Built for Demand That Hasn't Arrived

IBM just signed a multi-year deal with Together AI to build a large inference cluster on IBM Cloud, running on Nvidia's newest chips. The cluster doesn't go live until early 2027.

That gap matters more than the announcement itself. IBM is building capacity for the workload it expects to exist eighteen months from now, ahead of any workload sitting in front of it today, priced, contracted, and under construction well before the demand that justifies it has shown up anywhere except a forecast.

I've spent the past year telling clients that more AI capability doesn't automatically create more demand for what that capability produces. I call it the demand limit: the number of things AI helps an organization produce doesn't automatically grow the number of people who need what's being produced. Our own delivery data backs this up directly. Teams that pointed AI at well-scoped, high-stakes problems converted speed into value stakeholders actually cared about. Teams that pointed the same tooling at poorly scoped work generated volume nobody asked for and absorbed the cleanup cost. The demand limit is a strategy problem, and it sits upstream of any decision about compute.

IBM's cluster is that exact question, asked at infrastructure scale instead of team scale. A cluster is a bet that enterprise demand for cheaper, faster open-source model access will have grown into the capacity being built for it by 2027. That might be right. It might also be a well-capitalized version of the same mistake I watch teams make when they scale up AI usage before confirming anyone downstream needs what's coming out the other end.

The market gave a real-time verdict on how much patience is left for that kind of bet, in the same week's news. Nvidia posted a genuine earnings beat and got hit with a selloff anyway. SpaceX's debut earnings report told a similar story: strong revenue, overshadowed by capital spending growing faster than the revenue it's meant to produce. Investors are now checking whether the infrastructure spend behind a good quarter is on pace to earn itself back before they'll reward the quarter on its own terms, and increasingly deciding it isn't, or at least that nobody's shown the math yet.

That's the same discipline I ran on the viral $500-million-Claude-bill story back in May: don't accept the headline number, run the arithmetic, and check it against what the underlying infrastructure could actually support. The story didn't survive the math. Whether IBM's cluster is the right size for 2027's real enterprise AI demand is a question that deserves the same treatment, and right now nobody outside IBM has run those numbers publicly, including the market that just spent a week pricing in exactly this kind of skepticism against Nvidia and SpaceX.

This lands differently depending on the vantage point. For the technical leader watching this news and feeling pressure to expand their own AI footprint to match the industry's forward posture, the discipline is the same one I use on any spend story: instrument what the team is actually producing and who actually needs it before treating a hyperscaler's capacity bet as a signal about that organization's own roadmap. For the company making its own infrastructure or vendor commitments right now, a demand limit that hasn't been measured is a liability wearing a growth strategy's clothes, and IBM's eighteen-month gap between contract and capacity is exactly the window where that liability compounds unnoticed. For the industry, the Nvidia and SpaceX reaction is the tell: capital markets are no longer willing to fund the AI buildout purely on the strength of a forecast, and every capacity bet still being priced that way, IBM's included, is exposed to the same repricing the moment 2027's demand fails to show up on schedule.