AI Reindustrialization: M&A Teams Source Manufacturing Targets

Public AI stocks drew down while token demand accelerated. Gavin Baker's read on that gap is the most useful framing for M&A and corporate innovation teams that I've seen this year, not because it tells you which ticker to buy, but because it tells you which physical nodes to source.

Quick answer: The AI boom is reindustrializing America by driving demand for power infrastructure, semiconductor fabrication, copper, and skilled trades, not just software. For corporate M&A and innovation teams, the acquisition targets worth finding are the companies enabling that physical build: grid interconnect, behind-the-meter generation, modular data center construction, and the industrial supply chain behind them. Start with megawatts connected, not GPU slide decks.

The demand curve is not what the lab revenue prints suggest

Baker's sharpest observation is that heavy, paying AI users are probably fewer than 30 million people globally, possibly under 10 million. That number sits against roughly 1.5 billion knowledge workers who could theoretically use these tools. Old-economy companies spending on tokens typically allocate around 1% of compensation. AI-native companies run in the high single digits to 10%-plus. The highest-spending engineers individually consume 10 to 100 times the median user.

That spread tells a specific story: we are at the very beginning of the diffusion curve, and the supply side is already binding. If you cannot tell Baker one quantitative data point in your AI business that is getting worse—one customer metric, one leading indicator—his challenge to skeptics is that he has not found a single person who can answer it. The bear case requires that something measurable deteriorates. So far it has not appeared.

For corporate sourcing teams, that framing reorients the acquisition mandate. The interesting companies are not necessarily the model labs. They are the companies that make the build possible.

Physical constraints are running the table

The bottleneck in the AI build is not algorithm quality. It is wafers, copper, grid interconnect, gas supply, trained electricians, and financing for construction. Those constraints are slowing the boom, and Baker's position is that this is socially useful. A debt-funded buildout demands immediate ROI. A build funded largely from operating cash flows, as much of this one currently is, can absorb the long infrastructure timelines without forced liquidation.

Copper deserves more attention than it gets in AI coverage. If even 10% of the data center demand story lands, copper throughput requirements scale dramatically. The same applies to natural gas: US gas prices run roughly $2 to $3 per million BTU. Europe and Asia run $20 to $25. That gap is not a trade footnote. It is the underlying arithmetic that makes American reindustrialization a viable industrial policy claim, not just a slogan.

Payback periods on live megawatts are already measurable. Nebius-style builds are hitting roughly 9 to 10 months after customer prepay. SpaceX has moved faster because clusters come online quickly. The right unit of analysis for M&A teams evaluating this space is dollars per megawatt, not dollars per GPU. Hardware preference is nearly unreadable during a shortage. Deal structure and interconnect access tell you more.

All layers matter: the acquisition map is a layer cake, not a tournament

One framing error that costs sourcing teams real opportunities is treating the AI ecosystem as a competition between frontier labs, open-source models, and application companies, as if only one layer survives. Baker's read is that all of them can work simultaneously. Frontier models will do things that smaller models cannot. N-minus-one models will handle the volume. Open source will drive token costs down, which is good for Nvidia because cheaper tokens mean more tokens mean more GPUs. Application companies with genuine workflow integration will build durable businesses.

The implication for M&A is that the acquisition map is a layer cake, not a winner-take-all tournament. The question is which layer is undercovered relative to its strategic value. Right now that answer is infrastructure and enabling technology, the companies that are making the build happen in specific geographies. Those companies are less likely to appear in Crunchbase and more likely to show up when you map a region by active power procurement, construction permits, and industrial hiring.

That kind of geographic and sector-specific sourcing is where teams miss 40 to 60% of the market, according to FounderNest's 2026 Scouting and Deal Sourcing Report. The pattern holds especially for infrequent acquirers re-entering the market now after years on the sidelines. If your team is building a new mandate around AI infrastructure or reindustrialization targets, the startup sourcing mandate template we published is a useful starting point for scoping it precisely.

Data centers as local industrial policy

The political argument for data centers is being made badly. Baker's diagnosis: "stay ahead of China" is correct but ineffective because it is too abstract for most Americans to act on. The argument that works is local and wage-denominated.

Loudoun County, Virginia is simultaneously the richest county in the United States and the densest data center county in the world. That is not a coincidence. When a data center goes in, it transforms local tax revenue. In some cases it has been described as a 10x for a town's fiscal base. The labor demand is in the trades: electricians, HVAC technicians, plumbers, construction crews. The apprentice wages in those trades are already competitive with entry-level knowledge-work salaries. Baker's position is that the NPV on a trades apprenticeship is becoming favorable relative to a four-year degree precisely because of this build cycle.

