The AI Industry Just Signed Three Union Deals. Not One Is About What Happens After the Building Opens.
Organized labor's biggest wins of the AI era are all in construction — the one part of the AI economy that ends when the ribbon is cut. Here is what the paperwork actually says.
Back in March, when the news broke that OpenAI had signed something with America's building trades unions, I felt relief. Physical relief — the kind that arrives when somebody else picks up a problem you have been carrying around unopened.
I did not read the announcement. I read the headline and built a comfortable little syllogism on it: unions are at the table, unions are not naive about automation, therefore somebody with real leverage is finally asking the AI companies what happens to the people whose work the machines actually do. I said a version of that out loud more than once, to people who I suspect were glad to hear it.
Five months later the third deal landed, and I finally sat down with the paperwork. The comfort did not survive the reading.
Here is what I got wrong, and it is a very specific kind of wrong. I confused labor being at the table with labor being at that table.
First, what is actually on the paper
Three announcements in five months, one counterparty. On March 11, 2026, OpenAI and North America's Building Trades Unions — NABTU, the federation for the electricians, pipefitters, ironworkers and operating engineers who physically build things — joined forces to train construction workers. On August 10, BlackRock signed with organized labor to line up workers for data centers and energy projects — the latest effort, in Bloomberg's wire copy, to emphasize jobs created rather than replaced by rapidly growing AI. Two days later Meta announced a new partnership with NABTU covering skilled trades workers nationwide.
Not a small counterparty: NABTU speaks for more than 3.2 million skilled craft professionals, roughly 300,000 of them apprentices, under a president, Sean McGarvey, in the job since 2012. Trade press reading the run concluded NABTU has become a structural partner in how America builds its AI economy.
So far so good. Now the part I skipped in March: what kind of document is each of these?
None of them is quite the object the word "contract" puts in your head. BlackRock's is a memorandum of understanding; the two sides cited the benefits of responsible contractor programs and of project labor agreements — the sweeping contracts that set wages across a whole project. Citing the benefits of a thing is not signing one. OpenAI's commits the parties to constructive engagement on policy, workforce development and labor standards: useful, and not a wage floor. Meta's promises training, credentials and career pathways through America's Workforce Academy, launched in June with an initial $115 million first year investment and a guaranteed job for graduates. Those qualifiers — "initial," "first year" — are Meta's own; I keep them because Meta did.
One binding instrument exists, covering one site: on April 20, 2026, NABTU announced OpenAI's Michigan Stargate campus would be built under a building trades agreement, expected to employ more than 2,500 tradespeople and apprentices. One campus. Everything else is a pipeline.
And what does the pipeline teach? Affiliates will tailor training to high-voltage systems, cooling and fire suppression systems, and secure fiber networks. The reporter who wrote that up found nothing about AI and job loss anywhere in the announcement — because there is nothing to find, as a trade-press column the same week confirmed.
None of that is a scandal — it is a workforce agreement doing what workforce agreements do. The problem was the story I told myself about it.
The most important word in the file is "during"
Take the argument from the people selling the project. Oracle's press release for the Michigan campus, published June 1, 2026, says that during construction the project will create more than 2,500 union jobs, and thanks the 700-plus tradesmen and women currently building it. An honest number.
The release states no figure whatsoever for permanent operating jobs.
Not a trick — precision. Brookings researchers, updating their county analysis on August 10, 2026, found data centers do create local jobs, with caveats: at a typical treated county the estimates imply roughly 100-200 jobs, depending on facility type, with wages unaffected and home prices up 2% to 5%. Their plainer sentence is the one to keep: large projects "often promise only dozens to a few hundred permanent workers while the associated construction jobs are temporary."
CBS News found the same shape: a boomlet in blue-collar jobs, with experts noting most of it is construction work, and construction work is temporary. And the buildout is a modest slice of the whole: per a June Census Bureau report cited in Bloomberg Law's Daily Labor Report, data center construction accounts for 2.3% of all US construction spending, though private spending on it hit $50 billion in April 2026 alone. Small share, enormous check.
So the deals cover a workforce that is large, well-paid, genuinely skilled — and gone when the ribbon is cut. Nobody has signed anything covering the people who stay, or the far larger number who never enter the building and whose work it is built to do.
Before you decide the electricians sold you out
Here is where I argue against my own headline, because the trades have a better case than I gave them credit for.
Mark McManus, general president of the United Association, told Fortune and the AP that if his union had declared a moratorium on data centers because it did not believe they were right for America, the data centers would still be getting built. Refusing the work does not stop it; it just means somebody else does it, for less.
Then the baseline nobody enjoys saying out loud: 10.0 percent of wage and salary workers belonged to a union in 2025, and 5.9 percent in the private sector. Criticizing a union for winning a strong agreement for its members, in a country where nine in ten workers have none, is a strange use of one's outrage.
And the hardest objection: the displacement I am worried about may not be arriving on schedule. The R Street Institute, from the libertarian side, calls it the jobs apocalypse that wasn't, citing Federal Reserve Bank of Chicago work finding that exposure to AI "does not map mechanically onto job loss, at least in the short run."
