From Capture-Ready to Capture-Committed: Decarbonizing Natural Gas with CCS

Key Takeaways

• Community opposition to data center development has become a national conversation, and state legislatures are now responding with a broad spectrum of policy responses to address constituent concerns.  
• Relae’s policy experts have identified six broad categories of policy responses: tax incentives, disclosure requirements, transparency rules, cost allocation, incentive clawbacks, and moratoria. In 2026, incentive bills are stalling while disclosure, cost allocation, and clawback bills are advancing.
• The shifting action of state legislatures across these categories can guide developers on the relative political openness or opposition to data center developments across the United States, and should inform site selection and investment decisions.
• Every state legislature will be back in session in 2027, including major data-center markets like North Dakota, Nevada, and Texas that sat out 2026, so this picture will keep changing.

Which States Are Getting Harder to Build In

Community concerns have translated into opposition, delaying, derailing, or altogether denying projects. Local debates over land, water, and electricity prices have moved into statehouses as a result. 

If you're trying to figure out "Which states are becoming harder (or easier) to site a data center in, and why?", the state legislature is where that answer is being written. Last year, we tracked the initial wave of state bills on data centers. This year, the volume of bills has grown. The policy categories we used last year still broadly hold, but the distribution of emphasis across these categories has shifted significantly.

Relae’s policy experts reviewed the bills introduced in 2026 legislative sessions related to data centers and identified six categories of policy responses available to a state: tax incentives (“reward it”), disclosure requirements (“review it”), transparency rules (“reveal it”), cost allocation (“shape it”), incentive clawbacks (“revoke it”), and moratoria (“stop it”). These categories do contain some overlap, but together they show a range of what legislatures are trying. 

Each response answers at least one of three questions local communities have been posing for over a year now: Who bears the costs of new generation, transmission, and other grid upgrades? What are constituents entitled to know about data center developments in their region? Who decides whether a project goes forward?

For each policy response, we selected at least one representative bill that is either most consequential financially or most impactful to developer timelines. Notably, states with significant data center development pressure, such as North Dakota, Nevada, and Texas, did not have regular sessions in 2026, meaning the scope of policy responses reflected here is not exhaustive. 

Six State Policy Responses to Watch for Data Center Siting and Investment

Which policy response a state reaches for, and how that bill fares in the legislative process, signals how politically open or resistant it currently is to new data center development. That political landscape should directly inform data center site selection and investment decisions - and it’s changing quickly.

1. Reward It: Data Center Tax Incentives and Subsidies 

These bills try to attract data centers by giving them a financial reason to locate in a state. We call this response "reward it." It typically means property tax exemptions, sales tax breaks on equipment, or reduced utility taxes, usually in exchange for a minimum investment and a job-creation commitment.

West Virginia Senate Bill 623, the “West Virginia-Powered Data Center Incentive Act,” would have exempted data center equipment from property taxes and cut business taxes for the utilities supplying power to the data center. To qualify, a data center would have needed to invest at least $50 million, create 50 jobs, and run on coal-generated electricity for at least 80% of its operational capacity. The bill died in committee without a vote, which is emblematic of how this category of bills fared in 2026. 

Relative to the approximately 50 bills introduced (and 9 passed) in 2025 that provided tax incentives for data centers, this category declined significantly in the 2026 session. Part of the drop off could also be a product of the current tax laws of many states, which already include substantial tax incentives for data centers. 

For developers, this is the category to watch shrink: fewer states are actively competing to hand out new incentives, so existing programs may be close to the extent of what's on offer going forward.

2. Review It: Data Center Disclosure Requirements 

Under what we call the "review it" response, these bills require specific information to be shared with regulators or grid operators before or while the project is built, usually about power demand, interconnection plans, or backup generation.

Pennsylvania House Bill 1834, the “Data Center Act,” would direct the state’s Public Utility Commission (PUC) to oversee commercial data centers with a peak demand of 25 megawatts or greater. It would require disclosure of information relating to interconnection requests as well as planned backup generation. It would also fast-track interconnection for facilities that bring incremental clean firm resources. That disclosure helps grid planners in PJM spot duplicate projects in their interconnection queue, while also helping state leaders to understand if projects in the state are in jeopardy of being lured elsewhere. The House passed the bill 104-95, and is awaiting action in the Senate. 

This category, which also includes bills that require disclosure of information around environmental impacts and power or water usage, has grown in 2026. Pennsylvania has been at the center of many discussions around data centers and utility affordability this year. Governor Josh Shapiro's initiative on utility affordability identifies data center load as the pressure point while a wider discussion evolves in PJM over capacity costs. 

3. Reveal It: Data Center Transparency Rules and NDA Bans 

This is a new category in 2026. We call it the "reveal it" response: bills that stop public officials from signing non-disclosure agreements about data center projects. The goal is to inform residents about who is behind a proposed facility and how much power it will use.

Ohio House Bill 695 would bar county commissioners, township trustees, village mayors, and council members from entering into nondisclosure agreements (NDAs) related to their official duties. It directly targets the NDAs that have kept local officials from telling residents who is behind a proposed project or how much power it would consume. The Ohio legislature is in session until December 17, and the bill is still awaiting action in committee. A similar bill in Pennsylvania, House Bill 2359, would tie NDA prohibitions to a data center sales tax exemption. That bill has passed the House and is also currently pending in the Senate. Separately, in August, Gov. Shapiro banned NDA use outright on any data center project the state oversees.

4. Shape It: Data Center Cost Allocation and Ratepayer Protection 

These bills aim to allocate more of the grid infrastructure costs associated with upgrades to the data center. New power lines, substations, and other transmission costs fall on the data center, not on everyone else's electric bill, under what we call the "shape it" response.

California Senate Bill 886, the California Technology Innovation and Ratepayer Protection Act, was signed into law by Governor Gavin Newsom on September 21, 2026, as part of a seven-bill data center package. It passed the Senate 28-10 and the Assembly 49-7. It requires the California PUC to establish or update electricity tariffs by January 1, 2028, for data centers entering new transmission-level interconnection agreements on or after January 1, 2027. Under it, developers bear the cost of transmission upgrades and face a 10-year early termination fee if they back out or fail to reach capacity. This is a strong template for keeping the costs of serving large loads from impacting retail ratepayers, and similar bills should be expected in other states in 2027. 

5. Revoke It: Data Center Incentive Clawbacks

_These bills take back or narrow tax breaks a state already granted (the "revoke it" response), usually once lawmakers start weighing what those incentives are actually costing the state against what they bought.

Washington Senate Bill 6231, signed into law by Governor Bob Ferguson in April, will remove the sales tax exemption for replacing or refurbishing existing server equipment. The category of bills revoking previous tax incentives for data center development and operation has grown substantially as more states are starting to reconsider the large incentives they put into place at the beginning of the data center boom. Over 50 bills in more than 20 states have sought to restrict data center incentives in 2026, including New Jersey Senate Bill S4390, which was signed into law in August.

6. Stop It: Data Center Development  Moratoria

These bills seek to temporarily restrict certain new data center development outright (the "stop it" response), usually for a fixed period while the state studies the impact on the grid, water use, or electricity rates.

New York Senate Bill S10642/Assembly Bill A11560, the Responsible Data Center Development Act, would impose a permit moratorium on data centers of 20 megawatts or more. It passed the legislature on June 4th and is awaiting action by Governor Kathy Hochul. If signed into law, it would be the first moratorium enacted through a legislature. Separately, Governor Hochul has issued an executive order to establish a temporary moratorium on data centers, though her order only applies to facilities of 50 megawatts or more. 

This category did not exist a year ago within state legislatures, but as public concern about data centers has grown, several moratorium bills have been introduced. Maine’s LD 307 passed through the state legislature and would have done the same, but was vetoed by Governor Janet Mills in April. In Texas, Governor Greg Abbott has issued orders to state agencies to halt data center approvals in ERCOT until audits can be completed, and in September, he expanded the pause to also cover a review of environmental permits. 

