Texas Just Hit the Brakes on Data Centers
A 474GW pile of grid connection requests in Texas suggests AI's real bottleneck may not be GPUs but the power to run them.
How Big Is 474GW, Really
A gigawatt is a billion watts. Think of a large nuclear reactor as roughly 1GW of capacity, and 474GW starts to look like hundreds of large reactors' worth of nameplate capacity. That comparison has limits: a power plant's installed capacity and a data center's requested grid connection aren't the same thing. Plants run at partial utilization, and data centers don't draw 100% of their requested peak power every hour either.
A more useful comparison exists closer to home. ERCOT's record peak demand in 2026 sits at roughly 91GW. That means the connection requests now sitting in Texas's queue add up to more than five times the highest amount of electricity the entire state has ever drawn at once.
474GW The scale of large-load grid connection requests ERCOT is currently reviewing. This is not confirmed construction volume, and it may include large loads beyond data centers.
~91GW Texas's record peak grid demand. 474GW is about 5.2 times that figure.
5x+ The issue isn't that all 474GW will materialize. It's that the request volume has swelled so much that the grid operator now has to work out which of it is real demand.
Why You Shouldn't Take 474GW at Face Value
A developer building a single data center can file connection requests at multiple candidate sites. A company might request 1GW each in Dallas, Austin, and Houston, then pick only one location. Early-stage projects without finished financing or customer contracts are likely mixed into the total too.
That's why Governor Greg Abbott's first move in August wasn't a simple freeze on new connections but a call for verification and audit. The state asked whether projects are actually likely to be built, whether the grid can handle them, and asked for a look at water use and tax incentives too. On September 21, the state went further: the Texas Commission on Environmental Quality (TCEQ) halted permits for data center projects, and the state's announcement directed other state agencies not to advance related regulatory approvals until the needed information is in hand.
AI Scales at Software Speed. Power Scales at Infrastructure Speed
For the past several years, the bottleneck in AI infrastructure has been GPUs, high-bandwidth memory, and advanced packaging. Those constraints still matter. But the Texas episode adds a new question in front of all of them.
Can you actually turn the GPU on?
GPU generations turn over on roughly a one-year cycle. Data center buildings can go up relatively fast too. But large transformers, transmission lines, substations, gas turbines, and new power plants need far longer lead times for ordering, permitting, and construction. Reuters reported that lead times for high-voltage transformers in the US stretched to as long as 160 weeks in 2026.
That time gap is turning into a structural bottleneck for the AI industry. Compute demand spikes the way software does. Power infrastructure moves at the speed of land, construction, permitting, and equipment supply chains.
| AI Infrastructure Layer | Past Key Bottleneck | Emerging Bottleneck |
|---|---|---|
| Compute | GPUs, HBM, CoWoS | Compute efficiency per watt, actual time to operation |
| Facility | Server racks, cooling | Substations, transformers, switchgear, power and water for cooling |
| Grid | Relatively a background variable | Grid interconnection, generation source, permitting, community costs |
| Investment KPI | GPU count, announced GW | Energized GW, time-to-power |
The Fix Isn't Just Building More Power Plants
Adding generation capacity is necessary. AI demand is already widening the investment map to gas turbines, nuclear, transmission, batteries, transformers, and switchgear. But given supply chain and permitting timelines, power plants alone are unlikely to solve this in the near term.
That's what makes the AI Energy Management Alliance, launched by Nvidia, Google, and Emerald AI on September 16, worth watching. Their pitch is to turn data centers from fixed power consumers into flexible load. When the grid is tight, some training jobs get delayed, batteries discharge, on-site generation kicks in, or compute shifts to hours when the grid has slack.
The fix for AI's power problem is likely to land on two fronts: growing supply faster, and moving demand more intelligently.
What Investors Should Track Now: Energized GW
Imagine two AI cloud operators both secure 100,000 of the latest GPUs. Company A has finished its 2GW grid connection, substation, cooling, land, and permits. Company B has GPU purchase contracts but its grid connection isn't scheduled until 2029.
Same GPU count, different economic value. Company A can switch its GPUs on and turn them into tokens and revenue. Company B's assets may sit idle until the power infrastructure is ready.
That's why the next KPI for the data center industry is likely to be energized GW and time-to-power, not announced GW. What matters going forward isn't how many gigawatts a company announced, but when the power actually arrives, whether that power is secured for the long term, and who ends up paying for the added grid costs.
Insight Times Editorial Desk
