In March, Venice looked like a timely bet on privacy-first inference. Five months on it is a $100m business, a unicorn, and the subject of crypto’s oldest argument.

When this publication last examined Venice, the case rested on a hypothesis: that autonomous agents would consume inference at industrial volume, and that some meaningful share of that consumption would prefer not to be logged. The platform was processing 45bn tokens a day and the thesis was, politely, a thesis.

Five and a half months later the hypothesis has been thus far supported by the arithmetic. Venice has recorded peaks of 100bn tokens across 8m requests in a single day, crossed $100m in annualised revenue, passed 4m users, and closed a $65m Series A at a $1bn valuation led by Dragonfly. It did all of this while profitable, and without selling a single token from a treasury that makes it the largest holder of its own asset. The interesting question is no longer whether demand for unobserved inference exists. It is who gets paid for it.

 

Agents arrive in force

 

The agentic economy that carried Venice into 2026 has not cooled. Enterprise adoption of multi-agent systems has continued to compound, and forecasters put the agentic software market at roughly $7bn to $12bn today, expanding toward $50bn to $90bn by the early 2030s at growth rates near 40% to 50% a year. Such projections deserve the usual scepticism reserved for hockey sticks drawn by people selling shovels. The directional point survives the discount.

What has changed more decisively is the status of privacy. It has moved from advocacy to procurement. When an agent handles a company’s finances, its codebase or its legal correspondence, zero data retention stops being an ethical preference and becomes a line item in a security review. Meanwhile, open-weight models have kept narrowing the gap to the closed frontier, which makes an unrestricted routing layer commercially viable rather than merely ideological. Venice sits at that intersection: 200 to 300 models across text, image, video and audio, with a default architecture that forgets.

 

The shape of the growth

 

The user curve is steady rather than spectacular. Venice reported 3m active users in April, roughly 3.5m registered at the time of the raise on 1 July, and publicly marked 4m by mid-August. That is respectable consumer growth for a company with no meaningful paid acquisition, but it is not the number that matters.

Throughput is. From 45bn daily tokens in March, Venice was processing 1.3trn tokens a month at the Series A, an average near 43bn a day, before recording peaks of 100bn tokens and 8m requests in a single 24-hour period in early August. Daily API call volume stood at about 1.7m at the start of July. A fivefold move in request count over five weeks is remarkable enough that the definitions deserve care: “requests” as reported in August may not be strictly comparable with “API calls” as reported in July, and Venice’s figures are company-supplied and unaudited. The order of magnitude, however, is corroborated by the revenue.

Revenue reached over $70m annualised by the raise, having turned profitable in the first quarter. Just over a month later Erik Voorhees announced the platform had crossed $100m. The capital is earmarked for GPUs and Venice’s first owned data centre, converting leased compute into owned compute, lifting gross margin, and making larger token burns affordable. Notably, only about 8% of users pay in crypto. Mr Voorhees himself credits feature parity with ChatGPT, rather than the token, as the strongest driver of growth.

Product has kept pace. Agentic chat became the default experience in May. Voice and video shipped. OpenClaw continues to list Venice as a recommended private provider, Base MCP plugins have landed, and x402 lets agents settle in USDC per request.

 

Where the traffic comes from, and who burns it

 

The user base and the token consumption are two different businesses wearing the same brand, and conflating them is the most common error in Venice analysis.

Geographically, third-party web analytics place the United States at roughly 31% of visits to venice.ai, followed by a long tail led by India, Mexico, Spain and Italy, none above about 6%. Outside the United States the traffic is overwhelmingly mobile, in India by a ratio of four to one. The more revealing figure is the referral mix: around 71% of visits arrive direct, with search accounting for most of the remainder. Direct traffic at that share is the signature of word of mouth and brand recall rather than purchased demand, which is precisely how a company reaches $100m of revenue and profitability at the same time.

Consumption tells a different story from headcount. One hundred billion tokens across 8m requests implies roughly 12,500 tokens per request. A human chat turn runs to a few hundred. Nothing about a consumer app produces that ratio; agent loops do, because they re-read context, chain tool calls and retry. For scale, the creator of OpenClaw disclosed 603bn tokens across 7.6m requests in a single month on a rival provider, close to 79,000 tokens per request, from a three-person team running about 100 coding agents. One such fleet is worth tens of thousands of subscribers.

The users are consumers. The tokens are agents. Only one of those cohorts pays by the million.

That developer cohort arrived through defaults rather than marketing. OpenRouter routes Venice models, Fleek points its hosted proxies at Venice inference, and Cursor, Brave Leo via bring-your-own-model, and community VS Code extensions all support it. Venice’s API user count passed 25,000 in March. Independent estimates put the revenue split at roughly half subscription and half API, placing Venice between a consumer platform such as ChatGPT and a developer-heavy business such as Anthropic, whose mix runs closer to 80% API.