The strongest empirical read agrees, with a twist. Stanford's Digital Economy Lab, using ADP payroll data, reported in August 2026 that "we do not see widespread, economy-wide job displacement associated with AI." What they do see is concentrated: employment among workers aged 22 to 25 in highly AI-exposed occupations now sits about 19% below where it would be had it kept pace with similarly aged workers in less-exposed jobs — 15% by the same measure at the July 2025 data vintage — with the adjustment running through reduced hiring rather than increased separations. The authors are scrupulous: they call these descriptive patterns, not causal estimates and say they cannot yet answer the causal question definitively.
That changes my tone, not my point: there is a category of harm these three agreements do not address, nothing else addresses it at anything like this scale, and the celebration around the announcements has been filling that gap with a feeling rather than a clause.
We know what the other kind of agreement looks like
A few unions have already won the clause.
The Communications Workers of America puts the mechanism plainly: a union contract is backed by our power to grieve, arbitrate, and enforce, which is why AI language inside one has teeth a press release does not — bargaining units with the NewsGuild-CWA have ratified 58 contracts containing AI language, and members used it to win an arbitration against POLITICO management.
In Pennsylvania, the Governor's Office of Administration negotiated a binding "side letter" agreement with SEIU Local 668, which represents nearly 10,000 Commonwealth employees, and committed to a "human in the loop" process where human workers oversee, review and guide AI outputs. Read it carefully, though: the governor framed the rollout as having a goal of boosting productivity, not an agenda of substituting employees with AI. A goal, not a prohibition.
Europe carries the same hedges. Ireland's Financial Services Union and Bank of Ireland published a joint agreement in 2025 committing both sides to the human-in-control principle — while saying AI surveillance "should" be limited. Italian telecom unions negotiated a national agreement covering 200,000 workers that won them a seat at the table on AI and algorithmic management, and Konecta Italia signed Italy's first company-level AI agreement with SLC-CGIL.
Together, those flagship deals secured AI protections for roughly 30,000 members. NABTU alone speaks for 3.2 million.
Germany attached the right to the workplace, not to the project
Which brings me to the country that solved this by accident, decades before anyone needed it solved.
German employment law does not think in projects. It thinks in establishments — workplaces. Under the Works Constitution Act, a works council can be set up in all private sector workplaces with at least five employees, and its rights do not depend on who is building what.
In 2021, lawmakers went into the Works Constitution Act and added four words: the employer must inform the works council in due time of plans concerning working procedures and operations, "including the use of artificial intelligence," and consult early enough that the council's objections can still change the plan. Those clauses went into effect on June 18, 2021, under an act that also lets a council call in an outside expert whenever the employer wants to introduce and use AI.
A sharper right sits alongside it: works councils co-determine technical equipment suitable for monitoring employee behavior or performance — likely one of the most relevant works council's rights, German counsel say, even while conceding that legislators have otherwise written "hardly any AI specific regulations" for employment. And the trigger is capability, not intent: employer-side lawyers warn a tool falls under co-determination if it is merely objectively capable of monitoring, and decisions taken without consulting the council are not valid.
Then the provision the American pacts have no counterpart for. German law lists, among the workplace "alterations" an employer must disclose and negotiate, the introduction of entirely new work methods and production processes. Where that applies, the parties negotiate a social compensation plan — compensation for financial harm caused by the change — and if they cannot agree, a conciliation committee decides, its award taking the place of an agreement. Read the hedges in that same statute, though, because they are load-bearing: the regime reaches companies that normally have more than twenty employees with voting rights, and alterations which may entail substantial prejudice to staff.
So it is not an automatic trigger on every AI rollout. It is a named legal mechanism for the question the software changed the work — who pays for that? The United States has none.
The limits go further. In January 2024 the Hamburg labor court found no co-determination right where staff used ChatGPT voluntarily through private accounts at their own expense — while affirming the AI information rights still applied, and noting that a different assessment of the legal situation would be required if the employer provided company accounts. Narrow facts, narrow ruling.
The biggest limit is coverage. A works council is a right, not a guarantee: in 2023 works councils covered 41% of all employees and 34% in the private sector, down from 57% in 1996. Where one exists, the AI rights bite. Where one does not, a German office worker is closer to an American one than either side likes to admit.
And Germany did not opt out of the buildout: Berlin adopted a national data center strategy on March 18, 2026 targeting a doubling of overall capacity by 2030 — explicitly non-binding — while the boom is exposing the limits of the energy supply exactly as ours is. Same buildings, same grid fights, different labor architecture underneath: works councils there negotiate actual AI works agreements, which the Hans-Böckler-Stiftung's institute archives, and when a job really is restructured away, Germany pays a qualification allowance — 60% of net salary, 67% with children — to workers retraining because structural change reshaped their employer.
Not a fortress either. Just plumbing that exists.
Now run it forward to 2032
Speculating on purpose — take what is on the record and let it run.
