Announced Is Not Energized

This past Thursday I was driving outside El Paso, near the Texas border, when the desert opened up into a construction site bigger than anything I expected to see out there. Graded land running to the horizon, steel frames going up, fresh roads cut to places that do not exist yet. It looked like the early stages of a refinery or a port, not a software industry.

What got me was that this was my third trip to El Paso in the past year, and the first time any of it registered. Somewhere between my last visit and this one it went from nothing I noticed to impossible to miss from the highway.

It turned out I was looking at Meta's campus. The company announced El Paso in late 2025 as a roughly $1.5 billion data center, and in March raised the commitment to $10 billion for a campus built to scale to a full gigawatt of power. That is the thing I drove past, throwing up steel in the desert.

That is what AI looks like now, once you get past the model demos and the cloud dashboards. Land, transformers, fiber, and a line for electricity that stretches out for years.

Last week I wrote about Cerebras and the shift happening inside the chip, toward specialized silicon for different stages of inference. Standing on that roadside, a bigger version of the story came into focus. The constraint that matters most in AI is moving off the silicon altogether, and most of the people writing checks against this buildout have not repriced for it.

What's happening

Start with the spending, because the spending is genuinely staggering. The four largest US hyperscalers are now tracking more than $700 billion of capital expenditure in 2026, up from under $400 billion last year, and the figure has climbed with nearly every earnings call this season. Fold in Oracle, the neoclouds, the sovereign funds, and the model labs, and 2026 becomes the first year total compute capex passes a trillion dollars. There has never been an infrastructure cycle this large or this concentrated, and it is still rising into 2027.

None of that is in dispute. The demand is real and the money is real.

The read everyone is giving

The standard reading of all this is straightforward, and it is less wrong than out of date. AI needs compute, so whoever spends the most and locks up the most accelerators wins. The bottleneck is chips, the supplier is Nvidia, and you keep score in gigawatts announced.

That reading earned its keep. Through 2023 and 2024, GPU scarcity really was the whole game, and the firms that locked up supply early bought themselves a genuine lead. The spending now is a rational response to demand that has not let up.

The trouble is that it describes the market we just left. It quietly assumes that an announcement and a working data center are the same thing, and this year the two have come apart.

The cascade most people are missing

For most of 2024 and into early 2025, the thing you could not get your hands on was a GPU. That has flipped. The scarce thing now is power, and the unglamorous electrical gear that moves it from the grid into a rack.

High-voltage transformer lead times have stretched out to four years and beyond for the largest units. In the markets where most of the 2026 capacity is actually being sited, places like Northern Virginia, Phoenix, and Dallas, the wait just to interconnect to the grid now runs four to seven years. One write-up put it in a way that has stuck with me: electrical equipment is less than ten percent of a data center's cost and effectively all of its bottleneck. The chips are ready. The substation is the holdup.

The gap shows up in the project data too. Sightline Climate counted roughly 12 gigawatts of 2026 US data center capacity announced across about 140 projects, but only around 5 gigawatts actually under construction; much of the rest has no disclosed plan for power at all. Microsoft has said it is sitting on something like $80 billion in Azure orders it cannot fill because the electricity is not there. The grid is the constraint now, not the chip supply.

Meta's El Paso campus is the whole thesis in miniature. To energize that load on Meta's schedule, the local utility is rushing to stand up a dedicated 366-megawatt natural gas plant, around $473 million, built from 813 modular generators of the kind a grocery chain uses for store backup, wired to power Meta alone for its first five years before folding into the public grid. A company that publicly committed to renewable energy, in one of the sunniest cities in the country, is turning to on-site gas instead, mostly because the grid cannot energize the load fast enough. Announced renewable, delivered gas. The campus went public in 2025 and will not be fully powered until closer to 2027, and the distance between those two dates is the whole argument.

Power is only half the bottleneck. The other half is permission, and the word I kept hearing for it was NIMBY, not in my backyard. The El Paso gas plant is already drawing organized local opposition over air quality, water, and who pays. I watch the same fight closer to home in New Jersey, where more than sixty groups recently asked the governor for a moratorium on any data center drawing 20 megawatts or more, and the town of Millville just killed a 1.4-gigawatt project, the largest ever proposed in the state, over worries about power, water, and bills. A signed lease does not survive a county board that votes the project down, which is one more reason announced and energized capacity keep drifting apart.

I made a related argument a few weeks ago in the piece on the Adobe shareholder suit: approval is not oversight. A board can sign off on an AI strategy and still be on the hook, because approving something and building the system to monitor it are two different acts, and only the second one holds up when a court starts asking questions. The infrastructure version of that is the title of this issue. A press release commits a company to nothing. A signed power contract commits it to something. Capacity that is actually energized and running is the only version that pays for itself, and the market is mostly still reading the press releases.

The useful lens here is not a technology lens, it is a project finance one. Anyone who has underwritten a power plant or a toll road knows better than to confuse a groundbreaking with a cash flow. They check whether the contracted revenue, the financing tenor, the useful life of the asset, and the time to switch it on actually line up, because once those slip out of sync the asset does not merely underperform. It defaults. AI infrastructure has become that kind of asset, and it is mostly not being underwritten like one yet.

Investors are starting to work this out in real time. Meta's stock slid about eight percent the day it expanded the El Paso commitment, then dropped more than nine percent again a few weeks later when it raised its companywide capex guidance. Eighteen months ago a bigger spend number read as bullish. Now the market is grading each company on whether the spend is actually turning into revenue, and Amazon's trailing free cash flow has fallen by something like ninety-five percent as its capex ran ahead of it. That is the argument I am making, except it is showing up in the cash flow statement rather than in a newsletter. Quietly, the scorecard is moving from how much you announced to what you can show for it.

