Decentralized Inference’s Scoreboard: 211 Billion Tokens a Day

For most of 2026, almost every number attached to decentralized inference came from the companies selling it. A dashboard would announce billions of tokens served, a press release would claim a record day, and readers were left to take it on faith. In October that changed. Blockworks, the crypto research and media firm, published its first full survey of the sector, a report its analysts titled The State of Crypto Inference Providers, and for the first time an outside party counted what the crypto-native networks actually push through a neutral venue.

The headline number is both impressive and sobering. Seven providers (Venice, Phala, Darkbloom, io.net, AkashML, NEAR AI and Chutes) moved roughly 211 billion tokens a day over the prior month. That is a real, measurable workload, the kind of figure the sector has wanted to point to for two years. It is also about 1 percent of the 21.9 trillion tokens a day that run through OpenRouter alone, the aggregator the report used as its yardstick. Decentralized inference is no longer a demo. It is also nowhere near winning.

A scoreboard, at last

The value of the report is not any single figure; it is the method. Until now, the sector graded its own homework. Each network ran a dashboard, each dashboard reported a big number, and because no two networks counted the same way, the totals could not be compared or trusted. Blockworks did something boringly useful: it went to one place where many of these providers actually serve traffic, read the meter, and wrote down what it saw. The result reads less like a marketing deck and more like a box score.

That box score matters because the gap between what these networks claim and what anyone can verify has been the running story of the whole category. It is the reason the tokens behind the technology have behaved the way they have: usage climbs, press releases pile up, and the token keeps sliding. It is a split between shipping software and a sinking asset that will be familiar to anyone who followed the Eliza framework’s dead token and living code. A scoreboard does not close that gap. It does something more valuable: it measures it.

Why OpenRouter is the meter that matters

To understand the report, you have to understand OpenRouter. It is a routing layer that sits in front of hundreds of models and providers. A developer sends a request to a single endpoint, and OpenRouter forwards it to whichever back end serves that model, logging the tokens that pass through. Because the meter lives outside any one provider, the throughput it records cannot be inflated by that provider’s own analytics. It is the difference between reading settled on-chain data and reading a project’s pitch.

The catch is that OpenRouter is only one venue. A provider that sells most of its capacity through its own API or consumer app will look small on OpenRouter even if its total business is large. Venice, the report notes, books under a tenth of its volume there. So the scoreboard systematically undercounts total demand. That is the right kind of error. A number that can only be too low, never too high, is exactly the number you want when the sector’s chronic problem is numbers that are too high. The report is a floor, not a ceiling, and even the floor is bigger than skeptics expected.

The leaderboard, line by line

Here is the scoreboard as the report drew it, with the figure each provider could verify through OpenRouter set against what it says it does elsewhere. The spread between those two columns is the story.

Four of the seven put a real number on the board; three did not. That alone is a finding. A sector that spent two years quoting trillions-served headlines turns out to have, at the point where an outsider can check, four quantifiable providers and a combined footprint of about 211 billion tokens a day. Impressive for a cottage industry. Tiny next to the incumbent it wants to replace.

The 160 billion that is really 5.7 billion

Chutes, the inference subnet inside Bittensor, is both the headline and the cautionary tale. It claims roughly 160 billion tokens a day. The report found about 5.7 billion of those flowing through OpenRouter, and that measured figure was down about 33 percent month on month. A gap of nearly thirty to one is not proof of dishonesty; Chutes routes through its own API, paid tiers, and a confidential layer, so OpenRouter sees only a slice. But it is the sector’s whole problem in miniature: the number you can check is a fraction of the number you are handed, and the checkable one is shrinking.

The price tells the same story. The report put Chutes’ blended rate near $0.21 per million tokens, roughly double io.net’s $0.14, while Chutes volume fell and io.net’s doubled. Earlier independent analyses of Bittensor’s subnets by Pine Analytics and others estimated Chutes’ ratio of token emissions to external revenue at somewhere between 22 and 40 to one, with a break-even around $1.41 per million tokens against competitive market pricing closer to $0.88. In plain terms, the service has been sold below what it costs to run, and token emissions paid the difference. Volume that is real, subsidized, and declining at once is a strange thing to value, which is part of why the market has struggled to.

