The history of internet monetization is a history of indirection. You don't pay for Google — you pay with your attention, which Google sells to advertisers, who pass the cost to you through higher product prices. You don't pay for Facebook — same mechanism, different feed. The subscription model (Netflix, Spotify, SaaS) was supposed to be the correction: pay directly for what you use. But subscriptions have their own distortion. You pay the same whether you use the product every day or once a month. The heavy user subsidizes the light user, or more precisely, everyone subsidizes the median, and the product optimizes for the median, and the median is no one in particular.

There is a better model emerging, and it is built on tokens.

The Token as Unit of Value

In the context of large language models, a token is a chunk of text — roughly three-quarters of a word in English. When you use an AI service, you consume tokens: input tokens (your prompt) and output tokens (the response). The cost of serving your request is directly proportional to the number of tokens processed. This creates something that has been surprisingly rare in digital economics: a natural unit of value that is both measurable and meaningful.

The significance of this is hard to overstate. Most digital products have no natural unit of consumption. What is one unit of "using Gmail"? One email? One megabyte of storage? One login session? The arbitrariness of the unit forces artificial pricing — per seat, per month, per tier — that always maps poorly to actual value delivered. Tokens dissolve this problem. The cost is the computation, the computation is the tokens, and the tokens are directly correlated with the value received by the user.

The first time I saw a per-token pricing model, I felt the same clarity I felt when I first understood how electricity billing works. You pay for kilowatt-hours. Not "electricity subscriptions." Not "power tiers." Kilowatt-hours. A unit so natural it barely requires explanation.

What Crypto Got Wrong

It is worth pausing to address the elephant in the room. "Token economics" sounds like it belongs in a 2021 crypto whitepaper, sandwiched between a yield farming diagram and a governance DAO proposal. This association is unfortunate because it obscures a genuinely important idea.

What crypto got wrong was not the concept of tokenized value exchange. What crypto got wrong was the detachment of the token from any underlying utility. Most crypto tokens were, in economic terms, pure speculative instruments. Their value was circular: the token was valuable because people believed it would become valuable. There was no natural demand curve because there was no natural consumption. You didn't need the token to do anything except hope someone else would pay more for it later. This is not economics. It is musical chairs with a ledger.

The token economy I'm describing is the opposite. An AI token is consumed in the act of producing value. It cannot be hoarded, speculated on, or resold. Its price is set by the marginal cost of computation plus a margin, and its value is realized immediately in the form of useful output. The demand for tokens is derived demand — it exists because the output is valuable, not because the token itself might appreciate. This is the difference between a commodity and a collectible, and it matters enormously.

The Marginal Cost Revolution

The economics of token-based pricing are governed by marginal cost curves that behave differently from traditional software. In SaaS, the marginal cost of serving one additional user is approximately zero — the software is already built, the servers are already running. This is why SaaS margins are so extraordinary (70-90%)¹ and why the pricing is necessarily detached from the cost of service.

In token economics, the marginal cost is real and non-trivial. Each token consumed requires GPU computation, memory, energy². The marginal cost curve for an AI provider looks something like:

where $q$ is token volume, $c_0$ is the base cost per token, $c_1$ captures scaling inefficiencies, and $\beta > 0$ reflects the degree to which costs increase at scale (due to GPU saturation, thermal throttling, memory bandwidth limits). For most current providers, $\beta$ is small but positive — costs per token increase slightly at very high volumes.

This creates a natural pricing model: charge a percentage on top of token cost. If the base cost of processing a request is $c$ tokens at price $p$ per token, the platform charges $(1 + m) \cdot c \cdot p$ where $m$ is the margin. The user pays proportionally to value received. The provider's incentive is to reduce $p$ (through better hardware, more efficient models) because lower prices increase volume, and percentage margins on higher volume yield more absolute revenue.

The beauty of percentage-on-consumption is that it aligns every incentive. The provider wants the service to be cheaper (more usage). The user wants the service to be cheaper (more value per dollar). The platform wants both (more transactions to take a cut of). Compare this to advertising, where the provider wants you addicted, or subscriptions, where the provider wants you enrolled but inactive.

Price Elasticity and the Expansion of Use

The most exciting property of token economics is the price elasticity of demand. As the per-token cost drops — and it is dropping rapidly, roughly $50\%$ per year³ — new use cases become economically viable. Tasks that were too expensive at \0.01$ per 1K tokens become trivial at \0.001$ per 1K tokens. The elasticity coefficient $\varepsilon$ for AI tokens appears to be significantly greater than 1:

Empirical data from API providers suggests $\varepsilon \approx 2$ to $3$⁴, meaning a $10\%$ price decrease generates a $20$-$30\%$ increase in consumption. This is the signature of a market with enormous latent demand — people and businesses that would use AI extensively if it were cheaper, and who scale usage aggressively as prices fall.