It is important to note that many of the moratoria put in place at the state level, whether through the legislature or via executive action, are temporary, usually one year in duration. Often, the moratoria being considered include a plan to study the impacts of data centers on communities and whether utility rates may rise.

What to Watch in 2027

The most consequential policies in 2026 didn’t come through legislatures. They came from executive action by the governors of New York and Texas, which went further than any legislature with respect to their financial and commercial impact. The outcome of the 2026 midterm elections will shape the appetite of both the legislative and executive branches for additional action. In addition, every state legislature will be in session in early 2027, including North Dakota, Nevada, and Texas, which all sat out 2026, so we expect significant policy developments to continue at the state level next year.

In legislatures, the pattern across all six policy responses points in the same direction: recent legislative activity suggests that some states are moving from attracting data center development through incentives toward imposing additional conditions, oversight requirements, or restrictions on new development. Fewer bills providing new incentives for data centers are being introduced or are stalling in legislatures, while bills on cost allocation, disclosure, incentive reduction, and permitting are advancing more successfully.

Map Your State's Regulatory Risk

Site selection and investment decisions increasingly hinge on which of these six responses a state is pulling, and how fast that can change. Talk to our Policy Advisory Team about what a state's current and pending legislation means for a specific project or portfolio.

Key Takeaways

• Speed to operation is the value proposition. While grid interconnection can take 3-5 years in constrained regions, hydrogen fuel cell systems can be deployed in a fraction of that time—a large advantage for operators racing to meet data center capacity demands.
• Hydrogen is moving beyond backup. Fuel cells could supply 6-15% of incremental data center power, functioning as primary or near-primary power rather than emergency standby.
• Higher fuel costs can be rational. When the alternative is a multi-year wait for grid power and the development loss that comes with it, paying a premium for hydrogen is better than falling behind.

The Grid Can’t Keep Up With Data Centers

The data center infrastructure buildout is moving at a pace the US power grid was never designed to accommodate. Cleanview tracks about 63 GW of operating US data center capacity and another 392 GW in planning, roughly six times what exists today. Developers in high-demand corridors—Northern Virginia, Central Texas, Phoenix—are finding that the grid can't deliver new capacity fast enough.

Interconnection queues now stretch 3-5 years in many regions, and that’s before accounting for permitting, transmission upgrades, and substation buildouts. For operators that need power online faster than the grid can deliver it, this timeline gap is a fundamental constraint on the business.

This mismatch between AI deployment, generator bottlenecks, and grid timelines is why hydrogen is entering the conversation as a viable power option for a growing subset of data center deployments, alongside natural gas with carbon capture, geothermal, and long-duration energy storage.

Hydrogen Fuel Cells: From Standby to Primary Power

For most of its history in data centers, hydrogen has been discussed as a cleaner replacement for diesel backup generators. That framing undersells the opportunity. 

The shift happening now is more fundamental: hydrogen fuel cells are being evaluated as primary or near-primary power sources for behind-the-meter and facilities that can’t wait for grid connection. Notably, many modern fuel cell systems—including solid oxide fuel cells—can run on natural gas today and transition to hydrogen as clean supply becomes available. Bloom Energy's 2024 deal with American Electric Power to provide up to 1 GW of fuel cells for data center power illustrates this approach: the systems will initially run on natural gas, with the capability to switch to 100% hydrogen in the future.

Furthermore, fuel cell systems are modular, can be deployed incrementally as load grows, and provide continuous power output. This is a promising path, and one that will scale only as fast as the fuel cell supply chain can keep up. Unlike solar or wind, fuel cells aren’t weather-dependent. Running on hydrogen, they produce zero direct emissions at the point of use. Running on natural gas, they still emit CO₂, but their higher efficiency gives them a lower carbon intensity than a comparable gas turbine.

Goldman Sachs estimates that fuel cells could supply 6-15% of incremental data center power (the estimate doesn't specify fuel mix, though most near-term deployments would run primarily on natural gas). This is a meaningful share of new capacity, driven largely by operators who face years-long waits for both grid connections and new gas turbines. Bloom Energy, whose fuel cell products serve data center applications, has been a notable outlier in an otherwise struggling hydrogen sector, with its stock rising sharply on strong revenue performance and growing investor enthusiasm around data center power demand.

Speed to power depends on the fuel. Solid oxide fuel cells run on pipeline natural gas today, so operators do not wait for a hydrogen supply chain. Bloom states it can deliver onsite power for an entire data center in as little as 90 days, with the ability to add modules as load grows under certain circumstances. In all but a few select locations, hydrogen comes later when the same systems switch fuels as clean supply scales. Projects that need hydrogen from day one face additional challenges. Limited hydrogen availability and supporting infrastructure across much of the US can undermine the potential speed advantage of hydrogen-dependent projects. Nevertheless, locations with suitable hydrogen supply and supporting infrastructure may offer viable near-term opportunities.

The Hydrogen Premium: When Paying More Makes Sense

Hydrogen-generated power is more expensive than other forms of electricity generation in most markets. The logistics penalty (including compression, transport, and storage) can push hydrogen from $4/kg at the production site to $8-16/kg delivered, making the economics look unfavorable on a simple $/MWh comparison. In practical terms, hydrogen at $4-6/kg translates to roughly $120-180/MWh on an energy-content basis (LHV). At a fuel cell electrical efficiency of around 50%—typical for solid oxide systems—the effective cost of electricity rises to approximately $240-360/MWh for co-located production, compared to $40-60/MWh for natural gas fuel cells, $51-129/MWh for gas combined cycle, and $144-276/MWh for gas peakers. Operators can improve hydrogen economics by co-locating with production, avoiding the logistics penalty that can double or quadruple delivered costs.

While expensive, this cost comparison misses the central point. The relevant question is not whether hydrogen is cheaper than grid power, but whether it is better than having no access to power at all.

For an operator with signed contracts and a 100 MW facility sitting idle during a multi-year interconnection queue, the opportunity cost is substantial. At average US wholesale colocation rates (~$196/kW/month - a record high), a facility of that size represents roughly $235 million per year in potential revenue—meaning even a single year of delay can outweigh years of hydrogen fuel premiums. 

The Inflation Reduction Act’s 45V Production Tax Credit further compresses the gap, offering up to $3/kg for clean hydrogen production—but projects must begin construction by December 31, 2027, to qualify. That deadline adds urgency to what’s already a time-sensitive calculus.

Reframing the Cost Debate for Speed-to-Power

The debate about whether hydrogen is cost-competitive with grid power misses the strategic challenge facing data center operators today: the fastest path to reliable, clean power. Hydrogen can play a consequential role in that timeline in some markets.

Assessing that requires modeling the full picture: grid interconnection timelines, hydrogen supply chain costs at a specific location, the 45V credit’s impact on project economics, and how hydrogen compares to batteries and other bridging technologies for a data center’s particular load profile.

Relae’s advisory team helps data center operators and technology companies evaluate these tradeoffs with rigorous, independent analysis. If speed to operation is a factor in your energy planning, let’s talk.

Frequently Asked Questions

How much faster is hydrogen fuel cell deployment than waiting for grid interconnection?
Grid interconnection in constrained regions can take 3-5 years once permitting, transmission upgrades, and substation buildouts are factored in. Fuel cell systems, especially those that start on natural gas and then transition to hydrogen, can be deployed in a fraction of that time. They are modular and can be installed and scaled incrementally as a facility’s load grows, though the pace depends on equipment and fuel supply.

