I want to begin with a number that has been sitting in my drafts folder for three weeks, because I could not decide whether it was a death sentence or a mirror.
According to an Andreessen Horowitz survey, roughly two percent of American households pay for AI services. Two percent. There are about 131 million households in the United States, which means somewhere near 2.6 million of them have handed over a credit card for a chatbot, an image generator, a coding assistant, or whatever vertical wrapper the venture cycle has blessed this quarter.
The number is imprecise. a16z published no sample size, no methodology, no definition of "AI service," and no clarity on whether bundled subscriptions were counted. I am working from a single industry brief and a knowledge cutoff, and I will flag that limitation plainly rather than dress it up in false confidence. But even granting generous error bars, the figure lands like a cold hand on the back of the neck.
Because I have seen this number before. I have seen it wearing a different logo, in a different market, and I have watched what happens to the people who refuse to read it.
Let me be precise about what the two percent actually is, because the sloppiness of the reporting around it is itself a data point.
The dominant consumer AI subscription sits at roughly twenty dollars a month. ChatGPT Plus, Claude Pro, Gemini Advanced, Perplexity Pro — they cluster around the same psychological anchor as Netflix and Spotify, and that is no accident. The pricing teams looked at the entertainment streaming stack and copied the number consumers have already been trained to tolerate.
But the cost structure underneath is nothing like streaming. AI inference is expensive in a way a cached video stream is not. Every free conversation burns compute someone must pay for, and the free tiers — ChatGPT's, Gemini's, Meta AI's, Copilot's, Google One AI Premium's bundled access — are not loss leaders in the ordinary sense. They are structural subsidies funded by balance sheets that dwarf the entire consumer AI subscription market.
Run the arithmetic. 2.6 million households at twenty dollars a month is about 630 million dollars in annualized revenue. At thirty dollars, call it 940 million. That is the entire direct-to-consumer AI subscription business in the United States, generously estimated — a rounding error against the tens of billions flowing through enterprise contracts, API calls, cloud inference, and advertising. The consumer market is the storefront. The enterprise market is the vault.
To understand why the consumer number is so small, you have to look at where the AI revenue actually goes. The industry's real balance sheet is written in enterprise contracts — legal review, customer support, code generation, financial analysis — where a single seat can cost thirty dollars a month and the buyer is a company that measures return in hours saved. It is written in API calls, where developers pay per token and never see a user interface at all. It is written in cloud inference bills and in advertising, where AI is a feature of a product that was already monetized. The consumer subscription is the visible tip. The iceberg is corporate.
This is where the crypto reader should feel a jolt of recognition, because we have lived this exact story. For a decade we told ourselves that retail adoption was the metric that mattered, that the consumer would eventually pay directly for sovereignty, that the wallet would become the new browser tab. Then we looked up and discovered the actual money was never in the retail wallet. It was in the exchange spread, the infrastructure fee, the bundled distribution deal. The consumer was the demo. The enterprise was the customer. We did not notice the two percent until it was already ours.
So let me do what I actually do — take the number apart and see what it is made of.
If two percent of households pay, then the marginal consumer AI company is not building a subscription business. It is building a customer acquisition funnel for somebody else's bundle. And the somebody else is almost always a company that already owns a default surface: the operating system, the search bar, the productivity suite, the social feed. Microsoft bundles Copilot into Windows and Office. Google folds Gemini into Search, Android, and Workspace. Meta gives it away to keep the feed sticky. The standalone consumer AI app, no matter how elegant, is fighting for a sliver of attention the giants can subsidize into oblivion.
I recognize this shape because I have spent the last two years auditing the same pattern in crypto, and the parallel is not metaphorical. It is structural.
Take Layer 2. The entire scaling narrative of the past four years rests on the promise of decentralized sequencing — the idea that the rollup's ordering layer would eventually be shared, permissionless, credibly neutral. I have read the roadmap documents. I have sat on the governance calls. And the honest count is that "decentralized sequencing" has been a slide deck for two years while the actual sequencers running production traffic remain single, centralized nodes operated by a handful of teams. The value accrues to whoever controls the ordering — the invisible layer — not to the application on top. That is the same law that governs AI: the application is visible, the infrastructure is profitable. We chart the code, but the soul chooses the path, and the path the code walks always leads toward the layer that collects the fee.
Take the mining economy I have tracked since the fourth halving. When block rewards collapsed, the honest analysts admitted what the optimistic ones would not: hash power is consolidating toward a small number of pools, and a decentralization consensus that depends on three operators is a consensus in name only. The miners who survive are not the romantic solo operators. They are the ones with cheap power contracts and the scale to bundle. Again — the visible layer is the ideology, the invisible layer is the business.
