VVV rallied 34 percent on a document that refuses to make the accusation

Venice's VVV token rose 34 percent on Tuesday September 8 to $24.77 on a large volume spike, and kept going to a fresh all-time high near $27. The catalyst was a fight about mathematics, in which nobody has accused anybody of anything.

Worth being exact about that, because the trade depends on it.

The NYU mathematician Tristan Buckmaster published a statement describing what happened after he and Levent Alpöge produced finite-time blowup results for incompressible porous media, Boussinesq and 3d incompressible Euler, using large language models. He recounts calls with OpenAI on September 6, and one exchange in particular.

I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer.

He then writes, in the same document, "I have not seen OpenAI's proof. I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything."

That is a careful man being careful.

And it was enough. What moved the token was not a finding of misconduct. It was the discovery, by a great many people at once, that the question is answerable in principle and had not been answered. What would you conclude if a lab told you it did not look up your data and then went quiet when you asked about training on it?

David Hoffman put the trade in one line.

David Hoffman@TrustlessState

Frontier AI labs MUST steal user data, to gain an edge against their competition

The data theft is structural - it's apart of the arms race between OpenAI and Anthropic

If you don't steal your customers data, you lose to those who do

Anyway, $VVV up 40% today because @AskVenice can't steal your data

Venice is Erik Voorhees's project, running since January 2025 on the pitch that prompts stay on your device rather than on a company's servers, using open-source models with no content filtering and no account required. VVV is an access key rather than a governance token. Stake it and you get a proportional share of daily API inference capacity (which is roughly how a co-op sells you milk). A second token, DIEM, is minted by locking staked VVV and pays $1 a day of API credit in perpetuity. Emissions have been cut repeatedly, from 14 million a year at launch to 2.5 million as of September 1, with 2 million scheduled for October, and some revenue buys VVV on the open market and burns it.

So we went to look at the burn, because it is the only part of this that settles on a public ledger (and because Hoffman's pitch leans on onchain verifiability).

DefiLlama's series for Venice, which counts only the on-chain buy-and-burn and explicitly excludes subscription, API and credit revenue settling off chain, shows about $14,500 a day for a week and then $386,038 on September 8. Twenty-seven times the day before. Roughly half the trailing month in one session.

Be careful with that number before you spend it. A burn is a treasury action, and treasuries batch. A single large day may well be a scheduled sweep of a fortnight's off-chain revenue rather than a surge in usage, and DefiLlama's methodology note is honest that most of the interesting revenue lives somewhere it cannot see (card payments, Coinbase Commerce, Stripe-Bridge). Venice itself said in August that it passed a $100 million annualised revenue run rate, up from $70 million a month earlier, which is a claim on the off-chain side that nobody outside the company can check.

Our read is that this is probably a narrative trade sitting on top of a real business, which is a rarer combination in this market than the price action suggests, and that the narrative is the fragile half. Private inference is a genuine product with genuine customers. It is also a hedge whose value spikes exactly when a lab looks bad and decays quietly the rest of the time. We would expect Venice's on-chain burn to be back under $50,000 a day by the end of September. If it holds above that, the September 8 print was demand rather than accounting, and the token was cheap.

Buckmaster, for what it is worth, would rather be talking about the mathematics. He wrote that the significance of a mathematician and a model doing this work in a month "cannot be understated" and called it a Deep Blue-Kasparov moment. The crypto market read past all of it to one unanswered question about training data.