If hydrogen costs more than grid power, why would an operator choose it?
For an operator with a 100 MW facility sitting idle during a multi-year interconnection queue, the lost revenue—roughly $235 million a year at average US wholesale colocation rates—can outweigh years of hydrogen’s fuel premium. That makes the higher cost a rational tradeoff against a multi-year wait for grid power, not just an expensive alternative to it.

Do hydrogen fuel cell systems have to run on pure hydrogen from day one?
No. Many modern fuel cell systems, including solid oxide fuel cells, can run on natural gas today and transition to hydrogen as clean supply becomes available. Bloom Energy’s up to 1 GW agreement with American Electric Power follows this model: the systems start on natural gas, with the capability to switch to 100% hydrogen as the hydrogen supply chain matures.

Key Takeaways

• The electricity demand surge is real and accelerating. Just last year, data center load in the US was projected to increase from 25 GW to 120 GW by 2030. Today, Texas’ preliminary long-term load forecast projects over 187 GW of data center load by 2030, more than double today’s total peak demand of approximately 91 GW.
• Flexible loads that respond dynamically to policy signals are becoming a regulatory requirement. Texas Senate Bill 6 (SB6), signed into law in June 2025, is the clearest signal yet. It makes remote curtailment equipment a condition of interconnection for new loads of 75 MW or more, so utilities can disconnect them during declared firm load shed events, and separately creates a voluntary demand response program that those loads can elect to join. 
• Relae’s power system modeling puts a dollar value on what data center load flexibility is worth. Our ERCOT analysis shows that data center demand response can eliminate forced load shedding risk, even at 40 GW of data center buildout – preventing $5.5 billion in annual consumer welfare losses by curtailing an average of 5% of demand for under 1% of operating hours. 
• Flexible load curtailment and compute uptime do not need to be in conflict. In our modeling, demand response operates as a "ghost battery" at the data center's grid node, absorbing grid stress as a physical battery would. That construct has a direct real-world analog: on-site battery storage lets data centers draw from stored energy during grid stress events rather than curtailing workloads. Technologies such as those demonstrated by Emerald AI have shown 25% load curtailment at cluster scale while preserving compute service quality.

Data Center Electricity Demand is Testing Grid Limits

Electricity demand in the United States is growing at its fastest pace in decades. Leading the surge is a rapid buildout of data centers, driven by the expansion of artificial intelligence. ERCOT, the grid serving most of Texas, is projecting up to 187 GW of new data center load by 2030, against a total peak demand today of approximately 91 GW.

The scale of this shift extends across the country. According to NERC's 2025 Long-Term Reliability Assessment (LTRA), summer peak demand across the US bulk power system is forecast to grow by 224 GW over the next 10 years, 69% above the prior year’s 10-year projection of 132 GW, with data centers as the dominant driver. As previously explored by Relae, data center energy capacity in the US is projected to increase from 25 GW to 120 GW by 2030, characterizing it as the first wave of a longer demand surge that electrification of buildings and transport will reinforce.

Policymakers are responding in real time, and regulatory responses like Texas SB6 are already rewriting the rules for how data centers connect to the grid. The pace of change is fast, but the siting, infrastructure, and interconnection decisions made today will have lasting consequences: they will determine whether data centers are grid assets or grid liabilities, and the financial difference between the two is measured in billions.

New Solutions for Data Center Demand Response: Texas SB6 as a Test Case

The rapid nature of this new wave of load growth means traditional approaches to managing the grid may not be sufficient. In the past, lead times on new sources of electricity demand allowed utilities to procure supply-side resources in advance. The scale and immediacy of data center deployment is revealing limitations of this approach. States, utilities, independent system operators (ISOs), and regulators are actively pursuing novel approaches for grid management in response.

Texas has become a focal point of US data center expansion, with SB6 as a leading policy response. The bill, which took immediate effect upon Governor Greg Abbott's signature on June 20, 2025, is the most significant restructuring of large-load interconnection rules in ERCOT's history. It requires large loads over 75 MW to install remote curtailment equipment as a condition of interconnection, so grid operators can disconnect them during declared firm load shed events, and creates a separate voluntary demand response program, procured competitively with at least 24 hours’ notice, that those loads may elect to join. If data centers are adding substantial new load to the grid, this reasoning goes, they should also contribute to grid stability by demonstrating load flexibility—reducing power draw at times of peak demand.

While demand response programs currently exist, incentivizing voluntary curtailment from data centers is challenging. In an AI compute arms race, the value of uninterrupted compute time far exceeds any available curtailment payment, such as via PJM’s capacity market mechanisms. 

Approaches to bridge that gap are coming to fruition: EPRI’s DCFlex program is working with hyperscalers and utilities to develop the technical protocols, measurement standards, and contractual frameworks that would make large-load demand response a routine grid service. Innovators like Emerald AI have demonstrated a 25% power reduction across a 256-GPU cluster over three hours during an Arizona grid stress event, while preserving compute service quality, helping to bridge the valuation asymmetry between energy and compute. 

Hyperscalers are already putting a flexible load commercial strategy into action. In March 2026, Google announced  1 GW in demand response contracts with multiple US utilities, including Entergy Arkansas, Minnesota Power, and DTE Energy.

The conversation has shifted from whether data centers can be flexible to how that flexibility gets structured and deployed. 

Putting a Dollar Value on Data Center Load Flexibility

For developers, investors, and grid operators navigating data center growth, the challenge has been making high-stakes siting and interconnection decisions without a clear picture of what load flexibility is actually worth, what inflexibility costs the system, or how curtailment requirements will reshape the regulatory landscape. 

Prior research has established that flexible data center load can absorb substantial grid stress. For instance, research from Duke University’s Nicholas Institute found that 22 of the largest US balancing authority areas could absorb approximately 98 GW of new flexible load if 0.5% of that load’s annual energy is curtailed, or 76 GW at a stricter 0.25% curtailment level. 

Relae’s analysis goes further by quantifying the economic cost at each increment of flexible load growth, and the precise threshold at which that flexibility stops being optional. We zeroed in on ERCOT, a region with high data center load growth and immediate regulatory stakes, determining the value of implementing flexibility and, conversely, the system risk of failing to do so.

How We Modeled It 

Assessing the impacts of load growth and flexibility solutions requires a systems-level analysis, best achieved via power market modeling. At Relae, we deploy our in-house power system modeling framework to navigate this complexity. 

Our toolkit includes CD-PyPSA-USA, used for this analysis, which is built on the Python for Power System Analysis (PyPSA) platform. This grid model simulates how power networks operate and evolve over time by solving for the least-cost optimization of the entire power system. Critically, our model is customized to explicitly represent complex, real-world dynamics, including data center load flexibility, co-located generation, and various policy constraints.

For this analysis, we simulated ERCOT operations under a range of data center growth scenarios. We modeled loads from 5 GW up to 40 GW in 5 GW increments, pairing each with sufficient on-site gas generation to cover roughly 70% of data center energy needs–a conservative estimate on the approach developers are taking today. 

We ran each scenario under two conditions: no load flexibility (“flex00”, the baseline) and 25% emergency curtailment capability (“flex25”), consistent with solutions exhibited by Emerald AI. This approach allowed us to determine the system's response to step-changes in electricity demand.

What Our Analysis Reveals

Before doing the analysis, we expected to see load flexibility become more important as more data center load is added to the grid, preventing forced load shedding with high lost-load costs. But we didn’t know how large this effect would be, or what amount of new data center load would start to trigger it.

Our analysis found that without flexible load management, forced load shedding first appears at 30 GW, small in scale at first (4.4 GWh over 3 hours) but growing sharply as load increases. At 35 GW, we observe 50 GWh of shedding across 39 hours. At 40 GW, shedding reaches 158 GWh across 81 hours, equivalent to nearly three times ERCOT’s average hourly energy consumption, with an economic cost at VoLL of approximately $5.5 billion.¹

By enabling on-demand data center load flexibility, forced load shedding is eliminated in every scenario we tested. The same 40 GW case instead sees 165 GWh of controlled, short-duration curtailment spread across 86 hours, less than 1% of hours in a year. Further, the average demand response in these hours was less than 5% of the nameplate data center load, with the largest event reaching 14% of data center load. The grid stays balanced, consumers remain connected, and data centers deliver substantial value to the system via flexible loads.

