The AI Labs Sold 80%. Customers Got 16%. Now They Want to Slow Down.

Glitches and GainsAI IPO
Written By Curtis Wadsworth

Last weekend, Anthropic CEO Dario Amodei published an essay called We Must Pace the Frontier. His argument: AI capabilities are advancing so fast, and a recent rogue-agent incident was so alarming, that frontier labs should deliberately slow how quickly they improve their models. Within hours, Sam Altman and Elon Musk publicly agreed. Demis Hassabis and Satya Nadella followed.

We have seen this movie. In March 2023, one week after GPT-4 shipped, more than 30,000 people signed an open letter calling for an immediate six-month pause on training anything more powerful, a pause the letter said should be public and verifiable and cover every serious lab. What followed was not a pause. It was the largest capital expenditure cycle in the history of the software industry. Elon Musk signed that letter and started his own AI company within months.

The difference is what a pause now costs. In 2023, roughly $1.7 billion was spent on generative AI. The labs were private, pre-revenue, and answerable to almost no one. Restraint was free, which is why it was easy to sign and easy to ignore. Today Anthropic is reportedly running near $65 billion in annualized revenue and preparing a listing that could value it above $2 trillion. Every word in a safety essay now has a price tag attached to training budgets, compute contracts, and to a valuation built on numbers the public hasn't seen yet. A call to slow down no longer gets taken at face value.

The reaction split three ways: One camp takes the safety case at face value. A second calls it regulatory capture: incumbents writing rules that newcomers can't afford to meet. A third follows the money, noting that Anthropic is weeks from one of the largest IPOs ever.

All three are arguments about the CEOs — their sincerity, their regulatory strategy, their balance sheets. The more useful question in this Nerd Lawyer’s humble opinion is what the customers are saying. And the customers are telling a very different story.

The oldest rule in tech, broken

Under-promise and over-deliver. Every enterprise software company founder learns this early, because enterprise buyers have long memories and renewal cycles. The frontier labs did the opposite and very publicly.

The promise. In late 2025, Anthropic published research estimating that, across 100,000 real conversations, AI cut task completion time by 80%. It was covered as field evidence that the productivity revolution had arrived.

Anthropic’s methodology undercuts the veracity of this number. Time savings were estimated by Claude itself, grading its own conversations. Anthropic also acknowledged the estimate did not account for the time people spend checking and correcting the AI's work. If you are a lawyer, an engineer, or anyone whose output has to be right, checking and correcting is not a footnote. It is the job. An 80% time savings that excludes verification is a 30-second draft plus twenty minutes of review, reported as 30 seconds.

Now the delivery. At Gartner's annual Symposium in Australia this week, chief of research Daryl Plummer cited research finding that software vendors now pitch 50% productivity gains from AI, while customers report roughly 16%. He is not an outlier. The large 2026 enterprise surveys all point the same direction:

  • McKinsey (1,719 leaders): Only 37% of polled executives attribute any EBIT impact to AI, unchanged from 2025. Just 6% qualify as "high performers," also flat, despite record spending.

  • KPMG (2,145+ leaders, 20 countries): Only 7% report established AI ROI.

  • Deloitte (3,235 leaders): 66% report productivity gains, but only 20% see AI-driven revenue growth.

  • WRITER: 75% of executives admit their company's AI strategy is "more for show" than real guidance.

The labs' own research claimed 80%. Software vendors today pitch 50%. Customers report 16%. That is not a miss at the edges — it is a miss by a factor of five against the labs' own number, in an industry that knows better than anyone what happens to a vendor that overpromises.

As a result, buyers are changing what they measure. Futurum's survey of 830 IT decision-makers found productivity dropped, from 23.8% to 18.0%, as the top ROI metric as enterprises started demanding revenue or margin impact instead.

That shift tells you everything. Hours saved is a metric you can claim without proving anything: it relies on self-reported estimates, it assumes the saved hour gets redeployed into something valuable, and it never has to appear in a financial statement. Revenue and margin can't be self-reported. Buyers moved to a metric vendors can't estimate on their behalf, which is how you know the old question stopped producing believable answers.

The tell: customers are voting with their routing

Buyers haven't abandoned frontier models. They have done something more damaging to the economics: they decided most work doesn't justify frontier prices.

Open-weight models — Qwen, DeepSeek, Llama, Mistral, GLM — now sit alongside frontier models in most enterprise stacks. Morgan Stanley, citing McKinsey, reports 63% of enterprises using open models, generally in combination with frontier models rather than instead of them. The split is predictable. High-volume, repetitive work goes to the cheaper open weight models; the hard reasoning goes to expensive frontier lab models.

The spending data shows what that split is worth. On Vercel's gateway in June 2026, open-weight models processed 29% of tokens but accounted for under 4% of spending, because they were available at roughly one-tenth of the average token price. 

Meanwhile the quality gap narrowed to almost nothing on the work most companies actually do. This summer, the Artificial Analysis Intelligence Index — a 0-to-100 composite of about ten benchmark tests covering math, coding, science, and agentic tasks — found that the top open models trailed the leading frontier model by three points.

