The hyperscalers’ dilemma

The opinions expressed here by Trellis expert contributors are their own, not those of Trellis or its editors.

Last year, Wall Street’s consensus for 2026 capital expenditure by the major tech companies averaged $365 billion. Today, it’s $725 – $805 billion. Roughly three-quarters is for direct AI infrastructure: GPUs, racks, campuses, substations. Capex estimates for the 2025-2030 buildout have risen about a quarter since October alone. Big Tech capex in 2026 will approach 3 percent of U.S. GDP — comparable, as a share of output, to the peak of railroad construction in the 19th century or the run-up to Y2K. 

What’s different from those prior episodes is the financing. Through most of the post-ChatGPT cycle, hyperscalers funded the buildout out of retained earnings, sustaining what bond investors had come to regard as an unspoken contract: AI speculation would be borne by equity, not credit. 

That contract is now broken. In 2025 the five hyperscalers —Amazon, Microsoft, Google, Meta and Apple — issued $121 billion of bonds, against a five-year average closer to $28 billion. Estimates for 2026 investment-grade bond issuance run $300 to $400 billion in a market where AI-related debt is already the largest single segment of Investment Grade bonds. The Dallas Fed now treats this as a duration-supply phenomenon material to U.S. interest rates. 

Layer that against a U.S. balance sheet past any defensible capacity to absorb stress, and the picture is novel: the largest private capital cycle in modern history, debt-financed at the margin, in a fiscal regime with no remaining shock absorbers. It is happening at a scale no domestic grid was built for and no electorate has been asked to ratify. 

For sustainability professionals, this brings a range of new challenges and opportunities not seen since the oil shocks of the 1970s — and once again this upheaval is being accompanied by energy price spikes. 

AI as the dominant marginal load 

The International Energy Agency projects that data centers will account for nearly half of U.S. electricity demand growth through 2030. By the end of that period, the American economy will burn more power processing data than it does smelting steel, refining aluminum, making cement, and producing chemicals —  combined. After two decades of flat domestic power demand, one buyer —  AI data centers — has put the grid on a growth footing. 

There is a counterintuitive consequence. Even as Washington has retreated from a coherent climate posture, large investors who do not care about the politics are pouring money into geothermal, advanced nuclear and grid-scale storage. Big data centers need 24/7 clean firm power faster than gas turbines, interconnect queues and litigation can deliver. 

Google is signing enhanced geothermal offtakes in Nevada. Microsoft has restarted reactors. Amazon is anchoring pre-orders for small modular reactors. The bipartisan support for geothermal moving through Colorado, the Mountain West and federal energy and water appropriation isn’t climate policy. It’s industrial policy refracted through computing power —  and it has moved faster in the past 18 months than the climate movement managed in 30 years. 

That is the optimistic reading. There is a less generous one. 

The consent deficit 

From Virginia farmland to Pennsylvania exurbs to Georgia counties to Cascade Locks, Oregon, this buildout is colliding with the consent of the governed. In Q1 2026 alone, at least 20 proposed data centers were cancelled in the face of organized local opposition —  roughly $42 billion of capex and 3.5 gigawatts of demand erased before the first concrete pour. A three-year tally of cancelled or stalled projects exceeds $85 billion. Baird counts 188 active local opposition groups across 40 states. A Colorado poll found that 91 percent of Coloradans support tighter rules on datacenter growth. 

This opposition is neither anti-technology nor partisan; the groups skew rural, cross-ideological and taxpayer-focused. They have noticed what the financial press has been slow to recognize: Utilities are planning roughly $1.4 trillion of capex through 2030, and a meaningful share will be borne by residential ratepayers —  $700 billion in higher household bills, according to the Energy Information Administration. 

Don’t be surprised when citizens start recalling local officials and voting out town councils.    

Navigating the ripple effects 

Sustainability practitioners need to watch states and public utility commissions that are now the operational front line in negotiations for hyperscalers’ needs for clean firm power and permitting cover. The price being extracted has four components: 

Additionality. A proposed law in Colorado would have required large-load data centers to source 100 percent of their power from new renewable resources by 2031, not existing ones. The bill faltered in the final days of the 2026 session, but it won’t be the last. Virginia is moving along a similar path. The implication for procurement: Contracting against existing renewable supply is increasingly insufficient. New generation built because of the load is becoming the regulatory floor, not a sustainability aspiration. 

Sealed cost recovery. The Colorado framework would have required operators to pre-pay or sign 15-year contracts covering the incremental generation, transmission and distribution costs that their load imposes. That is, the load pays for its own infrastructure, not the household down the street. Expect this to be mainstreamed. 

Community benefit and protections for vulnerable communities. In disproportionately impacted communities, the Colorado bill required cumulative-impact reviews, public hearings and binding community-benefit agreements. Texas, Georgia and Oregon are weighing similar measures. This is the language of environmental justice, and it will likely be required in utility-scale procurement contracts regardless of what administration is in Washington. 

Load-following clean firm generation. A separate bipartisan bill in Colorado, which also failed this year, would require investor-owned utilities to solicit geothermal projects while clearing permitting friction for thermal energy. While some logistics remain to be worked out, legislators agreed that the megawatts those loads will need should come from beneath Colorado, and the upside should accrue to Coloradans. 

Although both datacenter bills stumbled this year, they will return in 2027, and they are already being studied by every public utilities commission (PUC) and statehouse with a material datacenter pipeline. 

How corporate sustainability leaders can respond 

Treat AI power demand as a Scope 3 emissions vector with first-order materiality. The carbon intensity of AI training and inference varies by an order of magnitude across regions and utility mixes. Procurement choices for AI services are now functionally energy-mix choices. 

Update power purchase agreements and cloud commitments to the rising bar. Additionality, 24/7 carbon-free energy matching and load-following firm clean supply are no longer leading edge; they’re the floor that a credible policy environment should codify. The reputational gap between “100 percent renewable” claims sourced from existing supply and what states will require is about to widen. 

Engage seriously at the state and PUC level. Federal climate policy is in retreat; state energy policy is accelerating. Most corporate sustainability functions are still organized around a federal-policy reflex that no longer fits the terrain. 

Treat community license as procurement risk. A datacenter contract, direct or indirect, with no community-benefit floor and no ratepayer firewall is a stranded-asset event waiting to happen. Every Virginia subdivision watching its bills rise to subsidize a data center campus it never agreed to is a future “no” vote on the grid investments that the energy transition requires. 

The AI buildout is the most powerful demand signal for clean firm electricity that has ever existed in the U.S. — doing more to commercialize enhanced geothermal and re-rated nuclear plants than three decades of climate policy. But data centers built without consent and underwritten silently by households, would be the most efficient machine in operation for destroying the social license of the energy transition itself.