Here is a number that should stop anyone pricing the "AI × crypto" narrative cold: $1.79M.
That is the combined annual recurring revenue of two AI products shipped by Meitu — Picchi at $1.11M and MVLAND at $0.68M — disclosed, according to the source material I'm working from, in an internal address marking the company's eighteenth anniversary. Meitu, ticker 1357.HK, is not a random name in this space. It is the first major Hong Kong-listed company to allocate corporate treasury into BTC and ETH, a program that once put roughly $100M of shareholder capital into digital assets.
So the anomaly is structural, not cosmetic. A firm whose single most consequential blockchain act was a treasury allocation now ships consumer AI apps with, as far as the disclosure reveals, zero on-chain surface. No token. No settlement layer. No provenance log. No verifiable inference. Just fiat-card subscriptions, denominated in USD, and a studio hit rate of two out of ten.
And the market is already pricing the opposite story. The consensus inside crypto is that AI and blockchain are converging — agents transacting on-chain, inference cryptographically verified, compute decentralized. Meitu is the natural test case for that thesis: a crypto-holding, AI-shipping company. If convergence were demand-driven, this is where it would show up first.
It doesn't. That is the finding.
Context
Meitu does not appear in most crypto research. It should, because it sits at an unusual intersection that the industry keeps claiming to care about but rarely examines with real data.
Founded in October 2008 in Xiamen, the company built its franchise on consumer image processing — beauty filters, retouching, segmentation — and accumulated a user base in the hundreds of millions. That heritage matters because it tells you where the technical moat is and, more importantly, where it isn't. Meitu's engineering depth is in applied vision: pixel-level manipulation, aesthetic tuning, and the product surface that wraps them. It is not, historically, in foundational model research.
The blockchain chapter came in 2021. Meitu became the first major Hong Kong-listed corporate to move treasury into digital assets, disclosing a program that put roughly $100M of shareholder capital into BTC and ETH. The move was read as a bet on crypto as a store of value. It was also, in hindsight, a bet on narrative. Meitu's share price has since behaved less like image-software equity and more like a leveraged proxy on crypto sentiment — a fact that matters enormously when we try to read the AI disclosure that follows.
The AI chapter is the one under audit. Meitu built and marketed a self-developed vision model, MiracleVision, and pivoted its product line toward generative capability. The disclosure I'm analyzing describes an internal "venture studio" model: ten studios incubated in parallel, each shipping a distinct consumer AI product. Two of them — Picchi and MVLAND — have reportedly reached product-market fit. The ARR figures attached are $1.11M and $0.68M respectively, denominated in USD.
Now, the sourcing, because it constrains everything downstream. This is an industry brief with extremely low information density — a handful of factual claims and almost no underlying data. Any claim about product form, business model, or architecture is inference, and I'll flag it as such. The report's own framing places the event around October 2026, extrapolated from the "eighteenth anniversary" language against an October 2008 founding. Picchi and MVLAND are not products I have reliable prior knowledge of. I treat their specifics as low-confidence.
What I can analyze with confidence is the structure: a listed company with a crypto treasury, running a portfolio model on AI consumer apps, reporting micro-scale ARR. That structure is legible, and it is the same structure the crypto industry keeps trying to build — usually with a token attached. The difference is the reporting currency.
Core
Strip the numbers to their substrate. $1.11M ARR at a hypothetical $10/month subscription implies 9,250 paying subscribers. $0.68M implies 5,667. Combined, roughly 14,900 paying users across two products.
That is the entire commercial footprint being celebrated. Two products, about fifteen thousand people, a combined $1.79M annualized. For a company whose overall revenue is measured in billions of RMB, this rounds to zero — under 0.1% of the top line by any reasonable estimate.
I've run this decomposition before. In 2020 I wrote a Python simulator to test flash-loan attack vectors across Uniswap V2 and Compound, and the discipline it taught me is simple: never accept a headline metric without reconstructing the unit. ARR is a headline metric. Reconstructing it here yields a picture of an early-stage portfolio, not a business.
