The Gigawatt Race: AI's Power Hunger

A gigawatt of AI infrastructure costs $100 billion to build. Jensen Huang said it at Computex last week, and the number landed like a brick. Not because it's shocking—hyperscalers have been spending at that scale for months—but because of what comes next: at that density, power becomes the product. "Throughput per watt is revenue," Huang told the crowd in Taipei, according to ServeTheHome's live coverage of the keynote.

The AI boom has triggered the fastest infrastructure scramble since the 1980s, and it's not playing out in chip fabs or data halls. It's happening in gas turbine factories, pipeline networks, and utility boardrooms. Natural gas turbine orders hit a 25-year high on data center demand, according to the International Energy Agency , which warned that AI infrastructure spending now threatens to crowd out traditional LNG buyers. The constraint isn't silicon anymore. It's electrons.

Can the Grid Keep Up?

GE Vernova reported that data centers generated more orders for electrification equipment in the first quarter of 2026 than in all of 2025 combined, with $2.4 billion in Q1 orders tied to data center customers , the company disclosed in its earnings call. The company signed 21 gigawatts of new gas turbine agreements in Q1, and its backlog climbed from 40 to 44 GW . Pricing on new gas turbine orders in the first half of 2026 is tracking 10 to 20 points higher on a dollar-per-kilowatt basis than fourth-quarter 2025 orders , CEO Scott Strazik told analysts.

The turbine shortage is acute. Gas turbine prices are expected to be up 195% over 2019 levels by the end of this year, and waitlists are stretching into the early 2030s , according to BloombergNEF. The focus on natural gas has already doubled the cost of natural gas turbines and increased wait times to up to seven years . Some companies are getting creative: ProEnergy is repurposing retired jet engines into power generators to bypass the backlog, Data Center Dynamics reported.

The scale is staggering. RBC Capital Markets forecasts natural gas consumption from data centers will reach approximately 6.1 billion cubic feet per day by 2030, roughly a 20% increase to baseline power demand from 2025 levels . Deloitte estimates incremental power demand could translate into as much as 12 Bcf/d of additional gas demand by 2030, up from roughly 35-37 Bcf/d today . That's enough to power Japan.

But there's a catch. Current filed plans for data centers represent ~125GW in the U.S.—a staggering 50% or more increase in gas demand that the distribution grid is not ready for , according to energy consultancy CleanEpic. Pipelines store gas by packing them with higher pressure; during the January 2026 winter storm, a 20% demand spike drained East Coast pipeline capacity and sent Henry Hub prices from $7.75 to $30 per MMBtu in a week. A permanent 50% increase isn't viable without new pipelines—and new pipelines are nearly impossible to build.

What About the Chips Themselves?

Huang's Computex keynote was a victory lap for Nvidia's Vera Rubin platform, the company's most ambitious rack-scale system yet. Vera Rubin is now in full production , Huang announced, with Rubin-based products available from partners in the second half of 2026, with AWS, Google Cloud, Microsoft, and Oracle among the first to deploy . The 336-billion-transistor Rubin R100 GPU promises 5x the inference performance of Blackwell at 10x lower cost per token , according to analysts at Goldman Sachs.

Nvidia stock jumped 6% on the day , CNBC reported, as Huang's message—"compute is revenue"—resonated with investors who'd grown skeptical of data center economics. Huang specifically mentioned Oracle, Nebius, and CoreWeave during his speech, and shares of those three companies jumped 9.9%, 14.5%, and 14% respectively .

But the chip supply chain remains brutally tight. Top-tier customers including Nvidia, AMD, Broadcom, and Google have locked in over 85% of TSMC's total CoWoS packaging capacity for 2026, leaving less than 15% for all other AI chip companies , according to industry analysis. Both HBM3E 8-hi and 12-hi memory stacks are fully allocated for 2026, with prices rising 15–22% year-over-year, and SK Hynix has indicated shortages may extend into late 2027 . TSMC's CEO stated on the April 2026 earnings call that it will not be until 2027 that "supply can meet demand" , according to a CNAS report.

The constraint has shifted from design to manufacturing. The principal risk to AI sector growth is no longer chip design but the physical manufacturing capacity for advanced semiconductor packaging and high-bandwidth memory, with TSMC's CoWoS technology and the HBM supply chain acting as the primary governors on AI hardware deployment through 2027 .

What Changed This Week

Huang's Computex keynote reframed the AI infrastructure debate around power efficiency rather than raw compute. The message: a gigawatt is a gigawatt, and whoever delivers the most tokens per watt wins. That shift is already reshaping capital allocation—GE Vernova's Q1 data center orders exceeded all of 2025, and gas turbine prices are up 195% since 2019. Meanwhile, the chip supply chain remains sold out through 2027, with TSMC's packaging capacity and HBM memory acting as hard limits on how fast AI infrastructure can scale.

What to Watch

GE Vernova expects to reach at least 110 GW of combined gas turbine backlog and slot reservations by year-end 2026. Nvidia's Vera Rubin volume production ramps in Q4 2026, with Microsoft's Fairwater AI superfactories among the first deployments. TSMC's CEO has said supply won't meet demand until 2027—watch for any updates on CoWoS expansion timelines. And keep an eye on natural gas prices: Henry Hub traded at $3.25/MMBtu on Friday, down 2.4% according to market data, but pipeline constraints could send prices spiking again this winter if data center demand continues to climb.