Four hundred seventy-four gigawatts.
That is how much electricity data centers and other large users have now requested from the Texas grid, according to a Reuters figure. Three years ago, the number was about 48 gigawatts — a tenfold jump in a state that already carries 15% of all U.S. data center capacity in megawatts.
Texas has now frozen new grid connections while regulators work through the backlog and figure out which proposals come with real money, real customers and a real chance of operating. The industry has a name for the rest: ghost demand.
Last week this brief followed AI's plumbing through model tooling and the Nvidia-Hugging Face layer. This week the same plumbing drops a level — from software to capital, from capital to electrons, and to Apple's reappearance as a would-be server vendor.
Start with the capital, because it is still moving faster than anything physical.
Crusoe raised $3.9 billion in a Series F, pushing its valuation to $30.9 billion. Ten months earlier, the company raised $1.38 billion at a $10 billion valuation. That is a tripling of the sticker price in under a year.
The round was co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures and TPG participating. Crusoe also added three board members: Cloudflare CFO Thomas Seifert, Primary Digital Infrastructure CIO Bill Stein, and JB Straubel, the Redwood Materials founder who already sits on Tesla's board and personally invested in Crusoe in 2021.
The company started in 2018 mining crypto on flared natural gas. It now leases data center space, rents out its own GPUs, and sells inference compute. It has Meta, Microsoft and Oracle as customers, a reported $13 billion five-year cloud contract with Jane Street, and meetings with Goldman Sachs and Morgan Stanley about an IPO.
CEO Chase Lochmiller's framing is the tell: AI's era of abundance requires "controlling the infrastructure from electrons to tokens."
Crusoe's answer to the electron problem is scale plus modularity. In Abilene, Texas, it is adding a 900 MW AI data center for Microsoft, bringing total planned capacity at the site to roughly 2.1 GW. The build includes a 900 MW behind-the-meter power plant with battery storage, which is the whole point: when the grid can't deliver on schedule, you bring your own generation.
"When you're talking about gigawatt-level demand, the grid simply can't absorb that cleanly in most markets," JLL managing director Andy Cvengros told Data Center Knowledge. His other line lands harder: "A few years ago, a 100 MW deal was a headline. Now we're seeing single-site programs that dwarf that."
Industry estimates cited at a recent forum suggest 50% to 65% of large-scale data centers will run behind the meter in the coming years. That is a structural admission that the public grid is no longer the default supplier for the largest AI loads.
Crusoe is also building Spark, a modular "AI factory" it manufactures in-house — small enough to move by truck, fast to deploy, and less likely to trigger the local backlash that gigawatt campuses attract. It is a hedge against both construction labor and NIMBY politics, which is a strange sentence to write about a company valued at $30.9 billion.
Meanwhile, the grid itself keeps slipping further behind.
Fortune reported that data centers are projected to consume nearly 12% of all U.S. electricity by 2030, roughly six times their pre-AI share in 2018. NERC's reliability assessment now projects North American summer peak demand growing by more than 224 GW over the next decade — 69% above what it projected a year earlier.
"There is a real disconnect there," Grid Strategies president Rob Gramlich told Fortune, noting that tech companies move fast while utilities "notoriously move very slowly." U.S. power projects that came online in 2025 took a median of five years from grid connection request to commercial operation. A single gigawatt can power 750,000 homes.
Marsh's Kathryn Burke calls power access "probably the number one, if not top five bottlenecks for data center development in the U.S. right now." Gramlich's summary is blunter: "There is scarcity on the grid. Not everybody's going to get the full level of service that they want."
J.P. Morgan Private Bank sizes the hole at roughly 110 GW of shortfall by 2033 — the equivalent of adding more than 20 New York Cities' worth of demand. It notes that $600 billion in transmission and distribution spending is projected through 2030, against about $3 trillion in sector underinvestment since 1999.
Which is exactly why the ghost-demand cleanup matters. Reuters found more than 700 GW of large-user power requests across parts of the Midwest, Mid-Atlantic and South — more than ten times the industry's estimate of current U.S. data center power use. Once utilities demanded skin in the game, the numbers collapsed: Exelon cut its high-probability data center demand by about 40% after imposing stricter collateral requirements; AEP's Ohio pipeline fell by more than half after connection-study fees that can reach $100,000. Pennsylvania has seen more than 100 data centers proposed and only 20 apply for the permits to proceed.
Once utilities demanded skin in the game, the numbers collapsed: Exelon cut its high-probability data center demand by about 40% after imposing stricter collateral requirements.
Big Tech's planned AI data center spending is expected to top $700 billion this year. Bad forecasts at that scale don't stay internal.
The supply side is not standing still, but it is solving a different problem. Solar panels that cost around $5 per watt at the start of the century now run about $0.12 per watt, almost entirely because of Chinese manufacturing now estimated at roughly 1.36 terawatts per year. Ember estimates African countries will bring over 17 GW of solar online in 2026.
In China, the issue is inverted: too much generation. "China's grid is a very strong and stable one," BTR New Material Group vice chairman Youyuan Huang said at the Fortune Leaders Forum in Macau. "But we've installed too much green energy." China will account for 60% of all installed renewable capacity through 2030, and in July solar overtook coal as its largest source of installed power capacity.
The proposed fix is AI itself — Jiangxing Intelligence's forecasting "brain" and autonomous robots that inspect and clean panels. "The supply of wind and solar energy is highly uncontrollable," founder Haitian Pang said. Separately, cheap panels are straining grids in the other direction: the 2025 day-long blackouts in Spain and Portugal have been linked to grids not built for two-way transmission.
And then there is Apple, which has nothing to do with any of this yet.
Per The Information, Apple is considering re-entering the enterprise server market after 15 years, with systems built around two or four M8 Ultra chips and Nvidia's NVLink Fusion interconnect. Target customers: AI developers, enterprises and governments that want to run inference on private infrastructure. Launch window: around 2029, possibly never.
The demand signal underneath it is real. Apple's Mac revenue hit $10.35 billion, up 29% year over year — the fastest-growing product line in the quarter, with total revenue of $109.4 billion. OpenAI has reportedly bought tens of thousands of Mac minis and Mac Studios for reinforcement-learning work; Anthropic rented Mac minis from AWS. CICC raised its Apple target price 10% to $340.
But analysts are treating the server plan as a "strategic call option," not a catalyst. It contributes nothing to FY2026-27 earnings, and the underwriting question is whether Apple can sell differentiated private AI compute at a premium — unified memory, efficiency, privacy — rather than becoming another buyer of expensive compute. Goldman Sachs puts the AI server market at $1.24 trillion by 2030, and Dell, HPE and Super Micro are already posting 30% to 90% revenue growth in it.
So: should you care?
If you're an investor, the story has flipped. Power generators, engineering firms and electrical equipment makers now carry technology-like demand growth, and the constraint is regulatory, not technological. The secondary capex cycle — the one that follows the GPUs — is where the money is being raised right now.
If you're an operator, the lesson from Texas is that announced demand is not demand. Utilities are finally asking for collateral, ownership disclosure and water plans, and the verified numbers are holding up at maybe half to two-thirds of the headline. Build your forecasts off the funded pipeline.
If you're just watching, the summary is this: AI can raise $3.9 billion in a week. It cannot raise a gigawatt in less than five years. That mismatch is the whole story, and it is nowhere near resolved.