the outside looking in.

A few specific questions for the community:

Would you use a compute token system if it meant your unused API credits retained tradeable value instead of expiring? Would entertainment integration with your existing subscription change your usage patterns or your willingness to stay subscribed? And what regulatory or technical blockers do you see that I’m missing?

Genuinely want to hear pushback on this. The idea is either important or wrong, and I’d rather find out which one here than nowhere.

Disclosure: I hold no position in PRL or any cryptocurrency mentioned. I’m a security researcher and open-source maintainer in the Codex for Open Source program. I’d prefer OpenAI to remain competitive enough that my tools keep working.

2

Thanks for your “its not about — it’s about” essay.

However, it is quite the opposite of what any subscription business with consumption costs would do.

  • There are high-cost subscribers to a service, and they can buy more beyond their usage limits with credits (which also expire, a profitable breakage), or are billed for overages;
  • There are low-cost subscribers - your bread-and-butter, who use less than they consume.

You propose:

  • What if the value that low-cost subscribers seem to be not maximally employing was given to high-cost subscribers to reduce their need to pay more?

Thus, your argument for your scheme for internal exchange of value, addressed to an imagined company representive (where, currently, even explicitly-purchased API credits are non-transferrable by policy), which results in the manufacture of free credits for top users, seems to have shortcomings in its prompt input. It doesn’t seem likely to appeal as a point leading to profitability and return in an IPO prospectus.

You even included the AI’s suggested category in your paste, “API”, while talking about subscriptions in the body.

What would actually be good IPO business (which companies do):

  • increase profitability by booting the top 0.1% over-consumers off the service every month via opaque appeal-less means, such as inscrutable content policy violations.
3

Fair pushback — let me address the core objection directly.

You’re right that breakage is profitable. Starbucks carries ~$1.6B in unredeemed gift card balances at any given time. Every subscription business with consumption costs loves the user who pays and doesn’t show up. If the proposal were simply “take breakage margin and hand it to power users,” I’d agree — that’s value destruction, not creation.

But that’s not what this is. The token doesn’t transfer value from low-usage subscribers to high-usage ones. It enters a market. The low-usage subscriber holds an asset they can sell, hold, or spend in an adjacent ecosystem. The high-usage subscriber can buy tokens at market price if they need more compute. OpenAI takes a cut on every transaction. No one gets free credits — everyone pays, just through a different mechanism.

The question isn’t “breakage vs. no breakage.” It’s: what’s worth more — the margin from breakage, or the margin from operating a compute marketplace?

Breakage margin is linear. You earn it once per subscriber per month, and it scales 1:1 with subscriber count. Marketplace margin compounds with network effects — transaction volume grows with the square of participants, and OpenAI clips a fee on every exchange.

More importantly: breakage causes churn. The user who pays $200/month and uses $40 of compute will eventually notice. When they do, they cancel. You got a few months of 80% margin, then you got zero. A token balance in their account changes the psychology — they’re not “wasting money,” they’re “accumulating an asset.” Retention goes up. LTV goes up. And LTV × retention is what IPO analysts actually model.

On the category tag — fair catch, I should’ve put this under General. The substance is cross-cutting. Happy to repost if mods prefer.

4

I have my own AI output to read, which is based on my requests and interests. Don’t need your predictive text product arguing on your behalf.

Don’t enter the sphere of representing AI output as from a genuine human, as that is one of those policies that are in the “don’t” area of the OpenAI terms and conditions.

5

@SOS2

12 month scenario: At EOY a user could accrue $1.920 of compute. If a market is created, then the compute token value could be greater than $1.920 which the user can sell at a profit. And OpenAI takes a commision. This no different than a futures market for compute.

FASB would jump on this and create an opinion for capital gains tax. If there is a loss, the user could write-down a max of $3,000 per year.