Load Flexibility Is High Value and Presents an Opportunity for Storage

Our modeling puts a dollar figure on what flexible load is worth. At a VoLL of $35,000/MWh, each hour of demand response in the 40 GW scenario delivers approximately $64 million in avoided consumer welfare losses. Over a full year, that adds up to $5.5 billion, achieved through an average of just 5% demand response across the 86 hours of curtailment needed to eliminate all forced load shedding.

That value points directly to an opportunity for storage. In our model, demand response functions as a “ghost battery” at the data center’s grid node, absorbing grid stress exactly as a physical battery would, without any electrons needing to flow. That virtual battery can become a real one. On-site battery storage allows a data center to dispatch stored energy during grid stress events rather than curtailing workloads, maintaining compute continuity while relieving grid pressure.

The implications point in two directions. 

• For the hyperscaler or data center operator, physical storage converts a compliance obligation into an uptime guarantee: the curtailment event becomes a battery discharge, with negligible impact to the compute stack. 
• For the storage developer, co-location with large data center loads represents a high-value deployment opportunity with a clear commercial case. The avoided welfare costs per curtailment hour our model quantifies is the value a well-positioned battery asset, co-located at a data center node, can credibly claim to preserve.

The Data Centers of Tomorrow Need to be Grid Assets

The data center buildout underway is large enough to reshape grid reliability across entire regions, and the regulatory environment is beginning to reflect that scale. Texas SB6 is the most prescriptive example to date: it requires new large loads above 75 MW to install remote curtailment equipment operable during firm load shed events.

Our modeling quantifies what load flexibility is worth across this landscape: data centers with curtailment capability can provide significant value and avoid billions in consumer welfare losses annually. For developers and investors, designing that capability in from the start can convert a compliance requirement into a long-term grid asset.

The federal picture has moved in the same direction since this analysis was published. In May 2026, NERC issued a rare Level 3 Alert on computational loads; in June, FERC ordered six RTOs and ISOs to revise or justify their large-load interconnection rules, and in July, FERC directed NERC to develop mandatory computational-load reliability standards by the end of the year. We covered what that means for grid modeling and interconnection in Inside NERC’s Level 3 Alert on data center loads. Flexibility is no longer only a Texas statutory question; it is becoming part of the federal reliability framework.

Frequently Asked Questions

What is data center load flexibility, and how does it work?

Load flexibility is a data center’s ability to reduce the power it draws from the grid on short notice, during the small number of hours when the system is under stress. In practice, that means shifting or pausing deferrable compute, drawing on on-site batteries or generation, or pre-cooling the facility ahead of a peak. The point is not to consume less overall—it is to move a thin slice of demand out of the hours when the grid can least afford it.

What does Texas SB6 require of large data centers in ERCOT?

Texas Senate Bill 6, signed June 20, 2025, makes remote curtailment equipment a condition of interconnection for new loads of 75 MW or more in ERCOT, so utilities can disconnect them during declared firm load shed events. Separately, it creates a voluntary, competitively procured demand response program those same large loads can elect to join, with at least 24 hours’ notice. The mandatory piece is the disconnection capability; paid participation in demand response is a choice.

What does data center inflexibility cost? How much curtailment avoids it?

Relae’s ERCOT modeling found that without flexibility, forced load shedding first appears at 30 GW of data center load and reaches 158 GWh across 81 hours at 40 GW—roughly $5.5 billion a year in consumer welfare losses at ERCOT’s $35,000/MWh value of lost load. Enabling curtailment eliminated forced shedding in every scenario we tested, at an average of under 5% of data center demand across 86 hours, less than 1% of the year. Each hour of demand response in the 40 GW case is worth about $64 million in avoided losses.

Is data center load flexibility proven today, and how does it compare to on-site batteries?

Data center load flexibility is past proof of concept and into commercial deployment: Emerald AI cut power to a 256-GPU cluster by 25% for three hours during an Arizona grid stress event without degrading compute service quality; EPRI’s DCFlex initiative is building the protocols and contracts, and Google has signed 1 GW of data center demand response with US utilities. 

On-site batteries reach the same result from the other direction—instead of curtailing workloads, the facility discharges stored energy, which is why our modeling treats demand response as a “ghost battery” at the data center’s grid node. For operators who cannot pause compute, storage turns the same compliance obligation into an uptime guarantee.

Key Takeaways

• Accelerated phase-out schedules for key clean energy and decarbonization tax credits will shorten the runway for project development, which could stall or cancel projects.
• Urgency is paramount, and qualified projects should expedite construction and operational timelines to secure eligibility for existing credits.
• A more complicated policy landscape requires concerted effort to navigate, including with the support of policy professionals.

Reconciliation Rolls Back Much of the IRA

On Friday, July 4, 2025, the President signed a sweeping reconciliation bill, H.R. 1, that will add at least $3.3 trillion to the national debt and marks a pivotal, contentious moment for US clean energy policy. The law was enacted through the complex legislative process known as budget reconciliation, requiring only a simple majority of votes in the House and Senate. The new law substantially modifies or terminates many of the Inflation Reduction Act of 2022 (IRA)'s clean energy incentives and has extensive implications for the economic viability of American energy and manufacturing projects.  

In the Senate, three Republicans crossed party lines to vote against the bill, requiring Vice President JD Vance to break the tie. In the House, only two Republicans broke ranks to vote against final passage. While some of the more complex provisions of the bill, such as new foreign entity of concern (FEOC) restrictions, will require more time to fully assess, we've prepared a rapid run-down of key alterations to IRA incentives for carbon management, hydrogen, and clean fuel technologies.

What Is the 2025 Reconciliation Bill?

While the 2025 reconciliation bill is staggering in length, scope, and severity, containing provisions to cut Medicaid, reduce nutrition assistance, raise the debt limit, and cut taxes primarily for the wealthy, some of the most drastic sections of the bill modify tax incentives and other public funding for clean energy and emissions reductions.

‍

Many of the incentives to deploy clean energy that were created or enhanced under the IRA will be phased out early or repealed altogether. Credits with accelerated phase-out schedules include the newly created 45Y clean electricity production tax credit, which will no longer support wind or solar projects after 2027, and the 45V credit for clean hydrogen production for which projects must now commence construction before Jan 1, 2028 (moved up from Jan 1, 2033). 

Since the passage of the reconciliation package, there has been active litigation on several provisions, including an order from a federal district court to vacate IRS guidance that would have prohibited certain wind and solar projects from securing safe harbor.  The table below provides a detailed breakdown of key changes to major tax credits between the original IRA, the draft that moved through Committees in the House, and the final text that was passed by the Senate and signed into law.

Major Tax Credit Changes in the Reconciliation Law

‍

How FEOC Restrictions Threaten Clean Energy Supply Chains

Many clean energy tax credits include ambiguous language restricting projects connected to FEOC, complicating supply chains and creating new problems for developers of clean energy projects. The law also introduces a complex matrix of new definitions, such as "Prohibited Foreign Entities," which includes both "Specified Foreign Entities" and "Foreign-Influenced Entities."

The FEOC restrictions embedded in the reconciliation bill represent a seismic shift for clean energy developers. These new rules, designed to limit the influence of Covered Nations (China, Russia, North Korea, and Iran), will disqualify projects from receiving tax credits if they source components, minerals, or intellectual property from entities tied to these nations. In other instances, the partial ownership or investment of an entity with financial ties to a Prohibited Foreign Entity may also disqualify a project from qualifying for tax credits.

