Here is the part of the AI boom that rarely makes the glossy renderings. The chips get the headlines. The campuses get the ribbon cuttings. What keeps those halls from going dark is a far less glamorous stack of transformers, switchgear, backup batteries, and optical gear. And a surprising share of that kit still comes from Chinese factories. I have been watching this corner of the market for months, and the gap between the political speech and the procurement reality is getting harder to ignore.
Why The Power Stack Suddenly Matters More Than The Chip Stack
For years the conversation about strategic risk in artificial intelligence stayed locked on semiconductors. That made sense. Advanced accelerators are scarce, expensive, and tightly controlled. Then the buildout accelerated. Hyperscalers started announcing multibillion-dollar campuses as if land and capital were the only constraints. Power turned out to be the constraint that does not negotiate.
A data center is not a warehouse with extra air conditioning. It is a high-density electrical machine that must stay online. Transformers step transmission voltage down to levels servers and cooling plants can use. Switchgear concentrates breakers, fuses, and isolation equipment so operators can route power without taking the whole site down. Batteries cover the ugly seconds between a grid event and generator start. Optical transceivers move enormous volumes of data as light instead of copper. Miss any one of those pieces and the most elegant compute cluster becomes an expensive room.
Washington has started treating that stack as a national issue, not a facilities footnote. An executive order last week framed foreign-made bulk-power equipment as an extraordinary threat and gave energy officials room to block or condition certain transactions. That language is not subtle. It is also not cost-free. In my view, the United States is trying to de-risk AI infrastructure before Chinese hardware is fully embedded in it. That instinct is understandable. The timing is brutal.
Scrutiny over China’s presence in the U.S. data centre power stack has only risen over the past eight months or so, this is a shift from the previous focus that centred solely on the compute stack.
– Industry analyst commentary
The Quiet Dependencies Inside A Modern Campus
Ask a construction lead what is late on an AI site and you will not always hear “GPUs.” You will hear “the transformer.” You will hear “switchgear.” You will hear that the battery containers are on a boat and the interconnection study keeps moving. Chinese suppliers have filled those gaps because they could deliver volume when Western shops were already booked.
Energy researchers put the mid-term exposure in a few buckets. Grid connectors sit at the center: transformers, switchgear, and batteries. China’s share of certain transformer and switchgear categories runs near 30%. Battery imports tell an even sharper story, with Chinese origin accounting for more than 40% of U.S. arrivals in key categories. That is not a rounding error. That is a structural habit.
There is a second layer people skip because it is dull. Upstream materials. Copper. Electrical steel. Battery cathode inputs. If the finished box is assembled somewhere else, the ore and the specialty steel still have to come from somewhere. I find this part more uncomfortable than the finished-goods maps, because you cannot tariff your way out of a mine that does not exist yet.
- On-site substation transformers that step transmission voltage down for hyperscale campuses
- Medium-voltage switchgear that isolates faults without collapsing a hall
- Backup battery systems that bridge grid interruptions
- Optical transceivers that carry east-west traffic between racks and buildings
- Upstream metals and electrical steel that feed all of the above
Optical technology deserves its own paragraph because it does not look like “power equipment” until you watch a cluster stall without it. Chinese vendors lead global data center optical transceiver revenue. Collectively they account for roughly two-thirds of global unit supply. Names you may never see on a billboard still sit inside American halls because the bits have to move, and they have to move now.
How Fast Capacity Is Growing, And Why That Makes Substitution Hard
Forecasts circulating this summer put U.S. data center capacity on a steep climb, from about 62 GW in March 2026 toward 152 GW by 2030 if high-density AI workloads keep landing as planned. Those numbers are not a vibe. They are a procurement schedule. Every extra gigawatt wants transformers measured in the tens of megavolt-amperes, switchgear that can be maintained safely, and optics that do not become the next single point of failure.
Western manufacturers are not sitting still. One major grid supplier committed about a billion dollars to U.S. production of critical infrastructure, including hundreds of millions for a new large power transformer plant aimed at AI demand. Another pledged a similar scale of investment in American production for grid and turbine equipment tied to data center growth. That is real money. It is also not a warehouse full of finished 100 MVA units sitting on a lot in Ohio next Tuesday.
Factories take years. Skilled winders take years. Electrical steel capacity takes years. I have found that public debate loves the announcement and underprices the commissioning date. A press release is not a nameplate on a transformer.
Washington’s New Posture On Bulk-Power Gear
The latest executive action treats bulk-power system equipment as essential to national defense. Transformers, transmission lines, conductors, substations, high-voltage circuit breakers, and power control electronics all sit inside that frame. The argument is straightforward. Advanced manufacturing, data centers, artificial intelligence, and defense production have raised the country’s need for reliable electricity. A successful attack or a supply disruption would hurt more than it would have a decade ago.
Officials also keep repeating a theme from earlier trade fights: reshoring items tied to economic and national security. Tax policy, tariffs, and faster permitting are the tools on the table. Whether those tools can outrun a 2026–2028 interconnection queue is another question. I am skeptical of tidy timelines. I am not skeptical of the political will. The will is there.
