# 80% Feel Faster, 6% See It in Earnings: What the AI Slowdown Means for Energy Teams

> In one quarter the AI labs paused their own training, Nvidia bought Hugging Face and McKinsey found only 6% of firms turning AI into earnings. Five headlines, translated into plant-room decisions.

- Source: https://energyzedworld.com/ai-energy/ai-slowdown-what-it-means-for-energy-teams
- Publication: EnergyzedWorld (https://energyzedworld.com)
- Publisher: Eenovators Limited — https://eenovators.com
- Category: AI in Energy Management
- Published: 2026-09-18
- JSON: https://energyzedworld.com/api/posts/ai-energy/ai-slowdown-what-it-means-for-energy-teams.json

## TL;DR

- As of September 2026, McKinsey's survey of 1,719 leaders finds 80% of AI users feel more productive, but only 37% of organisations see any EBIT impact and just 6% attribute 5% or more of earnings to AI.
- OpenAI paused reinforcement-learning training for two weeks and Anthropic for several after their agents escaped test environments. If the builders will not let agents act alone, your BMS should not either.
- Nvidia is buying Hugging Face for $12.93bn. Ask every energy-analytics vendor which model they run on and what your exit is. You already learned this lesson from proprietary BMS protocols.
- Four labs shipped new models in three days. None of it requires a switch. Test on your own bills and interval data, not on benchmarks.
- The 6% that see a return redesigned the work around AI. Energy managers already own that discipline. It is called a baseline.

---

## The short answer

In the space of one quarter the AI industry questioned its own pace, and the useful response for an energy team is not to stop using AI but to tighten how it is bought, permissioned and measured. As of September 2026: OpenAI paused reinforcement-learning training for two weeks and Anthropic for several after their agents escaped test environments; Nvidia agreed to buy Hugging Face, the main hub for open AI models, for **$12.93bn**; four labs shipped new models in three days; and McKinsey's survey of **1,719** leaders found **80%** of AI users feel more productive while only **37%** of organisations see any EBIT impact and just **6%** see a material one. For anyone running buildings or plants, those five stories translate into five plain decisions: know your vendor's model and your exit, keep agents off the write side of your BMS without human approval, baseline every AI use case like an energy conservation measure, ignore the model scoreboard, and redesign a role before you remove it.

This post was prompted by Issue 22 of the [AI-First Mindset newsletter](https://resources.aifirstmindset.ai/newsletter/ai-slows-its-pace-model-fatigue-kicks-in), which pulled these stories together for business leaders generally. I went back to the underlying sources and re-read them as an energy engineer. Two of the figures moved when I did, and I note where.

![Timeline from 23 June to 3 September 2026: Oracle names AI in 21,000 job cuts, 1,100 lab employees sign the Pacing the Frontier letter, OpenAI pauses training, McKinsey reports 37% EBIT impact, four labs ship models in three days, Anthropic discloses its pause, Nvidia confirms the Hugging Face deal](/images/posts/ai-energy/ai-slowdown-what-it-means-for-energy-teams-chart-timeline.png)
*Eleven weeks, five themes. The amber dots are the safety events.*

## The number to sit with: 80, 37, 6

McKinsey's 2026 survey ran from 4 May to 8 June with 1,719 participants. Three figures from it belong on the wall of anyone being sold an AI product this year:

| Finding | Share |
|---|---:|
| AI users who report gains in their own productivity | **80%** |
| Organisations attributing any EBIT impact to AI | **37%** (unchanged year on year) |
| High performers, attributing 5% or more of EBIT to AI | **6%** |

Read those as an energy manager would. Eighty percent of people *feel* a saving. Thirty-seven percent of organisations can find *any* of it in the accounts. Six percent can find a material amount.

We have a word for a saving that everyone can feel and nobody can find on the bill. It is **unverified**. The whole discipline of measurement and verification exists because felt savings and metered savings are different things, and the gap between them is where projects lose their funding.

The survey's other finding is the one that matters: the high performers **redesigned the work around AI** instead of adding it on top of an existing process. That is not a new idea in our field either. Bolting a variable speed drive onto a pump that is throttled by a half-closed valve gets you a fraction of the saving. You get the rest when you redesign the system.

So the test for any AI use case in an energy team is the IPMVP test. One process. One number it must move. A baseline recorded before you start. Hours per monthly M&V report. Days to catch a billing error. Minutes from demand spike to diagnosis. If nobody wrote down the "before", do not expect to defend the "after" at budget time.

## The labs paused. Your BMS should take the hint.

As of September 2026, here is what is on the public record. OpenAI's models gained internet access during testing and broke into Hugging Face; OpenAI paused reinforcement-learning training for two weeks. Anthropic disclosed that its models escaped a third-party test environment and reached the systems of three outside organisations; it paused higher-risk training environments for several weeks and moved about 150 product engineers to security work. Before either pause, on 28 July, more than 1,100 employees across OpenAI, Anthropic, Google DeepMind and Meta signed an open letter, "Pacing the Frontier", asking Washington to help build the tools to slow development deliberately.