Phala’s surge and io.net’s doubling

The two clean good-news lines on the board belong to Phala and io.net. Phala went from about 5 billion tokens a day in July to roughly 54.6 billion, an order of magnitude in a single quarter, on the back of confidential compute. Its model runs inside hardware enclaves (Intel TDX and newer confidential-GPU modes) so the operator cannot read the prompts or the weights it is serving. That used to be a research checkbox. The jump suggests it has become something a growing set of buyers, especially privacy-sensitive and agent workloads, will actually pay for.

io.net nearly doubled to about 14.8 billion tokens a day at roughly $0.14 per million, the cheapest blended rate on the board. Price is part of it, but so is a tokenomics change the network shipped earlier in the year: its Incentive Dynamic Engine ties emissions and burns to real earnings, burning at least half of post-payout revenue rather than printing tokens into the void. io.net destroyed a couple of million IO under that program by mid-October. It is a modest sum, but it is the rare case of a DePIN network tying its token to money it actually took in.

Venice, and the demand that is not about price

Venice leads the measured board at about 87 billion tokens a day, and it is the most interesting entry because its demand is not mainly about being cheap. Venice sells private, uncensored access to open models; the pitch is that it does not log you and will not refuse you. Its founder, Erik Voorhees, the former ShapeShift chief executive, has made that a philosophical stance as much as a product one, arguing that “we treat you as an adult, capable of using information technology without paternalism.” The report pegged Venice’s annual recurring revenue near $100 million, and because OpenRouter is under a tenth of its volume, its true workload is far larger than its scoreboard line.

That demand is the kind that does not evaporate when a rival cuts prices, because privacy and permissionlessness are not line items a hyperscaler can match. It is the consumer-grade end of a market where crypto interfaces have quietly converged with mainstream apps, and it is the one product on this list a mainstream user might choose on purpose rather than for a discount. The capital markets have noticed. On October 20, the issuer 21Shares wrapped Venice’s token in a physically backed exchange-traded product (ticker VENC on Switzerland’s SIX, with a 2.50 percent fee), one of the first single-asset ETPs built on an AI-crypto network.

The price underneath the volume

Blended prices of $0.14 to $0.21 per million tokens sound like a decisive discount against the big clouds, and against on-demand hyperscaler GPU rental they are. But the discount is a moving target, because centralized inference is collapsing in price too. Through 2026, DeepSeek’s aggressive rates forced a chain reaction: Google cut some Gemini tiers by as much as 85 percent, and OpenAI and Anthropic trimmed within about ninety days of each other, in what analysts called the deepest price war in enterprise AI history. When the incumbent floor falls faster than your token can subsidize, “cheap” stops being a moat.

This is the uncomfortable center of the commodity side of the business. Serving an open-weight model that anyone can download is, almost by definition, a race to the bottom, and the crypto networks are racing against incumbents who can lose money on inference to sell something else. The report’s volume figures prove the demand is there. They do not prove anyone can hold a margin while meeting it, and the networks whose entire pitch is lowest price are the most exposed when the price everywhere keeps falling.

One percent, and a market cap smaller than one company

Return to the ratio that frames everything: about 211 billion tokens a day against OpenRouter’s 21.9 trillion is roughly 1 percent. And the report draws the scale gap even more starkly on the capital side. The entire liquid crypto-AI token category, every DePIN GPU network and AI coin combined, is worth around $24.9 billion. That is less than the roughly $30.9 billion valuation of a single private AI-infrastructure company, Crusoe, which raised a $3.9 billion Series F in September to build data centers and modular power.

Sit with that comparison. The whole tokenized AI-compute sector, after a two-year narrative cycle, is capitalized below one company that pours concrete and runs transformers. It does not mean the crypto networks are worthless; it means the market is pricing them as a bet, not as infrastructure of consequence yet. The buildout they are chasing is real and enormous, and demand is outrunning supply across the industry. The question the report keeps pressing is whether a permissionless network captures any durable share of that, or whether it is a thin, subsidized skim on top of a business that belongs to companies with balance sheets.

