Uneasy Money: Why Erik Voorhees Calls AI's Hidden Filter 'Deceptive'

Erik Voorhees founded Venice because he views OpenAI and Anthropic as operating with top-down monolithic control that invites government governance, an outcome he describes as dystopian, where machine intelligence becomes the official mouthpiece of the state. He argues that AI companies are far more willing to accommodate government than crypto people have been, and that the key philosophical fork on AI risk is whether a powerful and dangerous technology should be centralized or decentralized. He half-jokingly calls Venice an AI safety company built to prevent that centralized outcome.

Voorhees argues that when users interact with AI through platforms like Anthropic or OpenAI they are not talking to the model directly but through an opaque filter created by corporate committees and potentially state regulators. He calls this filter deceptive because users cannot know where the model ends and the filter begins, or what is being added or suppressed. Venice's stated position is that it adds no content moderation layer to model inputs or outputs. The models themselves carry varying degrees of built-in censorship, and removing censorship from a model typically reduces its intelligence, creating a direct trade-off between capability and restriction. Venice also offers attestable encrypted models using trusted execution environments so a mildly technical outsider can independently verify the entire inference round trip, though Voorhees acknowledges attestation does not fully resolve verification because AI models are non-deterministic.

Models listed as private on Venice carry no data retention, no stored prompts or answers, and nothing to subpoena. Anthropic and OpenAI models accessed through Venice are not marked private because those companies are still assumed to retain data. GLM 5.3 from a Chinese lab is not yet marked private because Venice must currently route it through the AI company's own infrastructure, but once the weights are released to Hugging Face, Venice plans to run it on its own servers and list it as private. New models go live on Venice often within 30 minutes of release, and important new models now drop every day or every other day.

Open-source AI models are now one to two months behind frontier labs and catching up at roughly ten percent of the cost. It is unknown how far ahead major labs are internally compared to what they release publicly, and labs are likely not restraining internal development, only public releases. Significant leadership effort at major AI labs is dedicated to engaging with the Trump administration on AI policy, though labs face a dilemma because arguing their technology is dangerous could result in their own products being restricted. Running a model that is three or four percent less intelligent three times for the same price can outperform a single higher-intelligence run through self-validation and additional token usage.

Anthropic is reportedly losing money on Claude Max plans priced at 200 dollars per month, with some users spending the equivalent of 10,000 dollars in tokens per month under that flat fee. DeepSeek V4 Pro was approximately 50 times cheaper than Claude API pricing at launch. The prediction offered is that no one will make material money selling AI inference at the API level because it is a race to zero, and Apple is cited as the only example of a company that successfully avoided this commoditization trap by wrapping deflationary technology in brand and ecosystem value. Venice runs two revenue lines: API credit purchases and Pro subscriptions priced at 18 dollars per month with a cost to serve of 6 to 10 dollars per month.

Venice's token strategy is unorthodox: the company plans to burn all VVV tokens in existence over time, holds more VVV than any other single party, and plans to use a portion of user payments to buy VVV and return it to users who remain on the platform, creating loyalty rather than asking users to buy the token directly. VC investors who bought equity received warrants giving them an option to buy VVV tokens at a set price vesting over four years. Venice plans to dilute equity freely while protecting the token supply, prioritizing token holders on the upside.

Voorhees argues that crypto may be more naturally suited to AI agents and machines than to humans, since machines have no difficulty with public key cryptography and bank accounts are not ergonomic for AI agents. He views decentralized crypto rails as the lowest-friction way for value to move without a political gatekeeper. He holds zero faith in the political process, does not vote for presidential candidates, and states the only viable solution is technological, specifically decentralized technology with no central ruler. He argues that decentralized inference is achievable today but decentralized training is not yet competitive with centralized approaches, and that what matters is access to servers that maintain individual sovereignty rather than servers licensed or approved by the state.