Created At: Tue Jun 09 2026
The Wrong Bet?
Agentic IPOs vs Market Reality
OpenAI and Anthropic are preparing their IPOs. The pitch is straightforward: agentic AI will become the default automation layer for business operations worldwide. The infrastructure contracts to support this future have already been signed — at costs that require the future to actually arrive pretty fast.
It probably won't.
Uber burned through its entire 2026 AI tooling budget in four months. Microsoft started revoking Claude Code licenses as costs spiraled. JPMorgan published a note titled "AI Token Costs Are Eating Internet Profits Alive." These are not teething problems. They are the economics of the use case behaving exactly as the underlying logic would predict — if anyone had examined it closely before the debt was taken on.
The companies now going public built their valuation on a market that may not exist at the scale required. That is worth examining before the prospectus lands on your desk.
What the Bet Requires
The AI infrastructure buildout is not priced as software. It's priced as utilities — the assumption being that agentic reasoning will become the default automation layer across business operations, at a scale large enough to justify what's been invested.
For that to hold, one condition has to be true: most business processes, at most companies, are better handled by reasoning chains than by deterministic automation.
They aren't. Most operational tasks have defined outputs. The right answer is known in advance. The path to it can be engineered once and executed reliably without a reasoning loop, without a model subscription, and without a token bill that compounds with every execution. The businesses discovering this — after the budget is gone — are not outliers. They are the general case.
If the general case doesn't need agents, the volume assumption fails. And without the volume, the infrastructure doesn't pay for itself.
The Legitimate Cases
There are applications where high-compute agentic reasoning is genuinely the right tool. Work where the output cannot be specified in advance. Where exploration is the point. Where the value is in finding something that couldn't have been asked for directly.
Frontier scientific research is the clearest example. Drug discovery, materials simulation, hypothesis generation across bodies of literature no human team could process at the required scale. The compute is justified because the task requires it.
But this market is not large. The number of organisations running frontier research at the scale that demands this infrastructure is finite and small. It is a legitimate use case. It is not a volume business. It cannot carry the weight of what is being built around it.
The Robotics Question
The other candidate that holds up to scrutiny is the central coordination of large autonomous robot fleets — high-stakes, complex, genuinely open-ended decision-making at scale.
Except the physics work against it.
Real-time control requires low latency. Agentic reasoning chains running against large remote models over a network don't deliver low latency. That gap is not a temporary engineering limitation. It's structural — the round-trip time alone disqualifies the architecture for anything that needs to act in the physical world in real time.
What robotic control actually needs is small, efficient models running on-device. Close to the hardware. Fast enough to act. Independent of a remote compute dependency. The mobile chip, not the data centre.
Now consider the other direction: if compute density improves to the point where a powerful model can run on the device itself — embedded in the robot, untethered from centralised infrastructure — then the use case that might have justified the bet is served by a completely different architecture. The centralised model loses from both ends. Either latency makes it unusable, or on-device capability makes it unnecessary.
The Three-Way Trap
The economics only resolve cleanly in one of three scenarios — and each one forecloses another.
Volume at current prices requires the business automation case to hold at scale. It doesn't.
Volume through lower prices sounds plausible — inference costs are falling and will continue to. But cheaper tokens compress the margins that were supposed to service the infrastructure. And lower prices accelerate the viability of open-source models running on local hardware, pulling enterprise volume away from centralised providers at exactly the moment the unit economics start improving.
Sustained high prices keeps the margins intact but makes on-premise deployment economically rational for most organisations. The enterprise base migrates. The revenue that justified the infrastructure shrinks.
There is no path through this triangle that doesn't create a structural problem somewhere else.
Overshot, Not Mistimed
The charitable read is that the timing is wrong but the direction is right. Costs fall, use cases mature, the volume eventually arrives. This framing is not entirely without merit — the inference cost curve is real, and it does change the calculus over time.
But timing and market size are different problems. Cheaper reasoning loops only justify the infrastructure if there are enough tasks that genuinely require reasoning in the first place. The evidence so far — accelerating budget overruns, enterprise licence revocations, JPMorgan flagging token costs as a structural threat to internet profitability — points toward a market size problem, not a temporary pricing mismatch.
The use cases that justify high-compute agentic reasoning are real. They are just narrow, specialised, and structurally insufficient to support what is being built around them.
The infrastructure was priced on a volume that requires the wrong tool to be the right tool for most work. That condition is not going to be met. The bill is already arriving — and the companies paying it are starting to notice.