№ 024Deep DiveApr 30 20267 min read

StablecoinsMarket Structure

Rails Don't Capture Rent

Crypto and AI are proving the same thing: building the infrastructure is not the same business as monetizing it.

By Akram Ayyash

Tether made over $10 billion last year. The Ethereum Foundation holds roughly $1 billion in its treasury. One built on top of the rail. The other maintains the rail itself. That gap tells you almost everything about where the economics actually sit in crypto. It is about to tell you the same thing about AI.

Infrastructure exists to create value for its users. That is its purpose. And it follows that users, not the builders, end up capturing most of that value. The money flows to whoever removes the last mile of friction between the technology and the workflow it improves. Tether does not run Ethereum. It removes the friction between dollars and crypto rails. That is a $10 billion business. Running the rail is not.

The Data Is Already In

In its January 2026 attestation, Tether reported 2025 net profit exceeding $10 billion on $186.5 billion of USDT in circulation. In its 2025 10-K filed February 12, 2026, Coinbase reported $7.18 billion of net revenue and $1.26 billion of net income. In its February 2026 earnings release, Circle reported $2.75 billion of 2025 revenue, of which $2.64 billion, roughly 96 percent, came from reserve income. Circle lost $70 million for the year.

Start with Circle. It made $2.75 billion in revenue and 96 percent of it came from reserve income. That means the entire business is basically earning interest on other people's dollars. Circle built a compliance wrapper around what is functionally a money market fund, put it on crypto rails, and called it a stablecoin business. And it still lost money. Tether does roughly the same thing with a fraction of the overhead and made $10 billion. Coinbase monetizes the other side: distribution, custody, execution, and the institutional trust required to move traditional money into crypto. All three depend on crypto infrastructure. None of them is a protocol business.

Joel Monegro published the fat protocol thesis in 2016, arguing that in crypto, unlike the traditional internet, value would accrue at the protocol layer rather than the application layer. The original insight was a comparison: HTTP does not make money, but Google does. Blockchain protocols, he argued, would be different. A decade later, it is starting to feel like they are not. Tether, Coinbase, and Circle look a lot like Google in the original analogy: application-layer businesses capturing the economics while the underlying protocol remains socially essential and economically thin. The users are the ones with businesses to run, frictions to remove, and problems to solve. Structurally, they can capture value because they are closest to the workflow. Infrastructure providers capture value too, but nowhere near as much.

AI Is Running the Same Experiment Faster

As of January 2026, OpenAI said revenue grew from $2 billion in 2023 to $6 billion in 2024 and more than $20 billion in 2025. Impressive top line. But OpenAI has reportedly never turned a profit and was projecting $5 billion in losses for 2024 as of September 2024 reporting. Anthropic, which was approaching a $900 million annualized revenue run rate by late 2024, projected it could lose as much as $5 billion in 2025 as it scales compute and talent. Revenue is growing. Cash is on fire.

The model builders are not just failing to capture the value their technology creates. From a cash flow perspective, they are destroying value. Meanwhile, a firm paying $20 per month for ChatGPT Plus or Claude Pro can save multiples of that if it redesigns how work gets done. McKinsey's 2025 global survey sharpens this: as of November 2025, 88 percent of respondents said their organizations were regularly using AI, but only 39 percent reported EBIT impact at the enterprise level. The bottleneck is not access to the model. It is the organizational work required to make the model save real money. For the model company, value capture is limited. For the user, the sky is the limit.

The AI industry seems to be learning this lesson from crypto, and leapfrogging. Anthropic is hiring Solutions Architects and Partner Solutions Architects whose job descriptions read less like research lab postings and more like enterprise consulting: guide customers from discovery to deployment, integrate the model into real systems. OpenAI's enterprise language now emphasizes helping organizations build, deploy, and manage AI coworkers. Instead of waiting for customers to figure out deployment on their own, the way crypto waited and is still waiting, the labs are trying to own the deployment layer themselves. In my view, that is an admission. Producing a model is not a business. Deploying the model inside a real company and saving that company real money, that is a business.

Deployment Is the Product

I have sat across the table from enterprise buyers evaluating blockchain infrastructure. Not once has the conversation started with consensus mechanisms or block times. It starts with reconciliation costs, settlement delays, and operational risk. The buyer does not want a chain. The buyer wants a solved problem that happens to run on one.

Of course the chain has to work. It has to have fast consensus, it has to be scalable, and it has to settle cheaply. In 2026, that is table stakes. The market is not just demanding those things but expecting them. The key question businesses are asking about blockchain today is not whether the chain works. It is how the chain can work to make their business run better. The same is true in AI. We expect the models to be fast, coherent, and factual. In 2026, that is becoming table stakes as well. The question has moved from "can the model do this?" to "how do I embed this model into my operations and save money?"

That shift is what makes blockspace and model inference start to look like commodities. Every Layer 1 is converging toward the same destination: fast, scalable, low-cost. The differentiation that existed three years ago is compressing. AI models are on the same trajectory. GPT-4 was a moat in early 2024. By late 2025, Claude, Gemini, Llama, and a dozen open-source alternatives had closed the gap on most enterprise tasks. The infrastructure is not fully commoditized yet, but it is trending that way. And the closer it gets, the more pricing power migrates away from the infrastructure layer and toward whoever deploys it.

In crypto, the native token is literally called a gas token. The analogy is useful. Gas is essential; nothing moves without it. But nobody builds a durable business on the thesis that gas prices will sustain a premium forever. The businesses that captured the economics in crypto (Tether, Circle, Coinbase) are the cars, not the gas. They consume blockspace to deliver something a customer actually pays for. The same structure is forming in AI. The model is the gas. The deployment, the workflow it improves, the friction it removes, the headcount it replaces, is the car. And as blockspace and model inference trend toward commodity pricing, the rent keeps migrating to whoever is furthest from the commodity.

What to Watch

So what should we pay attention to?

In crypto, watch whether more infrastructure firms start packaging treasury, settlement, and tokenization workflows rather than chain features. Visa launched a stablecoin advisory practice and U.S. USDC settlement in 2025, a payments company that understood, before most crypto firms did, that the money is in the implementation layer. The question is who follows.

In AI, watch whether model providers keep moving toward solutions architecture, enterprise controls, and partner ecosystems. That motion is the clearest signal that they understand the endpoint.

The honest question I do not have a clean answer to: can the labs vertically integrate? If OpenAI or Anthropic become both the gas and the car, if they own the model and the deployment, the commodity thesis breaks. These labs have more capital and more talent than almost any startup in history. Whether they pull it off comes down to the investors and the management teams making those bets. I do know that every previous infrastructure wave, telecom, cloud, internet, rewarded the deployment layer more than the pipe.

I believe the pattern holds until someone proves otherwise. And if everyone now understands that workflow redesign is the economic choke point, the next question is who owns that layer: the infrastructure builder, the systems integrator, or the enterprise customer itself.

Drop me a line:

What do you think? Do you like this? Do you not like this? I would love to hear your thoughts, so please reach me at akram@span.blog

Enjoyed this? Subscribe for the weekly deep dive, or drop me a line at akram@span.blog.