The most consequential number in Adobe's most recent earnings cycle was one the company did not print. It was not revenue. It was not net new subscribers, not the annualized recurring revenue figure that analysts have recited like a liturgy since 2013. It was a ratio — generative credits consumed, divided by generative credits purchased.
Everyone in this industry has been building that meter for three years. Adobe shipped one, quietly, inside a walled garden, and almost nobody in decentralized AI noticed that the hardest part of the problem had just been solved by a company that has never written a line of Solidity.
Firefly no longer behaves like a feature. It behaves like a utility. Every Creative Cloud plan now ships with a monthly allotment of generative credits; when the allotment runs dry, the subscriber buys more. That is not a product tier. That is a tariff. And a tariff is the moment a capability stops being a marketing bullet and becomes an accounted commodity — with a unit, a price, a basis risk, and a line in the income statement.
The earnings report is the occasion. The meter is the story.
I should state my position plainly, because it shapes everything that follows. I am not a shareholder, and I do not trade the stock. I spent the last five years inside governance design — first auditing the machinery of DeFi protocols, now architecting voting systems for DAOs that manage real treasuries. When I look at a metering scheme, I do not see pricing. I see a constitution. Generative credits are a constitution for who gets to create, how much, and at whose expense. That is why a creative-software earnings call deserves attention from people who claim to be building the infrastructure of the open web.
Here is the backdrop in a single breath. Adobe is a company that has spent most of the past four years valued somewhere between a hundred and fifty and two hundred and fifty billion dollars, generating roughly twenty billion in annual revenue, spending close to a fifth of it on research and development, and defending a position in professional creative software that no competitor has seriously threatened in two decades. Its AI strategy rests on a model family trained predominantly on its own licensed stock library, a provenance standard it co-founded called Content Credentials, and a distribution channel — Photoshop, Illustrator, Premiere, InDesign — that reaches essentially every professional designer on earth.
The bull case writes itself. Office 365 added Copilot, raised prices, and expanded seats. Adobe could do the same: bolt AI onto an installed base, lift average revenue per user, and call it transformation. The bear case writes itself too. Generative models collapse the skill floor. When anyone can produce a competent image with a sentence, the moat around professional tooling erodes, and the AI inference bill lands on the vendor's margin rather than the customer's.
Both narratives are comfortable. Both are also slightly beside the point. What actually needs explaining is what a generative credit is, economically — and whether the thing crypto has spent three years claiming to build is the same thing, or a shadow of it.
Start with the unit. A generative credit is a claim on a fixed quantum of GPU time, denominated in a currency that only one issuer honors. Structurally, it is a gas token. It is bounded by a scarce resource, consumed per operation, priced by a monopoly, and non-transferable outside the network that issued it. The analogy is not decorative. When I audited Curve Finance's governance mechanics in 2020, I pulled more than four hundred thousand lines of simulation logs to model how voting power concentrated. I found the same structural signature in a dozen other protocols: once you introduce a metered unit, the unit becomes a political object. Credits are no different. They are votes on who gets to generate.
A gas token inside a walled garden is still a gas token; it simply has no secondary market to discipline its price.
That absence matters more than the price itself. Ethereum's fee market is brutal and legible — you can see the base fee in real time, you can route around congestion, and you can exit to a rollup or an alternative chain. Adobe's credit market has none of those properties. There is no observable spot price, no congestion signal, no exit. Subscribers cannot arbitrage a cheap video generation against an expensive one, because there is no venue in which to do it. The subscriber absorbs whatever basis risk the issuer chooses to retain.
And the basis risk is not small. This is where the engineering gets interesting, and where the earnings narrative tends to go soft. Not all generative operations cost the same amount of compute. A vector recolor is a rounding error. A background removal is trivial. A two-second clip of plausible video with consistent lighting and subject identity is a different order of magnitude — possibly two. Latent video diffusion is not a heavier version of image diffusion; it is a different cost regime, because you are denoising across a temporal axis, and the memory bandwidth requirements scale roughly with frames multiplied by resolution multiplied by the parameter count you are forced to keep resident.