The venture-studio hit rate is the more interesting variable. Two of ten studios reached PMF — a 20% conversion. That number is unremarkable in isolation and revealing in context. Twenty percent is roughly the base rate of a seed-stage venture portfolio, where most bets die and a minority carry the fund. Meitu, whether it frames it this way or not, is running a venture fund with internal headcount. The "studios" are the portfolio companies; the parent is the LP; the 80% that didn't clear PMF are write-offs that never appear in the announcement.
This is where the crypto parallel becomes exact. Crypto has its own venture-studio model — ecosystem funds, grant programs, launchpads — and it reports the same way. You hear about the grants that shipped and the tokens that listed. You rarely hear the conversion rate, because the conversion rate is usually bad and the incentive is to celebrate the numerator.
There's a harder question buried in the ARR figure, though, and it's the one the crypto industry keeps failing to ask: what is the marginal cost of a dollar of this revenue?
If Picchi and MVLAND are consumer image or video generation products — the reasonable inference given Meitu's heritage — then gross margin is dominated by inference cost. Every generated image or video frame consumes compute. If the underlying model is self-hosted, the cost sits in GPU amortization and utilization. If it's a third-party API call, the cost sits inside someone else's margin, and the product is a thin wrapper with a pricing problem.
The disclosure doesn't say which. That omission is the single most important gap in the document. A self-hosted model implies a defensible cost curve and a real moat. A third-party API implies a business that can be replicated in a weekend and repriced to zero by any competitor with cheaper compute. In my audit work, the difference between these two structures is the difference between a business and a marketing surface. I once spent forty hours dissecting the circuit constraints in Zcash's Sapling upgrade and found an edge-case failure in the large-field-element arithmetic that caused silent state corruption under load. The lesson carried over: the interesting failure is never in the headline. It's in the layer nobody discloses.
Now the missing settlement layer, where the crypto lens earns its place. Meitu's products settle in fiat. A user pays by card; the company recognizes USD revenue; the money clears through a payment processor. There is no on-chain leg. No token, no stablecoin rail, no provenance log, no verifiable inference.
For the "AI × crypto" thesis, this should be the loudest data point in the room. The most credible consumer-AI products in this disclosure — from a company that demonstrably understands crypto, that holds crypto on its balance sheet — have no blockchain surface at all. Not because Meitu couldn't build one. Because the product doesn't need one.
I've spent time on the other side of this question. In 2025 I worked with a Singapore-based AI lab to integrate zero-knowledge proofs into their reinforcement-learning pipeline, so that an agent's decisions could be cryptographically verified without exposing the proprietary weights. The project was real, valued at $200,000, and it worked. It also taught me exactly where verifiable computation belongs and where it doesn't.
Verifiable inference belongs where trust is the bottleneck: autonomous agents moving capital, models whose outputs carry legal or financial weight, systems where two parties who don't trust each other need to agree on what a model actually computed. It does not belong in a consumer filter app, where the user trusts the vendor by default and the only thing being verified is whether the output looks good. Meitu's users pay for aesthetics. They are not paying for proofs. The assumption that AI will need a verification layer is a supply-side fantasy dressed as a demand-side insight. It may become true at the agentic layer. It is not true at the consumer layer, and the consumer layer is where the paying users are.
There is one place where a chain could plausibly attach, and the disclosure is silent on it: provenance. AI-generated images and video are the highest-risk vector for deepfakes, and the regulatory response is converging on mandatory content labeling — the EU AI Act's transparency obligations, national AI-content-marking regimes, the C2PA content-credentials standard. The technical answer is a signed provenance chain: cryptographic attestation of origin and edit history, a manifest that travels with the asset.
That is a blockchain-shaped problem. A content-provenance log is, structurally, a chain of signed state transitions — the same primitive as a transaction log, pointed at media instead of money. The tooling exists. The demand is regulatory, not consumer. And Meitu — an image company, selling into multiple jurisdictions with USD-denominated revenue — is exactly the kind of firm that will eventually have to ship it.
The disclosure says nothing about it. Not a watermark, not a manifest, not a compliance strategy. For a company whose entire franchise is synthetic imagery, that's a conspicuous absence, and it's the gap I'd watch most closely.
