There is a persistent narrative that the artificial intelligence boom is simply a story of companies racing to buy AI—pouring money into models and infrastructure as fast as they can. But the actual spending data tells a more nuanced and far more interesting story. When you look at where corporate dollars are really flowing, the headline isn't that everyone is buying AI infrastructure. It's that they're buying what sits between them and the large language models.
The Rise of the Routing Layer
Across a corporate spend platform tracking roughly 40,000 companies—ranging from small and medium businesses up to enterprises, and including AI-native firms and next-generation startups—the two fastest-growing vendors over a recent six-month window were not model providers at all. They were routing layers: one called Fal AI and another called Open Router.
These products perform a deceptively simple but increasingly essential function. Open Router helps a company decide which large language model to use for a given task. Fal AI does the same thing for generative media models. In other words, businesses are no longer just buying raw intelligence—they're buying the orchestration that helps them deploy that intelligence wisely. The hard question is no longer "Do we have access to a powerful model?" but "Which model should we use, and for what?"
This is a meaningful shift. A year ago, almost no one knew what these routing platforms did. Today they are among the fastest-growing line items on corporate cards. That trajectory tells us something important about how the market has matured.
Why Companies Stopped Trying to Pick the Winner
The deeper insight behind this trend is that businesses have largely given up on choosing a single victor. Every week brings news of a new model from one of the major labs, each leapfrogging the last. In an environment of such relentless churn, committing fully to one provider feels less like a strategy and more like a gamble.
Consider how quickly the ground shifts. Six months ago, the coding tool Cursor was the darling of forward-leaning companies, growing at something like 1,000 percent. Today it has fallen out of the top ten by spend. This doesn't mean companies abandoned it—they simply stopped assuming it was the best tool for everything. Rather than crown a single champion, they began backing several horses at once, buying multiple tools together and letting middleware companies like Replicate, Fal, and Open Router help them sort out what to use and when.
The most accurate way to describe this behavior is hedging. Companies are effectively saying: let the model makers duke it out among themselves; we'll use these orchestration platforms to tell us what to deploy for each specific use case. They're declining to trust the billboard in the subway or the advertisement during the big game telling them to use one particular model for everything. And the evidence suggests this approach is working—the explosive adoption of routing platforms is the proof.
Crucially, this is not a sign of paralysis or indecision. It reflects a more sophisticated understanding of the landscape: different models genuinely specialize in different things, and the goal isn't to find a single winner but to match each task to its best-suited tool. This is smarter capital allocation dressed up as caution.
Vibe Coding and the End of Typing
A parallel shift is unfolding in how software itself gets built. "Vibe coding"—the practice of prompting an AI to write code rather than writing it by hand—has dramatically lowered the barrier to building software. The landscape here has three main players: Replit, Framer, and Lovable. The first two were founded roughly a decade ago, while Lovable is genuinely AI-native, founded in 2023. That recency hasn't held it back; Lovable went from zero to 100 million dollars in revenue in under a year, making it arguably the category leader.
It's worth being honest about the limits, though. The internet is full of breathless claims—someone tweeting that they "vibe coded" a replica of an established public software company over a weekend, declaring the incumbent dead. That's a misreading of what makes a real business. An established company isn't just its code; it has distribution, brand, customer relationships, and a host of other assets that can't be conjured from a prompt. Replicating an interface is not the same as replicating an enterprise.
That said, the legitimate use cases are real and valuable. Vibe coding is excellent for spinning up a first version of a product—a working draft that a team can later refine and eventually ship to customers. The spending data reveals a telling pattern here: users of Lovable tend to bundle it with other "stop doing the manual work" tools. They pair it with Gamma, which generates presentation decks, and WhisperFlow, which replaces typing with dictation. In essence, the people who have stopped coding by hand have stopped doing nearly everything that requires manual typing. They're leaning on AI for the first pass—the quick-and-dirty analysis, the rough draft that knocks out the first 80 percent of the work. This makes them dramatically more efficient, even if the final 20 percent still demands human judgment. The consequence is that how companies build, and how they hire, is fundamentally shifting.
Two Paths to Winning
None of this means the established giants are losing. The same period that has seen vibe-coding startups soar has also seen incumbents post enormous results—one major enterprise software company recently printed an 800-million-dollar quarter on its agentic AI product, while data-platform incumbents continue to thrive.
The reconciling principle is this: startups are winning by owning greenfield, while incumbents are winning by bolting AI into the distribution they already own. These are not contradictory outcomes. A nimble startup can capture entirely new use cases that didn't exist before, building from scratch on AI-native foundations. At the same time, an entrenched player can layer AI capabilities onto its existing customer base, brand, and reach—monetizing intelligence through channels it already controls.
In other words, you don't have to choose between these futures. We will continue to live in a world where vibe coders launch products overnight and where the large incumbents keep crushing it. Both stories are true simultaneously.
What the Money Reveals
The throughline across all of this is that corporate spending is a more honest signal than corporate hype. The hype says everyone is racing to buy AI models. The money says something subtler: businesses are buying the orchestration, the middleware, and the workflow tools that let them extract value from a volatile and rapidly evolving set of models—without betting the company on any single one of them. They are hedging intelligently, automating their first drafts, and matching tools to tasks. That is not the behavior of a market caught up in mania. It is the behavior of a market learning, in real time, how to actually put this technology to work.