The Platform Layer

The business model that emerges from this is what I think of as the "metered platform." You build a product — a writing tool, an analytics dashboard, a code editor, a research assistant — and behind the scenes, every AI-powered feature consumes tokens. The user pays for the product, and the cost of the product scales with how much AI they actually use. Light users pay less. Heavy users pay more. Everyone pays roughly in proportion to value received.

The platform operator's margin is a percentage of token throughput: transparent, predictable, and aligned with user value. This is structurally similar to how payment processors work (Stripe takes $2.9\%$ of every transaction) or how cloud providers work (AWS charges for compute consumed). It is, I believe, the natural pricing model for the AI era, and it will eventually displace both advertising and flat-rate subscriptions for any product where AI is a significant component of the value delivered.

The era of "pay \20$/month for unlimited AI" is already ending. Usage-based pricing is not a retreat — it is a maturation. The token is the kilowatt-hour of intelligence, and we are just beginning to build the grid.

The Decentralized Grid

But what kind of grid? The assumption embedded in the previous section is that the platform is centralized — that you use OpenAI's models, or Anthropic's, or Google's, and the platform operator routes your requests to one of these providers. This is the current reality, but it is not the inevitable one. And the forces pushing against it are accelerating.

Open-source models are getting very good. Not just good enough for hobbyists — good enough for production. Llama, Mistral, DeepSeek, Qwen⁵ — the gap between the best open model and the best closed model shrinks with every release cycle. More importantly, the hardware required to run these models is becoming accessible. A serious inference server can be built for the cost of a used car. A rack in a colocation facility costs a few hundred dollars a month. Cloud providers rent GPU instances by the hour.

What this means is that anyone can become an AI provider. Not at the scale of OpenAI — but at the scale of a specialty shop. A small operator who has fine-tuned a model on legal documents, or medical imaging, or code review for a specific framework, or translation for a particular language pair. The model might be a 7B parameter open-source base with LoRA adapters⁶ trained on proprietary data. It runs on two GPUs in a closet, or on a reserved instance on AWS, or on a rented node in a data center. The total cost of operation is a few hundred dollars a month. The model is genuinely good at its narrow task — perhaps better than the general-purpose frontier models, because specialization is a powerful force.

I keep reaching for the phrase "mom-and-pop model shops," and I think the analogy is more precise than it sounds. Before Walmart and Amazon, the American economy ran on specialty retail. The butcher knew meat. The baker knew bread. The hardware store owner knew which screw you needed for that particular hinge. These businesses were small, specialized, and embedded in a community of trust. They were displaced by scale economics — it was cheaper to buy everything in one place, even if the expertise was thinner.

The AI economy may reverse this pattern. Scale economics still apply to training — only a few organizations can afford to train frontier models. But scale economics apply much less to inference, especially specialized inference. Running a fine-tuned model on a narrow domain is cheap. The expertise is in the data and the fine-tuning, not the infrastructure. And the quality advantage of specialization is real: a model trained on ten thousand hours of contract law will outperform GPT-5 on contract review, just as a butcher who has been cutting meat for thirty years will outperform a Walmart deli clerk.

I imagine a future where my AI assistant doesn't route every request to the same provider. It calls my friend's legal-analysis model for contract questions. It pings a small outfit in Estonia that runs the best translation model for Baltic languages. It uses a local model on my own machine for private queries. Each call costs fractions of a cent, settled instantly, no subscription required.

The Discovery Problem

But this vision immediately collides with a practical problem. If there are thousands of specialized model endpoints scattered across the internet — running on personal servers, cloud instances, data center racks — how do you find them? How do you know which one is good? How do you pay them?

The subscription model cannot work here. You cannot have a subscription to every mom-and-pop model shop you might ever need. The administrative overhead alone would be absurd — fifty API keys, fifty billing dashboards, fifty terms of service. And you don't know in advance which endpoints you'll need. Your AI assistant encounters a Latvian legal document and needs a Latvian legal translation model right now. It can't stop and ask you to sign up for a monthly plan.

What's needed is three things:

1. A registry — a place where model operators can list their endpoints, describe their capabilities, and set their prices. 2. A discovery protocol — a way for AI agents to search this registry, find the right endpoint for a given task, and evaluate its quality. 3. A payment mechanism — a way to pay per call, instantly, without pre-registration, without subscriptions, without invoices.

The first two are hard engineering problems. The third is a problem that has been solved, though most people don't know it yet.

HTTP 402: The Missing Status Code

When the architects of HTTP designed the protocol in the early 1990s, they reserved a status code that was never implemented: 402 Payment Required⁷. The specification noted it was "reserved for future use." For thirty years, it sat there — an empty slot in the protocol, waiting for the internet to figure out how to handle native payments.

