The Big Four US hyperscalers revised their 2026 capex up after Q1 earnings, not down. Amazon, Microsoft, Alphabet and Meta now collectively guide to roughly $710 billion in 2026 AI infrastructure spend, with Amazon at ~$200 billion, Microsoft at ~$190 billion, Alphabet at ~$185 billion and Meta at ~$135 billion [24/7 Wall St., May 1, 2026; CNBC, April 29, 2026]. The earlier $630 billion figure carried by Al Jazeera reflected pre-revision Morgan Stanley estimates. None of that money is the bottleneck. The bottleneck is the grid.
What’s happening
- Sightline Climate’s May data center outlook tracks roughly 5 GW of US 2026 capacity under active construction against 16 GW planned. Tech Insider’s review of the same outlook frames the gap as a ~7 GW shortfall against the 12 to 16 GW operators had committed to deliver this year [Sightline Climate, May 2026; Tech Insider review of Sightline, 2026].
- Sightline tracks 190 GW across 777 announced data center projects globally since 2024. Its base case is that 30 to 50 percent of the 2026 pipeline will not come online before year-end [Sightline Climate, May 2026].
- Transformer delivery has stretched from 24 to 30 months pre-2020 to as long as 5 years today [Tech Insider, 2026 review of supply-chain reports]. Grid interconnection waits in Northern Virginia, Phoenix and Dallas are now reported at 4 to 7 years [Sightline Climate, May 2026].
- Wood Mackenzie reported that US energy grid limitations caused a 50 percent quarter-on-quarter drop in new data center projects at the end of 2025 [Al Jazeera, May 28, 2026].
- China’s East Data, West Computing strategy is set to lift installed data center capacity from 32 GW at end-2025 to 40 GW at end-2026, and to 60 GW by 2030 [Rystad Energy, May 20, 2026]. AI and HPC facilities will be 48 percent of that capacity by 2030.
The contrast is not chips. The contrast is electrons.
Brazil angle
While Washington argues about permitting, Brazil’s largest single IT infrastructure contract quietly hit its first milestone. Bloomberg reported on May 12 that Elea Data Centers completed phase one of its R$2.3 billion (about $470 million) data center project for Petrobras [Bloomberg, May 12, 2026]. The 17-year contract for a 30 MVA facility in São Bernardo do Campo, São Paulo was first awarded in October 2025 and will host Petrobras’s supercomputers [Data Centre Magazine; PR Newswire, October 2025]. The site was bid out as renewably powered from day one, with liquid cooling and water reuse engineered into the design [Data Centre Magazine, October 2025].
The contract is the largest IT infrastructure award ever made by a Latin American company. What makes it relevant to the watt-gap story is that the energy contract and the construction schedule were sequenced together. Brazil’s surplus renewable generation is the load-bearing fact underneath that sequencing. The same surplus is what makes the R$15 billion BNDES Sovereign Plan Credit Line presented to IBRAM in May a coherent industrial policy, not just a financing announcement.
US angle
Hyperscaler capex commitment is intact and accelerating. Microsoft’s April 29 earnings raised calendar-2026 capex guidance to roughly $190 billion (from earlier ~$120 billion run-rate estimates), citing ~$25 billion of component price inflation [CNBC, April 29, 2026; Microsoft FY26 Q3]. Amazon’s 2026 guidance is around $200 billion [24/7 Wall St., May 1, 2026]. The capital is there. The watts are not. Data Center Watch, the research arm of 10a Labs, counts 36 US data centers blocked or stalled between May 2024 and June 2025 [cited via Al Jazeera, May 28, 2026]. The opposition is rate-payer cost spillover, water draw, and grid stability.
Even tech leaders concede the point. Elon Musk, speaking at Davos in January, said the limiting factor for AI deployment “is fundamentally electrical power.” His framing, that very soon the US will be producing more chips than it can turn on, lines up with what Sightline is tracking [Al Jazeera, May 28, 2026].
China angle
China’s advantage is not just generation. It is regulatory speed. Huawei’s own FusionDC case studies (notably the China Mobile Shaanxi deployment, ~1,000 racks delivered) document end-to-end build times of 6 months for prefabricated modular campuses [Huawei Digital Power; FusionDC case study]. Capital Economics, citing the Huawei delivery model, has flagged the same six-month figure against at least a year for US equivalents [Leah Fahy, Capital Economics, via Al Jazeera, May 28, 2026]. China added more than 430 GW of wind and solar in 2025 alone, more than half of global additions that year [BloombergNEF, cited via Al Jazeera, May 28, 2026].
On May 12, Beijing’s SASAC announced the start of operations at China’s first large-scale renewable energy project directly linked to a data center, a 500 MW wind and solar plant in Ningxia feeding a China Datang cloud campus via a dedicated transmission line [Al Jazeera, May 28, 2026]. That is the model: gen plus transmission plus campus, sequenced as one project.
The caveats are real and need temporal context. China’s eight national computing hubs are now running at roughly 63 percent utilization as of 2024, up from 59 percent in 2022 [Liu Liehong, National Data Bureau, 2024 Big Data Expo]. The widely cited 20 to 30 percent figure applies to broader-fleet idle capacity under the original Eastern Data Western Computing rollout in the early 2020s, not the current hub fleet [Reuters; AspiStrategist]. SMIC co-CEO Zhao Haijun’s February warning about rushed AI capacity sitting idle was framed as a global caution, using China’s earlier overbuild as the cautionary precedent, not as a China-specific risk for 2026 [Bloomberg, February 11, 2026]. China has watts. China has fewer leading-edge chips.
What it means
The watt gap is the cleanest read for why TAI-P (power inputs) is up 13.6 percent year-to-date while TAI-M (materials) is up 6.4 percent. Capital is repricing the bottleneck. Grid equipment, uranium and nuclear utility equity are absorbing the marginal hyperscaler dollar faster than copper or rare-earth magnet exposure. Materials still matter, but the binding constraint in 2026 is delivery, not extraction.
For SDX, the read is harder. Brazilian producers are positioned as the long-cycle supply answer. Petrobras-Elea is the proof of execution. The equity market is not yet pricing the option.
What to watch
- PUCN ruling on the Google-Ormat 150 MW Nevada geothermal PPA, expected H2 2026. First test of whether US utilities can execute the gen-plus-transmission-plus-campus pattern China just demonstrated in Ningxia [Ormat / NV Energy, Feb 17, 2026].
- Applied Digital Polaris Forge 3 energization, targeted August 2027. The May 20 lease covers 300 MW IT load and 430 MW grid power under a 15-year $7.5 billion take-or-pay [Applied Digital, May 20, 2026]. Slippage is the canary.
- China 15th Five-Year Plan data center addendum. Whether the 80 percent renewables rule in the 8 national computing hubs holds, or whether interprovincial grid coordination slips it [Rystad, May 20, 2026].
The watts decide the cycle. Until execution catches commitment, the gap is the story.