Then there is the stablecoin yield complex, where I think the two-percent lesson turns genuinely dangerous. Products like sUSDe promise a yield that looks like consumer-friendly income, and the machinery underneath is a maturity mismatch with stacked risk — a structure that performs beautifully in a bull market and unwinds first when sentiment turns. These instruments are the consumer AI subscription of crypto: a product priced for a world that behaves, sold to people who need it to behave, and structurally fragile precisely when it matters most. The two percent tells us that even in the most hyped technology cycle of the decade, the consumer will not pay for a promise alone. They pay for something that disappears into their life.
Based on my audit experience — six months inside the security models of failing L1s, three of them quietly centralized in their consensus — I have learned that the gap between the narrative and the machine is almost always wider than the marketing admits. The same gap runs through the AI economy. The demo is decentralized and sovereign. The deployment is a cluster of servers behind a load balancer with a single company's name on the invoice.
And there is a deeper current here that I have spent this past year writing about inside a governance DAO: the question of sovereign data rights. When the value of an AI service is captured invisibly — in the model weights, the telemetry, the behavioral exhaust of every free conversation — the consumer is not merely failing to pay. The consumer is paying with something they cannot see on any invoice. Their data is the subscription fee they never agreed to. This is why the two percent matters to anyone who cares about digital autonomy: a free product that monetizes your behavior is not free, and a paid product that cannot survive on its price is not sovereign.
I have watched this play out from the inside. When I wrote a manifesto on sovereign data rights inside an AI-governance DAO, three regulatory bodies across Europe and Latin America cited it — not because the ideas were novel, but because the gap between what these systems claim to protect and what they actually collect had become impossible to ignore. Regulation follows the invoice. When the invoice is invisible, so is the harm.
Here is the insight I think almost everyone is missing, and it is why I keep circling back to this number. The two percent is not a story about AI at all. It is a story about the difference between a technology people admire and a technology people depend on. Streaming reached sixty-plus percent household penetration because it became furniture — invisible, ambient, unremarkable. AI has not become furniture. It has become a demo people watch once, admire, and then abandon for the free tier. Crypto, for all its ideological fire, has the same problem: it built a cathedral that most people visit as tourists and never move into. The tourists take photographs. They do not pay rent.
Now let me argue against myself, because the comfortable reading of the two percent is the one I distrust most.
The comfortable reading is: "AI is a bubble, consumers don't want it, the whole thing collapses." That reading is lazy, and it is wrong, and I have watched too many smart people make exactly this mistake with crypto.
The uncomfortable reading is that two percent is not a failure state. It is a stable equilibrium. The consumer subscription is a small, durable, premium niche — the way vinyl records and artisanal coffee are durable niches — while the real economy of the technology lives somewhere the household never sees. The AI revolution will not be monetized at the kitchen table. It will be monetized in the API call, the enterprise seat, the cloud bill, the ad impression. The consumer is the visible two percent. The invisible ninety-eight percent is where the margin lives.
And this is the blind spot that should worry every crypto founder reading this. We keep benchmarking ourselves against the wrong comparison. We compare crypto consumer adoption to AI consumer adoption and feel vindicated — "look, they're struggling too." That is the wrong frame. The right frame is the one the giants already understand: if you cannot become the default surface, you become the subsidized feature. The two percent is not a verdict on AI's potential. It is a verdict on the strategy of selling directly to a consumer who has already been handed a free version by a company that owns their home screen.
The second blind spot is subtler, and it is the one that keeps me up at night. Every infrastructure layer that captures value by being invisible also captures power by being invisible. The sequencer no one votes on, the pool that controls the hash, the yield structure no one fully reads — these are the two-percent problem inverted. The consumer does not pay because the value is hidden. But the value is hidden because someone made it hard to see. Sovereignty and opacity are not the same thing, and we have spent a decade confusing them. We chart the code, but the soul chooses the path — and too often we let the code choose for us while we were busy applauding.
So what do we do with a mirror?
We chart the code, but the soul chooses the path — and the path the market is choosing is not the one the pitch decks describe. The two percent is not telling us to abandon the consumer. It is telling us to stop mistaking attention for payment, and payment for trust. The next cycle will not be won by whoever ships the loudest chatbot or the flashiest rollup. It will be won by whoever builds the layer people actually live inside — the one they never notice, the one that earns its place not by asking for twenty dollars a month, but by becoming the thing without which everything else stops working.
The question for this bear market is not whether the technology is real. It is whether we have the discipline to build something a person would miss. And the answer, for now, is written in a number no one wants to read: two percent.