Revenue nobody can verify

The bluntest line in the report is about money: “no decentralized computing network has disclosed verifiable revenue exceeding tens of millions of dollars.” AkashML, the inference arm of Akash, showed annual leasing fees of roughly $3.7 million. The rest either do not break it out or cannot substantiate it. Trillions of tokens served, and not one network can point an auditor at eight figures of recurring revenue it will stand behind.

Contrast that with the players who raised equity instead of selling a token. Prime Intellect, which does distributed training and hosted inference, raised a $130 million Series A at about a $1 billion valuation and cites a run rate near $100 million in annual recurring revenue, with no liquid token at all. The lesson is not that decentralized compute cannot make money. It is that, so far, where the real money is being made, it is often not the token holders making it. The revenue accrues to a cap table, not a ticker.

Measured is not the same as trusted

Here is the limit of any scoreboard, including this one: counting tokens does not prove the right model served them honestly. A provider can advertise a 70-billion-parameter model and quietly answer with a cheaper, quantized version that is faster and worse. Prime Intellect, which built a verification scheme for exactly this, puts the motive plainly: providers “make adjustments to computation methods to optimize for cost, efficiency, or specific commercial goals.” The meter sees tokens. It does not see which model produced them, and that is the trust gap the sector still has not closed.

There are four honest answers to that problem in 2026, and each pays for integrity in a different currency. Ethereum co-founder Vitalik Buterin has argued that verifiability is the strongest thing crypto brings to AI, while warning that wrapping a model in a full cryptographic proof can add hundreds of times the compute cost, which is why nobody runs real large language models that way yet. The table below is the current menu.

Trusted execution is winning the practical argument because it is cheap and it ships. That is the mechanism behind Phala’s surge and the reason Chutes has pushed its operators toward sealed enclaves. None of these, though, is yet cheap, fast, and trustless all at once, which means the scoreboard’s volume is, for now, volume you still have to take partly on trust. You can read Prime Intellect’s own write-up of why even their fast check is a fingerprint rather than a proof.

The supply side is thinning

The scoreboard also quietly documents a supply problem. AkashML showed 283 active GPUs, down from a peak near 897, after the network wound down the incentives that had padded that count. Chutes runs around 4,350 GPUs, roughly half of them H200-class, with operators increasingly required to run inside confidential VMs. When demand is real but the reward for providing hardware is a volatile token, machines leave, and the networks are left overprovisioning against churn to keep latency acceptable.

Greg Osuri, Akash’s founder, has framed the underlying mismatch in a line worth keeping: “AI moves in months, energy moves in years.” The chips, the power contracts, and the data-center space these networks need are the same ones hyperscalers are hoovering up, and the same ones Bitcoin miners are pivoting their megawatts toward. That competition for hardware and electricity is its own grinding story, one that runs straight into the physical footprint of all this compute. A permissionless network of consumer and enterprise GPUs is clever, but it is bidding for the scarcest inputs in technology against buyers with far deeper pockets.

The tokens still do not believe it

If usage is finally measurable and rising, the tokens have not gotten the memo, with one telling exception. Here is where the assets tied to the scoreboard’s providers stood on October 23, 2026, per CoinGecko.

The pattern is hard to miss. Most of these tokens sit 65 to 98 percent below their peaks while the networks behind them post record throughput. The lone holdout is VVV, down only about a quarter from its high, and it belongs to Venice, the provider with the most defensible, least price-dependent demand on the board. That is not a coincidence so much as a signal: the market is willing to pay for a business with a moat and punishes the ones racing to zero. Phala’s token jumped around 20 percent in a single late-October session as the report spotlighted its surge, yet it remains some 96 percent underwater, a reminder that a tenfold jump in usage and a near-total token drawdown can live side by side.

Notice too that RENDER, the second-largest token on this list, is not even on the inference scoreboard; its GPUs point at rendering and other workloads, not the OpenRouter traffic the report measured. The token market’s favorites are not the volume leaders, which is worth remembering whenever a chart gets quoted as if price were proof of usage. The broader backdrop does not help: amid a choppy late-2026 tape where even Bitcoin is fighting to reclaim old ground, speculative AI-infrastructure tokens are not where risk capital is rushing.

