Speed to power: Why behind-the-meter power generation provides competitive edge in the AI era

Reading Time: 11 minutes

Power is becoming one of the defining constraints on AI growth.

For the past several years, enterprise AI progress has been shaped—and limited—by familiar infrastructure pressures: GPU availability, memory, storage, networking connectivity, and the ability to design environments capable of supporting data-intensive workloads. Given supply constraints, data centers haven’t been able to buy their way out of these limitations.

Those constraints have hardly disappeared. But as organizations move from AI experimentation to production, another limitation is moving up the priority list: access to reliable, scalable power.

AI-ready data centers demand far more electricity than traditional enterprise environments, especially when they support high-density GPU clusters for model training, fine-tuning, or large-scale inference. The challenge is not simply that AI uses more energy. It is that demand is growing quickly, concentrating in specific regions, and arriving faster than grid planning and interconnection processes can always support.

According to EPRI’s “Powering Intelligence 2026,” data centers could consume 9% to 17% of U.S. electricity by 2030, up from 4% to 5% today. Deloitte also reports that 72% of surveyed power and data center executives view power and grid capacity constraints as very or extremely challenging for data center infrastructure buildout.

For organizations trying to scale AI, this changes the infrastructure conversation. Power is no longer a downstream facilities concern. It now affects site selection, deployment timelines, AI architecture, workload placement, cooling strategy, and the pace at which organizations can turn AI investment into operational value.

In the AI era, power readiness is becoming a prerequisite for AI readiness.

Data center executives view power and grid capacity constraints as extremely challenging for data center infrastructure buildout.

Power is the new AI deployment constraint

The grid has long provided reliable, cost-effective power to data centers. Most data centers still rely primarily on grid power, supported by on-site backup systems for resilience. But AI is changing both the density and urgency of data center energy demand.

And AI-focused data centers consumed approximately 155 TWh in 2025, representing a 50% single-year increase and roughly 32% of total data center electricity consumption.

EPRI notes in its “Speed to Power Data Centers” executive summary that AI processors can have power densities up to 10 times higher per square foot than traditional storage-focused facilities. That demand has exposed a mismatch between data center development timelines and grid expansion timelines, making speed to power a critical factor in site selection.

That mismatch means data centers can often be built faster than utilities can complete the grid upgrades, substation work, transmission expansion, or interconnection processes needed to serve them. Deloitte notes that some grid connection requests now face waits of up to seven years.

For AI leaders, that delay becomes an IT and a business problem. An organization may have the business case, hardware roadmap, executive mandate, and budget to scale AI but if it cannot secure enough reliable power, the project can stall before it reaches production.

That is why “speed to power” has become a strategic metric for AI infrastructure. Enterprises need more ready access to power instead of waiting years for grid capacity to come online.

“Speed to power” has become a strategic metric for AI infrastructure.

What is speed to power?

Speed to power refers to how quickly an organization can secure, deliver, and manage the energy capacity required to support its data center workloads.

For AI environments, this includes more than access to megawatts. It also includes whether the facility can support high-density racks, advanced cooling, resilient power delivery, utility interconnection, and long-term growth.

The central question is no longer simply, “Can we get enough compute?” It is, “Can we power the infrastructure needed to ensure that compute provides return on investment?”

“The challenge is no longer simply generating more electricity,” wrote Susan C. Mcleod, vice president of data center market development at Hitachi in a blog on speed to power. “It’s delivering power quickly enough to keep pace with growth. This is why speed-to-power has become a strategic priority across the data center industry.”

This is where behind-the-meter (BTM) power is gaining momentum.

The challenge is no longer simply generating more electricity. It is delivering power quickly enough to keep pace with growth.–Susan C. Mcleod, VP of data center market development, Hitachi

What is behind-the-meter power?

Behind-the-meter power refers to electricity generated on the customer’s side of the utility meter, often at or near the data center itself.

Instead of relying exclusively on power delivered through the public grid, data center operators can supplement grid supply with dedicated generation sources such as natural gas, fuel cells, battery storage, solar, microgrids, or hybrid architectures that combine several sources.

This does not mean data centers will abandon the grid. In many cases, the strongest model is hybrid: use grid power where capacity is available, add behind-the-meter generation where it improves speed and resilience, and design the environment so power, cooling, and compute can scale together.