CoreWeave is the cleanest place to watch the new scorecard at work. Its latest quarter showed a $99.4 billion contracted backlog and roughly 3.5 gigawatts of contracted power, against just over 1 gigawatt actually live. Notice where the company put its emphasis: not on the backlog headline, but on how fast it is converting those contracts into energized capacity, on the $8.5 billion investment-grade facility it raised against the contracts and hardware, and on having no debt due until 2029. It is being valued on the distance between what it has signed and what it has switched on, and on whether its balance sheet can outlast the wait. That distance, 3.5 gigawatts on paper against a bit over 1 running, is the business, and closing it separates the providers that grow into durable platforms from the ones left exposed when scarcity pricing fades.

What this means in three years

Push the timeline out to 2029 and the questions that decide the winners barely resemble the ones being asked now. Most of it comes down to one thing: whether AI revenue catches up to the capex curve before investors lose their nerve. Almost everything else follows from how that resolves.

If revenue does catch up, the engine will be agentic workloads, and this is the part the skeptics keep underrating. A basic chatbot makes one call to a model. An agent that plans, pulls in data, calls tools, and checks its own work can fire off dozens of calls to handle a single request. As that becomes the normal way AI gets used, consumption climbs far faster than user counts suggest, and it climbs as steady inference demand rather than one-off training spikes. Cheaper models do not undercut this; they feed it, since every drop in the cost of a token makes another batch of use cases worth running. The capacity going up now gets absorbed, and the operators who lined up power, financing, and contracts turn into real platforms.

If revenue lags instead, the reckoning will not land evenly. A shortfall would bite hardest on capacity that only penciled out at shortage-era prices and was financed on the bet those prices would hold. Companies that look commanding today measured in announced gigawatts could look very thin three years from now measured in gigawatts that are energized, used, and paid for.

One thing holds in either case: power turns into the real moat. Those four-to-seven-year interconnection queues are already pushing the biggest operators to lock up their own generation. Meta's gas bridge in El Paso is one version of that. Oracle's Project Jupiter, a fully islanded microgrid with no grid tie at all, is a more extreme one. Over the next 18 to 36 months I expect bring-your-own-power to move from a workaround to a default design assumption for any frontier-scale site, with the data center business blurring into the power generation business. Whoever locked in firm generation early holds an advantage that money cannot conjure up quickly, whichever way the demand question breaks.

The thread running through all of it is the one I keep coming back to with boards. The lasting value goes to whoever turns silicon, power, financing, and genuine demand into capacity that is switched on, used, and billing, rather than to whoever announced the largest cluster.

What a decider should do

If you sit on a board, or run a PE portfolio with exposure to this directly or through a company that leans on it, the habit to break is treating capex announcements as proof of capability. Ask instead for one number: contracted capacity against energized capacity, in megawatts, today. CoreWeave's own gap, 3.5 gigawatts contracted versus a bit over 1 live, is the most revealing line in its results, and your portfolio companies ought to be able to hand you theirs. Then check whether the customer contracts, the financing, the hardware's useful life, and the time to firm power are all on roughly the same clock. An underwriter treats a mismatch there as default risk, not a rounding error, and on this asset class so should you. A backlog leaning on a few names deserves the same scrutiny.

If you are a CTO or a buyer placing a bet on a provider, start from the assumption that some of what you are being sold is scarcity-era capacity priced for a market on its way out. Push on the provider's power strategy specifically, including whether it can generate behind the meter, rather than counting its GPUs. And be honest with yourself about your own workload, because training is bursty and cluster-heavy while inference is steady and latency-sensitive, and signing for one on the terms of the other is how budgets blow up.

If what you really want to know is whether the cycle holds together, there is one signal worth more than any capex headline: the workload mix shifting from training toward production inference and agentic use. That is the tell that demand is becoming recurring consumption instead of a one-time build-out. When it turns up in a provider's revenue, the capacity is being absorbed. When the only thing growing is the announced gigawatt count, it is not.

And whatever AI roadmap has your name on it, the timing assumptions inside it are shakier than the announcements make them sound. When capacity arrives, and what it costs, now hinges on transformer and interconnection queues that no amount of capital can shrink to zero. Leave yourself slack in the schedule, and a fallback for when the power slips.

It is the same discipline that protects a board on AI oversight more broadly. Trust the evidence over the assertion, and trust the megawatt that is running over the one that has only been announced.

The bottom line

The first act was about whether the models could do the work. The one we are in now is about whether all this capital becomes capacity that is actually powered, connected, financed, and running. Writing the checks for GPUs was the straightforward part, and it is largely done. Turning them into live megawatts is the hard part, and it is only getting started.

Over the next 18 to 36 months the market is going to separate the capacity that exists on paper from the capacity that exists in production, and reprice the difference accordingly. The teams that saw the constraint move off the chip will look like they were paying attention. The ones still issuing gigawatt press releases will look like they were keeping score in a game that had already changed.

Announced is not energized. The faster that shows up in how you underwrite, how you buy, and how you build, the better placed you are for the act that is actually beginning.

P.S. Pick one piece of AI infrastructure you actually depend on. A vendor your roadmap is built around, a portfolio company spending to build or buy capacity, a cluster someone has promised you access to. Ask the one question that matters: how wide is the gap between the capacity it has announced or contracted and the capacity it has actually energized and put to work, and who is carrying the risk in the space between the two. If the answer that comes back is basically a press release, you have just found the thing the next three years are going to reprice.