Three companies, $GOOGL, OpenAI, and Circle, announced this week that they are building shared infrastructure for agentic payments. At almost the same moment, market analysts started warning that SpaceX and OpenAI going public could signal a top. Most people read those as two separate stories. They are the same one, and reading them together changes what the second one means.
The short version of the argument is this: the capital is moving because the infrastructure phase is ending, not because sentiment has peaked. Agents are getting the ability to hold and move money, the rails to do it on are being standardized right now, and the business models that will monetize all of it are resolving into a small number of recognizable shapes. That is the moment a company wants permanent capital rather than venture money. Everything below is the evidence for that reading, and the places where it could still be wrong.
Google, OpenAI, and Circle Build the Rails
The agentic payments partnership joining Google, OpenAI, and Circle is not a headline about three logos aligning for a press cycle. It is about who controls the rails when an AI agent needs to spend money on your behalf. Agentic payments, meaning systems where software acts, authorizes, and transacts without a human confirming each step, require infrastructure that looks nothing like a credit card checkout flow. A checkout flow assumes a person, a session, and a moment of consent. Agents have none of those in the same shape. What they need instead is programmable, auditable, and composable money movement, and blockchains turn out to be unusually good at that specific thing.
Matthew Ball wrote in The Metaverse that blockchains “will not become the dominant means for storing data, computing, payments, LLCs,” but that they “will become key to many experiences, applications, and business models.” Jensen Huang, from a very different vantage point, called the technology “a fundamental new form of computing.” Hold those two claims next to each other and the useful position sits in the gap between them. Not dominant, not the whole stack, but key, in the way TCP/IP is key. You do not think about it, and everything depends on it. Bitget Wallet’s onchain payments layer routing between banks, card networks, and blockchains is a bet on precisely that version of the future. Boring infrastructure, extraordinary leverage, and no requirement that anyone abandon the existing system for it to matter.
Four Business Models for AI Agents
The companion question, what an AI agent business even sells, is finally getting specific. The reporting on how agent business models are splitting four ways identifies open-source infrastructure, token distribution, SaaS, and acquisition as the live options. The four-way split is not chaos. It is terrain being carved before anyone knows which ridge line will hold.
Each one defends a different asset. Open-source infrastructure builds defensibility through ecosystem lock-in, which is slow to establish and very hard to dislodge once it is. Token distribution turns users into stakeholders, essentially issuing your own money alongside your product, which aligns incentives and imports an entirely new category of risk. SaaS is the known quantity that enterprise procurement already knows how to approve, which is worth more than it sounds when the buyer is a large company with a budget cycle. Acquisition is what happens when a larger player decides it is cheaper to buy the capability than to build it. All four will survive, because they are not really competing for the same position. The question is which one wins at which layer of the stack, and that is a question about where the switching costs land.
The Training Corpus Decides the Default Stack
Here is the piece that is not getting enough attention, and it sounds narrower than it is: the agents themselves are becoming better at Java than at Erlang. In Vibe Coding, Gene Kim, Steve Yegge, and Dario Amodei describe a real enterprise considering a migration away from Erlang, a language famous for resilient concurrent systems, to modern Java, specifically because AI coding tools perform better with Java. Not because Java became technically superior for the workload. Because the frontier models were trained on far more Java than Erlang, and the assistance gap is now large enough to outweigh the architectural fit.
Sit with the implication. Technical decisions are being shaped by the distribution of the training corpus, which is a thing no architect ever put on a decision matrix. The winning language in an AI-first development environment is not the most elegant one. It is the one the model knows best, and the model’s knowledge is a historical accident of what was publicly available to scrape. Now scale that observation past languages. The winning payment rail in an agentic economy is not necessarily the most elegant one either. It is the one the models are trained to use, the one the APIs are standardized on, and the one the tutorials teach.
Which is why Ben Hylak’s note that OpenAI is throwing an autoresearch hackathon this Saturday with Raindrop and Modal deserves more weight than a weekend event usually gets. The tooling in that room, agents and multi-model orchestration, is the tooling that will define how the next cohort of developers builds. These are not academic exercises. They are curriculum-setting. Whoever writes the tutorials shapes the default stack, and the default stack is what ends up in the training corpus that shapes the next round of defaults. The loop closes on itself, and it closes fast.
What the IPO Timing Actually Signals
That brings the argument back to the warning it started with. The case that mega-offerings from SpaceX and OpenAI could mark a market top rests on a familiar and mostly sound observation. When the biggest private companies finally decide to go public, the reasoning goes, they are timing the window. Liquidity is good, narratives are clean, and institutional appetite is peaking, which historically is often the point at which sentiment has already priced in everything good that can be priced in. The concern is not wrong. It is incomplete, because it treats the timing as a read on sentiment and ignores what the money is actually building.
Put the two stories in tension and a different reading emerges. These companies are not going public because the market is peaking. They are going public because the infrastructure phase is ending and the application phase is beginning, and that transition is exactly when a company wants permanent capital rather than venture money. The window is open not because sentiment is frothy but because the narrative is finally coherent. Agents can transact. The rails are being laid. The business models are resolving into four recognizable shapes. An investor buying that story is not buying the top of a cycle so much as the beginning of a new cost structure for every industry that moves money.
Whether the valuation is right is a separate question, and nothing above answers it. Whether the infrastructure earns its price depends entirely on what gets built on top of it, and that part is genuinely undecided. What is not undecided is where the deciding happens. It is happening this Saturday, at a hackathon, among people who will never appear on a market-top list, and in the routing tables of a payments layer nobody reads about. The top is not where they are building. The top is what gets built.
Leave a Reply