Half the GPUs are idle. That's the whole AI story right now.

Two numbers from the last few months tell the same story, and it isn't the one the keynotes are telling.

The first: an MIT NANDA report found that 95% of enterprise generative AI pilots are producing no measurable impact on P&L. Not "still ramping." Not "early days." No measurable impact.

The second: in a recent survey of enterprise AI buyers, more than 80% reported their GPUs are running at half capacity or less. Companies bought the infrastructure. It's sitting there.

Those two data points are the same data point. Enterprises provisioned for demos, not workloads. They bought AI capacity before they knew what work it was doing.

The order of operations matters

The standard playbook the last two years went like this: pick a model, buy the credits or the hardware, stand up a pilot, then go looking for a workflow to plug it into. That order is backwards, and the failure rate reflects it.

In our experience with smaller clients, the ones getting real value did it the other way around. They picked a single annoying process. Something like: routing inbound leads, categorizing support tickets, extracting fields from PDFs, drafting first-pass proposals. Then they added a model to the specific step where a model helps. The infrastructure question came last, and usually the answer was "an API call."

None of that shows up in a keynote. It also doesn't require a GPU cluster.

What the 5% actually do

Forbes reported earlier this year on the narrow set of AI agent deployments that are making it to production. The pattern is consistent: narrow scope, clear success criteria, a human still in the loop for the parts that matter, and integration with the systems the business already runs on.

That last part is where most pilots die. A chatbot that can't read your CRM, write to your ticketing system, or pull from your data warehouse is a demo. It stays a demo until someone does the integration work, and the integration work is 80% of the project.

The model is the easy part. The plumbing is the job.

What this means if you're evaluating AI right now

A few things we'd tell a founder or ops lead who asked:

- Don't start with the model. Start with a workflow that's slow, expensive, or error-prone, and where you can measure what "better" looks like. - If your pilot needs a dedicated GPU cluster to prove out, it's probably the wrong pilot. Most useful business automations run on an API call and a queue. - Keep a human in the loop wherever a wrong answer costs real money. Augmented beats autonomous almost every time in the current generation. - If you can't describe the process on a whiteboard, AI won't rescue it. Fix the process first.

The 95% failure rate isn't a story about bad models. It's a story about buying tools before defining the job. The boring part, the process work, is where the wins are. It's also where we spend most of our time.