Power inverters made abroad, the devices that help stitch generation and storage into a site, were added to a federal covered list of equipment judged to pose unacceptable national security risk. Separately, reports have circulated that the administration is drafting restrictions on new Chinese optical transceivers. Those devices are the quiet plumbing of AI networking. Ban the cheap, high-volume versions and you do not just shuffle vendors. You shuffle lead times and bill of materials costs across every campus still in design.
The rapid growth of advanced manufacturing, data centers, artificial intelligence, and defense production has increased the Nation’s dependence on abundant, reliable electricity and magnified the consequences of a successful attack or supply disruption on the bulk-power system.
Optics, Photonics, And The Illusion Of Instant Substitution
American photonics houses are not weak. Several of them design excellent modules. A few even received multi-billion-dollar strategic investments from a leading chip firm earlier this year. Design is not the whole problem. Volume is the problem. Cleanroom capacity, automated packaging, and yield at scale take longer to clone than a slide deck implies.
Research notes from the optical market have been blunt. Western competitors may have the photonic designs. They currently lack the factory choreography required to absorb the unit volume of the large Chinese specialists inside a 12-to-24-month window. That sentence should be taped to every board packet that assumes a ban is a one-quarter event.
If restrictions land, hyperscalers will hunt local alternatives. Costs go up. Qualification cycles stretch. Dual-sourcing becomes a religion. None of that is tragic in the long run. In the short run it collides with campuses already poured in concrete. Perhaps the most interesting aspect is how little of this debate happens in public until a lead time slips from 60 weeks to 90.
Transformers Are Already Tight. Policy Can Make Them Tighter
Power transformers and substations were already looking short this year, on the order of a 15% gap for transformers and about 8% for substations in some market estimates for 2026. Restrictions on Chinese-made units would not invent that shortage. They would concentrate it.
The painful slice is the large end of the market, the 100 MVA+ class. That is exactly where data center developers have been turning to Chinese shops to compress lead times. If those doors close, the projects that used import units as a schedule hedge lose the hedge. I do not love that outcome. I also do not pretend the security concern is imaginary.
| Component | Why campuses need it | Near-term pressure |
| Large transformers | Step grid voltage down on site | Already scarce in 100 MVA+ class |
| Switchgear | Isolate faults and route power | Long lead times, limited Western slots |
| Batteries | Ride through grid events | High import share from China |
| Optical transceivers | Move AI traffic at speed | Two-thirds of global unit supply from Chinese firms |
| Inverters | Tie generation and storage in | Rising security screening |
A bulk of the near-term pain lands on operators who used Chinese units specifically to stay on schedule. That is not a morality play. It is a calendar problem. Developers who locked domestic suppliers early look smart. Developers who optimized for speed look exposed. Both groups were responding to the same incentive: get the hall live before the model generation turns over.
The 5G Analogy Nobody In This Industry Wants
There is a useful, slightly uncomfortable comparison with earlier telecom rip-and-replace campaigns. Removing embedded foreign gear after it is installed is slower and more expensive than screening it at the door. Network rollouts slipped. Costs jumped. Lawyers got rich. The physical plant did not become safer overnight just because a statute said so.
AI campuses are heading into a similar fork. De-risk now, while designs can still swap vendors, or de-risk later, when the concrete is cured and the interconnection agreement is signed. Political appetite in Washington is tilting toward the first path. I think that is the correct strategic instinct. I also think markets will hate the quarterly optics of it.
Have we learned anything from that earlier episode? A little. Qualification labs are more mature. Buyers ask harder questions about firmware and remote access. Still, a transformer is not a radio unit. You cannot air-freight a spare 200 MVA machine because a committee changed its mind on Friday.
What Hyperscalers Can Actually Do In The Next Two Years
Grand strategy is easy in a paragraph. Procurement is not. Teams that want to stay online without betting the campus on a single origin country have a short list of unglamorous moves.
- Map every critical electrical and optical SKU by country of origin, including subcomponents, not just the badge on the cabinet.
- Pay for dual qualification now, even if the second source looks expensive on a spreadsheet.
- Reserve factory slots at Western transformer and switchgear plants instead of shopping the spot market in 18 months.
- Treat optics as a strategic commodity, not a catalog item, and lock multi-year capacity with more than one photonics house.
- Build battery and inverter architecture that can accept multiple chemistries and control stacks without a full redesign.
- Work interconnection queues with utilities as if power equipment lead times were the binding constraint, because they often are.
None of that sounds like a keynote. All of it sounds like adults running a multi-gigawatt program. In my experience, the projects that slip are rarely the ones that lacked vision. They are the ones that assumed a catalog price in 2024 would still be a catalog price in 2027.
Investors Should Watch The Unsexy Tickers
If you only follow accelerator vendors, you are watching one layer of a three-layer stack. Grid manufacturers, electrical steel producers, battery integrators, and optical module houses sit closer to the new bottleneck. Policy that squeezes Chinese units is, on paper, a demand gift for Western suppliers. Capacity is the catch. A gift you cannot ship is just a backlog.