The Michigan campus opens. The 2,500 tradespeople finish and do what tradespeople have always done: follow the work to the next site, on the strength of an apprenticeship the AI industry helped pay for. Genuinely good. A caretaker crew stays — call it the low end of the Brookings range.
Meanwhile the techniques mature: more prefabrication, more automated layout, more repetitive work done by machines a data center company has every incentive to fund. And the trades discover the coverage gap they helped normalize is pointed at them. An agreement promising pathways says nothing about the pathway getting shorter.
Two towns over, an insurance claims department stops replacing the people who leave. No layoffs, no announcement — just a hiring pattern that quietly matches the entry-level shortfall Stanford is already measuring. Nobody in that building has a works council, a side letter, or a grievance procedure. No document, so no argument, so no story.
The honest history of this period may not be "AI destroyed the jobs." It may be duller and worse: the institutions that could have bargained over it were busy, successfully, bargaining over something else.
I would like to be wrong. The way to make me wrong is to write the clause.
What the people who study this for a living are saying
Organized labor already knows. Bloomberg Law's Daily Labor Report found the buildout exposing a tension within organized labor over how the technology will hit members' jobs. Anne Lofaso, a law professor at the University of Cincinnati and a former National Labor Relations Board attorney, put it starkly: the AFL-CIO always carries this tension because its unions have different interests, "but this truly is existential. This is about jobs." NABTU's chief of staff, Mike Monroe, answered for the trades — these are not the only things being built, "but this is where a lot of the action is" — and Todd Vachon of Rutgers supplied the arithmetic: without construction, members are on unemployment until it runs out.
The political flank is loud and unresolved. Senator Bernie Sanders and Representative Alexandria Ocasio-Cortez announced the AI Data Center Moratorium Act on March 25, 2026, and Sanders made the working-class case in Fox News's opinion pages — the coalition lines here are not the ones you expect. At a separate event in April, about a half-dozen AFL-CIO-affiliated leaders appeared alongside Sanders, and only two, from National Nurses United and the Association of Flight Attendants, explicitly said they supported a slowdown. Even the opposition is not one thing. Outside labor it is scaling fast: a tracking report identified 430 local data center opposition groups organizing on Facebook by late July 2026, up from 268 in April — which is why a building trades endorsement is worth real money at a zoning hearing.
And the researchers are making the clause writeable. CWA published union principles for artificial intelligence in December 2023, demanding language against invasive surveillance, unfair automated employment decisions and the reduction of union members' work. UC Berkeley's Labor Center built an inventory of technology clauses drawn from more than 500 union contracts; UNI Europa and partners built a database of clauses on AI and algorithmic management in European agreements. And Virginia Doellgast at Cornell's ILR School studies precisely this seam: digitalization, AI and job quality across North America and Europe.
The language exists. Somebody has to put it into a document that can be enforced.
What does this mean for you?
Learn the nouns. MOU, partnership, project labor agreement, side letter, collective bargaining agreement — different objects, different consequences. When the next AI-and-labor headline lands, find the noun before you decide how you feel. It took me five months.
If a union contract covers you, find out whether it says anything about AI. Not whether your employer has a policy — whether your contract does. A policy is revocable; contract language is grievable.
If you work in the public sector, Pennsylvania is the template. A binding side letter with a joint labor-management group is modest, unglamorous, and entirely copyable.
If you are one of the nine in ten with no contract, ask the German questions anyway. What systems are being introduced, what do they measure, who reviewed them, who decides when the output is wrong? You have no statutory right to an answer — asking in writing, as a group, still changes the temperature.
If a data center is proposed near you, ask for two numbers in writing. Construction jobs and permanent operating jobs, separately. If the second number is missing, that absence is the answer.
If you hire, or you are 22 to 25, watch the entry level and hold the hedge. The clearest signal so far is reduced hiring in exposed occupations, not mass layoffs — and the researchers will not yet call it causal. Both halves matter.
The lesson, as I see it
Nobody here behaved badly. NABTU bargained hard and won something real: training, apprenticeships and a seat at the table of the biggest construction boom in a generation. If I were a 24-year-old electrician I would take that deal on the spot, and be right to.
My error was arithmetic dressed up as optimism. I heard "labor won" and stopped counting who "labor" meant. It meant the people who build the box — never the people the box is pointed at, and it never pretended otherwise. The press releases are honest, which is exactly why reading them was uncomfortable.
The German lesson is not that Germans care more; their coverage has fallen to 41% and their buildout strains the same grid ours does. The lesson is architectural. They attached the right to the workplace instead of the project, long before anyone imagined this technology, and when it arrived they added a clause rather than starting over. Durable rights are boring, procedural and built early — and that trick is available to us.
So when the fourth deal is announced — and it will be, probably before winter — read past the headline to the noun, then look for the sentence about what happens after the building opens. If it is not there, do not call it a betrayal. Call it a vacancy.
Vacancies get filled by whoever bothers to show up with a draft.
Reading the actual paperwork instead of the press release is most of what the HAIA Foundation does with its week. If that is useful to you, subscribe — and forward this to someone who will be in the room the next time one of these gets signed.