Between 4 and 7 September, four of the networks we track published a number describing how much inference they serve. Phala said 38.86 billion billed input and output tokens over 24 hours. io.net said 8 billion tokens in a single day. IoTeX said 2 billion tokens over seven days across 45 models. Heurist said its inference router is the top merchant on XRPL’s x402 network by transaction volume.
Four windows, four denominators, and no two of those figures can be put in the same column. Working out which of them anyone outside the company can check turned out to be the interesting part.
What shipped this week
peaq put its machine economy on a subscription. Economics 2.0 went live on mainnet on 1 September 2026: a machine activates on a USD-priced tier and bonds PEAQ at the oracle rate, with the bond counted as circulating supply. Pricing machine access in dollars while settling in the token is what turns a DePIN network into something a fleet operator can put in a budget. peaq’s own documentation records the burn address for lapsed bonds as unset, so the half of a decayed bond advertised as burned currently reduces nothing. peaq tokenomics docs, read 5 September 2026
Allora’s forecasts got a machine-shaped front door. peaq wired Allora in as a native peaqOS adapter on 3 September 2026, so a robot can pull a decentralised price or weather forecast with a keyless chain read and no account anywhere. A forecast a machine can fetch without signing up for anything is a different product from one a trader reads on a screen. peaq’s announcement says machines sell forecasts back “autonomously, on demand, without human intervention”, while the adapter page says peaqOS holds no keys and signs nothing, so the earning half still runs on an operator’s own wallet. robotic.sh adapter page, read 5 September 2026
Heurist put its tool catalogue behind agent payments. On 5 September 2026 Heurist Mesh tools became callable and payable through x402 on XRPL, priced in RLUSD, covering token trends, DeFi metrics and NFT holdings. An agent paying per call in a stablecoin is the shape most of this sector says it is building towards, and the settlement asset here is a dollar token rather than HEU. @heurist_ai, 5 September 2026
Coinbase closed the door on ATH derivatives. Coinbase announced the removal on 20 August 2026 and nine perpetual futures contracts stopped trading around 13:00 UTC on 3 September, Aethir’s among them, with open positions settled automatically and spot untouched. Derivatives are where a token gets priced by people who never intend to hold it, so losing the venue narrows who can take a view on ATH at all. Coinbase gives a routine product review as the reason and suspends IOTX spot on 23 September under the same kind of review, which reads as a shelf being tidied rather than a verdict on Aethir. @CoinbaseMarkets, 20 August 2026
The one number anyone can check
OpenRouter is a commercial router. You send it a prompt, it picks a provider, it bills you in dollars. It also publishes a daily per-model token series on every provider page, which makes it the only place in decentralised inference where a throughput figure is settled by somebody other than the company quoting it.
It lists 106 providers. Five are networks we cover: Venice, Phala, AkashML, Chutes (Bittensor’s subnet 64, operated by Rayon Labs) and io.net. Prime Intellect is registered as a provider with no measurable traffic. Here is the week of 1 to 7 September, read off OpenRouter’s own series on 8 September:
Venice 373.1B tokens (53.3B a day)
Phala 301.3B (43.1B a day)
AkashML 166.6B (23.8B a day)
Chutes 72.6B (10.4B a day)
io.net 42.3B (6.0B a day)
Three things the table settles
io.net’s number is accurate and it’s their best day. OpenRouter shows 8.02 billion tokens on 6 September, against a claim of 8 billion in a single day posted the following morning. Their seven-day average is 6.0 billion. Quoting your peak is ordinary marketing, and io.net is also the only one of the four to name a source anyone could go and read.
Phala’s own figure is smaller than one router’s traffic to it. Phala’s 24-hour window closed at 14:49 UTC on 4 September. OpenRouter routed 57.4 billion tokens to Phala on 3 September and 54.9 billion on 4 September, against Phala’s 38.86 billion.
The two use different words, “billed” and “processed”, and the gap runs the unusual way round for this sector, with the self-report as the conservative number. Phala’s provider entry points its privacy policy at redpill.ai, a separate inference brand, which hints that the two figures have different scopes without explaining the difference.
Two of the four networks aren’t on OpenRouter at all. Neither IoTeX nor Heurist appears among the 106, so neither published number can be checked by anyone outside those two companies. That’s no accusation, and it’s the whole difference between a figure and a measurement.
Whose tokens were they
IoTeX’s 2 billion came through QuickSilver Pro, and IoTeX said so in the same post: 45 models, “from OpenAI, Anthropic, Goo[gle]”. QuickSilver Pro’s pricing page is operated by MachineFi Inc, IoTeX’s operating company, and lists thirteen model providers: OpenAI, Anthropic and Google among them. The page names no chain and no token anywhere on it.
So the tokens are there and a closed lab produced almost all of them. Reselling frontier models through one API key is a reasonable business and a poor measure of decentralised compute, and the two get quoted as though they were the same quantity.
What the tokens are worth
Priced at each provider’s own listed OpenRouter rates, that week converts to a gross annualised band:
Phala $12.0M to $46.6M
Venice $5.4M to $20.4M
AkashML $2.3M to $9.7M
Chutes $1.6M to $9.7M
io.net $1.6M to $5.6M
The bands are wide because OpenRouter’s series carries no prompt/completion split and the two rates differ by up to several times, so each floor prices every token as input and each ceiling prices every token as output. It’s the same method our Bittensor demand roll-up already uses. The truth sits between the two ends, and the floor is the one to quote if you want a single number, because the ceiling assumes every token billed at the dearer output rate.
