Tech Giants and Startups Launch Dueling Fixes for America's AI Power and Chip Crunch

AI companies can't get enough computing power, and they can't get it connected to the grid fast enough either. Two industry coalitions launched in recent weeks to attack both halves of that problem, and neither one involves Washington writing a check.

The first is a compute-sharing effort reported by Axios and led by Anjney Midha, which aims to pool idle AI chips scattered across labs, cloud providers and data centers into a single shared network. The consortium says single-customer data centers use less than 15% of their computing capacity on average outside of active training runs. The rest sits idle while smaller players go begging.

"We have a lot more compute capacity than people think," Midha told Axios. "We just need to connect and coordinate it."

The network has lined up roughly 760 megawatts of connected or potentially available capacity, with a goal of hitting 2 gigawatts by 2030, according to Axios. A shared scheduling system will let researchers, startups, universities, public-sector employees and national labs see what's available, where, and at what cost.

Sam Singh, head of AI at robotics firm 1X, told Axios that smaller companies simply can't compete with OpenAI and Anthropic for compute. Those giants can offer more money and lock in long-term contracts. Pooling idle capacity is meant to give the little guys a shot.

This effort is distinct from the AI Infrastructure Coalition, a separate group formed November 19, 2025, with members including Google, Microsoft and Andreessen Horowitz. That coalition does policy lobbying on data centers, energy and chips. It does not operate a shared compute grid.

The Power Bottleneck

The second coalition, the AI Energy Management Alliance, tackles a different choke point: getting data centers connected to the electrical grid at all. AEMA launched September 16, 2026, with founding members Emerald AI, Google, Nvidia and Anthropic, joined by utilities AES, Constellation, National Grid and NRG Energy, according to Complete AI Training and Engadget.

The pitch, laid out by Emerald AI CEO Varun Sivaram in a Fortune article cited by Engadget, is simple. New data centers can wait a decade or more for a grid hookup because utilities size capacity for the worst afternoon of the year. But the grid runs at only about 50% utilization on average. If data centers agree to throttle back during peak demand hours, Sivaram argues, the country could unlock 100 gigawatts of capacity on infrastructure that already exists.

A Goldman Sachs study put the more conservative number at 76 gigawatts nationally, according to The AI Insider. The Brattle Group estimates every 10% gain in grid utilization translates to roughly a 3.4% rate cut for ordinary electricity customers, according to AI Chat Daily.

The regulatory groundwork is already shifting. The Federal Energy Regulatory Commission directed the six regional grid operators it oversees in June to accommodate large customers willing to limit their draw in exchange for faster connections, AI Chat Daily reported. Texas's grid operator is finalizing similar rules, and Silicon Valley Power, a California municipal utility, has launched what it calls the nation's first flexible-load interconnection program.

Nvidia, Digital Realty and Emerald AI are set to switch on what they describe as the world's first power-flexible AI factory, a nearly 100-megawatt site in Virginia, later this year, according to AI Chat Daily. Emerald AI's chief scientist, Ayse Coskun, has said the technology can reduce but not eliminate the industry's need for new power plants.

The Fair Criticism

Engadget raised a legitimate point worth taking seriously: AEMA doesn't address the pollution, noise, and community disruption that have fueled local backlash against data centers nationwide. The alliance's stated goal is narrower—faster grid connections in exchange for flexible power use, not environmental mitigation. That gap matters for residents living next to these facilities, even if it's outside what AEMA set out to fix. Engadget also flagged that AEMA functions as a lobbying operation aimed at governors, regulators and federal policymakers. The group's own materials describe pressing its case "in state capitals and in Washington."

Why Industry Is Moving, Not Government

Both coalitions are not federal mandates. They're industry groups solving a market problem because the alternative is data center developers simply building off-grid campuses with their own power generation. That's a more expensive, less efficient path, and it's already happening in places where utilities can't move fast enough.

The existing federal option, the National Artificial Intelligence Research Resource, has supported more than 600 research projects and 6,000 students as of 2026 and is transitioning from pilot to full operation, according to Crypto Briefing. It's a smaller-scale, taxpayer-funded version of the same idea these private coalitions are now trying to do at industry scale, without asking Congress for more money.

None of this changes the global math. China's "Eastern Data, Western Computing" initiative reported 1,882 EFLOPS of intelligent computing capacity by March 2026, according to Crypto Briefing, built through state-directed coordination the U.S. can't replicate. India is rolling out its own National AI Compute Grid in phases through 2026 in partnership with state governments. Pooling idle American chips and speeding up grid connections buys efficiency. It does not manufacture a single new chip or power plant, and the chip shortage that started this race is still the one thing nobody's coalition can fix.