Now stack the two balance sheets. On one side, roughly $100M of corporate treasury in BTC and ETH. On the other, $1.79M of annualized AI revenue. The ratio is about 56:1.
This asymmetry is the actual story of the disclosure, and it's a story about narrative substitution. Meitu's crypto treasury made it, briefly, a crypto stock. When that narrative cooled, the company needed a new one, and "AI" was the available substitute. The venture-studio disclosure does the narrative work the treasury announcement once did: it signals that the company is positioned for the next thing.
The problem is that the two narratives are decoupled. The treasury is a passive allocation — Meitu isn't building on-chain infrastructure. The AI products are a consumer play — they don't touch a chain. The company holds crypto and ships AI, and the two facts have no mechanical relationship to each other. They're both inputs to a valuation story, not a business.
I've seen this pattern in token markets repeatedly. A project announces a treasury allocation to signal conviction, then announces a product to signal utility, and the two never connect. The token trades on the composite narrative. When the narrative cracks, there's nothing underneath but the numbers — and the numbers here are $1.79M.
Compare the models directly. A traditional venture studio incubates companies internally, takes large equity, and measures success by exits. Meitu's version incubates ten products internally, measures success by PMF, and — based on the disclosure — doesn't sell them. They stay inside the parent.
Crypto's version is the ecosystem fund or the DAO studio: capital deployed into a portfolio of projects, often with a token attached to each, success measured by listing and liquidity. The measurement is financial and immediate. The failure mode is that the token becomes the product.
Meitu's model is structurally healthier on this dimension. There is no token to front-run the product. PMF has to be earned through subscriptions, not speculation. The 20% conversion is a real filter, not a listing event. In 2021, while the market chased digital art, I forked the OpenZeppelin library to prototype a gas-optimized ERC-721 variant and cut minting costs by 40% through calldata compression — the kind of engineering that only matters when a product has to survive on real unit economics rather than a mint. Meitu's studios are forced into that discipline. Crypto's launchpads mostly aren't.
But the model has its own failure mode, and it's the one the disclosure hides: fragmentation. Ten studios in parallel means ten codebases, ten design systems, ten go-to-market motions. Unless they share a substrate — a common model, a common billing layer, a common data pipeline — the portfolio doesn't compound. It disperses. The disclosure gives no evidence of a shared substrate. If there isn't one, the 80% failure rate isn't a portfolio effect; it's a coordination failure.
Reached PMF" is doing a lot of unexamined work in this disclosure. Product-market fit has no standard definition. In some shops it means a retention curve that flattens; in others a paid conversion rate above a threshold; in others a founder's gut. The disclosure provides none of these. It provides the label.
This is the same failure mode as TVL in DeFi. Total value locked is a headline number that aggregates deposits, incentives, and double-counted recursive positions into a single figure that looks like adoption. It isn't adoption; it's a measurement artifact. "PMF" here has the same texture. Without churn, cohort retention, LTV/CAC, or payback period, the label is a press release, not a metric.
The crypto industry has a specific allergy to unit economics because its early winners never needed them — token appreciation substituted for margin. AI consumer products don't have that luxury. If Picchi's users churn at 8% monthly, its $1.11M ARR is a leaky bucket, not a business. The disclosure can't tell us, and the omission is the signal.
The choice to report ARR in USD rather than RMB is a strategic disclosure, not a formatting one. It points the products at overseas markets — higher ARPU, weaker price competition, and a payment stack that clears in hard currency. Meitu already runs large overseas consumer apps, so the distribution channel exists.
But USD denomination also imports a compliance surface. Selling AI-generated media into the EU, the US, and other jurisdictions means inheriting their content rules: the EU AI Act's transparency duties, provenance requirements, and platform-level liability for synthetic media. The revenue is cleaner offshore. The regulatory load is heavier. That trade is worth making for a $1.11M run rate. It gets more complicated at scale.
