Google, Nvidia, and Anthropic Form a 100-Gigawatt Alliance to Untangle AI's Grid Problem

The constraint isn't compute anymore. It's kilowatts—and who gets access to them, and when. On September 28, 2026, Google, Nvidia, and Anthropic formalized that reality by joining grid-software company Emerald AI to launch the AI Energy Management Alliance. The target: 100 gigawatts of capacity for new AI-computing facilities. That number alone signals how seriously the industry's largest players now treat power as a first-order infrastructure problem, not an afterthought for facilities teams.

What the Alliance Actually Does

The alliance's core mandate is coordination—specifically, aligning AI data-center demand with electrical-grid operations in real time. Emerald AI, the organizing entity, supplies the software layer for managing flexible data-center loads. The operative word is flexible. Traditional data centers run at relatively flat, predictable power draws. AI workloads—training runs, large-scale inference clusters—are anything but. They spike, they pause, they scale unpredictably. Grid operators struggle to accommodate that profile without costly overprovisioning or reliability risks.

Emerald AI's software is designed to make data-center demand legible and responsive to grid conditions. The alliance gives that software a mandate to operate at a scale that no single company could negotiate with utilities alone. By pooling Google, Nvidia, and Anthropic under one coordination framework, the initiative creates negotiating weight and, more importantly, a shared technical standard for how AI facilities communicate their load profiles to grid operators.

The Constraint Driving the Urgency

The formation of this alliance isn't strategic positioning for a future problem. AI infrastructure expansion is already running into hard power and grid-connection limits. New data-center projects—particularly those sized for frontier model training—require grid interconnection agreements that can take years to finalize. Utilities in high-demand regions are managing queues measured in gigawatts of pending requests. In some markets, power availability has become the primary site-selection criterion, outranking land cost, fiber access, and even tax incentives.

The 100-gigawatt target the alliance is chasing isn't a forecast—it's the scale at which coordinated demand management starts to matter to grid operators. Individual company commitments are negotiated bilaterally and produce inconsistent outcomes. An industry-wide alliance targeting 100 gigawatts of AI-computing capacity gives utilities and grid operators a single interface for managing what will otherwise be a fragmented and destabilizing demand surge.

For Nvidia specifically, the stakes extend beyond its own facilities. Nvidia sells the hardware that goes into these data centers—GPUs that sit idle whenever a facility can't get the power it needs. Grid friction is a direct revenue constraint for a company whose growth depends on customers being able to deploy at scale. For Google and Anthropic, operational AI infrastructure is core product infrastructure. A training cluster that can't connect to the grid on schedule isn't a planning failure; it's a competitive setback.

Why Software Is the Lever

The instinct when facing a physical infrastructure shortage is to build more of the physical thing—more transmission lines, more generation capacity, more substations. That's happening. But it takes a decade and hundreds of billions of dollars. Software-defined demand flexibility is faster and cheaper, and it compounds across participants.

If a data center can signal to the grid that it will curtail a percentage of its load during a stress event—and do so reliably, automatically, and in coordination with other facilities—it earns grid operators' trust and, in many regulatory frameworks, financial compensation. Emerald AI's position in the alliance is to make that coordination technically feasible at the scale of 100 gigawatts. The software manages flexible loads: it decides, in real time, which workloads can be deferred, throttled, or shifted to off-peak windows without breaking SLAs.

This is demand response, a concept utilities have used with industrial customers for decades. What's new is the scale, the speed of the load changes involved in AI workloads, and the sophistication required to manage them without disrupting the underlying compute jobs.

The Bigger Shift

The AI Energy Management Alliance is evidence of an industry crossing a threshold. For years, the dominant narrative around AI infrastructure was about compute—chip architectures, cluster sizes, interconnect bandwidth. Power was a cost line, not a constraint. That framing is over. The companies building and supplying AI infrastructure have accepted that grid access is now a gating factor, and they've responded by building an institution to manage it collectively.

The move to 100 gigawatts of coordinated AI-computing capacity—anchored by the industry's most consequential labs and hardware supplier, organized around a software layer purpose-built for grid flexibility—marks the moment AI infrastructure strategy became, unavoidably, energy policy.