For M&A and innovation teams, this geographic signal is an early indicator of where the next infrastructure-linked acquisition targets will cluster. The towns seeing behind-the-meter generation projects, new substation permits, and data center construction are producing suppliers and integrators that will not appear in standard databases for 12 to 18 months. Reading that layer—active construction, local hiring, power procurement deals—is how you find the company before your competitor does.

The same methodology applies to the broader reindustrialization story. We mapped this pattern in the Japan-to-North-America green-tech corridor and it holds: the physical infrastructure signals precede the fundable company by long enough that teams starting from a directory are perpetually late.

The enterprise model ownership question

Baker's enterprise stack prescription is specific. Take a strong open-source or large-foundation base model, run reinforcement learning and supervised fine-tuning on your own proprietary data so the company owns the weights, then route to one or two frontier models for complex planning while cheaper models handle execution volume.

The strategic logic is risk reduction. If intelligence is a core input into your business, the way a specialized database or a pricing model was a decade ago, then sharing your core enterprise context with a frontier lab is a financial-health risk, not just a compliance issue. Kirkland and Ellis committing roughly $500 million to build in-house validates the size of the prize, not a law firm's ability to run a continuously updated model factory. Whoever becomes the abstraction layer of intelligence for an enterprise is in the most valuable seat in that stack.

For M&A teams, this creates a sourcing question worth filing now: which companies in your sector are building proprietary fine-tuning infrastructure or sector-specific model deployment tooling? Those are the targets that will be strategic in 36 months and overpriced in 60. Coding and legal are the clean early examples because the work is documented and verifiable. The other 1.5 billion knowledge-work seats are harder, messier, and less covered, which is where the interesting acquisition pipeline sits.

What the Nvidia ecosystem signal actually tells you

Baker's Nvidia framing is worth taking literally. Nvidia's moat is not only chips. It is land, power, shells, financing relationships, open-source tooling, and the entire supply chain that has to coordinate for a cluster to come online. Open source benefits Nvidia because it drives token volume, which drives GPU demand. The ecosystem is self-reinforcing.

For corporate sourcing, the investment thesis is to find a way to plug into that ecosystem at the supplier and infrastructure layer, not to compete with it. That means the targets worth finding are companies that hold scarce inputs: interconnect rights, behind-the-meter generation capacity, specialized construction expertise, or proprietary data on power availability by region. Those are not on the typical venture-funded startup track. Many are private, regional, and not listed anywhere a standard directory would surface them.

Reading deal structure and interconnect agreements is a more reliable signal than reading hardware preference during a shortage because hardware preferences shift as supply fluctuates. Power agreements and interconnect positions are long-term and harder to replicate.

FAQ

What does reindustrialization mean for corporate M&A teams right now?

Reindustrialization in the AI context means the physical infrastructure build, data centers, power generation, grid interconnect, copper supply chains, and specialized construction is creating a new category of acquisition targets in geographies that M&A teams have not historically sourced from. The relevant companies are often off-market, regional, and not captured in standard databases, which makes geographic and sector-specific sourcing methodology the differentiator.

How should corporate teams evaluate AI infrastructure targets versus software targets?

Corporate teams evaluating AI infrastructure targets should use dollars per megawatt and payback on live capacity as primary metrics, not GPU count or model benchmark scores. Software targets in this space should be evaluated on whether they hold a proprietary data position or model weights because the enterprise value in AI applications is increasingly in the fine-tuned model trained on company data, not the interface on top of it.

Why are trades and construction relevant to an M&A sourcing mandate?

Trades and construction are relevant because the AI data center build is constrained by licensed electricians, HVAC technicians, and specialized construction capacity more than by capital. Companies that control those labor pipelines or hold established relationships with utilities for behind-the-meter generation are sitting on scarce inputs that will become acquisition targets or strategic partnership targets as the build cycle extends.

How do you find companies in the AI infrastructure and reindustrialization stack before they appear in directories?

Finding companies before they appear in standard databases requires reading physical signals: active power procurement filings, substation permits, construction activity, and local hiring patterns in target geographies. The district-level evaluation methodology we outlined in the innovation district scorecard applies directly here. Measure pilots running and megawatts contracted, not press releases about planned investment.

Is the AI boom a bubble that changes this analysis?

Every major technology cycle has produced a bubble, and an overbuild often follows. The current build is notably different in that a significant share is funded from operating cash flows rather than debt, which removes the forced-liquidation pressure that collapses debt-funded overbuilds quickly. The scarce physical inputs—power, land, interconnect, copper—are real constraints that will persist regardless of which application companies succeed. That means the infrastructure layer of the M&A map stays relevant across boom-and-correction scenarios.

If your team is sourcing against an AI infrastructure or reindustrialization mandate and needs a vetted short list of active, relevant companies rather than a directory to comb through, Innovation Scout is the fastest way to start.