Earlier this week, a colleague forwarded me a link with a single line of commentary: "Menlo just said the quiet part out loud." The piece was thin โ a market brief, barely three data points, not a single hard figure quoted from the underlying report. But the headline carried a phrase that stopped me mid-scroll: enterprises had "badly underestimated" their AI spending. Coming from Menlo Ventures โ a firm whose portfolio is woven into the very fabric of the AI application economy, an early backer of Anthropic and a dozen others riding the same thesis โ that adjective does more work than any spreadsheet could.
I have spent enough years inside protocol economics to know that when a stakeholder publishes a warning about their own value chain, the warning itself is the story. The numbers are secondary. What matters is why someone with every incentive to keep the narrative buoyant chose to puncture it โ however gently, however briefly.
Let me be clear about what this brief actually contains, and what it conspicuously does not.
The article reports that Menlo Ventures has observed enterprise AI adoption accelerating rapidly, while simultaneously finding that companies have systematically underestimated what that adoption costs. That is the entire content: adoption is fast, bills are higher than expected. No multiples. No sample size. No industry breakdown. No before-and-after comparison. A research report that presumably contains granular spend-versus-budget data has been compressed into an adjective. "Badly."
This is not a failure of journalism so much as a signal about the genre. Industry briefs exist to transmit a directional signal without the burden of evidence. And the direction here is unmistakable: the era of enterprise AI as a faith-based purchase is ending. We are entering the accounting phase.
To understand why this matters, you need only look back at cloud computing's own adolescence. In the late 2000s and early 2010s, enterprises migrated to AWS and Azure under a seductive premise: pay for what you use, and you will spend less than on owned infrastructure. Then the bills arrived. Finance departments discovered that cloud costs rose rather than fell, that usage sprawled across teams without attribution, that the marginal cost of a workload was never actually zero. The result was the birth of an entire discipline โ FinOps โ and a cohort of companies like Cloudability and Apptio built specifically to help enterprises see and control what they were spending. IBM eventually bought Apptio for $4.6 billion. The cloud bill shock did not kill cloud. It created a new industry layer on top of it.
AI spending is following the same trajectory, but the underlying cost structure is far more unforgiving. In a traditional SaaS model, the marginal cost of serving one more seat approaches zero. You budget per user, you forecast per user, and the math holds. AI inference does not work this way. Every API call carries a real, non-trivial GPU cost. Every token processed consumes physical compute. And the multiplier effects are brutal: a single agentic workflow can consume fifty to a hundred times the tokens of a simple chat interaction โ a rough estimate from my own protocol design work, but one that aligns with what I hear from infrastructure teams. Long-context windows, multi-step reasoning chains, RAG pipelines that retrieve and re-inject documents on every turn โ each layer compounds the bill in ways that a per-seat budgeting spreadsheet cannot capture.
This is the structural root of the underestimate, and it is the one thing the Menlo brief never explains. The technology is not expensive because it is badly engineered. It is expensive because it works by consuming compute, continuously, at a rate that scales with sophistication rather than headcount. Enterprises budgeted for software. They got a utility.
The second-order effects deserve closer attention than the brief gives them. When an enterprise discovers that its AI bill is three or five or ten times projections, the response is rarely to abandon the technology wholesale. It is to ask harder questions about return. That is healthy in isolation โ but it arrives at a moment when a great many AI application companies have built their entire sales pitch on the promise of cost reduction. "Use our AI to do more with fewer people." That thesis collapses if the AI itself is a major line item. You cannot credibly sell savings when your product's own operating cost is the thing blowing up the budget.
The deepest risk here is not that enterprises will stop adopting AI. It is that the ROI math will fail at precisely the layer where valuations are most stretched โ the application layer โ while the infrastructure layer continues to collect.
This is the contrarian reading that the brief, in its haste, completely misses. A cost overrun is a transfer. Every dollar an enterprise spends beyond budget is a dollar of revenue recognized somewhere upstream โ at the model provider, at the cloud platform, at the GPU supply chain. In the near term, the Menlo finding is bullish for infrastructure and bearish for application-layer discipline. It is the kind of divergence that makes quarterly earnings confusing and venture portfolios lopsided.
But there is a subtler signal buried in the source itself, and it is the one I keep returning to. Menlo Ventures is not an outside critic. It is a participant. When a participant publishes a caution, you should read it as expectation management โ a gentle, pre-emptive softening of the ground before a reset. This is not cynicism; it is pattern recognition. The same instinct that led me, during the 2022 collapse, to stop trading narratives and start auditing zero-knowledge circuits tells me that the most reliable signals come from people with skin in the game who choose to speak plainly anyway.
What does the practical response look like? The cloud analogy offers a map. Enterprises that survived the cloud bill shock did so by building cost observability โ tagging, attribution, chargeback, the unglamorous plumbing of financial accountability. AI needs the same, at token-level granularity. Model routing that sends simple queries to cheap models and reserves frontier models for complex reasoning. Caching strategies that avoid reprocessing identical prompts. Inference optimization โ quantization, speculative decoding, KV-cache reuse โ that converts engineering efficiency directly into budgetary relief. None of this is exotic. All of it is underbuilt. And the firms that build it will occupy the same structural position that FinOps vendors occupied a decade ago: indispensable, unsexy, and increasingly expensive to replace.
The open questions are the ones the brief should have answered. How large is the underestimate, really? A factor of two suggests budgeting friction; a factor of ten suggests a business model problem. Which industries are most exposed? My instinct says mid-sized enterprises are the fragile middle โ large enough to run real AI workloads, too small to have dedicated FinOps teams. And is the pricing model itself about to shift? If enterprises cannot forecast per-token costs, the market will migrate toward per-seat or per-outcome pricing, transferring cost uncertainty back to vendors. That migration, when it comes, will be the clearest tell that the accounting phase has fully arrived.
I have watched this industry long enough to be wary of both the true believers and the doomsayers. The truth usually lives in the unglamorous middle: the technology is real, the costs are real, and the gap between them is where fortunes are made and lost. Menlo's brief, thin as it is, marks a threshold. The question is no longer whether AI is worth adopting. It is whether anyone can actually afford to run it the way they planned. And the answer, like every honest bill, will arrive later than expected and larger than budgeted.
