The Real Constraint Is Supply

Power

Every data center needs electricity, and a large AI data center can use as much power as a small city. That single fact runs through almost every constraint we found this quarter, so it's worth separating into two parts.

Part one is the grid. Most companies want to plug their data center into the public electric grid — the same network of power lines and substations that powers homes and offices. The problem is that utilities can't hook new, very large customers up fast enough. The line to get connected (called an "interconnection queue") is long, and the wires needed to carry that much power (transmission lines) take years to build. $WULF TeraWulf, a company that builds data centers for AI computing, put it plainly on its call this quarter: "The constraint on AI infrastructure is not demand. It is power."

Part two is the workaround — building your own power on-site, often called "behind-the-meter" power. Instead of waiting on the grid, a company installs its own generation equipment right next to the data center, so it never has to wait for the utility. That sounds like a clean fix. This quarter's calls show it isn't.

$GEV GE Vernova makes gas turbines — the machines that burn natural gas to spin a generator and produce electricity. Its turbine backlog now stands at $176 billion, and the company says it's essentially sold out through 2030. $CAT Caterpillar, which makes large generators, told investors that new orders are now looking at delivery "towards the back half of 2028 and into 2029." $USAC USA Compression, whose engines power natural gas equipment, said lead times on large engines have stretched to roughly 200 weeks — close to four years. And $BE Bloom Energy, which makes fuel cells (devices that turn natural gas or hydrogen directly into electricity without burning it, a quieter and more efficient process than a turbine), has its own backlog stretching about four years out. Its CEO didn't sugarcoat what that means: "We think a 4-year backlog is not a trophy. It's a concession of constrained supply."

So both paths to power — the public grid and the private backup plan — are backed up. And the timing doesn't match. A data center building itself can be built in about 18 months. The turbine meant to power it can take three to five years to arrive. The building is often finished long before the thing that powers it shows up.

People

Equipment is only half the story. Somebody has to actually install it, and that requires skilled labor — specifically, electricians.

$STRL Sterling Infrastructure, a company that builds the physical sites data centers sit on, reported its backlog of signed work grew to $5.2 billion, up 131% from a year ago. Its CEO was direct about the limiting factor: "if we had 1,000 or 2,000 more electricians, we'd be growing it even faster." $WCC WESCO International, a company that distributes electrical equipment to contractors, described the same bottleneck from a different seat: "Demand's outstripping supply across the value chain. Starts with power. Followed by labor."

$EME EMCOR, a mechanical and electrical contractor that does the wiring and plumbing work inside data centers, gave us a number that shows the gap growing. Its "remaining performance obligations" — essentially, the pile of signed work it hasn't finished yet — grew 44% this quarter. Its actual revenue, the work it completed, only grew 20%. The promises are piling up faster than the company can deliver on them.

$FIX Comfort Systems, which does modular mechanical construction for data centers (building parts of the system off-site in a factory, then assembling them on location), said it is fully booked and is actually encouraging competitors to enter the market — not because business is bad, but because there simply aren't enough contractors to go around.

One more example worth noting: $AMRC Ameresco, an energy services company, has started leasing land on U.S. military bases — Pearl Harbor and Naval Air Station Lemoore among them — to build data center power projects. Federal land is already permitted, which skips months or years of local approval. Ameresco's pipeline of these projects went from 60 megawatts to over 1 gigawatt in a single quarter (a gigawatt is roughly enough electricity to power around 750,000 homes).

None of this fixes quickly. You can't train an electrician in a quarter, or even a year. This is a slow-moving constraint by nature.

Components

Between a finished computer chip and a working AI server, there's a layer of parts most people never think about. This quarter's calls showed that layer is stretched thin too.

Start with circuit boards — the flat green panels that chips and other parts get soldered onto. A laptop uses a circuit board with maybe 8 to 12 layers of wiring stacked inside it. An AI server needs a board with more than 100 layers, some pushing past 140, because it has to carry both huge amounts of power and huge amounts of data through a tiny space. At that density, the wiring that carries power starts to interfere electrically with the wiring that carries data. Only about four companies in the world can reliably build boards this complex. $TTMI TTM Technologies, one of them, developed a design approach that physically separates the power layers from the data layers to solve the interference problem. Its revenue crossed $1 billion for the first time this quarter, up 37%, with data center-related revenue up 91%. Notably, $TTMI TTM licensed part of this technology to competitors — a sign that demand is so far ahead of what any single company can supply that growing the whole market matters more than keeping the technology exclusive.

Next, substrates — the base material a chip is actually mounted onto before it goes into anything else. $AXTI AXT, which makes these materials, said: "Customer demand continues to outpace supply no matter how fast we add capacity." The company is turning down orders it simply can't fill.

Then there's packaging — the process of sealing a finished chip into a protective casing so it can connect to a circuit board. $TSM TSMC, which manufactures chips for companies like $NVDA NVIDIA, $AAPL Apple, and $AMD AMD, said: "Our packaging capacity is so tight that now it's limiting my customers' growth." That's a chip manufacturer saying the bottleneck isn't making the chip anymore — it's what happens to the chip after it's made.

$QCOM Qualcomm summed up the mood in four words: "Everything is at 100% utilization, everything."

Connectivity

Once the power is on, the building is staffed, and the chips are packaged, everything still has to be physically connected — data centers to each other, servers to servers, chip to chip. This quarter's calls pointed to real physical strain here too.

$GLW Corning makes the glass fiber that carries data as pulses of light between data centers. Its message was simple: "if we could make more, we could sell more." $ANET Arista Networks, which makes the networking switches that route data between servers inside a data center, reported deferred revenue (money customers have already committed but Arista hasn't delivered on yet) jumped $600 million in a single quarter to $6.9 billion. Its multiyear customer commitments nearly tripled to $9.7 billion.

$APH Amphenol, which makes the physical connectors and cables that link servers, racks, and networking gear together, posted record orders of $10.7 billion this quarter, with organic order growth of 63%. And $AAOI Applied Optoelectronics, which makes optical transceivers — small devices that sit at each end of a fiber cable and convert electrical signals into light and back — said actual customer demand is running $1.4 to $1.5 billion higher than what it can currently manufacture, a gap of roughly $300 to $400 million.

Every layer we looked at — power, people, components, connectivity — tells a version of the same story from a different seat at the table.

What the lead times tell you

Line up the delivery times companies gave this quarter and a pattern shows up. Gas turbines: three to five years. Large electrical transformers: two to three years. Big industrial engines: close to four years. Training an electrician: multiple years, with no shortcut. New chip packaging capacity: roughly 18 months to bring online. A data center building shell: also about 18 months.

Many of the things that go inside a data center take longer to produce than the building meant to house them.

Meanwhile, the demand side of this isn't a forecast — it's already committed. $AMZN AWS alone reported a backlog of roughly $496 billion in signed customer commitments. That's money customers have already agreed to pay, waiting on infrastructure that hasn't been built yet.

If the buildout keeps going at this pace, these constraints don't clear up in a quarter or two — the lead times themselves rule that out. What that means is the companies sitting in the middle of this chain — the turbine makers, the circuit board builders, the electrical contractors, the fiber and connector companies — have unusually clear visibility into demand that's already locked in, years out.

That's what this quarter's earnings calls showed us. We're not predicting anything here — just passing along what the companies said, in their own words.

All quotes are from public Q2 2026 earnings calls. This is not investment advice. Built using agentic AI tools, which can be incorrect. Do your own research. thebuildout.co