$VVV Community Call | September

Venice’s reported daily LLM-token volume rose from 100 billion on June 25 to 250 billion on September 16. John said the chart measures fundamental usage, including LLM tokens but excluding image and video activity; growth can reflect more users, more use per user, or models becoming better or cheaper. Extrapolating the curve points to 1 trillion daily tokens in December, while a linear projection suggests late January. John said December was possible but January or February more likely. Venice had expected 15–20% monthly growth in a strong scenario, but this year’s pace was closer to 35–40%, beyond its expectations.

Token growth is a useful business indicator, John cautioned, but does not directly reveal revenue or VVV burns. Image and video use is also increasing, though those models consume fewer tokens than LLMs and cost more per turn. VVV annual emissions fell from 3 million to 2.5 million and were scheduled to drop to 2 million on Thursday. Staking yield depends both on emissions and the amount staked, and John said 2 million annual issuance may still be too high for VVV to become deflationary. Venice had made no further decisions and would monitor market and staking responses, potentially communicating more by mid-October.

The DM supply target rose to 40,000 after staying unchanged for a year after launch, and had increased for six consecutive weeks. John said VVV locking accelerated after target increases, then eased in the prior week amid price changes and unstaking; it would take more time, possibly months, to judge the change amid other market forces. He rejected the claim that a higher target was “inflating DM,” noting that DM’s price had risen as supply increased. Holders need DM to unlock VVV and may need to buy it when they want to unlock, while favorable VVV and DM prices can encourage minting DM and selling VVV. The host noted VVV had risen from about $1–$2 to around $30 over the prior six months to a year, including a rise from $18 to a $30 peak on September 8.

Venice burns VVV when users buy subscriptions and uses 5% of API-credit spending to buy and burn VVV. Programmatic amounts vary by subscription tier, roughly reflecting subscription costs; Pro subscriptions burn $2 in VVV, up from $1 after an April change. The team may revisit amounts, including if subscription prices change. Discretionary burns depend on factors such as revenue, growth, expenses, programmatic burns, and the share of free cash flow available to return through burns. The team uses a continuous TWAP rather than one large market purchase. Programmatic burns eclipsed discretionary burns for the first or second month, though discretionary burns continued to grow; Venice expects to keep increasing them for now and may rely more on programmatic mechanisms over time.

The speakers discussed allegations that OpenAI may have used mathematician Tristan’s Codex notes in claiming a solution to a $1 million Millennium Prize problem. The host said two mathematicians had publicly said about 12 hours before OpenAI’s claim that they were nearing a solution using AI tools, including Codex. Tristan insinuated that OpenAI may have accessed his notes, and the transcript says the solution used logic similar to his work. OpenAI did not explicitly rule out influence from his work and said it could not rule out the information entering training data. John said frontier providers may collect submitted information without knowing every item in their training sets; he questioned how reliably users can verify “don’t train on this” settings and said competition creates incentives to collect expansive data. He viewed the episode as a warning about AI privacy and public trust.

A judge ordered OpenAI to turn over 20 million anonymized ChatGPT logs in a consolidated AI copyright case after rejecting its privacy argument. John said litigation can reveal what companies retain, while Venice’s approach is to keep no user data it could turn over in response to a legal request. Venice has deliberately avoided training its own models, which John described as reducing cost and stress while allowing the company to integrate new models rather than compete in a training race where a model can be surpassed in about two weeks. Training remains a possible future option, not a current focus. He said privacy concerns and refusals by OpenAI and Anthropic can bring users to Venice, and that more capable models can increase usage and revenue. Venice is also betting that specialized models and agent orchestration will remain valuable: its Mines beta lets creators package models and agent instructions into task-specific experiences, including combinations that may be cheaper or more efficient. The first Venice data center was expected before the end of October, possibly slipping into November; any margin gains and resulting impact on burns will take time to assess.