On December 9, the City of Starbase will hold a public hearing on tax abatements for a project SpaceX has described to local officials as a gas-fired power plant and super-compute facility on roughly 444 acres east of Brownsville, with an estimated capital investment of about $24 billion.
Read that figure as a power plant with a computer attached, not the other way around. The $24 billion is the company's estimate "for initial phases," covering a generating station and the shells to hold compute. It is not a bill of materials for GPUs, and the dollars are incidental to whether any of it gets built.
SpaceX is not short of money. In late October, Elon Musk said the company had closed $40 billion of debt financing to buy GPUs for its AI operations. In its third-quarter report it recorded $19.59 billion of capital spending, $16.44 billion of it on AI.
Capital is no longer the constraint; permits, tenants, and final investment decisions are what stand between announcements and energized steel.
What stands between that capital and 748 megawatts of energized steel is a municipal vote, pending air and water permits for 35 turbines, and the posture of a new Texas governor whose data-center policies are stricter than his predecessor's.
That is the shape of the AI power crunch in late 2026. The industry's headline gigawatt figures are options on future capacity, and the binding constraint has shifted from turbines and grid queues to the permitting, tenancy, and final-investment-decision gates that convert announcements into energized steel.
The UK offers the cleanest example of how wide the gap can yawn. Drax's stock exchange announcement describes a partnership with Brookfield and Radiant, Brookfield's UK AI infrastructure platform, for a data centre on the Drax Power Station estate in North Yorkshire. The site has about 4GW of grid access and 2.6GW of dispatchable generation. Headlines said £20 billion.
Drax's own stock-exchange filing is more careful. Brookfield "estimates that total potential investment, including semiconductor chips, could exceed £20 billion." The BBC described the project as having "secured £20bn backing." The company's own press release says "£20bn partnership."
What has actually been signed is a development partnership, and what has actually been scoped is smaller still. Phase 1 is "up to 100MW" using grid power, targeting commercial operations from early 2028 — and it is subject to planning approvals, regulatory and network approvals, "securing an acceptable tenant for the project," commercial agreements, and each party taking a final investment decision. Radiant will try to secure anchor customers before construction starts. Public consultation on the planning application is expected to begin this month.
Phase 2 is an "ambition" for 500MW between 2028 and 2031, behind the meter under a long-term PPA, subject to consents. Phase 3 comes after 2031. Together they could reach gigawatt scale, or they could not.
The £20 billion headline is an estimate of potential spending if every phase is built, including the chips. Strip the optionality and the committed number is 100MW with no tenant and no FID.
This is not a criticism of Drax so much as a description of how the sector now sells itself. Reusing a power station's existing grid connection is the smartest available answer to the UK's connection queue — CNBC notes reports that Microsoft pulled out of anchoring an Nscale site near London amid grid-connection delays. But the queue is only the first gate. The tenant is the second, and the FID is the third.
Anthropic's approach shows what it costs to buy past gate one while gates two and three remain shut. The company agreed to spend $3.4 billion securing 1.7GW of gas-fired generators, taking over equipment orders originally placed with Caterpillar by Texas supplier OnePWR Solutions. The orders were placed roughly two years ago but required additional financing; Anthropic is not buying OnePWR itself, it is stepping into its commitments. The deal also covers engineering and construction work to install the equipment.
The timing arithmetic is the point. Global turbine orders have reportedly exceeded 100GW against manufacturing capacity estimated at 60GW to 70GW. Average US delivery times have reached six years, with some orders extending into 2030 or 2031. By taking over existing orders, Anthropic avoids starting from scratch and waiting for Caterpillar to slot a new one into its production schedule.
That buys queue position. It does not buy an air permit, a water permit, a tax abatement, a tenant, or a final investment decision. The generators arrive; the site that will host them still has to be consented. xAI's eight converted GE jet turbines in Mississippi — up to 400MW combined — have already drawn a challenge from the Southern Environmental Law Center over the site's emissions permits, which is the shape of the fight that follows every fast gas deployment.
Widen the lens and the pattern holds. Australia's data centre investment has grown around 10-fold over the past three years, according to the Reserve Bank of Australia, with forward-looking indicators suggesting it stays significant. But the RBA's own note cautions that the direct contribution to domestic activity is considerably smaller than headline spending suggests, because roughly 70 per cent of the machinery and equipment is imported — typically from Taiwan and South Korea.
Sydney and Melbourne absorb most of the build because they have grid power, water, fibre, skilled labour and proximity to end users. The RBA's liaison contacts have raised concerns that data centres are diverting land away from industrial and residential construction. The investment is real; the domestic industrial base it implies is not.
The political system has started to price the bottleneck. The Senate voted 73-25 to take up bipartisan permitting-reform legislation aimed at speeding reviews for energy projects that now face yearslong delays getting to the grid — the first major action since the midterm elections. That vote is the backdrop determining whether the option-to-megawatt conversion rate improves; it does not fix any individual project.
The project-level picture is messier than the legislative one. At the Infocast Projects & Money conference in New Orleans in January, EDP Renewables North America CEO Sandhya Ganapathy described developers facing "one brick after another." Arevon CEO Kevin Smith said at least half his development pipeline will have to deal with some kind of Interior Department issue — transmission lines crossing a stream, wetlands borders — even though Arevon has no projects on federal land. He expects the delays to show up more pronouncedly in the last six months of 2026, and the supply-demand gap to widen further into 2027.
Lotus Infrastructure's Himanshu Saxena described developers racing to get projects under construction by July 4 to lock in federal tax credits for projects that cannot be completed by 2027. KeyBanc's Andy Redinger called last year the firm's best ever, because the rush to beat deadlines accelerated rather than slowed transaction volume.
That is the paradox worth holding onto: policy hostility has so far accelerated deal flow while lengthening the horizon on which any of it energizes. The pipeline is being financed against a permitting system whose throughput nobody can forecast.
What follows for anyone underwriting this sector is a change in what counts as a competitive advantage. It is no longer generation capacity, which can be ordered and is increasingly rationed by a six-year queue. It is permitting and offtake engineering: the ability to site a project where consent is achievable, to secure an anchor tenant whose contract survives the FID, to sequence approvals so that equipment orders and construction capital are not stranded.
The industry is staffing up accordingly, even if it does not say so. Smith's team at Arevon is doing community engagement at the state and local levels specifically so that Republican officeholders can be pulled in when a project needs help. That is permitting work, not engineering work, and it is now a line item in the development budget.
Investors should apply the same discipline. Discount announced gigawatts by the probability of consent. A 4GW site with a 100MW Phase 1 and no tenant is a 100MW option, not a gigawatt campus. A 748MW gas plant awaiting a December 9 vote is a vote, not a plant. A 1.7GW turbine order is a queue position, not a permit.
The money has been raised. The equipment has been ordered. What remains is the slower, less legible work of turning options into steel — and that work runs on signatures, not checks.