That is why plant announcements matter more than slogans. A new large transformer hall in the United States is a multi-year object. Same for automated optical packaging lines. Same for cathode material that does not rely on a single processing corridor. I would rather own a dull factory with a booked slot book than a narrative stock with a slide about “AI exposure” and no winding machines.
Costs will leak into power purchase agreements, lease rates, and the all-in price of training runs. Not overnight. Not in a straight line. Enough to matter for anyone modeling hyperscale capex as if electricity and iron were free.
Materials, Mines, And The Part Of The Story That Refuses To Trend
Let’s talk about the boring stuff that decides whether the shiny stuff ships. Electrical steel is not a meme commodity. It is a specialized product with few qualified mills. Copper is already spoken for by grids, vehicles, and heat pumps. Cathode materials sit inside a processing map that still runs through a small number of countries. You can reshore final assembly and still import the guts.
That is why “Made in America” labels on a finished transformer can hide a more complicated bill of materials. I am not saying the labels are fake. I am saying origin is a stack, not a sticker. Policy that only watches the last factory will miss the first three.
Exposure stack, simplified: Finished cabinets and tanks Magnetic steel, copper, insulation Control electronics and firmware Raw concentrates and refined metals
If you want a practical test for any new industrial plan, ask where the electrical steel comes from and how long a second mill would take to qualify. If the room goes quiet, you have found the real timeline.
Costs, Lead Times, And The Mood On The Ground
Talk to people actually buying this equipment and the tone is less ideological than the hearings. They want gear that works, arrives, and can be serviced without a diplomatic incident. Many of them used Chinese units because the alternative was slipping a campus by three quarters. That is not treachery. That is a Gantt chart.
Restrictions change the math. Local alternatives can lift prices. They can also lift certainty, which sometimes is worth more than the extra dollars on a transformer. Grid certainty is already a pricing variable for data center land. Equipment certainty is about to join it. Sites with a clear path to domestic or allied gear will clear higher. Sites that assumed an open import window may need a rewrite.
Is some of this overdone? Maybe at the margin. Not every circuit breaker is a spy novel. Firmware-rich inverters and remotely managed optical modules are a different category from a tank of mineral oil and copper coils. Lumping them together makes for stronger speeches and sloppier rules. Good policy would separate dumb iron from smart electronics. I am not convinced every draft will bother.
What “Winning The AI Race” Looks Like If The Lights Flicker
There is a temptation to treat AI supremacy as a software story with a chip subplot. Power makes that story look thin. A country can design the best models on earth and still stall if it cannot energize the halls that train them. China understood industrial capacity as strategy a long time ago. The United States is relearning that lesson in public, at the price of a few awkward quarters.
Allied production can close part of the gap. Canada, Europe, Japan, and South Korea are not theoretical friends in this category. They have transformer shops, steel know-how, and photonics labs. Diversifying toward that circle is not autarky. It is adult risk management. Autarky would be pretending every winding machine must sit inside one border by Christmas.
I keep coming back to a simple picture. A dark data hall is not an abstract national security memo. It is idle accelerators, delayed models, and a utility scrambling to re-energize a bus. The equipment that prevents that picture is now political. That will raise costs. It may also prevent a nastier surprise later. Both things can be true.
A Practical Checklist For Readers Who Have To Make A Call
If you sit on a board, a credit committee, or a municipal planning desk, you do not need another slogan. You need questions that survive a staff meeting.
- What share of on-site transformation and switchgear is already contracted from a single country of origin?
- Which optical modules inside the fabric can be dual-sourced without a six-month requalification?
- Do battery and inverter controls create a remote-access surface that a future covered-list decision would freeze?
- Has the project reserved domestic or allied factory slots, or is it still shopping the residual market?
- What happens to commercial operation if large transformer deliveries slip by two quarters?
- Are interconnection studies using realistic equipment lead times, or last year’s brochure?
Answer those poorly and the model still trains in a slide deck. Answer them well and you may look conservative for a year. I will take the second look. Markets punish delay. Grids punish fantasy.
The Part I Keep Turning Over
Every boom invents a bottleneck it did not want to see. In cloud’s first wave it was fiber and cheap power in a few counties. In this wave it is dense power, specialized iron, and the optics that stitch racks into something that behaves like one machine. China built the factories that make that iron affordable and available. The United States is now deciding how much of that convenience it can still live with.
There will be higher invoices. There will be projects that slip. There will also be new plants on American soil that would not have been funded in a quieter year. That mix is messy. Industrial policy always is. The mistake would be pretending the power stack is a side quest while the models keep scaling.
So here is the uncomfortable close. The AI race is not only a contest of algorithms and accelerators. It is a contest of transformers that arrive on time, batteries that start, and light that keeps moving through glass. If that sentence sounds too physical for a software decade, good. Physics was always going to collect the bill.
