$VVV Community Call | October

Venice passed 5 million users in October, having taken 13 months to reach its first million and just 68 days to grow from 4 million to 5 million. John said users, usage, revenue, and subscriptions had accelerated through the year, particularly in the prior four to six weeks, and expected Venice to reach 6 million faster than it reached 5 million unless conditions changed dramatically. Daily tokens served rose from about 250 billion at September’s end to 500 billion and then more than 1 trillion in October, crossing John’s previous year-end estimate about two months early.

John linked the recent usage inflection to replacing Kimmy K25 for agentic chat and GLM46 for classic chat with an uncensored DeepSeek default, later upgraded to DeepSeek 101.7. He said testing showed a strong improvement and users responded positively, while noting that benchmark performance alone cannot establish whether a model’s style or tone will suit users. Agentic chat is a web-based mini-harness that interprets intent and selects models or tools, including image and video generation, so users can request an outcome without coordinating each tool. Venice expects to combine agentic and classic chat over time, but plans to retain manual model selection for power users.

Monthly VVV burns passed $1 million on October 28, with daily burns averaging about 1% growth over roughly three months. John said monthly burns had grown around 16–20 times since December, and programmatic burns began outpacing discretionary burns in either the prior month or the month before. He forecast that burns could rise tenfold within a year if the business continued growing at even half its current rate. He also said falling inference costs let Venice offer Pro users intelligence for $18 per month that would have been too expensive a year earlier, and could deliver more inference value per dollar to DM holders. He sees sustained cost declines as potentially relevant to DM’s long-term market value, while treating short-term price movements as noise.

Video was likely the largest revenue-side growth driver over the previous three or four months, according to John, as video models improved and video calls remained more costly than LLM calls. Venice Studio contributes about 25–30% of company revenue. The community cited Black Thunder, a nearly 20-minute, 1980s B-movie-style AI film and Lumara Film Festival winner, as evidence that users can produce work beyond “AI slop.” The speaker was unsure whether the project received $1 million plus another $1 million in Venice credits; the film reportedly used about $2,000 in credits, while comparable production was said to have previously cost millions.

Venice’s first exchange-traded product, VENC, was listed on a Swiss stock exchange by 21Shares. John said Venice neither funded nor participated in the product and that he learned of it from the public announcement. He called the listing an IPO-like milestone because it could give VVV access to investors unable to participate directly on-chain, while reiterating that Venice has no plans to IPO. He expected similar products first in Europe over the course of the year, with possible US products next year if conditions go well. Another speaker said VVV might appeal to public-market investors seeking AI exposure, but the discussion of a possible Anthropic IPO and valuation was speculation.

John said prompt injection can lead an AI agent browsing websites to disclose user data after encountering malicious instructions in page content or metadata, potentially without the user knowing what the agent saw. Venice’s security principle is to avoid possessing or retaining user data, since data held by a service can potentially leak. He warned that companies should assume stored data could eventually be exfiltrated and that AI-driven vulnerability discovery will increase exposure and require rearchitecture. He cited a PHP vulnerability that could give attackers “God mode” on affected websites and estimated about 70% of websites were vulnerable, noting that older, poorly maintained sites may remain exposed. In his view, crypto infrastructure has faced high-value exploit risks for years and is better maintained against them.

John said Venice runs DeepSeek as an open-weight model without sending inference data back to DeepSeek, but he was unsure whether a Department of Homeland Security ban covered only direct use of DeepSeek services or independently hosted versions too. He said preventing direct data transfers could address a real risk, while a broader ban might reflect misunderstanding of the announcement. He described “obliteration” as fine-tuning or post-training to remove restrictions and refusals; Venice offers an obliterated DeepSeek version that defenders can use for cybersecurity tasks less likely to trigger refusals. One speaker felt open-source models had narrowed the frontier-model gap for much of the year, though progress seemed to plateau in the past month; this was a personal impression, and Kimi K3.1 Pro appeared close to frontier models on benchmarks.