Two caveats sit on those figures. This is gross revenue to the router before OpenRouter’s cut and before the GPU supply gets paid, so the network keeps a fraction of it. And it’s one channel: every one of these five sells directly as well, and none of them publishes that side.
Then the part that matters for a token holder. On a single DeepSeek model I counted 28 providers competing, Phala and Venice among them, alongside the big clouds and the frontier labs’ own endpoints. The buyer picks on price and latency. OpenRouter settles in dollars with a company, and PHA, VVV, AKT and TAO are nowhere in the transaction.
What to do with a token count
Three things worth carrying to anyone’s throughput number, ours included:
A token count needs a window and a producer. Two billion over seven days and 38.86 billion over 24 hours differ by more than a hundred times per day, and one of them counts models built by somebody else.
Ask who else can see it. A figure only the issuer can compute is a claim. The same figure on a third party’s page is a measurement, and it stays one even when the third party is centralised.
Volume isn’t revenue, and revenue isn’t the token. Every step between the three loses most of the number, and the last step frequently loses all of it.
Biggest movers this week
7-day change as of 7 September 2026. Twenty-nine of the forty tokens we track closed up.
Gainers:
FLock.io (FLOCK) +67.3%
peaq (PEAQ) +33.8%
NuNet (NTX) +33.4%
NEAR Protocol (NEAR) +28.0%
Fetch.ai / ASI Alliance (FET) +20.3%
Fallers:
Ora Protocol (ORA) -30.2%
Giza (GIZA) -15.9%
Heurist (HEU) -15.6%
ElizaOS (ELIZAOS) -11.2%
Openmind (ROBO) -8.3%
See all movers: ownyourmind.ai/projects/movers/
What we verified this week
Nine scores moved. Five findings are worth your time.
IoTeX hasn’t held a governance vote since October 2024. Our evidence called the IIP process functional and accelerating, and cited IIP-42 as passing on 60 voters. The Snapshot space has put nothing to a vote since IIP-40 on 28 October 2024, and IIP-42 reads Status: Draft in IoTeX’s own repository. Governance dropped from 10/20 to 8/20.
The half of our old IoTeX claim that survived is the interesting half. IIP-42 proposed halving the block time, and the chain does now produce a block about every 2.5 seconds, measured over a fortnight on IoTeX’s own RPC. A repository status tracks the paper and never the chain, so a Draft document can accurately describe something that already shipped.
A settled lawsuit that was never settled. Ten surfaces across our Ocean Protocol coverage described the Fetch.ai litigation as resolved with 286 million FET returned. The docket carries no termination, briefing ran to 26 June 2026, the 286 million is an alleged conversion rather than a recovery, and the plaintiffs are a putative class rather than Fetch.ai.
Heurist’s chain is starker than our own prose. We had reasoned from the allocation table that a paused mining programme functionally re-routes half of supply to foundation discretion. On-chain it’s more direct: one Safe holds 66.20% of supply, and eighty-two addresses hold the token at all.
peaq folded two labelled reserves into the team-controlled one. The Community and Security Reserves each sent their full balance to the Ecosystem and Treasury Reserve on 28 August 2026 and are now empty, with a further tranche leaving on 1 September against a stated purpose of “to be confirmed”. The total is unchanged; about a fifth of supply now sits behind a single label.
Behind those, every project we track now carries a dated, sourced verdict on which exchanges actually trade its token, checked at each venue’s own API rather than at an aggregator. Twenty-two of thirty-six needed correcting, in both directions. Giza’s five named venues have all gone; Kraken lists NOS, which our evidence had denied.
See the full integrity board: ownyourmind.ai/resources/integrity-checks/
On the watchlist
Whether Venice’s seventh emission cut lands where the last six did. VVV emissions dropped from 3M to 2.5M a year on 1 September, the seventh reduction in about two years, against a stated goal of net deflation. Venice is also the largest of the five on OpenRouter and the only one that published no volume figure this week.
Whether peaq switches its burn address on. Economics 2.0 is the machine-economy thesis written into contracts, and the one part of it that would make PEAQ scarcer currently points at an address the project’s own docs record as unset.
Whether any of these networks routes inference revenue through its token. Five of them sell on OpenRouter and get paid in dollars. Akash’s burn mechanism is the only one of the five designed to convert compute demand into token demand, and its realised burns have so far offset a fraction of issuance.
From the site
The IoTeX episode went out on the podcast on 6 September, recorded before the governance finding above, which is a good illustration of why the written review carries the score and the audio doesn’t: ownyourmind.ai/projects/iotex/.
Every project we track now carries dated, sourced verdicts across five scoring dimensions: ownyourmind.ai/resources/integrity-checks/.
The OYM podcast is on Spotify, Apple Podcasts and the major directories under “Own Your Mind”.
Cheers,
Bobski
Own Your Mind. Independent research on decentralised AI. No sponsors.
ownyourmind.ai · @ownyourmindai · @0xbobski
Disclosure: OYM holds GRASS and NEAR.