The Debrief: July 2026

The platforms are buying the inputs

Once a month, I sit down with some of the brightest minds in the industry and argue through the construction news that actually mattered. The episode is the conversation. This is the analysis.

The month in one line: the platforms are buying the inputs. Procore paid $845 million in cash for DroneDeploy and its 20 trillion square feet of jobsite imagery the week we recorded. New York decided which data centers get to exist by rationing permits. Suffolk decided it would rather build its own intelligence layer than rent one. Even the market data got contested, with the industry’s two starts trackers describing the same month in opposite directions. Every story this month is a fight over who controls the raw material, whether that raw material is visual data, megawatts, workflows, or the numbers themselves.

This briefing synthesizes the month’s construction and construction tech news with practitioner commentary from the July Reviewed as Noted roundtable. Tyler Campbell of Fieldproof, who advises contractors on AI and data strategy, returns. Joining him this month: @AJ Waters, four weeks into a new role at Procore, and Victor Muchiri , recently off the GC side, who reads deals through a capital allocation lens.

---

1. Procore Buys the Input Device: $845 Million in Cash for DroneDeploy

The news: On July 30, Procore announced it will acquire DroneDeploy, the reality capture and robotics platform, for $845 million in cash, expected to close later in 2026. DroneDeploy operates in more than 180 countries and holds roughly 20 trillion square feet of visual jobsite data, tens of millions of user annotations, and a record of more than 100,000 labeled safety issues. Procore brings nearly 400 million photos, more than 126 million drawings, and over 10 million RFIs, submittals, and inspections recorded annually. The deal follows Procore’s January acquisition of Datagrid and the July 23 launch of its Digital Coworker agent packages, making 2026 the year Procore assembled an AI stack piece by piece: first the agent team, then the agent catalog, now the capture layer that feeds it.

Why it matters: Procore has been explicit that it wants to be a system of action, not just a system of record, and an agent can only act on what it can see. Everyone at the table opened this year saying the same thing about jobsite data capture: the input device is still broken. Buying DroneDeploy is buying an automated input method, plus the orchestration layer for drones and, eventually, ground robotics. It also lands ten months after Procore’s API terms began prohibiting marketplace partners from bulk downloading platform data for commercial purposes, including training large language models. Read together, the sequence is hard to miss: the terms fenced the data in, and the acquisitions are stocking the fenced area.

From the roundtable: The table’s first drone stories date to 2017 and 2018, a dam project measuring earthwork and a data center too big to photograph on foot. Eight years later that capture layer is an $845 million strategic asset, and the panel split on what the money actually bought. Victor’s read is capabilities and pole position: Autodesk has reached rough product parity, the big platforms are buying net new modalities rather than doing R&D, and against robotics startups raising hundreds of millions at seed, “the 845 number is actually not that crazy” from a venture math perspective. He also flagged the sleeper asset: training data for robotics, the kind of business Scale AI built for the LLM companies, sitting on top of data Procore already has. Tyler read the price itself as the tell: “I look at that as DroneDeploy probably just sold a lot of data on a lot of different job sites to Procore.” Put in the seat of a $300 million GC, his response to the deal was blunt: “I’m looking at my Procore bill every year, and I’m going, how much of this data am I putting into other people’s hands?” His answer is to pull data back internally and build an ontology layer that reads his whole business. AJ, corporate hat removed, defended the acquisition on context rather than code: “I can go build an app in my basement in four weeks, but if I’ve never set foot on a job site, I still have no context. AI is a tool, just like an arc welder is a tool.” He then put the hard question to Tyler’s build-it-yourself plan: is your data actually more secure in a homegrown stack with no security team, and pointed to the retail mega-breach that famously came in through a contractor. Nobody at the table fully answered that one.

2. New York Freezes Hyperscale Permits, and Offers the Trades a Deal

The news: Governor Hochul signed Executive Order 62 on July 14, making New York the first state to pause permitting for data centers capable of drawing 50 megawatts or more while the state prepares a Generic Environmental Impact Statement. The order sets no fixed end date; expiry is tied to the GEIS, whatever the “up to one year” framing in the announcement says. Roughly 12 gigawatts of data center load sits in the state’s interconnection queue, more than 8 gigawatts of it added in 2025 alone. The buried lede is the labor trade: the order directs a Community Investment Framework prioritizing prevailing wage, project labor agreements, local hire, and apprenticeships. Two days later, EPA moved the opposite direction, telling developers that islanded on-site power plants with no grid connection fall outside the Acid Rain Program. One arm of government slowed the grid path while another cleared the build-your-own-plant path, in the same week.

Why it matters: The stakes are quantified in AIA’s July 20 forecast: data center construction spending up 33% in 2026 and 25% in 2027, roughly 8% of all nonresidential building spending by 2027. Strip data centers out and the commercial forecast becomes a 1% decrease this year. A single asset class is the entire difference between a sector that grows and a sector that contracts, which is exactly why a single state’s permit action is now a national construction story. And because a moratorium is not a cancellation, the pause quietly consolidates the pipeline toward the developers with the deepest balance sheets, the ones who can wait out a GEIS or build their own power and skip the queue entirely.

From the roundtable: Nobody at the table read this as a win for the industry. AJ’s scorecard was that the politicians won and everyone else lost, at least at this juncture: the jobs inside these facilities are undercounted, the public underestimates how much of their daily life runs through them, and the prevailing wage provision, whatever its merits, means data centers will pull even harder on labor markets that were already losing people to them. Victor found the winners out of state: “The winners are the other states. If you’re trying to build a data center in New York and now you can go to Pennsylvania, why would you not?” As an owner, the regulatory headache is itself a site selection criterion. My own addition: whatever delay New York thinks it is imposing, triple it, because the labor that leaves for neighboring states during the pause does not come back on the state’s schedule.