This FEOC language matters for developers and investors because of the resulting global supply chain disruptions, investment uncertainty, and compliance burdens. The clean energy sector is deeply reliant on global supply chains, especially for solar panels, batteries, and wind components, industries where China currently dominates. The IRA intended to counter this by moving the manufacturing and production of these supply chains to the US. Project developers must now thoroughly review their supply chains and capital providers, and may need to quickly pivot to compliant resources. 

In February 2026, the IRS released interim guidance on the FEOC provisions to provide safe harbor guidance for clean energy manufacturing, investment, and production credits to help taxpayers gauge whether material assistance was provided by a prohibited foreign entity. 

Other Major Rollbacks to the IRA

Beyond clean energy tax credits, the reconciliation package also repeals and rescinds many other IRA provisions. This includes a full rescission of all unobligated IRA appropriated balances at the Department of Energy's Loan Programs Office, and several other programs, including:

• The Tribal Energy Loan Guarantee Program
• Greenhouse Gas Reduction Fund
• Transmission Facility Financing

A complete list of rescissions of energy-related funding is outlined in Sections 60001-60024 and 50402 of the law. These rescissions represent tens of billions of dollars in lost climate investments made under the IRA, which would have provided funds to state, local, and Tribal governments, federal agencies, non-profits, and commercial project developers to reduce emissions and update critical infrastructure.

What Can Project Developers and Other Companies Do?

Developers will need to act quickly to meet updated commence construction and place into service requirements, though circumstances are technology specific (e.g., safe harbor updates to 48E and 45Y). Tax credits generally have advanced commence construction and operational deadlines, resulting in a strong first-movers advantage. Companies should also review their supply chains and revise equipment and material procurement sourcing plans as necessary to address restrictions presented in the reconciliation bill.

An executive order from President Donald Trump issued on July 7 will further complicate how companies proceed. In the EO, the President directs his administration to "strictly enforce the termination of […] 45Y and 48E […] for wind and solar facilities." The Administration will likely issue extremely strict interpretations of "commence construction" clauses and FEOC requirements in forthcoming tax credit guidance issued by the Treasury Department, though these moves are quite likely to face litigation.

The new restrictions being proposed by the Administration, including specific details on FEOC, qualified equipment, commence construction, and other reporting requirements, will require additional guidance from the IRS and provide an opportunity for engagement through public comment. It is important that impacted companies weigh in during these public comment periods, not only to help inform and influence the final rules issued by the Administration, but also to build an administrative record that could support litigation efforts to strike down the final rules.

Staying Ahead of Policy Changes

Given the rapidly shifting landscape of energy policy, it's paramount that companies stay abreast of the latest changes and dedicate resources to understanding how they may be affected. Policy professionals, including the experts at Relae (formerly Carbon Direct), can support organizations as they engage in the regulatory process, anticipate and prepare for new legislation, and navigate the requirements to access essential tax credits and incentives. Even under new constraints, expert guidance can help maximize impact and minimize disruption.

Frequently Asked Questions

How does the reconciliation bill change the timelines for major clean energy tax credits?
Most clean energy tax credits saw their windows shortened relative to the original IRA: 

• The 45Y and 48E credits now terminate entirely for wind and solar facilities placed in service after December 31, 2027, with a separate phase-down (75% in 2034, 50% in 2035, 0% after) for other technologies. 
• The 45V clean hydrogen credit's "commence construction" deadline moved from December 31, 2032 to December 31, 2027. 
• The 45Z clean fuel credit now ends on December 31, 2029 (versus 2027 in the original IRA, but bonuses for SAF have been removed and new emissions-calculation methods favor corn ethanol). 
• Notably, the 45Q carbon capture credit saw little change and retained transferability, with credit values for enhanced oil recovery and utilization raised to match secure geological storage.

What are the FEOC restrictions, and why do they matter so much for developers?
FEOC ("Foreign Entity of Concern") restrictions disqualify projects from tax credits if they source components, minerals, or intellectual property from entities tied to China, Russia, North Korea, or Iran. Even partial ownership or investment ties to a "Prohibited Foreign Entity" can trigger disqualification. The definitions are complex and still being clarified through IRS guidance, meaning developers need to review supply chains and capital providers carefully and may need to pivot to compliant sourcing.

What should project developers do now in response to these changes?
Developers should move quickly to meet the earlier "commence construction" and "placed in service" deadlines, since credits now benefit early actors. This includes reviewing and potentially restructuring supply chains and procurement plans to address FEOC restrictions, and closely monitoring forthcoming IRS/Treasury guidance. 

Key Takeaways

• AI's carbon footprint has two distinct parts: embodied emissions from building data centers and operational emissions from running them, both accelerating as global data center electricity use is set to double by 2030, and AI-focused use to triple.
• Managing that footprint will require deliberately steering technology architecture, power sourcing, and materials choices, instead of leaving them to react to demand after the fact.
• Eight concrete strategies, from smarter chip design to firm clean power and carbon removal, can cut AI's footprint today, without waiting on new regulation.
• US data centers used 4% of the USA's total electricity in 2024, and are projected to use as much as 15% by 2030.

Introduction

The rapid growth of artificial intelligence (AI), particularly large-language models (LLM) and generative AI, has taken many by surprise. This surge has led to escalating electricity demands at data centers and raised concerns about the strain on the power grid. It has also sparked the construction of new, larger data centers, resulting in growing embodied emissions tied to building and maintaining AI physical infrastructure.

Managing the risks of increased greenhouse gas (GHG) emissions from AI requires investment, expertise, and new approaches to building and operating many aspects of AI operation and supply chains. The immediate task is to understand these risks, gather the necessary information, and to avoid poor outcomes by proactively managing construction, operation, and emissions associated with the growth in AI. In parallel to that work, it's important to recognize that AI can itself be a real force to reduce emissions incrementally and dramatically across a wide range of sectors.

What Is the Carbon Footprint of AI?

The carbon footprint of AI consists of two main parts: "embodied" emissions that come from manufacturing IT equipment and constructing data centers, and "operational" emissions that come from electricity consumed by servers, memory and networking equipment as they perform AI-related calculations. Both of these aspects of emissions are growing as more data centers are built and existing data centers increase their share of power-hungry AI applications like generative LLM searches, AI agents, and AI image generation.

Understanding Electricity Demand for Data Centers

Today, the electricity demand from AI-specific applications is estimated to be less than 1% of global electricity use. To understand this number, it helps to start with the electricity consumed by the 12,000+ data centers worldwide, which was about 1.5% of global electricity consumption in 2024. (This excludes another 0.4% from cryptocurrency mining.) However, most of the computation at these data centers is not AI; instead, it's more conventional applications like e-commerce, video streaming, social media, and online gaming.

The amount of AI-based computation at data centers is hard to determine, but AI-dedicated accelerated servers consumed about one third of overall data center electricity in 2025, or roughly 0.5% of global electricity. Notably, this is projected to grow at 30% annually, much faster than conventional (non-AI) data center electricity use. However, that electricity use results in a relatively small share of greenhouse gas emissions: about 0.5% of global fuel combustion emissions, with AI data centers representing only a small portion of that value.

Still, the demand for AI applications is rapidly growing, and this is likely to drive up the electricity used by data centers and the associated greenhouse gas emissions. The most important implications of this trend are in the US, which hosts about half the world's data centers. Currently, data centers use about 4% of US electricity, but projections for the future range from a low of 9.5% to a high of 15.3% in 2030.

[cta]

How Electricity Sources Impact Data Center Emissions

A large increase in electricity use doesn't necessarily result in a similarly large increase in greenhouse gas emissions. Currently, a significant portion of the electricity powering data centers comes from zero-carbon sources such as wind and solar. This is partly because of large, corporate power-purchase agreements (PPAs) signed by leading data center operators, particularly Amazon, Meta and Google. 