Is this what you are taking about? Yet another whoHa for efficient markets (ECON 101). The big question is about risk profiles for users - which will be quite different from traditional investors…

6

You’re tracking exactly where this goes, and the futures market analogy is the right one.

On the $1,920 scenario — yes, and the key insight is that the token’s market price isn’t just “face value of unused compute.” It prices in future demand for inference. If the market expects compute demand to rise (more AI agents, more enterprise adoption, more entertainment use cases consuming tokens), the token trades above the nominal compute value today. That’s the futures premium. Users aren’t just saving unused compute — they’re holding a position on the growth of AI usage. That’s a much more compelling reason to stay subscribed than “I might need it later.”

The FASB/tax point is sharp, and I’d argue it’s actually a feature rather than a complication. The moment compute tokens get capital gains treatment, they’re legitimized as an asset class by the regulatory framework itself. That’s the thing every crypto project has been trying to achieve and mostly failing at — regulatory recognition that this isn’t play money. A compute token backed by redeemable inference from a pre-IPO company with $25B in revenue? The IRS would want to tax that. And the $3K annual write-down on losses gives retail users a familiar risk framework — same as stocks, same as any other investment they already understand.

Your question about risk profiles is where this gets really interesting. You’re right that subscription users are not traditional investors. But that’s the point — they don’t need to be. The risk exposure is naturally capped: worst case, your tokens go to zero and you’ve “lost” compute you weren’t going to use anyway. You already paid the subscription for the primary service. The token upside is a free option attached to a product you’re already consuming. That’s a fundamentally different risk profile from someone buying crypto on an exchange with money they could’ve put in an index fund. The user’s downside is capped at breakage they would have lost regardless; the upside is unbounded. That asymmetry is what makes it work for a mass consumer base rather than just for sophisticated investors.

The “efficient markets ECON 101” framing is fair — but I’d push back slightly. This isn’t just another market for an existing commodity. It’s creating a market for something that currently doesn’t trade at all. Unused subscription compute today has a market price of exactly zero — it expires silently. Giving it a nonzero price is the paradigm shift. Everything after that (futures dynamics, tax treatment, risk pricing) is just the market doing what markets do once an asset exists.

There’s actually already a primitive version of this in the wild. Xiaomi’s MiMo Token Plan — an AI API subscription — credits unused compute from the current period against your renewal price. Users have reported renewal costs showing as negative numbers because their unused balance exceeded the next month’s fee (the platform doesn’t refund cash, but renewal drops to ¥0.01). It’s a closed-loop, non-transferable, platform-internal version of exactly this idea — the platform is acknowledging that unused compute has residual value and letting users recapture it, just only through renewal.

The jump from “credit against renewal” to “tradeable token” is a big one, but the underlying economic admission is the same: unused subscription compute is not worthless. MiMo is pricing it internally. The proposal here is to let a market price it externally.

7

The only problem I see is that this is based on OpenAI’s current subscription models which would only give rise to a thin market. Nobody likes wide bid | ask spreads. What is needed is a market that provides enough liquidity for this to work - which means subscription models for Enterprise customers as opposed to pay-as-you-go? Errr.. ahh.. don’t know well how that would go over. But then again, who I’m I to speculate?

8

You’ve identified the real engineering challenge. A thin market kills the whole idea — if the spread is 40% nobody trades and the token is effectively illiquid, which means it’s not really an asset.

Two things solve this.

First, the token doesn’t have to launch into a thin market. OpenAI has 900M weekly active ChatGPT users. Even if only 5% of subscribers generate tradeable tokens, that’s a massive participant base compared to any crypto launch in history. The cold-start liquidity problem that kills most token projects doesn’t apply when you’re bolting a market onto an existing product with hundreds of millions of users already paying.

Second — and this is why the entertainment integration isn’t a side feature but a structural necessity — gaming users are token sinks. They consume tokens for AI-powered RPG sessions, interactive fiction, character roleplay. That’s constant buy-side demand. Productivity users who don’t exhaust their compute are token sources. You need both sides for a liquid market. Subscription-only would be all sources and no sinks. The entertainment layer is what turns a thin one-sided market into a two-sided one with natural flow.