3. Suffolk Embeds AI Engineers Inside Project Teams

The news: Suffolk is placing AI engineers directly inside project teams on mission-critical, healthcare, and higher-education work, running Claude and ChatGPT against RFI creation, change order documentation, procurement integration, and executive dashboards. The strategy is explicit: buy standard applications, build only bespoke tools, because Suffolk is “a construction company,” not a software firm. It is the strongest adoption signal of the month from a top-tier contractor, and it is also, roughly a year into the program, entirely unquantified. Every public statement remains forward-looking. Worth noting: Suffolk is one of the GCs named as a Trunk Tools user cut off when Procore revoked API access last fall, which makes the decision to build in-house read a little differently.

Why it matters: This is construction’s version of the forward deployed engineer, and the org design is the actual news. An engineer sitting in the trailer, hearing the real complaint at the real moment, is a structural choice any mid-size GC can imitate at a fraction of Suffolk’s cost. The model choice is the easy part and changes every six months. The discipline question is the hard part, and it is the one the panel spent twenty minutes fighting about.

From the roundtable: This was the sharpest disagreement of the episode. Tyler, who builds AI systems for contractors, came out skeptical: most of what he sees are process problems wearing AI costumes. He described a client losing an estimated $600K in revenue capacity because two tools were never connected with an out-of-the-box connector, no AI required. “If I throw AI at a crappy process, I’m getting crap faster.” He also warned where undisciplined mandates end: six-figure token bills with no verifiable return. Victor pushed back hard with the R&D math: “If you just imagine there’s five people there, 200K a head, so that’s a million bucks, double that for travel and expenses, that’s two million. If they find any sort of operational efficiency across the five billion that they do, that’s 100 percent worth it. You’ll do that trade every single day.” His position: chalk it up as R&D, scale it to ten people, and stop demanding attribution at the individual level, because the ROI shows up at the org level when a person can cover three jobs instead of one. AJ split the difference: he likes that Suffolk put the engineers where the work happens, and worries about the two things the press release skips, what these people cost and what happens when every project insists on doing things its own way. I land closer to Victor. Having lived at the project level, the zero-to-sixty on value is not hard to imagine if the people are senior enough, and Suffolk runs a tight enough ship that I would invest in this experiment. But Tyler’s closing caution deserves to hang over every one of these programs, and we will come back to it in the Outlook.

4. Two Data Vendors Described June in Opposite Directions

The news: On July 21, Dodge reported June construction starts down 19.9% to $1.42 trillion annualized, with nonresidential building down 9.1% and nonbuilding down 37.7%. Six days later, ConstructConnect reported June nonresidential building starts at a record $100.3 billion, up 32.8% from May and 35% above the twelve-month moving average, with offices and data centers accounting for more than a third of it. Same month, same country, opposite conclusions, and because there is no government series for construction starts, there is no referee. Both are proprietary trackers with different capture rules, seasonal adjustment, and scope definitions.

Why it matters: The reconciliation is probably scope, not scandal. Dodge’s headline includes nonbuilding infrastructure, which cratered, while ConstructConnect’s record is nonresidential building only. Both can be arithmetically true at once, which is somehow worse than one of them being wrong. The practical rule: name the vendor every time you cite a starts number, and when the headline and your own backlog disagree, trust your backlog. ConstructConnect’s own caveat, that the gains were “sharply concentrated rather than broad-based,” is the month’s thesis restated in vendor language.

From the roundtable: The ground truth at the table was unanimous and it was not Dodge’s. Tyler’s market read from his contractor clients: blowing up, people cannot keep up. Victor, from Ohio: data centers, manufacturing, everything going up at once, nobody can find help, boomtown. AJ has made a habit of interrogating the gloomy headlines: “Every time I see a headline that says construction starts are down, I instantly try to go figure out where... because it sure doesn’t feel like it anywhere.” His supporting evidence reached beyond data centers, pointing to Dulles airport’s newly announced $20 billion capital program as proof the big numbers are not one asset class. Tyler’s working theory for the split is project mix, that the two trackers are seeing different slices of the market flow through their systems, which is as good an explanation as any in a world where the industry’s picture of itself is assembled from vendor releases nobody audits.

Lightning Round: The Next Check

We closed by calling shots on the next big construction tech acquisition before year-end. My pick: one of the emerging scheduling and planning platforms goes, because it is past time Oracle faced a true challenger and the big platforms need a planning layer to feed scenario planning and the orchestration story. Victor agreed pre-con is the weakest flank of the big suites, then picked ClearStory: universally liked, real revenue, the kind of acquisition nobody argues with, assuming the founder sells. Tyler called his shot with more confidence than evidence, by his own admission: Trunk Tools gets acquired by a construction company, not a software company, because a large GC building its own ontology and agents gets a running start, and the platform fight has left Trunk Tools without a natural home among the majors. AJ abstained on specifics, week four and all, but confirmed the rumor mill is active and put a number on it: at least one, maybe two more major acquisitions before the year is out.

Outlook

The gap to watch is between who generates construction’s data and who gets paid for it. Contractors produced the 400 million photos, the 126 million drawings, and the 20 trillion square feet of imagery that made this month’s $845 million deal make sense, and the month’s biggest strategic moves, Procore’s acquisition, Suffolk’s build-in-house bet, Tyler’s clients pulling data behind their own walls, are all answers to the same question: is the data a contractor generates on its own jobsite an asset it controls or an input it donates?

Full Episode Here: