A neocloud can raise $500 million, buy a room full of H100s, and still end up with dead weight. The GPUs arrive. The invoices are paid. Then the company goes looking for powered space and gets turned away by a landlord that would rather sign a tenant with an investment-grade balance sheet than a startup with a burn rate.
That is the nastiest surprise in the AI buildout. For two years, most coverage treated power as a supply problem. More turbines, more substations, more transmission, more patience. Fine. But that misses the commercial reality. Scarce power has given data center owners the right to choose their customers. And they are choosing survival odds, not ambition.
Power scarcity is now a financing filter. Call it Credit-Gated Capacity: when energized land is scarce, landlords stop chasing the highest bidder and start screening for the tenant least likely to blow up halfway through a long lease. Venture funding does not clear that bar. Credit ratings do.
Latency Is Dead, Power Is King: How AI Flipped Real Estate Valuation
AI training changed the map. These workloads can sit far from users, but they cannot live without huge blocks of firm power. JLL said it plainly in its 2024 outlook: access to power now drives data center location and valuation more than latency does.1 For years, the premium sat near major fiber hubs because milliseconds mattered. A training cluster running for weeks cares far more about a substation than about being close to Ashburn.
West Texas is the obvious example. So is parts of Wyoming. Neither market mattered much to classic low-latency data center logic. They matter now because they offer acreage near generation and transmission.
LandBridge's S-1 says the quiet part out loud. It markets surface acreage as attractive for data center development because it sits near power generation and transmission infrastructure.2 That tells you where value is moving. In this cycle, dirt near a substation can matter more than software.
The split between inference and training is real, but messy. In Northern Virginia, both still coexist because the ecosystem is too dense to separate cleanly. Still, the center of gravity has moved. Training wants remote, power-heavy sites. Inference stays closer to users. That shift forces every serious model builder into the same hunt for a small set of training-capable parcels that are already energized or close enough to be believable. Brokers and developers keep talking about gigawatt campuses, but the list of sites that can actually move on a useful timeline is short.
The Credit Gate: How Power Scarcity Became a Financial Filter
The mechanism is simple. Scarcity gives landlords room to discriminate. They use it to screen for tenant quality, and tenant quality now means balance sheet strength.
Scarcity does not reject the neocloud at the door; it hands the landlord a balance-sheet checklist and lets the credit rating do the rejecting.
Digital Realty has been explicit that tenant credit quality drives its leasing: it treats investment-grade profiles as critical when it commits to long-term, capital-intensive builds.3 That is rational behavior. If a landlord is going to spend years and serious capital on a powered build, it wants a tenant that will still exist when the lease starts paying. A venture-backed neocloud with thin margins and heavy cash burn is the exact profile they avoid. CBRE's H1 2024 data shows why they can afford to be choosy: power constraints pushed primary-market vacancy to record lows and asking rates up more than 20% year over year.4
So venture-backed tenants are losing leases even after they secure GPUs. Silicon and power are now separate procurement battles, and the gatekeepers are different. Dylan Patel of SemiAnalysis has made the same point: the race is no longer just about securing H100s but about securing the megawatts to turn them on, and neoclouds keep getting squeezed by hyperscalers that can buy power at a scale they cannot match.5 The predictable outcome is stranded silicon. Expensive hardware. No place to energize it.
The Energy Desk as a Weapon
Hyperscalers saw this early. They did not just buy chips. They started with land and queue positions, then moved upstream into power procurement itself. Microsoft says it is "pioneering new approaches to energy procurement and grid integration."6 Translate that into plain English: a software company is building the instincts of a utility.
Microsoft and Amazon are signing firm-power deals before smaller cloud operators even enter utility negotiations. Amazon paid $650 million for Talen Energy's Cumulus campus to secure behind-the-meter nuclear capacity. Talen called it "a template for how we value behind-the-meter capacity."7 Microsoft contracted the full output of a revived Three Mile Island, which Constellation said shows the premium value of firm, clean baseload nuclear power for AI.8 Every deal like that shrinks the pool of available firm power. A neocloud is not matching a nine-figure campus purchase, and it is not jumping a five-year interconnection line.
| Power source | Profile for AI training | Who can secure it |
|---|---|---|
| Firm nuclear (baseload) | 24/7 high-utilization, commands a premium | Investment-grade hyperscalers |
| Intermittent renewables | Cheaper but unreliable for high-uptime clusters | Broader field, but insufficient alone |
| New grid interconnection | 5+ year queue before commercial operation | Anyone patient and well-capitalized enough |
The nuclear premium gives the game away. Training clusters need round-the-clock utilization. Intermittent renewables help on cost and optics, but they do not solve for uptime on their own. Firm baseload does. So old nuclear plants that looked marginal a few years ago now look like prized AI infrastructure.
Are Grid Queues a Permanent Moat? Transient Versus Structural Scarcity
Most bottlenecks do not matter equally. Ignore temporary equipment shortages. Pay for grid access.
"First, there was a neural net chip shortage. Then, the next shortage will be voltage step-down transformers. You've got to feed the power to these things."
- Elon Musk, Bosch Connected World 2024
Transformer shortages will clear. Interconnection queues will not. Musk is talking about the temporary layer. Transformers and switchgear are backed up today, but manufacturers are expanding output. Eaton told investors its backlog for electrical equipment had grown sharply and that it was expanding capacity to meet multi-year data center demand.9 GE Vernova reported "unprecedented demand" for turbines and electrification gear.10 Factory bottlenecks are painful. They are not permanent.
The real moat sits in interconnection and transmission. Lawrence Berkeley National Lab says the average wait from interconnection request to commercial operation now runs past five years.11 No factory fixes a permitting fight or a transmission right-of-way dispute. Dominion Energy's 2024 resource plan says Northern Virginia load is growing at an unprecedented rate and will require major new generation and transmission infrastructure.12 Read that as a warning from the utility itself: the best-known AI market in America is pressing against the grid's limits.
Regulation Is the Ceiling on the Power Grab
There is a limit to how far this can go, and FERC just showed it. When hyperscalers try to bypass the grid with behind-the-meter arrangements, the argument shifts to cost shifting. Who pays to maintain the broader system if the biggest new loads peel off and cut private deals?
FERC rejected the expanded Talen-Amazon nuclear arrangement, saying it "fails to demonstrate that the specific co-location arrangement does not result in unjust and unreasonable cost shifts to other grid customers."13 Utility Dive called it a setback for tech companies trying to secure large amounts of power quickly through co-location at existing plants.14 That matters for venture-backed neoclouds, wholesale colocation tenants, and enterprise model builders. The clean off-grid escape hatch is not as open as the market hoped.
You can see the anxiety in where Big Tech is placing its longer-term bets. Sam Altman said at Davos that future AI energy demand will require a breakthrough and that it pushes OpenAI to invest more in fusion.15 Mark Zuckerberg said he expects power constraints to hit before compute constraints.16 If the buyers with the strongest balance sheets are already talking like energy hawks, current interconnection timelines are nowhere near enough for multi-gigawatt training demand.
What Builders Should Do Before the Window Closes
Start with power. Everything else is secondary.
For a neocloud without an investment-grade rating, the best move is to pair with a creditworthy anchor tenant that can get a landlord comfortable. Second best is pre-energized land in less crowded markets like parts of Indiana, New Mexico, or Wyoming, where hyperscaler demand is rising but not yet total. Last resort is speculating on queue positions and hoping politics, utilities, and local opposition all break your way.
The next financing round that matters will not buy chips. It will buy a substation slot.