Adobe prices these operations in the same credit denomination, with adjustments that are directionally correct but not obviously calibrated to marginal cost. That is a deliberate choice, and I think it is the right one commercially. Uniform credit pricing across wildly heterogeneous compute costs is not a pricing model; it is a hedge against the vendor's own uncertainty about its inference economics. It cross-subsidizes the cheap operations to make the expensive ones psychologically affordable, and it preserves optionality in case video generation turns out to be a loss leader for a decade.
But a hedge has a counterparty, and in this case the counterparty is the subscriber. If you buy a plan primarily to produce short video, you are subsidizing someone else's typography. That is not a scandal — it is how every bundled utility in history has worked. It is, however, an argument that has been made in crypto for eight years, and it deserves to be stated with the same rigor on this side of the fence.
The more consequential structural question is what the credit system does to the creator's relationship with the work. On paper, credits democratize access. In practice, metering always concentrates. I have watched this happen at protocol scale. Emissions flow to whoever can afford to hold the position longest. Credits flow to whoever can afford the highest tier. The mechanism is neutral; the distribution is not. Metering does not remove gatekeeping; it relocates it from capability to balance sheet.
Now consider the asset Adobe holds that the market consistently undervalues, and which nobody writing about the earnings call seems to mention at all: Content Credentials.
Content Credentials is the consumer-facing implementation of C2PA, a provenance specification that binds cryptographically signed metadata to a media file — who created it, with what tool, and what was changed since. Adobe co-founded the effort. It is embedded in Photoshop, Lightroom, and Firefly output. On a balance sheet it is worth approximately nothing. In a courtroom, it may be worth more than the model weights.
The reason is that the central legal question hanging over the entire generative industry is not whether models can generate; it is whether anyone can prove what went into them. Getty Images' litigation against Stability AI is the loudest instance, but it is one instance of a category. Discovery in these cases is a nightmare precisely because training data provenance is a forensic problem, not a query. If every ingestion event produced a signed manifest, the question — was this image used to train this model — would collapse from eighteen months of expert testimony into a database read.
That is a ledger problem. It is the one ledger problem in the entire AI stack that is genuinely unsolved and genuinely valuable, and it is being worked on by a handful of crypto-native teams — Numbers Protocol anchoring C2PA manifests on-chain, Story Protocol formalizing IP registration and licensing as programmable primitives — while the dominant creative software vendor in the world runs a federated trust list instead.
I do not think that choice was stupid. Federated trust is faster, cheaper, and legally cleaner to govern than a public chain, and Adobe answers to a corporate board, not a token-holder vote. But it is a governance choice, and governance choices have consequences. A federated list is amendable by its stewards. A public ledger is amendable by nobody. Silence is the only consensus that never forks — and Adobe has chosen, deliberately, to remain the only voice that matters in its own trust graph.
The parallel to the data-availability debate is almost exact, and it is where the crypto side of this conversation has been getting the framing wrong for two years. I have argued repeatedly that dedicated DA layers are overhyped — that the overwhelming majority of rollups do not produce enough data to justify a dedicated availability market, and are paying a premium for an abstraction they will never saturate. I stand by that. But I have also learned to be precise about what I am dismissing.
The claim that dedicated data availability is overhyped as a market does not imply it is unnecessary as a primitive. The same is true of licensed training corpora. Most models do not need one. The frontier ones, the ones that will be litigated, absolutely do. Adobe Stock is a dedicated DA layer for training data, and it is the single most defensible asset the company owns in this cycle — more defensible than the model, more defensible than the interface, and completely invisible in a revenue log.
Which brings me to the competitive layer, and to what I think is the most misread dynamic in the entire story. The threat to Adobe is not Midjourney, and it is not Canva. It is the shape of its own pricing architecture.
