Frontier labs were built by startup and research people — Y Combinator, academic pedigrees, founder DNA. Their world runs on speed, narrative, and scale. Enterprise IT was built by operators — decades of surviving budget cycles, procurement gauntlets, and migrations that had to work on day one. Their world runs on resilience, process, and trust.
They’re different machines, optimized for different things.
The labs are moving at the pace of possibility. Enterprises are moving at the pace of responsibility. When those two tempos meet, it can look like one side is failing — and which side looks like it’s failing depends on your perspective. From the labs, enterprises look too slow; from inside the enterprise, the labs look like they’re innovating too fast and losing money too fast to be trusted. But really, two sophisticated systems are colliding, each doing exactly what it was designed to do. Today, those two tempos are badly out of phase, and the gap is widening: the frontier labs are advancing faster than enterprises are adopting. That mismatch is the primary reason the AI bubble has become a runaway train — and it will keep running until there’s a major paradigm shift in how either side operates, or until it bursts. Judging by the strong beliefs and institutional momentum on both sides, I now assign the highest probability to a burst.
I think we’re going to see a lot of pain in AI. Two problems aren’t just feeding each other — they’re compounding: the category has taken on far more investment, and far more losses, than it can digest; and the gap between how fast the labs ship and how fast enterprises adopt keeps getting wider. Here’s my hypothesis: the longer this goes on, the more money AI companies (neoclouds, AI labs, AI startups) lose, and the riskier they look to enterprises as something to bet on. Enterprises don’t care how big you are or what your top-line revenue is. They care about your ability to deliver your service and live up to your contracts, which means viability, and profitability once you’ve grown past what VCs can absorb.
First, as a category, AI tech companies and their investors have already bitten off more than can get digested by enterprise consumers in the near future. There’s too much investment that isn’t going to meet ROI at all — or at least not anytime soon. I’m not going to publicly name which specific companies I think are over-leveraged (I have my private beliefs), but the over-leverage is real at the category level.
Second, the pace of innovation is faster than the pace of enterprise adoption, and that gap is widening. Enterprises will either cherry-pick whatever’s latest and roll with that, or they’ll intentionally stay behind until a more stable, lower-cost version comes out. The fundamental issue is that, for the overwhelming majority of non-tech enterprises, AI use cases are optimizations. Sure, everyone wants to take advantage of optimizations that improve the bottom line, but very few non-tech enterprises face an existential crisis because of AI. In other words, they can afford to wait, because the impact is a point or two of profit, not a do-or-die threat that forces action.
That’s where the asymmetry gets brutal: enterprises, as a category, will survive. Individually, some will fail because they get disrupted by new tech — think Blockbuster. But the AI labs, as a group, cannot afford to wait. Status quo is a strategy that requires zero innovation for enterprises — the labs can’t stand still without losing to each other, and categorically, because of the massive losses they’re taking, they’re forced to keep sprinting even when the market isn’t catching up. With losses that large and the first signs of capital running dry already showing, “slow down” stops being a strategic choice and becomes a survival question. That leaves the AI labs with a near-impossible choice: slow down innovation without losing the race.
Which is exactly why you’ve seen so much desire for an external party, like the federal government, to be the reason innovation slows down — and all the BS about “AI escaping the lab and posing a threat to the world,” with narratives like “it’s too powerful to release.” Not because enterprises need permission — they’ve been cautious on their own for decades. Because the labs need cover. No AI lab wants to declare first that it’s slowing down when it’s short on cash. Someone else to blame. Someone else to move first.
In the long run, I believe all the capacity gets used and enterprises will adopt AI, much like every other major infrastructure buildout cycle — it makes too much sense not to. We’re going to see enterprises that operate drastically more efficiently than they do today. In my opinion, the problem is purely timelines. And that timeline gap is what makes you question which current vendors survive, which ironically means less investment in the short term until the shakeout happens. Demand-side hesitation starves the revenue that would justify the supply-side burn, which makes investors more nervous, which tightens capital further. A self-reinforcing squeeze. I know this because I meet IT leaders every day who are genuinely excited to leverage AI, but they all share the same fear: betting on the wrong vendor and being stuck with that decision for much longer than they’d like, while eroding their credibility.
Unlike startups, which can switch providers on a dime, enterprises have to budget and go through approvals, security reviews, procurement, RFPs, and more. The cost of changing your mind is much higher than the cost of waiting for a confidence signal — especially when the opportunity at hand is an optimization and the status quo is the existing operation.
IT leaders also face a psychological problem: the operations and business teams often don’t understand and appreciate how fast AI tools are evolving, or how likely it is that the first vendor they select will need to be replaced before it even reaches production. For operations, constantly re-tooling and re-training the workforce is a major disruption. For finance, it’s an unexpected cost, and it casts a lot of doubt on whether IT knows what it’s doing or is just experimenting in production at a very high operational cost. And remember, the cost isn’t just the spend on tech. It’s the internal implementation cost required to deploy a technology into production, and that cost is often many times greater than the cost of the tools. That means no “discount” or “early termination clause” will act as an equivalent offset.
Optionality is one of the top things our customers look for, especially in AI, and they want PolyAPI to give them optionality against vendors. We do this by implementing the backend services for AI — memory, tool calls, knowledge hub, and integration tooling — using the architectural principle of loose coupling, with our middleware platform as the abstraction layer. That lets multiple agents and AI providers interface with their systems while they keep full control, data ownership, and data consistency across AI channels, plus the ability to swap out AI models and operational systems without impacting one another. We’re helping them adopt at a time when they’re very nervous about adopting. Most AI vendors are asking enterprises to bet on them. We’re offering a hedge against betting on anyone (including ourselves). In the upcoming shakeout, that’s not a nice-to-have — that’s the thing that keeps you alive while everyone else is stuck with their bad decisions.