Backlogs just hit their lowest reading since January everywhere except data centers, and Fluor's stock jumped on a bet about which layer of the boom actually pays.

THE EXECUTIVE BRIEFING
THIS WEEK’S KEY TAKEAWAYS

Full episode write-up at the bottom↓

Senior AEC leaders at a closed London roundtable agree hourly billing is cracking. Big clients want the same scope for fewer hours, priced on outcomes, and nobody at the table had a working answer for what replaces the hours model.

ABC's Construction Backlog Indicator fell to 8.0 months in July, the lowest since January. Data center contractors sit on 11.4 months of backlog. Everyone else has 7.5, a gap that widened from 2.5 months in June to 3.9.

Fluor's shares jumped 17.4% in a day after CEO Jim Breuer said the firm would chase power ahead of data centers themselves, betting the safer money sits one layer beneath the boom.

7 THINGS WORTH YOUR ATTENTION
ON THE RADAR THIS WEEK

POWERED BY:

In late July, we sat in a closed room in London with some of the most senior people we know in this industry: Atul Khanzode, DPR Construction's chief technology officer, alongside tier-one leaders, engineers, investors, and founders. The only house rule was to say what was actually keeping you up at night. We're sharing some of what came out of that room today. Some of it stays with us.

What followed felt strangely well-timed. Within two weeks, the numbers confirmed what half that room already suspected: American contractors posted their shortest backlogs since January, the US lost jobs while construction gained them, and one of the biggest engineering firms on earth watched its shares jump nearly 18% for saying it would avoid building data centers directly.

The house rule at the roundtable was to say what you actually think. The first thing the room agreed on was that the pricing model construction that has run on for decades is starting to crack. Most big contracts get priced by counting hours: estimate how many people a job needs and for how long, then build the fee on top.

The biggest clients on the planet are now telling their supply chains, politely and otherwise, to deliver the same work with fewer hours because of AI, priced on the outcome instead. Nobody at the table had an answer for what replaces it. Everyone agreed it lands in design first, then construction.

Second, the near-term AI exposure sits in the office, where field labor is scarce, robots remain years from most site tasks, and roughly half the people at a large contractor never lift a tool, only estimating, reporting, and handling admin, which is exactly where the pressure lands first.

Third, no crisis is forcing anyone to change, and the room saw that as the real danger: the biggest contractors on both sides of the Atlantic just had record years, full order books included, and winning teams rarely pull their best people off live jobs to rethink the work, because those are exactly the people pinned to the hardest projects.

Fourth, data is the one asset this industry holds that the technology world can't copy, and there are no shared standards for it yet, so most firms feed AI tools without any policy governing it, which means the tools sold back are trained on whatever gets handed over. Atul has talked before about DPR's rule that innovation only survives contact with a real project budget, testing everything before it touches DPR's tech investment, and that same discipline filled the room in July.

Three weeks ago we asked whether the data center pipeline was pausing or peaking. Part of the answer arrived last Tuesday, in the backlog.

Associated Builders and Contractors surveys members monthly on backlog, the industry's runway of signed but unbuilt work. In July, the average backlog fell to 8.0 months, down 0.8 on the month and year, the lowest since January, falling everywhere: every category, region, and company size.

The 12% of contractors holding data center contracts sit on 11.4 months of work. The other 88% are down to 7.5, a gap we flagged back in March that's now widening fast: 2.5 months in June, 3.9 in July. Anirban Basu, ABC's chief economist, didn't mince words: "The data center boom masks the depth of this weakness." Firms doing $30 million to $50 million a year just recorded their lowest backlog since March 2020.

Hold that against the jobs report from the same week, when the US economy shed 23,000 jobs in July while construction added 22,000, almost all specialty trades on nonresidential work, nearly canceling out an entire economy's losses on its own. Contractors' own confidence readings still sit above the growth line, so the same firms reporting a shorter backlog still expect sales to rise.

Then Fluor reported revenue of $4.3 billion, up 9%, new awards of $6.1 billion, backlog of $26.9 billion, and shares up 17.4% in a day, after chief executive Jim Breuer put where the firm plays in one line: "Power, number one. Data centers, number two." Fluor will only build data centers where terms fit its risk appetite, betting the money sits one layer down, on contracts where someone else carries the surprises, and the market repriced the company nearly 20% for saying so.

Two results from the same fortnight tempered that enthusiasm: AECOM took a $337 million hit on a 2019 project, on terms it says it wouldn't accept today, and Balfour Beatty, sitting on a £22.9 billion order book, reported a UK margin of just 3.4%, a reminder that booked work is not the same as banked profit.

Certain developments in AI this week bears directly on this industry. Meta and Nvidia both released open-weight models, and twenty-five American tech companies, Microsoft among them, signed a letter asking Washington not to restrict them, notably without OpenAI or Anthropic, who sell closed models and are both preparing to float on the stock market.

Here's the plant-hire analogy regulars of this show will know: a closed model is like a hiring kit with the supplier's operator riding in the cab, where you pay by usage and never touch the machine, while open weights mean you download the machine itself and nobody meters your hours. The open-weight race has so far been led by Chinese labs, DeepSeek, Kimi and Qwen, with Meta and Nvidia's releases this week marking the American answer.

There is a definite reason why a construction executive should care that the price of intelligence is collapsing toward zero. GPT-4 level capability cost around $30 per million tokens in 2023, and today it's fractions of a cent, with Epoch clocking the decline at roughly fiftyfold a year and Gartner projecting running costs to fall another 90% by 2030. Venture firm N47 found enterprise spend per token fell 67% in a year, and open-weight models now cover close to half of production traffic, up from 11% two years back.

Sovereign AI is spreading through venture capital as intelligence gets cheap. Sequoia pulled about eighty founders into a room to teach them one thing: own your intelligence. Their slogan: "Not your weights, not your product." Owning means training an open-weight model on your own data until it beats the general-purpose giants at your job. Harvey, the legal AI firm, can't train on client files, so its lawyers build fake deal rooms instead. Cursor trains its own models, citing running costs five to ten times lower once it owns the model.

Ownership is evident in construction too. Larsen and Toubro, one of India's biggest engineering groups, took an order this week to host ten thousand Nvidia chips at its Chennai campus, calling it their entry into the AI factory business, a builder wanting to own the computing as well as pour the concrete. We went deeper on that data ownership gap in an earlier briefing.

On pricing: the hours model is already cracking in design, and it's only a matter of time before it reaches the field. If your fee structure hasn't had this conversation yet, it will soon.

On backlogs: the topline number matters less than which segment produced it. A contractor sitting on eleven months of data center work and a contractor sitting on seven and a half months of everything else aren't living in the same economy, even if their invoices look similar.

On ownership: Atul's test at DPR, whether a technology survives contact with a live project budget, is worth pointing out for every AI vendor on your books, the hours, the power, the model, before someone else applies it for you. One place to start: ask which model actually runs underneath the product, and let the answer, or the silence, tell you who you're really buying from.