The Bottleneck Moved From Silicon to Substations, and Texas Is the Proof

Gartner puts a date on when AI hardware pulls ahead of everything else drawing power in a data center: 2027, after an 84% jump in 2026 alone. That crossover is getting the headlines, but I find the more telling number to be the growth rate of conventional servers, which was less than 1% in 2025 and are on track for just 1.2% in 2026.

What's actually growing, and what isn't:

→ Global data center electricity climbs 26% in 2026 to 565 TWh, up from 447 TWh in 2025, and tops 1,200 TWh by 2030, per Gartner's June forecast - modeled projections, not measured outcomes; the 2030 figure carries the most uncertainty.

→ AI-optimized servers: 95 TWh in 2025 to 175 TWh in 2026 (+84%) to 258 TWh in 2027, the year they overtake conventional hardware.

→ Conventional servers: essentially flat, landing at just 200 TWh in 2027, barely above where they sit today.

→ Cooling electricity climbs 22.6% in 2026 to 195 TWh from roughly 159 TWh in 2025, a base larger than what AI servers drew that year, though its 36 TWh increase falls well short of AI servers' 80 TWh gain.

→ More than 75 data center projects worth $130 billion were blocked in the first three months of 2026 amid local opposition over power and water costs.

→ Texas halted new interconnections on August 3. Governor Greg Abbott directed the state utility commission and ERCOT to audit every data center in the interconnection queue before any proceeds, denying grid access to non-compliant projects. ERCOT's queue has gone from 233 GW in January to 474 GW, roughly 90% of it data centers, more than five times the state's peak demand, though queues are notoriously inflated by speculative filings and developers holding multiple interconnection points, so not all of it is firm demand.

The flat line:

The 2027 crossover is being treated as news, but it was inevitable arithmetic: put an 84%-a-year line next to a 1%-a-year line and they cross whenever you like. The number that actually says something is the flat one: conventional compute isn't shrinking, it's been sidelined, as growth consolidates into a single, power-hungry category. That concentration matters more to a grid operator than any percentage: AI growth arrives as a handful of enormous single interconnection requests, each needing new transmission and often a new substation, while the same growth spread across existing conventional sites needs no new connection at all. Cooling makes the same point from the other side: a larger base, growing more slowly in absolute terms than AI itself.

Silicon to substations:

That concentration is why the bottleneck has rotated from chips to grid connections, and Texas just demonstrated it unambiguously. A year ago the constraint was GPU allocation, something a chip roadmap could eventually fix. Now it's interconnection and permitting, which run on timelines Nvidia can't speed up. The state that spent a decade selling itself as the path of least resistance for large loads is now checking paperwork instead of approving connections: power plans, water use, cooling design, tax incentives, ownership.

The regulators are reaching for the same lever from two directions. FERC has given all six of its regional grid operators until roughly August 17 to justify or rewrite the tariffs governing how loads above 50 MW connect to the transmission system. ERCOT wasn't on that list; Texas sits outside FERC's jurisdiction, which is why its version arrived as a governor's directive instead of a federal one.

Gartner's own advice to operators - prioritize efficiency upgrades, lock in grid access, and invest in high-efficiency cooling and edge computing - reads more like a utility research note than hardware guidance, itself a signal of where the chokepoint moved.

The 2029 problem:

The obvious bridge is gas, and it's arriving in a different shape than the behind-the-meter projects that drew attention earlier this year. NRG announced principal commercial terms on August 4 for a 1.2 GW plant in Texas backed by an unnamed global hyperscaler, expandable to 2.4 GW, but grid-connected and customer-funded, a meaningfully different bet that adds supply to the system instead of stepping around it. It also arrives in late 2029. Meta's 6.6 GW of nuclear contracts with TerraPower, Oklo, and Vistra mostly land in the 2030s. That's the whole problem in miniature: the fixes on the board arrive on the far side of the decade, and the demand curve is here now.

From here, keeps tabs on whether the Texas audit becomes a template. The utility commission takes up ERCOT's suspended interconnection classifications on August 20, and whatever standard emerges is the one other states will borrow. Either way, the flat line stays flat; what's actually in contention is how much of the AI curve above it ever gets the power to run.