Run the arithmetic on the cost side. A consumer image generator at a $10/month subscription might allow a few hundred generations per user per month. At a realistic inference cost of a fraction of a cent to a few cents per image, depending on resolution and model, the compute line can consume 20-60% of revenue. Add payment processing at roughly 3%, app-store fees of up to 30% if distributed through mobile, and infrastructure, and gross margin compresses fast.
A self-hosted model amortizes GPU cost across all ten studios and improves with utilization. A third-party API model pays someone else's margin on every call and has no lever to pull. The disclosure doesn't specify, and the difference between these structures is, again, the difference between a business and a wrapper.
Here's the arbitrage that should bother anyone pricing AI-crypto assets. Picchi and MVLAND together generate $1.79M in real, fiat-denominated subscription revenue from roughly 14,900 paying humans. That revenue is verifiable in the weak sense: it's disclosed, it's denominated in a hard currency, and it corresponds to people choosing to pay.
Place that next to the population of "AI agent" tokens. A meaningful number have either no revenue or revenue that is circular — denominated in their own token, paid by their own treasury, or generated by emissions. Several carry fully diluted valuations in the hundreds of millions.
Suppose Picchi were a crypto project instead of a Meitu studio. It would have launched a token, and the token would have been priced against a narrative, not a revenue line. A project with $1.11M ARR and a plausible "AI agent" story has, at various points in the cycle, supported fully diluted valuations from $50M to $500M — a multiple of 45x to 450x revenue. Meitu's $1.79M of hard revenue supports no such multiple, because listed equity is priced on earnings, not vibes.
That gap isn't a claim that the token projects are fraudulent. It's a claim that the two instruments measure different things. One is a claim on cash flow. The other is a claim on attention. Both can be rational. They are not the same asset, and conflating them is how retail capital gets misallocated. The crypto market measures narrative velocity. Meitu, because it's listed, is forced to measure the other thing. That's why its number looks so small. It isn't that the AI isn't working. It's that the market that reports in dollars has a different denominator than the market that reports in narratives.
There's a further asymmetry the disclosure hides. Meitu's BTC and ETH position is marked to market. Its AI revenue is not marked to anything — it's just a run rate. In a bull market, the treasury appreciates and the AI products look like rounding errors. In a bear market, the treasury draws down and the AI products become the "real business" story. The company gets to switch narratives depending on the cycle.
That optionality is the actual product. Not Picchi. Not MVLAND. The ability to point at whichever balance sheet is working.
Contrarian
Now the counter-intuitive part — the blind spot the convergence crowd keeps missing.
The consensus inside crypto is that AI and blockchain are converging: agents transacting on-chain, inference verified, compute decentralized. Every conference panel assumes it. Every roadmap gestures at it.
Meitu's disclosure is a clean counterexample, from the least likely source — a company that already holds crypto. The most commercially legible AI products in this dataset have no chain, no token, no verification layer, and no plan for one. And they work. Users pay. PMF is reached, twice, without a single on-chain transaction.
If convergence were demand-driven, you'd expect the crypto-holding company to be first in line. It isn't. That tells you the convergence thesis is, so far, supply-driven: builders who want to use the tools they have, looking for problems that justify them. The demand side — the fifteen thousand people paying $10 a month — has no idea any of this exists, and doesn't care.
Composability isn't a checkbox you tick after shipping. It's a ecosystem condition, and ecosystems form around incentives, not architectures. Meitu's incentive is to sell subscriptions. There is no seat in that loop for a chain. We don't get to assume the incentive will appear just because the tooling did. That's not analysis; it's projection.
Takeaway
Watch the next disclosure, not this one.
The signal to track is whether any of the remaining eight studios ships with an on-chain component — a provenance manifest, a stablecoin rail, a verifiable-inference layer, anything. If none do, the "AI × crypto" thesis gets its cleanest natural experiment: a crypto-holding, AI-shipping company that chose, studio after studio, not to touch a chain.
My forecast is that it won't, and that the gap between $1.79M and the $100M treasury keeps widening. The vulnerability isn't in the products. It's in the narrative that prices them — and narratives don't survive the arrival of real numbers.
$1.79M is a real number.