AI stopped being a capital project
Issue 01 said realised value went down. A bigger survey published the same week says it went nowhere at all. I was wrong about the direction, and the thing both surveys do agree on now has a price attached.
By Mark Rhys Thomas
The Lead
First, a correction to Issue 01.
I wrote that realised value from AI went down this year, and I put it in the line at the bottom of the issue. Accenture supported it. The share of companies reporting widespread and sustained value from AI fell from 32 percent to 23 percent inside a single year.
McKinsey published its own survey the same week, and it is bigger. 1,719 respondents across 97 nations, weighted by each nation's contribution to global GDP, fielded through May and June. It asks a slightly different question, whether AI has contributed to EBIT at all rather than whether the value is widespread and sustained, and it finds 37 percent. Essentially unchanged on last year.
Flat, not falling. I overstated the direction. Both surveys agree on the part that matters, which is that enterprise value did not rise while investment did, but went down was a stronger claim than my evidence carried, and I should have written flat.
What McKinsey found instead is more useful anyway.
Eight in ten respondents say AI has improved their individual productivity. Half say it has improved their decisions. And 37 percent can point to any EBIT contribution at all, with the share of genuine high performers stuck at 6 percent for the second year running.
That gap is the Pilot Trap stated as arithmetic. The productivity is real. It is being felt by real people doing real work. It is simply not arriving anywhere a CFO can see it.
Now the part that is new this week. The bill has started to matter.
Twenty percent of McKinsey's respondents say AI operating costs, token costs included, are now constraining their use of AI. Twenty-eight percent are spending more than a tenth of their entire enterprise ICT budget on it. Michael Chui, writing up the survey, uses the word tokenomics and makes the point plainly: the price per token has fallen, and the number of tokens consumed has risen faster.
Gartner puts a floor under this from an entirely independent direction. In 2026, for the first time, global spending on inference, which is what you pay every time a model is used, overtook spending on training, which is what you pay once to build it. $23.3 billion against $19 billion. Fifty-five percent of AI-optimised infrastructure spend now goes to inference, rising to 59 percent next year.
AI has stopped behaving like a capital project and started behaving like a utility bill.
Pharma is not set up for this. Commercial technology in pharma is bought the way software has always been bought. A business case, a capital approval, a multi-year licence, a renewal conversation in year three. An agent that runs an MLR pre-check ten thousand times a day is not that. It has a unit cost and it has a volume, and almost nobody in a pharma commercial organisation can tell you either number for any system they are currently running.
You cannot manage a variable cost with a fixed-cost governance model. That is a plant problem, and it is the argument for shared production infrastructure rather than a prototype shop bolted onto every brand team.
Pharma spent three years asking whether AI works. It never got round to asking what it costs to run.
Also on the Line
COMPUTE
Your inference bill is not about to fall on its own. Nvidia will cross $100 billion in quarterly revenue next quarter, putting it near $432 billion annualised. Anthropic signed a roughly $45 billion cloud agreement with Nscale for about 460 megawatts in West Virginia, live at the end of 2027. Nvidia is in talks to buy Hugging Face, the main neutral host for open models, at around $13 billion against a $4.5 billion valuation three years ago. That is committed capital which has to be earned back out of what you pay to run models. Waiting for the price to drop is not a strategy.
DATA
Ireland's AI Office can compel disclosure from December. The Regulation of Artificial Intelligence Act 2026 was signed into law on 21 July. The AI Office of Ireland now exists and has a chief executive. Until December it cannot compel information from an organisation that has not voluntarily engaged with it. From December it can. For Irish-domiciled pharma that is roughly three months to get documentation of what your AI systems do, and what they were built on, into a state that survives being asked for it. Most could not produce that today.
PROCESS
The headcount reductions people forecast do not arrive. McKinsey asked the same question two years running. Last year, 32 percent expected AI-related headcount declines within twelve months. This year only 14 percent report that declines actually occurred, less than half the prediction. And 39 percent are now forecasting them for next year, a higher number than the forecast that already failed. Meanwhile mid-level managers and individual contributors report AI-related strain at 47 percent against 31 percent of executives. The strain is real and it is landing a layer below the people booking the savings.
What Did Not Meet Spec
Every week I read roughly sixty items and publish four. This is the in-process control.
Rejected this week: nine sponsored placements written to read as editorial, the newsletter's own recruitment ads and referral promos, and the usual tooling minutiae, CUDA on RISC-V, speculative decoding, embedding model releases, which matter only if you are writing code this week. From the general technology feed, the rocketry, the foldable phones and a browser exploit walkthrough.
Two rejections worth naming.
The first cuts against my own lead. McKinsey reports that 93 percent of enterprises deploying AI are running over their AI budgets. It is a perfect number for the argument I have just made, and I am not using it, because it comes from a separate McKinsey FinOps survey of 75 respondents rather than from the 1,719-respondent survey I led on. Both are McKinsey. Only one of them can carry a claim in front of a pharma board.
The second I wanted more than anything else I read this week. Meta ran an internal programme modelling 60 percent headcount reductions replaced by AI agents, executed the first round of layoffs, then cancelled the second, with Mark Zuckerberg reportedly conceding that agentic development had not moved the way the company expected. It is the best counterexample I have seen this year and it supports everything above. I read it in a newsletter summarising a news site, which puts me two removes from anyone who did the reporting. It goes in the hold file until I have the original in front of me.
Showing the batch record means showing the rejects. It also means showing the corrections.
The Line
AI stopped being a capital project this year and became a utility bill. Nobody in pharma owns that bill.
Sources
- 1. McKinsey QuantumBlack, The state of AI in 2026: On the road to ROI, published 25 August 2026. Dan Tinkoff, Lieven Van der Veken and Michael Chui, with Tara Balakrishnan. Online survey fielded 4 May to 8 June 2026. 1,719 respondents across 97 nations, 36 percent from organisations above $1 billion revenue, weighted by each nation's contribution to global GDP.
- 2. Gartner, Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% in 2026, press release, 10 August 2026. Inference $23.3 billion against training $19 billion. 55 percent of AI-optimised IaaS spending supports inference in 2026, rising to 59 percent in 2027. Taken from Gartner's own release, not from the secondary coverage that rendered it as 55 cents of every AI cloud dollar, which uses a wider base than Gartner does.
- 3. Accenture, Pulse of Change: What's top of mind for today's leaders, research report, 26 July 2026. Cited here only for the 32 percent and 23 percent figures carried in Issue 01.
- 4. Department of Enterprise, Tourism and Employment, Regulation of Artificial Intelligence Bill 2026, June 2026. Act signed into law 21 July 2026. AI Office of Ireland established. Compulsory information powers commence December 2026.
- 5. Nvidia quarterly revenue, the Anthropic and Nscale agreement, and the Hugging Face acquisition talks are reported figures carried by CNBC and associated coverage rather than company filings, and are presented here as reported.
Every figure in this issue was checked against the primary source. Where a number originates with a third party, it is attributed to them and not to the firm quoting it. Where I got a figure wrong in a previous issue, it is corrected in this one.
The list
Subscribe to Batch Record
Weekly, every Monday. One issue, one layer, one line worth quoting.