Quick question. How do you power six gigawatts in the Arizona desert, next to a dry lake bed? According to the permit application, with a lot of batteries. This is The Data Center Daily, and it's Wednesday, November 11th. Four stories today: SpaceXAI's filing for a six gigawatt campus in Cochise County, Latitude Intelligence's count of how little carbon-free supply utilities have lined up for signed data center load, a fresh Morgan Stanley note that slides five gigawatts of US builds into 2029 but leaves its capex forecasts alone, and a Nature Energy paper on who ends up paying for the grid.
Data Center Digest, reporting Tuesday:
SpaceXAI, through its affiliate Tombstone Resources LLC, has applied to Cochise County for a special development permit covering 747 acres near the Willcox Playa. The plan calls for up to nine buildings of 493,000 sq ft each, brought online in four 1.5GW stages, with construction starting in March 2027 if zoning is approved.
The power terms, as the Digest reports them: the company says it will pay for grid upgrades, including 3.6 billion dollars of power infrastructure, 500 million of which is needed for phase one, and it will install a 24 gigawatt-hour battery system. The county's zoning decision is expected at a meeting on Wednesday, December 9th. Six gigawatts. Arizona's whole grid had about 32 gigawatts of net capacity in 2024, per the EIA figures the Digest cites. That's one campus asking for nearly a fifth of a state's fleet. The battery is the shield, there to keep demand swings off the state grid. And I'll give them this: developer-paid upgrades plus onsite storage is exactly the template everybody started demanding after last week's ballot results. Some caution on what this is. It's an application for a special development permit, not an approval, and it reaches us as a summary. The Digest gives no filing date and names no generation source beyond the grid upgrades and the storage. It also flags that getting enough accelerators to fill the site may be a bigger obstacle than power. Now put it next to the rest of the pipeline. We've spent weeks on Memphis and Southaven. TechRadar, writing up a SemiAnalysis assessment on Tuesday, says Musk wants to add ten gigawatts in 2027, taking SpaceX from roughly three gigawatts to more than ten, and the analysts think the schedule is achievable. Colossus 2 reportedly has 89 gas turbines good for about two and a half gigawatts, but environmental limits will eventually cap permitted turbine output at about 1.24. And Mississippi's temporary permission for the extra turbines reportedly requires phasing them out by July 2027. So the Arizona design, grid plus storage from day one, reads like the lesson from Southaven. For a sense of the money behind it, Morgan Stanley's model, which we'll get to shortly, has SpaceX spending 236 billion dollars on capex in 2027, according to one summary of the note.
Nick Zenkin, head of research at Latitude Intelligence, writing in Latitude Media:
If the data centers with signed contracts at those utilities draw 65% of their contracted capacity, which is close to the reported median rate by utilities, the 69 GW of gas plants currently planned or under construction would cover about 45% of the electricity their signed data centers will need in a year. Every carbon-free project these utilities have proposed — counting solar, wind, nuclear, geothermal, and hydro as carbon-free — would cover about 30%.
Those are the vertically integrated utilities Latitude covers. The remainder, about 120 terawatt-hours a year or a quarter of the total, has no planned supply at all. And that assumes every proposed carbon-free project gets built. Count only what regulators have approved, and the carbon-free share falls to 15 percent while the unplanned share grows to 40. Watch what signed means here. Latitude only counts load with an executed agreement to take power. Dominion told investors on October 30th it had 54.9 gigawatts of data center contracted capacity in Virginia. Just 12.2 of that is in electric service agreements. The rest is construction and engineering letters. Contracted, in the loosest possible sense. On Latitude's definition, AEP Ohio has signed more than twice as much as Dominion, even though the headline totals say the reverse. Just over half the utilities it tracks have signed less than a tenth of the load they report, and nearly a quarter have signed none. Georgia Power has 12.8 gigawatts under signed contract out of an 80.1 gigawatt pipeline. The renewables math is the rough part. Over a year, a gigawatt of data center takes about 1.9 gigawatts of wind or 2.8 of solar at US average output. Then there's the peak. In PJM's 2027-28 capacity auction, every 100 megawatts of solar counted for 8 at peak, against 74 for a combined-cycle gas plant. Cover the peak with batteries and you need 2.3 gigawatts of wind plus storage, or 4.1 of solar plus storage, per gigawatt of load. And 63 gigawatts of the solar, wind and batteries in these plans hasn't gone out to bid. Meanwhile, data center customers plan 73.7 gigawatts of their own generation, almost all of it gas, and 59 percent of signed load sits in states that don't require utilities to add carbon-free power. Markers to watch: PJM's filing this month on data centers that bring their own power, final resource plan orders in Nevada and North Carolina by year-end, and Georgia Power's quarterly large-load report later this month. Zenkin's framing is that the officials who just won office inherit those contracts.