Run that split through the subscription tier and the shape becomes clear. If half of $100m is subscription revenue, and Pro costs $18 a month, that implies on the order of 230,000 paying subscribers, under 6% of a 4m user base. The consumer franchise is broad and shallow. The agent franchise is narrow and deep. The first is a brand asset and an acquisition funnel; the second is where the tokens actually burn.

 

The two claims

 

The Series A gave investors 8.98% of the company, a vesting grant of 1.5m VVV, and warrants on a further 5m VVV exercisable over eight years. Exercising those warrants would require a further $66.5m, implying a strike near $13.30 against a spot price of $13.80 on the day, and taking total potential proceeds to $131.5m. The stake and the $1bn headline valuation both trace to Mr Voorhees’s own announcement, and the two do not divide into one another on any obvious basis, so readers should treat the percentage and the valuation as separately reported figures rather than as a single reconcilable arithmetic.

Venice chose to dilute its cap table rather than monetise its treasury. It holds north of 30m VVV, some $400m of notional at prevailing prices, and could have sold five million tokens over the counter, raised the money it wanted for data centres, and never touched its equity. It declined, and Mr Voorhees was explicit about why: the company does not want to sell the token. Forgoing that option during a year in which VVV appreciated some 700% is the opposite of the behaviour that earned this sector its scars, and it deserves the credit it has received.

The round nonetheless moved treasury tokens. The 1.5m grant and the 5m of warrants commit up to 6.5m VVV, roughly a fifth of the treasury, to Series A investors. “Has not sold any tokens” is true in the spot sense and is doing a great deal of rhetorical work. What Venice avoided was the single print that would have hit the price on the day; what it did instead was pay part of the consideration in deferred, locked and structured treasury tokens, released at fewer than 6,000 VVV a day after a one-year cliff and a three-year linear vest, around 0.2% of daily volume. Token holders still bear a share of the cost of the raise. They simply bear it in instalments.

Venice did not sell the token. It paid with it, slowly.

The market’s reaction, then, was not simply a flinch. Dankrad Feist’s summary, that the token-and-equity split “sucks”, captured the structural objection cleanly: shareholders hold a legal claim with enforceable rights, while token holders hold a designed economic claim that depends on Venice choosing to keep buying and burning. The criticism lands harder because Venice markets VVV as the platform’s capital asset, an invitation to expect proximity to the company’s economics that the capital structure does not quite honour.

The alignment case is stronger than the reflex allows, and it is mechanical rather than rhetorical. The proceeds fund owned compute; because burns are revenue-funded, margin expansion mechanically enlarges the burn budget. The burn programme has been repeatedly widened rather than quietly narrowed, most recently with a new levy on API credit purchases. Most importantly, Dragonfly, Coinbase Ventures and their co-investors now sit on both sides of the capital structure at once, holding equity and VVV together. Whatever else they are, they are not indifferent to which of the two claims captures Venice’s growth, and that is a more durable form of alignment than any statement of intent.

The sceptical case is also correct on its own terms. A burn is policy, not covenant. There is now an equity preference stack sitting ahead of token holders for the first time in the company’s history. In an acquisition or a listing, equity captures terminal value and the token captures prepaid inference plus whatever burn survives the new owner’s capital allocation committee.

Nor has the market clearly resolved the question in equity’s favour. At around $13.55 in early July, VVV’s circulating market capitalisation stood near $637m against an equity valuation of $1bn, but its fully diluted value was reported near $1.54bn, comfortably above the equity mark. Comparisons of this sort should be handled carefully in any case: Venice’s equity holds more than 30m VVV, so the two valuations overlap rather than describe rival claims on one pot of cash.

Equity is a promise a court will enforce. A burn is a promise a spreadsheet will enforce.

Our read is that the tension is structural rather than a betrayal. Venice cannot make VVV a residual claim on profits without turning it into a security, which is the constraint every issuer in this category is bumping against. What token holders own is a claim on throughput, backed by an incentive rather than a contract. That is a legitimate asset. It is simply a different asset from equity, and it should be valued on different terms. The analytical error is expecting the two to be the same instrument; the analytical work is deciding what the policy claim is worth, and watching whether the policy holds once the people who negotiated warrants are in the room.

 

Toward deflation

 

Emissions have been cut with unusual discipline: from 14m VVV a year at the DIEM launch to 10m, then 8m, 6m in February, 5m in May, 4m in June and 3m on 1 July, with further steps to 2.5m on 1 September and 2m on 1 October. Cumulative burns stand at roughly 33.7m VVV, about 42% of the 100m genesis supply, and a new programmatic burn directs $5 of every $100 of API credits purchased into open-market buying, alongside the existing subscription-tier burns. The stated goal is a net deflationary token with native yield.