US technology companies have been buying renewable energy for years. Global corporate clean energy procurement hit a record 62 GW in 2024, then fell to 55.9 GW in 2025, the first annual decline in nearly a decade, as elevated power prices and policy uncertainty made even large buyers more selective. Meta, Amazon, Google, and Microsoft still accounted for roughly 49% of global clean energy procurement in 2025, with Meta and Amazon alone securing 20.4 GW combined, including 4.7 GW of nuclear power.

The use of low-carbon power means that the net emissions from these data centers is smaller than the electricity consumption numbers might suggest. Of course, a crucial consideration is whether this low-carbon power is truly "additional," meaning that it is being added to the grid and not simply taken away from other uses. Data center operators are also expanding beyond their traditional wind and solar PPAs by exploring novel approaches to try to meet this standard, including geothermal projects in the US and Taiwan.

However, the projected electricity demand from AI applications at data centers will be difficult to meet entirely with low-carbon power, at least in the near term. Despite installing over 43.2 GW of wind, solar and battery projects in the US in 2025, these generators face a long wait for interconnection approval in many parts of the country. Geothermal and hydro power, which offer steady ("baseload") low-carbon electricity, remain constrained in the near term. And the interest in scaling up nuclear power, from restarting full-scale reactors to novel small modular reactors (SMRs), faces significant regulatory, cost, and supply chain hurdles.

One important source of low-carbon electricity that has not received enough attention is natural gas-fired power equipped with carbon capture and storage (CCS). This technology has the potential to significantly reduce emissions from existing power plants and enable new projects to achieve near-zero emissions.

The Role of Embodied Emissions in Data Center Construction

Embodied emissions include all emissions associated with the extraction, production, transportation, construction, and disposal of materials used in construction.

The embodied emissions from constructing data centers are substantial, and include concrete, steel, and IT hardware. Scope 3 GHG emissions for data centers—which include embodied emissions—range from approximately one-third to two-thirds of overall lifetime emissions. At Microsoft, Scope 3 emissions made up about 86% of the company's total FY2025 footprint and grew roughly 12% year over year, with capital goods driving most of that increase. In FY2024, capital goods alone accounted for about 41% of Microsoft's Scope 3 emissions, and purchased goods and services (including IT hardware) accounted for another 34%. In response, Microsoft has started using wood in some data center construction to reduce this impact. While using wood offers a partial solution, it cannot fully offset the emissions of even a single facility, and wood supply chains remain limited.

Major data center operators are working hard to address this challenge, including emphasizing the need for standardized emissions measurements and disclosures for key building materials. Ultimately, achieving deeper decarbonization will require further action to address both operational and embodied emissions.

Eight Strategies to Reduce the Carbon Footprint of AI

1. Adapt Technology Architecture

Efficiency is the foundational strategy in any clean energy approach. As such, chipmakers are developing ways to cut energy use from the outset, such as incorporating more memory directly onto computer chips or hard-wiring basic calculations. These innovations have already reduced energy consumption in new computer chips substantially, in some cases a 96% improvement. Examples of this include NVIDIA's Blackwell platform and the company's newer Rubin platform, launched in 2026, continues that trajectory. Likewise, servers are being designed with new architectures that minimize internal data transfers, delivering additional efficiencies. Even more efficiency gains may be possible with emerging technologies like photonic computing.

2. Optimize Training Geography

There are also significant opportunities to manage AI's energy use through time and space optimization. For example, a large portion of the energy consumption for LLMs occurs during the training phase, prior to a model's deployment for inference. Because these training tasks are not location-dependent, they can be carried out in regions with abundant, low-cost, low-carbon electricity, as part of broader efforts to dynamically move computing tasks to reduce emissions, known as carbon-aware computing. Additionally, server requests for generative AI tasks, like ChatGPT searches, can potentially be routed through systems powered by low-carbon electricity. Although this may add only a few milliseconds of latency, it could substantially reduce emissions from computing operations.

3. Select Appropriately-Sized Models

Not all generative AI tasks, like ChatGPT queries, are equal in terms of energy demand. Leading AI companies are increasingly focusing on using smaller, more efficient AI models to perform these tasks, achieving nearly equivalent quality for far less energy consumption. A notable recent test of that idea came in January 2025, when China's DeepSeek released a model with competitive performance that was trained using less powerful chips and far fewer computing hours than its established rivals. Similarly, many AI applications, such as digital twinning and satellite-based pattern recognition, consume far less electricity than generative LLMs, because of their specialized, relatively efficient models. This can even save energy compared to non-AI approaches: for example, some of the most advanced AI-driven weather prediction models require far less energy than traditional weather simulations, running on a laptop rather than a supercomputer.

4. Address Fugitive Methane Emissions

As data center operators increasingly plan on using natural gas for new electricity supply, reducing upstream emissions from gas production and transmission will be crucial. In the U.S., the Environmental Protection Agency (EPA) 2024 Methane Rule was designed to cut these non-carbon dioxide greenhouse gas emissions by approximately 80%. However, Congress repealed the rule's methane fee in 2025 and barred the EPA from collecting it until 2034. The EPA has since extended compliance deadlines and loosened flare and vent-gas requirements, with litigation over those changes still ongoing. Meanwhile, tools from companies like Kayrros and organizations like Carbon Mapper help detect methane leaks and attribute them to specific operators. The best actors in the industry emit minimal methane, less than 0.5% of what is produced. This standard is achievable for nearly all gas producers.

5. Use Carbon Capture on Power Plants

For both new and existing natural gas-fired power plants, carbon capture and storage technology offers the potential for generating firm, low-carbon power. While many plants currently in operation continue to emit unchecked, this doesn't have to be the case: their emissions can be captured and securely stored geologically. Hyperscalers and project developers should pursue new investments and business models for CCS to reduce existing emissions by 95% or more. For new generation projects, options like NetPower, Arbor, and CES will soon enable emissions abatement of 100%, or even more if combined with biopower to deliver carbon dioxide removal as well. Achieving this will require the development of carbon dioxide pipelines, barges, and storage facilities, which face their own challenges, such as permitting and community approval, that must be addressed directly.

6. Add More Zero-Carbon Power to the Grid

Roughly 8,200 solar, wind, and battery projects in the U.S. are seeking grid interconnection. By the end of 2025, the interconnection queue held roughly 2,060 GW of proposed generation and storage across thousands of projects, and its composition shifted meaningfully. Solar, wind, and storage volumes in the queue all declined year over year (although remained at high absolute levels) while natural gas capacity in the queue grew by 86%. Our blog post, The $5.5 Billion-Dollar Case for Enabling Data Center Load Flexibility, covers one way hyperscalers are working around the wait rather than simply enduring it. These delays need to be addressed, and permitting reform remains an unresolved, live debate. The Manchin-Barrasso bill, which once looked likely to pass, was tabled in December 2024 and never became law. As of 2026, no comprehensive federal permitting law has replaced it. One potential innovation is to use AI to accelerate the development of power flow models and streamline the paperwork required to complete the regulatory process.

7. Invest in Low-Carbon Building Materials

While wood is a promising low-carbon building material, we'll also need glass, concrete, steel, aluminum, and computer chips with minimal embodied carbon emissions. Hyperscalers currently face significant challenges accessing low-carbon versions of these materials, which will eventually be produced using low-carbon hydrogen, carbon capture and storage, and low-carbon electricity. However, these systems require significant investment, workforce development, and permitting to be built. Without these advancements, the embodied emissions from data centers will increase rapidly and significantly in the US, Europe, and globally.

8. Increase Carbon Dioxide Removals

It's already clear that AI applications at data centers will generate emissions from electricity use and embodied carbon that cannot be avoided in the near term. Estimates of current greenhouse gas emissions exceed 300 million tons per year and are likely to grow this decade. These emissions should be measured using full life-cycle analysis and then offset through high-quality carbon removal projects, preferably those with high durability.