Adam Button, writing in InvestingLive on a new Morgan Stanley note:
In a report titled Buy the Drop: AI Investment Opportunities Post the US Midterms, the bank says that the AI investment cycle remains intact, even if political opposition to data centers pushes some construction projects into 2029.
Yesterday we had the bank's Friday note putting up to 12.4 gigawatts of Texas and Ohio capacity at risk through 2030. This follow-up goes national. The bank expects 20 percent of US projects planned for 2028, about 5 gigawatts, to slip to 2029, and it assumes nothing already under construction is touched. A fuller summary from Druckfin gives the path: 20.7 gigawatts of new US critical IT capacity in 2027, 19.9 in 2028, then 27.6 in 2029. That makes 2028 a four percent decline in US additions, followed by 38 percent growth. The capex forecasts don't move: about 1.6 trillion dollars in 2027 and 1.8 trillion in 2028, with some downside risk flagged on the later year. So per the bank, demand is fine and the buildings are what's short. Druckfin says its power model now shows a 44 gigawatt gap in powered capacity in 2028, up 15 percent. The bottleneck moves to the powered shell. Lovely if you already own energized megawatts. Not so lovely if you sell the gear. The note itself warns of a 2028 air pocket for suppliers: slowing orders, margin normalization, earnings downgrades, even with AI demand intact. Druckfin's 2027 split by company: Alphabet 375 billion, Amazon 355, Microsoft 269, SpaceX 236, Meta 225, Oracle 97, CoreWeave 46 and Nebius 37. The bank expects new builds to shift toward states like Tennessee and Louisiana, and about 20 gigawatts that can't be built in the US to spill over into Asia and EMEA. It also expects more scrutiny on new construction loans and neoclouds. Here's the soft spot, and Druckfin flags it. Morgan Stanley's 20 percent exposure is well below other estimates. Goldman says states likely to add restrictions hold about 45 percent of planned US capacity. Bloomberg puts about 40 percent in states whose new governors want a pause. And the bank concedes that bans on onsite fossil generation, along the lines of Virginia's approach, would turn a delay into a cut. It's a forecast that holds only if governors write rules, not bans. The signals the note points to: the first formal rules from the governors-elect in Texas, Ohio, Georgia and Michigan, Nvidia's results later this month, and whether hyperscalers start naming sites in friendlier states.
Cheng Feng and Fengqi You, in a paper published Monday in Nature Energy:
In wholesale markets, cumulative price increases attributable to data centres exceed 10% in the most exposed hubs by 2025. In regulated tariffs, residential users are allocated a median of 47% of costs despite contributing far less to demand growth.
The method, per the abstract, is causal analysis of price records alongside transmission plans and regulatory cost-allocation dockets. The facility inventory comes from S&P Capital IQ Pro and 451 Research under commercial license, and the processed data and code are posted on Zenodo. Forty-seven percent to residential, at the median, for growth they mostly didn't drive. That number is going straight into consumer advocates' filings. And the distribution is worse: rural communities see burden increases 76 percent larger than urban ones, and the lowest-income households face extra energy burdens of up to 28 percentage points. The fiscal finding is the one site selection teams should read. The authors find local economic returns are almost unrelated to spending, a spending elasticity of 0.16, and the states absorbing the largest burden increases capture the smallest returns. That's last week's ballot results in one line. And it arrives while the Senate bill we've been tracking would have new data centers of 20 megawatts or more pay full grid-upgrade costs, and while every RTO works through FERC's large-load orders. If your pitch is that you pay your own way, expect to get asked about this paper. One limit: the abstract doesn't name the hubs that crossed ten percent, so read it as a national pattern rather than a forecast for any one utility's bill.
If this show saves you some reading during the week, a rating or a short review in your podcast app helps other grid and compute people find us. Two sentences is plenty. Bonus points for naming the docket we made you read.
Dates to hold: Nvidia reports Tuesday the 17th, comments on NERC's data center registration petition at FERC are due Friday, November 20th, and Cochise County takes up zoning on December 9th. We're back tomorrow, Thursday, November 12th. Bring a battery. A big one. The Data Center Daily is a Lantern Podcast. From Sarah and Matt, until tomorrow.