Coinbase has filled that slot.

The protocol is called x402⁸, and its elegance lies in its simplicity. It works like this: a client makes a request to an API endpoint. If payment is required, the server responds with HTTP 402 and a machine-readable payment envelope — the amount, the currency (USDC), the network (Base, Solana), and the recipient address. The client constructs a cryptographic payment proof, includes it in an HTTP header, and retries the request. The server verifies settlement on-chain and serves the response.

No API keys. No OAuth flows. No billing dashboards. No invoices. No subscriptions. Just: request, pay, receive. The entire transaction settles in the time it takes to process a normal HTTP request.

The choice of USDC — a dollar-denominated stablecoin — is critical. This is not speculative crypto. USDC is pegged to the dollar, fully reserved, and regulated⁹. It has the stability of fiat currency with the programmability of blockchain. A model operator in Estonia can receive payment from a client in São Paulo without either party involving a bank, a payment processor, or a foreign exchange desk. The payment clears in seconds and costs fractions of a cent in transaction fees.

HTTP 402 sat dormant for thirty years because the internet had no native payment layer. Credit cards require merchant accounts. Bank transfers require correspondent banking relationships. PayPal requires both parties to have accounts. USDC on a fast L2 chain requires nothing except a wallet address. The missing payment layer was not a software problem. It was a money problem. And stablecoins solved it.

The Bazaar

x402 solves the payment problem. But payments without discovery is like having a credit card in a city with no street signs. You need to be able to find the shops.

This is where the Bazaar¹⁰ comes in — a discovery layer built on top of x402. It functions as a machine-readable catalog of every x402-enabled endpoint on the internet. Model operators register their services: what the model does, what it costs per call, what inputs it accepts, what quality guarantees it offers. Clients — whether human developers or autonomous AI agents — can search the Bazaar, filter by capability and price, and start making paid requests immediately.

The registration is free. The pricing is set by the operator. The discovery is open. Think of it as a Yellow Pages for AI endpoints, except the phone call also handles the payment.

The implications are significant. A researcher in Nairobi who has trained the best Swahili-English translation model in the world can register it on the Bazaar, set a price of $0.001 per request, and start earning money from anyone on the internet who needs that capability. No venture funding. No app store approval. No payment processor integration. Just a model, an endpoint, and a wallet address.

The Agentic Economy

The real transformation happens when the customer is not a human but another AI. This is where the economics become genuinely new.

Consider the workflow: you ask your AI assistant to analyze a contract written in German, compare it to US case law, and draft a response. Your assistant doesn't do all of this itself. It routes the German translation to a specialized translation model. It sends the legal analysis to a fine-tuned legal reasoning model. It uses its own general capabilities for the synthesis and drafting. Each sub-call is a paid x402 transaction. The total cost might be $0.03. The entire workflow completes in seconds.

Your assistant is not just a model — it is a broker. It knows your preferences, your budget constraints, your quality requirements. It discovers specialized endpoints on the Bazaar, evaluates their price-performance trade-offs, and routes sub-tasks to the optimal provider. This is economic agency in the literal sense: an agent making purchasing decisions on your behalf, in real time, at a granularity that no human could manage.

where each sub-task $i$ has its own margin $m_i$, token count $t_i$, and per-token price $p_i$. The total cost is the sum of individually priced micro-transactions, each settled independently. This is the compute equivalent of a supply chain — specialized producers, coordinated by an intelligent buyer, with payment flowing proportionally to value contributed.

We are building an economy where the transaction costs are so low that it makes sense to pay a stranger three-tenths of a cent for a single inference call. When transaction costs hit zero, the optimal firm size approaches one¹¹. Every model operator is a firm. Every fine-tuned adapter is a product. Every API call is a transaction. Adam Smith's pin factory, but the pins are tokens and the factory is the internet.

Building the Grid

So the grid is not a centralized utility. It is a distributed network of specialized operators, connected by a payment protocol that settles in seconds, discoverable through a machine-readable catalog, and orchestrated by AI agents that act as intelligent brokers. The token is still the unit of value — but the token now flows through a genuinely decentralized marketplace, not a walled garden.

This is what crypto was supposed to build. Not speculative tokens detached from utility, but a programmable payment layer that enables new kinds of economic activity. The irony is that the killer app for blockchain was never decentralized finance or digital art. It was paying for a function call. The most boring possible use of a distributed ledger — and therefore, almost certainly, the most important one.

The mom-and-pop model shop is coming. The butcher, the baker, the fine-tuned legal reasoning model maker. They just need a way to hang a shingle on the internet and accept payment for their craft. The tools now exist to let them do exactly that.