To effectively reduce the environmental impact of AI, all eight strategies discussed must prioritize the communities most affected: frontline communities near new infrastructure, consumers facing price increases, and tribal authorities with limited legal protections. Our own research on community opposition to AI data centers found that transparency, not cost or environmental impact alone, is the dominant driver of pushback across 46 stalled or blocked projects. We explore this concern further in our blog, Who Pays for the AI? The Hidden Costs of Rising Data Center Demand, including how ratepayers, not just data center operators, often absorb the cost of new grid infrastructure. Planning should begin by understanding the needs of these communities, ensuring that efforts focus on minimizing harm while maximizing benefits. Equity and justice must be embedded in every stage of planning, production, and permitting across all strategies.

AI's Power Demand Is Indicative of Broader Electricity Demand

AI is just one part of a broader trend of rapidly growing electricity demands, including from electric vehicles, heat pumps, industrial electrification, green hydrogen, and various e-fuels. The challenges AI presents to hyperscalers, communities, regulators, and investors serve as a preview of the complex, far-reaching impacts emerging in other sectors. The same questions keep recurring. Who secures reliable, affordable power fast enough? Who ends up carrying the cost and emissions burden of getting there the wrong way?

Managing AI's power demand will require building the technology architecture, clean firm power supply, and materials strategy to meet that demand deliberately, rather than reactively. AI's carbon footprint underscores the critical need for expertise in clean electricity, grid management, decarbonization, and carbon removal—expertise that will become increasingly vital as more companies realize the complexity and cost of the journey ahead.

Fortunately, AI itself can be part of the solution. With applications in grid management, material science, and advanced manufacturing, AI has the potential to play a powerful role in the climate response.

Read the full 2025 report: ICEF Sustainable Data Centers.

Frequently Asked Questions

How much electricity do AI data centers actually use? 

AI-specific computation likely accounts for around 0.04% of global electricity use today, but data centers overall (most of it non-AI computation) used about 1.5% of global electricity in 2024. In the US, which hosts roughly half the world's data centers, Lawrence Berkeley National Laboratory puts current usage at 4% of US electricity, projected to reach 9.5-15.3% by 2030 as AI-specific demand grows.

Will more efficient AI models like DeepSeek reduce data center energy demand? 

Not necessarily. DeepSeek's 2025 debut showed that competitive models can be trained with less powerful chips and fewer computing hours, but whether that translates into lower total energy demand is contested. Historically, efficiency gains in computing have tended to get absorbed by increased usage rather than reducing total consumption, so the honest answer is that it depends on whether demand growth outpaces the efficiency gained.

What is being done about the embodied emissions from building AI data centers?

‍Embodied emissions, from concrete, steel, and IT hardware, can account for one-third to two-thirds of a data center's lifetime emissions. Strategies include using lower-carbon materials like wood where feasible, developing low-carbon concrete, steel, and chips, and standardizing emissions disclosures for building materials so operators can compare and choose lower-footprint options.

Key Takeaways

• Agentic inference has changed the economics of AI. Tokens are becoming units of work and the economic driver is now the work produced, not token generation. Per-token costs are falling and the willingness to pay for work produced is rising; these two trends compound. This tailwind enhances AI economics and has spillover impacts on all layers of the AI stack.
• The AI infrastructure question has shifted from whether demand will show up to whether the physical stack can scale quickly enough. That stack includes power generation, grid capacity, interconnection, compute, memory, networking, cooling, siting, and community acceptance.
• Carbon Direct Capital and Relae (formerly Carbon Direct Inc.) have a differentiated view because the two entities work across both sides of the constraint: Relae advises hyperscalers and energy buyers on power and grid bottlenecks, while Carbon Direct Capital invests in the technologies that relieve those bottlenecks.
• Carbon Direct Capital sees better risk-adjusted returns investing in the physical foundations of AI, including clean firm power, energy system efficiency, data center efficiency, and inference-optimized compute, rather than chasing late-stage AI application valuations.

A Better Question Than "Is AI a Bubble?"

The most important development in AI economics is agentic AI turning tokens into work, a shift that reframes the bubble debate which dominated investor conversations, sell-side notes, and Chief Information Officer surveys through early 2026. Hyperscalers spent approximately US$380 billion on capital expenditure (capex) in 2025 and have guided to approximately US$720 billion of capex in 2026.¹ Carbon Direct Capital and Relae have worked together to build project-level models for both training and inference facilities to demystify the numbers and understand financial and technical sensitivities. The core finding was that the assets could be bankable using standard assumptions and that the binding constraints were physical, not financial. That conclusion has been reinforced in recent months by new developments.

Concretely, AI is moving from single prompts and answers to multi-step workflows that plan, reason, call tools, verify outputs, and keep state. This shift to inference is the structural successor to training in the initial AI capex cycle; it changes power requirements, time to power, and compute architectures all at once. Goldman Sachs estimates that agentic AI could drive a 24-fold increase, relative to a 2026 baseline, to roughly 120 quadrillion tokens per month globally by 2030 as per-token costs continue to fall. SemiAnalysis makes the same point from another angle: the value of frontier tokens has risen as agentic workflows become useful, while hardware and software improvements have reduced the cost of producing each token.

This does not mean every AI company is attractive, every data center project works, or every valuation is justified. It means the easy bubble framing is missing the more investable question. If token demand is compounding and the unit value of work produced is rising, the scarce resource is not abstract enthusiasm. It is the physical infrastructure required to turn that demand into work produced.

The Data Center Model Still Matters, But the Box is not a Black Box

Our internal modeling for an illustrative 167-megawatt inference data center using Nvidia Blackwell graphics processing unit (GPU) servers suggests the potential for high-teens percent equity returns under a defined set of assumptions.² We built a bottom-up underwriting, beginning with the number of users served per inference data center, assuming approximately how many tokens they will demand daily, and translating that token demand into compute needed based on industry-standard quantization and utilization rates. We then inferred the number of GPUs and servers needed to achieve the desired compute, which ultimately drove the total invested capital and power demand based on assumed thermal power designs and power usage effectiveness (PUE). On the revenue side, we used GPU-as-a-service rental rates as one proxy for the market value of compute capacity. A hyperscaler would not rent scarce compute externally if it had higher-value internal demand for that same capacity. As we will detail below, GPU rental prices have been steadily increasing on the back of inflecting inference demand.

This model is not the entire argument; it is the starting point. An important lesson is that power cost alone does not break data center economics. Electricity is slightly over 10% of total costs in our model: a 50% increase in power price reduces equity-level returns by less than 2%. Access to power, speed of interconnection, and equipment availability matter more. In other words, the economics of the model facility are workable, but only if the facility can be built and powered on the timeline customers need.

That is where most AI commentary remains too superficial. It treats the data center as a black box: capex goes in, tokens come out. That misses the bottlenecks inside and around the box. AI racks are moving far beyond traditional cloud power density. Cooling is shifting from air to liquid and two-phase systems. Networking and high-bandwidth memory become binding constraints in inference architectures. Grid interconnection queue wait times stretch to years. Communities can and do block projects. The technical, physical, and political constraints are increasingly the drivers of potential returns.

Inference Makes the Constraint Structural

While training is episodic, inference is recurring. A training run can be delayed, accelerated, or redesigned. Inference happens every time a user asks a question, a developer runs an agent, a business automates a workflow, or an application calls a model in the background. Agentic inference multiplies that load because one user action can become many model calls, validation loops, and memory reads; industry benchmarks show that agentic systems consume 5–30 times more tokens than a standard chat interaction.