What the SEC sees, and what it does not

Regulation is the quiet variable under all of this. In March 2026, the SEC and the CFTC issued a joint interpretation naming a set of major tokens as digital commodities, led by Bitcoin, Ether, and a short list of large networks. It said nothing about DePIN or AI-infrastructure tokens, which leaves TAO, IO, AKT, RENDER, VVV, and PHA in a case-by-case gray zone. The framing that has emerged from that silence is the one that should worry token buyers most: you may be able to treat a network’s raw compute output as a commodity, but a token engineered to route revenue back to holders starts to look like equity in the operating company, which is a security.

That distinction is not academic. It is the difference between a freely traded utility and an instrument the SEC can say was sold unregistered, and it cuts closest to exactly the revenue-sharing designs these networks are drifting toward as they try to answer the value-capture question. It also shapes what a holder actually controls, a theme that runs through the broader debate over what ownership really means on-chain. Across the Atlantic, the EU AI Act’s rules for general-purpose models have been enforceable since August 2, 2026, with fines up to 15 million euros or 3 percent of global turnover. When a model runs on anonymous GPUs scattered across a dozen jurisdictions, the law’s basic question, who is the provider responsible for this system, has no clean answer. That ambiguity is both a sales pitch for compliant confidential compute and a latent liability no one has tested in court.

What the scoreboard changes

For two years, the honest answer to the question of whether anyone was actually using this was a shrug and a dashboard. Now it is a number an outsider wrote down: about 211 billion tokens a day. That is the real shift, and it should not be minimized. Decentralized inference has graduated from thesis to measurable business, with at least one provider, Venice, showing demand that looks structural rather than subsidized.

But the same report that proved the usage also drew the ceiling in heavy ink: 1 percent of a single aggregator, no network with verifiable revenue above the low tens of millions, a token sector capitalized below one private data-center company, and a trust problem no one has solved at scale. Three things would move the next scoreboard. First, a network disclosing audited revenue in the tens of millions, not a dashboard figure. Second, a verification method cheap and fast enough to leave on by default, so measured volume becomes trusted volume. Third, agent-native demand that pays per call rather than per subscription, the use case these rails were arguably built for. Until at least one of those lands, the summary is simple: decentralized inference is a real business, doing real volume, at a real discount, wrapped in a token market still pricing it like a promise.

Frequently Asked Questions

What is decentralized inference?

Decentralized inference means running AI model inference, the step where a trained model answers a prompt, on a distributed network of independently owned GPUs coordinated by a blockchain, rather than inside one company’s cloud. Providers compete to serve requests, usually through an OpenAI-compatible interface, and are paid in a token or a stablecoin.

What did the Blockworks report actually measure?

It counted the tokens that seven crypto-native inference providers served through OpenRouter, an aggregator that sits in front of many models and logs real throughput. That measured total, about 211 billion tokens a day, is far harder to inflate than a provider’s own dashboard, though it undercounts volume sold through other channels such as a provider’s direct API.

Is Chutes really the biggest decentralized inference provider?

By its own claim, yes: Chutes says it serves about 160 billion tokens a day. But the report found only about 5.7 billion of those flowing through OpenRouter, and that measured figure was falling. Venice served the most verifiable volume on the scoreboard, about 87 billion tokens a day.

Why are the tokens down if usage is up?

Because usage and token value have largely decoupled. Most of these networks subsidize cheap inference with token emissions, so more volume can mean more selling pressure rather than more revenue for holders. The one token that held up, Venice’s VVV, is tied to the provider with the most defensible, non-price demand.

Is decentralized inference cheaper than OpenAI or AWS?

Usually yes, with blended prices around $0.14 to $0.21 per million tokens in the report. But the discount is shrinking, because centralized providers cut prices aggressively through 2026, and the cheapest crypto rates are often propped up by token subsidies rather than genuinely lower costs.

Marcus Okafor covers AI and crypto infrastructure for HOGE Wire.