Let that sink in: Three points of capability, ten times the price. If frontier models delivered the transformation that was advertised, enterprises would not be routing the bulk of their volume to models that cost a tenth as much. They are paying the premium only for the narrowing slice of work where the premium is visible, and treating everything else as a commodity.

Put the Promise in the Contract

The vendor pitched 50%. Your team got 16%. Inside the Revenue Contracting Suite™, I draft the MSAs, SaaS terms, and SOWs that tie AI claims to measurable performance — whether you're buying the tool or selling it. Your contract should know the difference.

Book a 20-Minute Call→

A skeptic will point to Menlo Ventures. Surveying 495 U.S. enterprise decision-makers, Menlo Ventures found open-source share of enterprise usage falling from 19% in 2024 to 11% in 2025. However, Menlo measured usage in estimated dollars spent, derived from what respondents said about their workloads. Money is the wrong yardstick here. When open-weight models cost a tenth as much per token, a category holding 11% of the dollars can account for a far larger share of the work — exactly what the token data shows.

A warning inside the savings. Cheap is not free, and the terms vary more than the benchmarks do. If you are routing client data, regulated data, or your own trade secrets through the cheapest available endpoint, the savings are real and so is the exposure. Read the terms of service before you route.

Now stand where the labs are standing. They are committed to compute contracts measured in tens of billions. The capabilities they ship get matched by free alternatives within months. Their customers report gains that don't reach the income statement. And their pricing power survives only in the shrinking band of work where three points of capability is worth ten times the price.

That is not a company at the top of a technology curve. That is a company that needs the next chapter of the story to be about something other than productivity.

Which brings us to the timing of "pacing the frontier." If you promised a revolution and your customers are reporting a rounding error, a safety pause is an elegant way to change the subject. The next model doesn't have to be a leap. It just has to be responsible.

Follow the money, not the manifesto

On September 13, the Financial Times reported that Anthropic told shareholders its adjusted operating income will be positive for a second consecutive quarter. Second-quarter revenue reportedly topped $11.5 billion with roughly $559 million in adjusted operating profit, and annualized revenue reached about $65 billion by late July. The company has reportedly selected Nasdaq for a listing that could value it north of $2 trillion.

The operable word in this profit figure is adjusted. The reported gross margin is calculated before the cost of training the models, and before the revenue shared with distribution partners such as Amazon. Training is by far the largest expense every frontier lab has. It is the reason these companies raise tens of billions and sign multi-year compute commitments. Reporting a margin that excludes training is like a restaurant reporting margins before food and labor. The number isn't false, but it describes a business that doesn't exist.

Now line up the calendar. Anthropic's listing is now reported for November, targeting a valuation as high as $2 trillion. Slowing the pace means spending less on training runs — the cost the 80% margin figure leaves out. In addition, Anthropic is weighing whether to ship a new model before the listing, according to Reuters. What complicates that decision, per the same reporting: its own call for a slowdown and the need to show profitability.

None of this proves bad faith. Sincerity and strategy are not mutually exclusive. Safety concerns and financial convenience can point the same direction at the same time. But when a story arrives that simultaneously lowers expectations for the next model, reduces the biggest cost line, and arrives as its own listing slips to November, the ordinary rules of diligence still apply: read the filings, not the essay. The S-1 will say more about the state of AI than any manifesto published this month. When the loudest voices for slowing down are the ones who most need to reset expectations, the customer data deserves more weight than the press release.

What this means for builders

Whether you are buying AI or building it, the promise gap is a legal infrastructure problem, not just a strategy one.

  • If you're buying AI, the productivity claims on the vendor's website are not in your contract. Tie pricing and renewals to measurable outcomes, define acceptance criteria, and keep clean exit rights. If the tool delivers 16% instead of 50%, your contract should know the difference.

  • If you're selling AI, your productivity claims are going to be analyzed by enterprise buyers and investors who have read the same surveys. Claims you can't substantiate are a sales problem today and a liability problem tomorrow. 

  • If you're routing to open models to cut costs, the license and the data-handling terms are the contract. Confirm the license permits your commercial use at your scale, confirm whether the provider trains on your inputs, and make sure your own client and vendor agreements permit sending that data there at all.

  • Either way, measure before you deploy. A use case that goes live without a baseline metric can never prove it worked. Build the baseline into the contract and the rollout plan from day one.

This is exactly the work our Revenue Contracting Suite™ is built for: MSAs, SaaS terms, and SOWs that tie AI promises to measurable performance, on both sides of the deal. If you're signing or selling an AI contract this quarter, talk to us about the Revenue Contracting Suite before the promise gap becomes your problem.

→ Book a 20-minute call and let's see where you stand.

Curt Wadsworth, J.D., Ph.D. is the founder of Nerd Lawyer Entrepreneur Services, LLC, an AI-native corporate and IP law firm serving founders and growth-stage companies. He's based in Pittsburgh and bar-admitted in Pennsylvania and the USPTO.

Sources

This post is for general information and is not legal advice.

© 2026 Nerd Lawyer Entrepreneur Services. All rights reserved.

Curtis Wadsworth
Previous

What Anthropic's S-1 Teaches Entrepreneurs About Fundraising
Next

Your AI Agent Just Became a Compliance Artifact