EPRI’s “Speed to Power Data Centers” framework describes several emerging strategies, including grid-connected flexible models, bridge-to-grid approaches that use temporary or permanent generation while awaiting full grid access, and islanded models that operate off-grid permanently.

For organizations pursuing AI at scale, the value is clear. Behind-the-meter power can help accelerate deployment, reduce exposure to grid delays, improve resilience, and give data center operators more control over infrastructure buildout timelines.

It is not a shortcut around planning. It is a practical response to a market where AI infrastructure demand is moving faster than traditional power delivery models can always accommodate.

 

How behind-the-meter power reduces grid pressure

BTM power can reduce pressure on the grid, but the impact depends on how the system is designed.

If a data center can generate a meaningful share of its own electricity, it draws less power from the public grid during normal operations. That can reduce the immediate burden on local utility infrastructure, especially in constrained regions where new data center demand would otherwise require major upgrades.

Hybrid models can also help. A facility might use grid power for part of its load, behind-the-meter generation for additional capacity, and battery storage or microgrid controls to manage resilience and peak demand. This can make the site less dependent on utility timelines while still preserving the benefits of grid connection.

In practice, BTM power is a way to reduce grid dependence, not eliminate the grid altogether. Most organizations will still need to account for backup power, interconnection requirements, permitting, emissions, fuel supply, and long-term growth.

But that does not diminish its value. For AI deployments where waiting years for utility capacity is not commercially viable, behind-the-meter power can be the difference between moving forward and staying stuck in planning.

Why behind-the-meter execution is important

Public perception of data center energy projects depends heavily on execution.

Communities are more likely to support new infrastructure when they can see clear economic value, responsible planning, and credible efforts to reduce strain on local systems. Behind-the-meter power can support that case when it helps avoid placing all new demand on the public grid, improves resilience, or enables investment in local infrastructure.

Concerns arise when projects lack transparency or when local communities do not understand the power strategy, environmental controls, or long-term benefits. Natural gas generation, for example, can provide fast, reliable, dispatchable power, but organizations still need to address questions surrounding emissions, air quality, cost and supply,  and long-term energy strategy. Renewable and storage-heavy approaches may align more closely with sustainability goals, but they may not provide sufficient 24/7 power alone for large AI workloads.

The answer is not to treat BTM power as inherently good or bad. The answer is to choose the right mix of power sources for the workload, location, timeline, and business objective — then communicate that strategy clearly.

When executed well, behind-the-meter power strengthens the case for AI infrastructure because it shows that the organization is investing in the capacity needed to support growth rather than simply adding unmanaged demand to an already constrained grid.

Kinds of behind-the-meter power generation

AI training and inference environments require large amounts of reliable, always-available power, often measured in tens or hundreds of megawatts. At the same time, utilities in many regions face grid constraints, long interconnection queues, and growing competition for available capacity.

That is changing how organizations evaluate power. Cost still counts, but it is no longer the only factor. Speed-to-power, reliability, scalability, and AI suitability are now central to infrastructure planning (see Table 1).

In the near term, natural gas and hybrid architectures are among the most deployable options. Fuel cells and microgrids offer strong resiliency advantages. Solar plus storage can support sustainability goals as part of a broader energy mix. Nuclear co-location and small modular reactors represent longer-term pathways for large-scale, low-carbon AI infrastructure.

The right answer depends on timing, location, workload requirements, risk tolerance, and long-term growth plans.