Inference demand is also resilient in both directions. If efficiency gains lower the cost per token, more workflows become economic and total token consumption rises - the classic Jevons Paradox. However, token prices do not necessarily need to fall for inference spend to grow. As the economic unit shifts from tokens generated to work produced, customers may pay more per token when an agent delivers work produced that is worth more than the inference cost. Regardless of token price, tokens must all route through the same physical bottlenecks and we are seeing an increase in inference demand.

The architecture of inference is also changing. Some workloads will prioritize low-latency answers. Others, especially agentic work without a human waiting on every token, will prioritize memory, state, context, and cost per completed task. That means the AI infrastructure stack will become more heterogeneous, not less: XPUs (specialized AI accelerator chips), custom silicon, photonics, memory hierarchies, and edge or regional deployment models will all matter. The pricing data shows demand for more AI infrastructure overall: on-demand GPU rental capacity is effectively sold out across all chip generations in early 2026, with one-year Hopper H100 contract pricing rising 15–20% month-on-month through March 2026 and Blackwell B200 rental rates up 23% in March alone. When rental rates rise into a wave of new chip supply, supply is not catching up to demand.

Power Is Not One Constraint, It Is Several

Saying "AI is power constrained" is true, but not specific enough. The real problem has several layers. First, data centers need more electricity than many local grids can deliver on hyperscalers' timelines. Crucially, some grids can supply sufficient power but not continuously for 8,760 hours per year, conflicting with traditional assumptions about service reliability and leading to novel strategies around flexibility and intermittent self-supply. Second, the grid must be able to absorb large, fast-moving computational loads without creating reliability risks. Third, customers need energy procurement strategies that satisfy cost, reliability, climate, and public-acceptance requirements. Fourth, projects must get built in real communities, through real interconnection processes and real permitting fights. Power is not simply a commodity to purchase. It is an infrastructure development problem.

This is where Relae is directly relevant. Relae has assembled a team of scientific, engineering, and market experts to support a paying power and energy advisory practice serving hyperscalers, energy buyers, and power producers. Its work answers the questions customers are asking before the market prices them: how to get more capacity out of existing physical grid infrastructure; how to assess the costs and value of load flexibility through advanced modeling capabilities; how to make clean firm generation bankable; how to reduce data center energy intensity; how to validate "bring your own power" and "bring your own compute" structures; and how to build projects that communities will accept. 

In the last twelve months alone, Relae has supported hyperscalers on bankability assessments for next-generation geothermal, scoped load-flexibility programs for multi-hundred-megawatt, single-customer sites, and modeled the carbon and reliability profile of "bring your own power" configurations against grid-tied baselines. Carbon Direct Capital leverages our network of technical experts at Relae, including power engineers, geologists, and electrochemists, to conduct credible technical diligence and to gain insights into early stage market trends and emerging preferences.

What Carbon Direct Capital Is Investing Behind

Our investment focus follows the bottlenecks. On the power side, we are investing in technologies that can deliver reliable power on AI timelines. Sage Geosystems is a next-generation geothermal platform with hyperscaler buy-in; Carbon Direct Capital co-led its US$97 million Series B with Ormat Technologies. We could not have made this investment without the deep expertise of the Relae research team which analyzed Sage's technical results to date to help underwrite future project feasibility. ION Clean Energy is a company that retrofits carbon capture technology onto natural gas combined cycle plants to create "blue electrons"; Relae is in active dialogue with multiple large power users on this topic. Carbon Direct Capital is also actively evaluating the enabling picks and shovels around geothermal, nuclear, fuel cells, and more.

On the data center efficiency side, we are investing in technologies that reduce the amount of power required for a unit of AI work. While it is encouraging to see incremental annual gains in chip efficiency, these are scaling far more slowly than compute demand, driving the need for more innovative technological solutions. As one example, a team at Relae helped us understand the fundamental energy consumption requirements of a standard complementary metal-oxide-semiconductor (CMOS) chip, and the potential of all-optical computing as an alternative. This led to Carbon Direct Capital investing in Neurophos, a photonic compute company targeting step-function gains in energy efficiency per chip that are beyond those achievable by existing GPUs. Carbon Direct Capital joined the company's US$110 million Series A alongside Gates Frontier, Microsoft's M12, Aramco Ventures, Bosch Ventures, and others. More broadly, we are studying other layers of the data center technology stack including networking, memory, cooling, and inference-optimized architectures because the next phase of AI infrastructure will not be solved by simply buying more of yesterday's hardware.

The Bear Case Deserves to Be Taken Seriously

There are real risks to the AI boom: Hyperscaler free cash flow can compress if capex grows faster than revenue. Model efficiency gains can reduce the amount of compute required for a given task. Training demand may be more episodic than the market assumes. Local opposition can slow or cancel data center and power projects. Some new data center capacity could become expensive cloud infrastructure competing on price if AI revenue disappoints.

Those risks are why Carbon Direct Capital frames this as an investment in constraints, not in AI enthusiasm. If efficiency improves, inference use cases expand and the bottleneck shifts to deployment, memory, power, and cost per unit of work produced. If training demand slows, inference and enterprise agents still require recurring capacity. If local grids cannot absorb load, technologies that unlock power, reduce energy intensity, or improve flexibility become more valuable. If some AI applications or model developers fail, the upstream physical bottlenecks remain for the rest.

The Investment Conclusion

The AI infrastructure opportunity sits at the intersection of frontier technology risk, project-finance economics, and energy-system engineering. Underwriting this opportunity well requires addressing all three at once; Carbon Direct Capital is built to do just that. The technical team at Relae has a pulse on emerging stakeholder preferences and scientific breakthroughs, understands novel technologies deeply, and is highly experienced in conducting detailed technical diligence to ensure that projects are viable and scalable. Carbon Direct Capital combines these market and technical insights with our commercial underwriting to facilitate new investments. We are not picking AI winners. We are not picking pure energy assets. We are investing in the companies and technologies that have to exist for AI to sustainably scale.

Frequently Asked Questions

Is the AI capex boom a bubble? While valuations vary, token demand and physical infrastructure needs are real and compounding. The correct question to ask is not whether AI is a bubble, but what the binding constraints are. Our modeling shows that constraints are physical, not financial. 

What are the real constraints on AI infrastructure growth right now? AI infrastructure growth is constrained by power availability, grid capacity, and interconnection speed, not capital availability.

Why does inference matter more than training for long-term AI power demand? Inference is recurring and grows with AI usage, it is not episodic like training runs.

What is Carbon Direct Capital investing in, and why? We are investing in clean firm power, energy system efficiency, data center efficiency, and inference-optimized compute—the physical bottlenecks rather than application-layer valuations.

Disclaimer

Carbon Direct Capital Management LLC is an investment adviser registered with the US Securities and Exchange Commission (SEC). Registration as an investment adviser does not imply any particular level of skill or training. Additional information about Carbon Direct Capital Management LLC, including our Form ADV Part 2A Brochure, is available on the SEC's website at adviserinfo.sec.gov.

This content is provided for informational purposes only and should not be construed as or relied upon as investment, legal, tax, or other advice. You should consult your own advisers regarding legal, business, tax, and other matters related to any investment. Any projections, estimates, forecasts, targets, prospects, or opinions expressed are subject to change without notice and may differ from opinions expressed by other employees of Carbon Direct Capital Management LLC, its affiliates, investors, portfolio companies and other individuals, groups or entities. Certain information contained herein may have been obtained from third-party sources believed to be reliable; however, Carbon Direct Capital Management LLC makes no representations about the accuracy or completeness of any such information or its appropriateness for any given situation. Any investments or portfolio companies mentioned are not representative of all investments made by funds managed by Carbon Direct Capital Management LLC, and there can be no assurance that any investment will be profitable or that future investments will have similar characteristics or results. Past performance is not indicative of future results. The content speaks only as of the date indicated. This content does not constitute an offer to sell or a solicitation of an offer to buy any security. Any such offering will be made only pursuant to formal offering documents.