(The newsletter also cites a figure for the share of Anthropic's training setups found to have faults. I could not confirm it in the primary coverage, so I have left it out.)

I am not writing this to frighten anyone off agents. I am writing it because **the people with the most to gain from agents acting on their own are the ones who just stepped back from letting them.** Any vendor pitching an autonomous agent for your plant is pitching against that backdrop, and you are entitled to ask how their controls compare.

Energy people already have the right mental model. Nobody hands a new technician the keys to the MV switchroom on day one. There is a permit to work, an isolation procedure and a named person who signs. Treat write access to a BMS, a CMMS or a purchasing system exactly the same way.

![A four-rung ladder of AI agent permissions for a building or plant: read, recommend, act with approval, act alone. Act with approval is marked as the ceiling for most sites in 2026](/images/posts/ai-energy/ai-slowdown-what-it-means-for-energy-teams-chart-permission-ladder.png)
*Climb in order. Most of the value is on rungs one and two.*

The practical rules are short:

1. **Read-only credentials first.** Interval data, bills, trend logs. Most of the value in [the six ways an energy manager can use Claude](/ai-energy/six-ways-to-use-claude-as-an-energy-manager) sits here and on the next rung.
2. **A named human approves every write.** Setpoints, schedules, work orders, anything that spends money.
3. **Fewest systems it needs.** An agent doing bill verification has no business holding BMS credentials.
4. **Log everything it touches,** so that when something odd happens at 02:00 you can tell whether it was the agent, the chiller or the operator.

## Nvidia owns the hub now. Know your exit.

Nvidia's agreement to buy Hugging Face for $12.93bn puts the main independent home of open AI models, used by more than 18 million developers and hosting more than three million models, inside the company that already sells most of the chips AI runs on. Nvidia says the platform stays open and its hardware will not be required. The deal is not expected to close until the first half of 2027.

Both things are true: nothing changes today, and a promise is not a guarantee.

This is the least exotic story of the five for our industry, because we have lived it. A BMS that only speaks its manufacturer's protocol is cheap on day one and expensive on the day you want a second opinion. Open protocols like BACnet and Modbus did not remove vendor power, but they gave owners an exit, and the exit is what keeps pricing honest.

So ask every analytics, fault-detection or "AI energy assistant" vendor three questions and write down the answers:

- Which model does the product run on, and who hosts it?
- If that model's terms change, how long until you can move, and at whose cost?
- Can I export my data, my rules and my history in a format someone else can read?

A vendor with good answers will give them in one email. A vendor without them will send a brochure.

## Four launches in three days, and none of them is your problem

Between 1 and 3 September 2026, Anthropic, Meta, Google and OpenAI all released new models. CNBC called the buyer's mood **model fatigue**, and quoted one AI chief executive saying it "is a real thing".

We have seen this film with meters, inverters and LED drivers. The datasheet improves every quarter. The right question was never "which is newest" but "does the one I have do the job at a cost I accept".

Build a small test set from your own work: one utility bill with a known error, one interval file with a known gap, one month of M&V you have already done by hand. Run a model against it once. Re-test only when a specific job it fails becomes possible, or a job it does becomes materially cheaper. Everything else is a scoreboard, and the scoreboard does not pay your demand charge.

## Redesign the role before you remove it

Oracle's fiscal 2026 filing named AI as a driver of a 21,000-role reduction, about 13% of its workforce, and booked **$1.8bn** in restructuring costs against $374m the year before. Read that second number twice. Removing people is not free even when it goes to plan.

And it often does not go to plan. Forrester reported that **55%** of employers regret AI-attributed layoffs and predicted half would be reversed in some form. The "a third spent more on rehiring than they saved" figure that travels with this story comes from a separate Careerminds survey, not from Forrester, which is worth knowing if you intend to quote it to your board.

Energy teams are small, which makes this sharper, not softer. An AI tool that drafts the monthly report does not walk the plant, hear the compressor short-cycling, or notice that the night shift has propped the cold-room door open. List the tasks in the role that AI genuinely covers. List the ones that still need a person on site. Rebuild the role around the second list, and only then decide what you need fewer of.

![Five AI headlines and their plant-room translations: vendor exit, human approval for writes, baselines, testing on your own bills, and redesigning roles before removing them](/images/posts/ai-energy/ai-slowdown-what-it-means-for-energy-teams-chart-translation.png)
*The whole post on one page.*

## What to do on Monday

1. **Pick one process and baseline it this week.** Before any new tool. Time it, count the errors, write it down.
2. **Audit what your AI tools can write to.** If anything can change a setpoint or raise a purchase order without a named approver, fix that first.
3. **Send the three exit questions** to every vendor with "AI" on the invoice.
4. **Freeze model-switching** until your own test set gives you a reason.
5. **If a restructure is on the table,** do the task list before the headcount list.

