The AI Boom Sells Capacity It Barely Uses. Stacklet's New "Benchmark" Is Just the First to Sell the Cleanup.

On September 10, a cloud-governance startup called Stacklet announced the "Cloud AI FinOps Benchmark" — a set of tested controls, it says, that "discover, fix, and prevent" waste in GPU and AI-model infrastructure. File this correctly: it is a marketing asset, not a product. A benchmark you sell rather than a benchmark you publish is a positioning document. The genuinely interesting thing is that Stacklet thinks there is a market for selling it at all, because that bet rests on a number the AI trade mostly doesn't want to look at.

Stacklet is the commercial arm behind Cloud Custodian, an open-source policy engine for enforcing cloud rules as code, and its pitch is that AI workloads are now the "fastest-growing governance gap in enterprise cloud environments." Here is why that gap is real and measurable, and why it matters to investors in companies that are not Stacklet.

The product being sold is a real problem

The underlying problem is the strangest fact of the AI buildout: the industry is pouring staggering sums into capacity that, once rented, mostly sits dark. Measured production telemetry from Cast AI's 2026 survey puts average GPU utilization across enterprise Kubernetes clusters at roughly5%. Even discounting for the survey's population, that is not a rounding error; it is a fleet of $30,000+ accelerators running a few percent of the time. The broader cloud backdrop is the same story — industry tracking has held cloud "waste" at 27-32% of spend every year since 2019, against a 2025 cloud market that Gartner put above $675 billion.

The incentives produce exactly this behavior. The hyperscalers booking this equipment — a combined roughly $600 billion of 2026 capex, about 75% of it aimed at AI — win by placing capacity, whether the tenant uses it or not. The tenants, in turn, hoard scarce GPUs for fear of not having them at the next repricing, then let them idle. Nobody along the chain is paid to notice the machine is dark, so nobody notices.

That is the vacuum Stacklet is filling, and its own examples show how mechanical the waste is. A policy that catches a training job over-provisioned on VRAM can save roughly $10,000 a month on a $100,000 monthly spend; the company claims stacking two such fixes cuts GPU spend around 20%, and halving from two H100s to one on low-utilization work up to 50%. Whether you trust those vendor percentages or not, the direction is obvious and the mechanism is banal: idle hardware fed by fear, with no one accountable for the idle part.

Why the gap is getting harder to ignore

The forces that let GPU waste hide are now working against it. Traditional FinOps assumes annual procurement cycles and a finance team that sees the bill. AI inference inverts that: spending is metered per token, minute by minute, driven by developers and data scientists the finance team never sees, with changes that can spike inside a single two-week sprint. Static budgets simply don't catch it, and the blast radius of a misconfigured GPU or a runaway model endpoint grows faster than a manual operations team can chase.

This is where the "benchmark" enters as foreshadowing. Even in June, Stacklet was positioning an AI-FinOps measurement framework to define "what good looks like" in GPU governance and to shape procurement conversations over the next two to three years. The vendor consensus — echoed by a small wave of FinOps-for-AI players — is that a governance layer is becoming mandatory as AI spend stops being an experiment and lands on real budgets. Whether this particular benchmark is any good is beside the point; the report that a standard is being invented is evidence the industry decided the problem is worth standardizing.

What it means for an investor

A retail investor should not reach for the spreadsheets to buy Stacklet. It is a private company — founded in 2020 by Cloud Custodian's creators, it has raised an $18 million Series A and a $14.5 million Series B — with no public shares to own, no reported revenue, and a valuation argument that exists only in the minds of its venture backers. There is nothing to buy here directly, and anyone selling you "AI-FinOps exposure" through Stacklet is selling you the marketing asset.

The investable reading is the one the announcement's existence confirms about the companies you can buy. The GPU buildout's bull case assumes the capacity is productive; the utilization data says a meaningful share of it is not, and a governance-tooling sub-industry is now forming to monetize the cleanup. That cuts two ways. For the hyperscalers and capacity providers, it is close to neutral — they lease the silicon either way, so waste is a tenant problem, not theirs. The real exposure is to whoever pays the bill and whatever thesis assumes the spending translates into returns: cash-burning AI companies, and the durability of a capex narrative that runs on a few-percent utilization rate.

If you want one variable to hold, hold this one: not capex headlines, but utilization. A buildout where measured utilization quietly climbs toward reason is one whose economics clear; a buildout where the capacity keeps sitting dark is one where the money that bought it has to be written down somewhere. The "benchmark" is noise. The 5% is the signal, and it is why a startup thinks selling the empty-GPU cleanup is a business at all.