A company that earns the majority of its revenue from seat licenses cannot ship a product priced purely by consumption without cannibalizing the thing that made it. This is the innovator's dilemma expressed as a constitution: the incumbent's pricing model was frozen at a moment when generation cost nothing, and amending it requires a supermajority of internal stakeholders whose compensation depends on the old text. When I helped design a quadratic voting mechanism for a five-million-dollar community treasury in 2024, our design constraint was never mathematical. Identity verification, sybil resistance, and the eligibility snapshot were all solvable. The hard part was convincing three core developers and a handful of large delegates that a mechanism that reduced their relative influence was nonetheless worth adopting. We shipped it. Participation rose roughly thirty percent. But the win came from trust built in a small room, not from the elegance of the formula.
Adobe has the formula. It does not obviously have the room. That is the same failure mode Tezos promised to solve in 2017 with self-amending governance, and the same failure mode that made me spend six months of my teenage years reading whitepapers instead of sleeping — the belief that a protocol could rewrite its own constitution without a rupture. To govern the future, we must debug the present, and the present constraint on Adobe is not compute. It is the fact that its pricing model is a constitutional document with no amendment procedure.
Here is where I diverge from nearly everyone in my own corner of the industry.
The prevailing crypto thesis about AI is that compute is the choke point, and therefore compute must be decentralized. Bittensor emits tokens to subnet miners competing on model quality. Akash runs reverse auctions for GPU capacity. Render denominates rendering jobs in a burn-and-mint equilibrium. io.net and a dozen others aggregate idle silicon. The intellectual energy is enormous, and the premise is that whoever controls the FLOPs controls the future.
Adobe is a natural experiment against that premise, and it fails the thesis. Here is a company with no token, no distributed validator set, and no permissionless participation, and it has successfully metered AI consumption across a customer base in the tens of millions — with pricing power intact, without a governance crisis, and without anyone needing to agree on a consensus mechanism. Decentralization was never the prerequisite for metering. It is the prerequisite for contestability, which is a different and more expensive good.
The binding constraint on AI commercialization turned out not to be compute at all. It is liability. It is the question of who owns the output, who paid for the input, and who is on the hook when a generated image resembles a photograph that someone licensed for a different purpose. That is not a FLOPs problem. It is an attribution problem, and attribution is the one thing a public ledger does natively and a database does not do convincingly.
So the decentralized AI sector has spent three years optimizing the bottleneck that Adobe proved you can route around, while leaving the bottleneck that Adobe cannot route around almost entirely untouched. The code is law, but the humans are the bug — and the bug in this case is a licensing contract, not a kernel.
The honest counterargument is that I am comparing an incumbent with a twenty-year distribution advantage to startups with neither distribution nor a legal department. That is fair, and it cuts deeper than I would like. Federated trust works for Adobe because Adobe is the trust anchor. Numbers Protocol cannot be the trust anchor for Getty, and Story Protocol cannot compel a licensing counterparty to appear. This is the cold reality that every provenance startup eventually collides with: the ledger does not create the counterparty. It only makes the counterparty's commitments verifiable once they exist.
I still think the primitive is right and the timing is early, which is a different statement from the primitive being wrong.
What I would watch over the next eighteen months, and what I think actually determines whether this earnings cycle was a footnote or an inflection, is not the credit consumption rate. It is whether Adobe exposes a Firefly endpoint with per-call pricing that a third party can settle. The day a design agency's procurement system pays for a generation through a metered API rather than an annual seat, the walled garden acquires a door — and the meter stops being a private tariff and starts becoming a market.
There is a second signal, quieter and more decisive. Content Credentials is currently a signature. If Adobe ever anchors those manifests to a public, neutral, append-only substrate — even a permissioned one with public verifiability — then training data provenance stops being a discovery process and becomes a query, and the entire litigation overhang across the industry compresses. That moment would do more for decentralized infrastructure adoption than a decade of compute-market incentives, and it would arrive not as a philosophical victory but as a legal convenience.
In the void, we found our own gravity. The walled garden found its own meter instead, and it works. The question worth sitting with, as the tape chops sideways and everyone waits for direction, is not whether decentralized AI can out-compute Adobe.
It is whether anything we build will ever be as useful as the thing Adobe built by accident — a bill that tells the truth about what a generation costs.