Table 1: Energy delivery sources rated by speed-to-power and AI suitability

Energy delivery approach	Speed-to-power	Reliability	Scalability	AI suitability
Natural gas turbines and reciprocating engines	High: One of the fastest mature options available today. Many data center operators are pursuing gas because grid interconnection timelines can stretch for years.	High: Provides firm, dispatchable power independent of weather conditions.	High: Can support deployments ranging from tens of MW to gigawatt-scale campuses.	Excellent: Currently one of the most practical solutions for large AI training and inference environments where uninterrupted power is critical.
Fuel cells	Medium to high: Faster to deploy than major utility upgrades and can be added modularly.	High: Valued for continuous, resilient operation and high availability.	Medium: Effective for campus-scale deployments but less proven at hyperscale AI-factory levels.	Very good: Strong fit for organizations prioritizing resiliency, efficiency, and lower-emission operations.
Solar plus battery energy storage	Medium: Can often be deployed faster than new transmission infrastructure but still requires site development and permitting.	Medium: Reliability depends on weather conditions, battery duration, and supporting power sources.	Medium: Scales well in regions with available land but becomes more challenging for dense metropolitan deployments.	Moderate: Well suited as part of a hybrid architecture but generally insufficient alone for large, always-on AI workloads.
Nuclear co-location	Low: Limited by permitting, site availability, regulatory considerations, and project timelines.	Excellent: Provides stable baseload power with very high availability.	Excellent: Can support some of the largest AI campuses being planned.	Excellent: One of the strongest long-term matches for high-density AI infrastructure because of its scale, reliability, and low-carbon profile.
Small modular reactors	Low today: Commercial deployment remains limited and timelines are uncertain.	Potentially excellent: Intended to provide reliable baseload power similar to traditional nuclear.	Potentially excellent: Designed to scale through modular deployment.	Potentially excellent: Often cited as a future solution for multi-gigawatt AI demand, but commercial maturity remains the key challenge.
Hybrid grid plus behind-the-meter architecture	High: Allows organizations to bring additional capacity online while waiting for full utility connections or grid upgrades.	High: Diversifies power sources and reduces dependence on a single point of failure.	High: Supports phased expansion as AI capacity grows.	Excellent: Increasingly viewed as the most practical near-term architecture for balancing deployment speed, resilience, and long-term growth.
Microgrids /bring-your-own-power models	High: Designed to accelerate deployment in power-constrained regions.	High: Can continue operating during grid disruptions.	Medium to high: Scalability depends on the generation mix supporting the microgrid.	Very good: Particularly attractive where utility constraints would otherwise delay AI infrastructure deployment.

Speed to power creates competitive advantage

Power is no longer a secondary utility decision. For AI infrastructure, it now helps determine which projects move forward, where they are built, and how quickly they reach production.

Organizations that secure reliable power faster can accelerate AI initiatives, bring high-density infrastructure online sooner, and avoid delays caused by grid constraints or long interconnection timelines. Those that wait too long may find their AI roadmap limited not by compute strategy, data readiness, or executive ambition, but simply by access to electricity.

That does not mean every organization should pursue the same power model. The fastest option may not always be the right long-term fit. Natural gas, fuel cells, renewables, storage, microgrids, nuclear co-location, and hybrid architecture each solve different parts of the problem.

The opportunity is to design a power strategy that supports near-term deployment and long-term growth.

For many organizations, that will mean using BTM power to supplement grid capacity, improve resilience, and create a more predictable path to AI scale. In that sense, behind-the-meter power is not simply an energy workaround. It is becoming an AI infrastructure strategy.

Behind-the-meter power is not simply an energy workaround. It is becoming an AI infrastructure strategy.

How SHI can help organizations achieve speed to power

As power availability becomes a critical determinant of AI success, organizations need to understand whether their infrastructure, facilities, power and cooling setups, and long-term capacity plans can support future AI growth.

SHI helps organizations align AI infrastructure strategy with real-world deployment requirements. From data center assessments and high-density infrastructure design to AI-ready architecture planning and deployment support, SHI helps identify constraints before they become roadblocks.

SHI also works with industry partners, including EPRI, to better understand how AI-driven energy demand is changing infrastructure planning. SHI’s participation in EPRI’s SAFERai.power initiative reflects the growing convergence of AI strategy, data center readiness, and energy planning.

As the AI bottleneck shifts from compute availability to power availability, organizations need to plan for both. SHI helps customers build resilient, scalable AI environments designed for the next wave of growth.

AI readiness now depends on energy readiness

The next phase of AI adoption will not be limited only by model performance, data quality, or access to GPUs. It will also depend on whether organizations can secure the power, cooling, and infrastructure needed to run AI reliably at scale.

Behind-the-meter power gives organizations more options. It can reduce dependence on constrained grids, improve resilience, and help data center projects move forward when utility timelines do not match business demand.

The organizations best positioned for the AI era will not simply be those that buy the most compute. They will be the ones that understand the infrastructure dependencies that determine whether AI can scale.

In the years ahead, energy readiness will become one of the clearest measures of AI readiness.

NEXT STEPS:

Why AI demands a new kind of data center infrastructure 

Learn more about SHI data center facilities management

Speak to an SHI expert