And here’s the part that makes this an OpenAI-specific advantage rather than a generic idea anyone could copy: OpenAI already ships both GPT-5.5 and GPT-image-2 inside the same subscription. A tabletop RPG session powered by OpenAI isn’t just text — it’s text plus generated illustrations of scenes, characters, maps, items, all inline, all within the same conversation context. That’s a multimodal entertainment experience that Anthropic literally cannot replicate — Claude has no image generation model. Gemini is the only real competitor here (Imagen, video, Gemini 3.5 Flash), but Google has shown zero appetite for gaming platform partnerships.

Now here’s where it gets really interesting. Millions of RPG sessions generating richly illustrated campaign logs — text interwoven with contextually relevant images — is a training dataset that doesn’t exist anywhere else. No one has large-scale data of an AI proactively deciding “this moment in the narrative needs a visual” and generating one that fits the context. Once that data enters the fine-tuning pipeline, the behavior generalizes. You ask Codex to write a research paper? It proactively generates Figure 1, Figure 2 — not because you asked for images, but because it learned from millions of RPG sessions that certain moments in a document call for visual support. You’re building a presentation? It inserts diagrams before you think to request them.

Other LLMs can generate images — but only when explicitly prompted. An OpenAI model trained on this entertainment data would learn when to generate images, not just how. That’s the data flywheel: entertainment teaches the model proactive multimodal behavior, which feeds back into productivity tools, which makes the productivity experience so much better that users stay subscribed, which generates more tokens, which funds more entertainment sessions, which generates more training data.

You’re right that enterprise pay-as-you-go doesn’t fit this cleanly. But it doesn’t need to — the token economy sits alongside enterprise billing as a consumer-tier mechanism. Enterprise customers keep their negotiated contracts; consumer subscribers get the token flywheel. Two pricing tiers, two economic models, same infrastructure.

9

I think i’ll skip this topic. I liked the idea a lot of getting an amount of token every mounth that you can keep to use another month. That sounds so great to users but I think there is no way they implement it as lately the keep lowering what we get for the same amount of money we input.

They are currently trying to gain MORE money by becoming more and more predatory. Your idea sounds great but they will NEVER get more money by takings fees on a parallel market compare to keeping ALL the money anyway for all the unused token from i don’t know how much percent of people use their PLUS subscription to have a better google.

And most off all with the speed of reply and the permanent circle back to topics unrelated to the previous points such as the gaming community (never put unrelated topic in a single tread if you want to push your point), it really feels like you were talking to GPT about your idea and now you just copy the new answers from people into your chat and paste whatever he answers with very little twicks.

10

The reason OpenAI keeps tightening limits is exactly because the current subscription model is losing money — they’re burning $14B this year on $25B revenue. The subscription as it stands is underwater.

Now look at what they do with Codex: Altman personally resets usage limits every time it hits a million-user milestone, and promised to keep doing it up to 10 million. Why would a company that’s tightening everything else give away more compute on Codex? Because Claude Code already proved that AI-assisted coding converts free users into paying subscribers at a rate that justifies the compute cost. Coding is a use case where more usage = more retention = more revenue. The math works.

That’s exactly what the gaming/entertainment point is about. It’s not an unrelated tangent — it’s identifying the next use case where the same math works. A user who builds a 200-hour RPG campaign with AI-generated illustrations isn’t canceling their subscription. Ever. That’s Codex-level retention in a consumer context. And that’s why it belongs in a discussion about OpenAI’s profitability — it’s not about gaming, it’s about finding the next category where OpenAI would want to reset limits.

11

To put it in one line: You pay monthly. Then you produce, you play, you earn — all in one subscription. That’s the whole proposal.