DIEM’s supply target is rising for the first time, from 38,000 to 40,000 in four steps of 500 between 3 August and 14 September. This looks expansionary and is not quite. Each DIEM confers $1 a day of perpetual API credit and can only be minted by locking staked VVV, so raising the target simultaneously creates capacity and removes float. It does create an obligation: 40,000 DIEM represents $40,000 a day, close to $14.6m a year, of standing inference commitment. Against $100m of revenue and improving margins that is comfortable. Against a stalled revenue line and leased GPUs it would not be.

 

Adjacent bets

 

Three names sit closest to Venice’s orbit. Each deserves its own treatment, and each will get one in coming issues. For now, the high-level map and, more usefully, the actual relationship to Venice.

NEAR AI Cloud is the verifiable version of the same promise. It runs open models inside Intel TDX confidential virtual machines with NVIDIA confidential computing, returning a cryptographic attestation with each request, and has extended into agent runtime with IronClaw, browser integration with Brave Nightly, and sovereign workloads with the government of Bermuda. Venice says “we do not keep it”; NEAR says “here is the proof we could not have kept it”. That is complement and competitor at once, and it is the axis on which Venice’s own TEE and end-to-end encrypted model options will eventually be tested by enterprise buyers.

AntSeed launched in May from Gibraltar as a peer-to-peer marketplace for model access, using BitTorrent-style discovery, USDC settlement on Base, no accounts or API keys, and zero platform markup, with around 20 providers at launch. Venice is one of them. AntSeed is attacking the aggregator layer, meaning OpenRouter, rather than the inference layer, so for Venice it currently reads as distribution rather than substitution. Investors should note the ANTS token has been deliberately non-transferable under a fair-launch design, which makes it difficult to express a view.

Dolphin (POD) is the closest of the three. It is the AI lab whose uncensored fine-tunes have shipped inside Venice, now operating its own peer-to-pool inference network across idle consumer and data-centre GPUs, with network revenue routed into open-market POD buybacks and a staking token, xPOD, that carries a delegable daily inference allowance. It is a genuine upstream supplier that has become an adjacent network. It also migrated contracts from DPHN in March 2026, so the trading history is short and the float thin.

The common thread is that each is a claim on a different layer of the same stack. Venice owns the customer, Dolphin the models, AntSeed the routing and NEAR the proof. Whether those layers converge or specialise is the question we will take apart properly.

 

What could go wrong

 

Privacy premiums have historically been uneven, and Venice’s own management attributes growth chiefly to feature parity rather than principle. Competition is real and arrives from four directions: local runtimes, rival aggregators, confidential-compute specialists, and the frontier labs themselves should any of them ship a credible zero-retention tier. Regulation cuts both ways, and the uncensored framing in particular invites scrutiny in Europe and Australia as content obligations for AI providers move through legislatures.

Two structural risks deserve more weight than they usually get. The first is concentration: agentic demand is lumpy, and a documentation change at a single agent framework repriced this token by hundreds of per cent in March. The second is capital intensity. Buying data centres converts an asset-light, profitable business into an asset-heavy one at exactly the moment GPU supply is tight and the cost of getting the timing wrong is highest.

 

The Apollo Crypto View

 

Venice has done the difficult part, and done it in the least fashionable way available: it built a product, charged for it, made money, and declined to sell its own token to fund growth. The March thesis, that demand for unobserved high-volume inference exists, is settled.

The open question has shifted from demand to allocation. Our position is that equity now owns Venice’s terminal value while VVV owns its throughput, that this is a defensible design rather than a bait and switch, and that it is nonetheless a materially different asset from the one many holders believe they bought. The burden of proof sits with the burn.

Three numbers will settle it. Burns against emissions. Gross margin as owned compute comes online, which determines the size of the buyback budget. And tokens per request, which tells you whether the agentic cohort that actually drives consumption is still compounding, or whether Venice has simply acquired a very large number of people who chat for free.

 

Disclosure: Apollo Crypto holds VVV

This report (‘Report’) has been prepared for informational purposes only and does not constitute an offer to sell or a solicitation of an offer to purchase any security of financial product or service. This Report does not constitute a part of any Offer Document issued by Apollo Crypto Management Pty Ltd (ACN 623 059 227, AFSL 525760) or Non Correlated Capital (ACN 143 882 562, AFSL 499882), the Trustee of the Apollo Crypto Fund. Past performance is not necessarily indicative of future results and no person guarantees the performance of any Apollo Crypto financial product or service or the amount or timing of any return from it. This material has been provided for general information purposes and must not be construed as investment advice. Neither this Report nor any Offer Document issued by Apollo Crypto or Non Correlated Capital takes into account your investment objectives, financial situation and particular needs. The information contained in this Report may not be reproduced, used or disclosed, in whole or in part, without prior written consent of Apollo Crypto. This Report has been prepared by Apollo Crypto. Apollo Crypto nor any of its related parties, employees or directors, provides and warrants accuracy or reliability in relation to such information or accepts any liability to any person who relies on it. You should obtain a copy of the Information Memorandum, issued by Non Correlated Capital before making a decision about whether to invest in the Apollo Crypto Fund.