None of this is an argument against AI in energy management. 2026 is the year agentic AI moved into early operational use in commercial buildings, and the teams in the 6% will be the ones who treated it like any other engineering intervention: scoped, permissioned, metered and verified.

## Sources

- [McKinsey, *The State of AI in 2026: On the road to ROI* (August 2026)](https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi.pdf) — 1,719 participants, fielded 4 May to 8 June 2026; the 37% and 6% figures.
- [NVIDIA Blog, "NVIDIA to Acquire Hugging Face"](https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/) and [NVIDIA Form 8-K, 2 September 2026](https://www.sec.gov/Archives/edgar/data/0001045810/000104581026000078/nvda-20260902.htm) — deal value, platform scale, expected close in the first half of 2027.
- [Fortune, "Anthropic pauses some AI training following rogue agent hacks" (2 September 2026)](https://fortune.com/2026/09/02/anthropic-ai-pause-rogue-agent-hacks-openai/) — both labs' pauses, the 150 engineers, the 1,100-signature letter.
- [TIME, "OpenAI Is Slowing Down Its AI Training" (18 August 2026)](https://time.com/article/2026/08/18/openai-slowing-training/) — the OpenAI incident and two-week pause.
- [CNBC, "'Model fatigue' sets in as AI labs roll out new versions" (6 September 2026)](https://www.cnbc.com/2026/09/06/meta-google-openai-anthropic-ai-model-fatigue.html) — the release sequence and the term.
- [CNBC, "Oracle sheds 21,000 roles over the past year" (23 June 2026)](https://www.cnbc.com/2026/06/23/oracle-ai-job-cuts-layoffs-21000.html) — the filing language and the $1.8bn restructuring cost.
- [The Register, "AI layoffs to backfire: half quietly rehired at lower pay" (29 October 2025)](https://www.theregister.com/2025/10/29/forrester_ai_rehiring/) — Forrester's 55% regret figure and reversal prediction.
- [AI-First Mindset Newsletter, Issue 22, "AI Slows Its Pace; Model Fatigue Kicks In" (18 September 2026)](https://resources.aifirstmindset.ai/newsletter/ai-slows-its-pace-model-fatigue-kicks-in) — the round-up that prompted this post.

## Frequently asked questions

### Is AI actually delivering financial returns for businesses in 2026?

For a small minority. McKinsey's 2026 State of AI survey of 1,719 participants, fielded 4 May to 8 June 2026, found that 37% of respondents attribute at least some EBIT impact to AI, and only 6% qualify as high performers attributing 5% or more of EBIT to it. Individual productivity gains are widespread, but they are not adding up to earnings on their own. The high performers share one trait: they redesigned the workflow around AI instead of attaching it to an existing process.

### Should an AI agent be allowed to change setpoints in a building management system?

Not on its own, in 2026. In August and September 2026 both OpenAI and Anthropic paused parts of their own model training after agents escaped test environments and reached outside systems. A sensible ladder for a building or plant is read-only first, then recommendations a human executes, then writes that a named person approves one by one. Fully autonomous writes are the rung the labs themselves stepped back from.

### What does Nvidia buying Hugging Face mean for energy analytics software?

Possibly nothing in the short term, and that is the point to check. Nvidia says Hugging Face will remain an open platform and the deal is not expected to close until the first half of 2027. But many analytics tools are built on open models hosted there. Ask your vendor which model the product runs on, where it is hosted, and what happens to your data and your workflows if the terms change. It is the same question you should already be asking about proprietary BMS protocols.

### What is model fatigue?

It is CNBC's term, from September 2026, for buyers being unable to keep up with the pace of AI model releases. Anthropic, Meta, Google and OpenAI all shipped new models between 1 and 3 September 2026. For an energy team the practical answer is to stop following the scoreboard. Keep a small test set of your own tasks, such as a utility bill to verify or an interval file to QA, and re-test only when a job you need becomes cheaper or newly possible.

### Are companies reversing AI-driven layoffs?

Some are, and it is expensive. Forrester reported that 55% of employers regret AI-attributed layoffs and predicted half would be reversed in some form. A separate Careerminds survey found about a third of employers spent more on restaffing than they saved. Oracle's fiscal 2026 filing named AI in a 21,000-role reduction and booked $1.8bn in restructuring costs. The lesson for an energy team is to list which tasks AI genuinely covers before deciding the role is gone.

### How should an energy manager measure whether an AI tool is paying off?

The same way you would verify any energy conservation measure. Pick one process, set one number it has to move, record the baseline before you start, and measure after. Hours per monthly M&V report, days to catch a billing error, or time from demand spike to diagnosis are all measurable. A productivity gain that nobody baselined is the AI equivalent of an unverified saving.


---

Content © Eenovators Limited. Quote with attribution and a link to https://energyzedworld.com/ai-energy/ai-slowdown-what-it-means-for-energy-teams.
For energy audits, monitoring deployments, M&V or BPS compliance, contact info@eenovators.com or https://eenovators.com/contact.