Timely analysis

How Musk’s Colossus Is Turning AI Power Into a Policy Fight

November 7, 2026

How Musk’s Colossus Is Turning AI Power Into a Policy Fight

Elon Musk’s Colossus buildout is no longer just a story about a giant AI cluster in Memphis. It is becoming a live test of who pays for the power, who gets slowed down by permitting, and how far regulators are willing to go when AI infrastructure starts behaving like critical utility infrastructure.

For builders and investors, the stakes are immediate: frontier compute is now constrained by substations, transmission, fuel, and interconnection rules as much as by chips and models. For utilities and policymakers, the question is whether the grid can absorb these loads without shifting costs onto ratepayers or letting temporary workarounds become the new normal. That tension is showing up first in a small set of high-profile cases, but the policy response is moving quickly. 1, 2, 3

The old AI bottleneck was compute. The new one is power.

The Colossus effect starts with a simple mismatch: the industry can buy GPUs faster than it can deliver electricity to use them.

That is not just theory. Business Insider reporting described Amazon accumulating “dark GPUs” that had been delivered but could not yet be activated because the power infrastructure was not ready. The broader point is that hardware acquisition does not automatically translate into usable AI capacity. 4

Hardware and deployment specialists are making the same distinction. In 1 Minute Signal coverage of a16z, Pat Gelsinger’s framing is that the real bottleneck has moved from chip design to deployment, with power availability embedded in that lag. 5

That is the policy problem Colossus exposes. If AI clusters can only scale by securing their own power, then energy policy is no longer a background utility issue. It becomes part of the infrastructure strategy.

Colossus is not just big. It is vertically integrated infrastructure.

xAI’s Memphis and Southaven buildout shows how far a frontier AI company can push into physical infrastructure. Public reporting indicates the project has relied on temporary gas turbines, battery storage, substations, and a planned dedicated power plant to support the cluster. The company also says it is working with utilities on grid stability while transitioning toward more permanent generation. 6, 7, 8

The scale is the point, but it should be read carefully. One estimate put the Colossus system at 352.4 MW of power consumption, roughly comparable to the daily electricity use of a city of about 250,000 people. 9 More recent records-based reporting says SpaceX’s Super Intelligence segment reached 1.7 GW of installed nameplate compute capacity as of September 30, 2026. That figure is not the same thing as actual electricity consumed. 6, 10

That distinction matters. As the records analysis notes, a permit, an installed turbine, and an operating power plant are different milestones. Installed capacity is not the same as delivered power. 10

So the Colossus story is not simply that Musk is buying a lot of compute. It is that xAI is building a private power-and-compute stack to move around the grid’s normal pace.

"A permit, an installed turbine and an operating power plant are different milestones."

— Raj Mitra, Tesorb 10

The local backlash is forcing the issue into public view

The infrastructure strategy is colliding with local communities and regulators. Reporting on the Memphis and Southaven sites describes lawsuits over noise and air pollution, plus federal intervention tied to national-security arguments. 11, 12

xAI’s own materials push the opposite narrative: that company-funded substations, batteries, and grid support can stabilize the local system. 7, 8 That may be true in some operational sense. But the policy tension is obvious. If a private AI campus is large enough to require bespoke substations and temporary turbines, local planning stops looking like ordinary zoning. It starts to look like energy policy by another name.

That is why Colossus matters beyond Memphis. It is an early example of how a single AI deployment can force questions about who pays, who waits, and who gets priority when power is scarce.

"The conflict over Colossus 2 highlights the friction between the industrial-scale electricity and land-use demands of modern AI infrastructure and the local communities living in its shadow."

— 1 Minute Signal coverage of The Wall Street Journal 11

Regulators are moving from permissioning to burden-sharing

The most important policy shift is not that governments are suddenly pro-data-center. It is that they are beginning to formalize how much of the burden should sit with the AI company rather than the public.

FERC has directed grid operators to reform the interconnection process for large loads and to create rules that prevent cost shifting onto power consumers. It also wants more transparency around spare capacity and more consideration of behind-the-meter solutions. 1, 2

That is a meaningful change in posture. Instead of treating data centers as passive new demand that the grid must absorb, regulators are increasingly asking whether large loads should bring their own generation, pay their own interconnection costs, or accept curtailment when the system is stressed. Reuters reported that PJM plans to require new large users either to bring their own generation or enter a “connect and manage” framework with early curtailment. 13

The same logic is showing up in state-level deals. OpenAI’s Georgia project is being structured so the company pays infrastructure costs and accepts demand-response obligations. OpenAI’s own flexibility pledge frames that as a way to protect other ratepayers, while also acknowledging that demand response is not a substitute for new generation and transmission. 14, 15

For now, this looks less like a settled national doctrine than a fast-moving response to a handful of highly visible projects. But the direction is clear: the more power-hungry the AI buildout becomes, the more utilities and regulators will try to push more of the cost and risk back onto developers.

Energy policy is starting to look like industrial policy

The federal government is now treating AI infrastructure as strategically important. DOE’s Speed to Power initiative explicitly links faster generation and grid buildout to winning the AI race. 16 NERC has gone further, proposing new registry categories for computational load owners and operators, with thresholds that would capture most projected computational load by the end of 2027. 17

That does not mean the policy framework is complete. It means the old framing—where data centers were just another commercial load—is breaking down.

The economics are changing too. A Dallas Fed paper found that data centers already increased national wholesale electricity prices by 6.9 percent by 2025, with larger effects possible under heavier utilization. It also found the impact is nonlinear because transmission constraints and the supply curve steepen as demand rises. 18

In plain English: the first wave of AI demand can be absorbed, but the next wave gets expensive fast.

That is why the policy conversation is moving from “Can we connect it?” to “Who absorbs the risk if the load does not behave as promised?” PJM’s revised backstop procurement plan is one answer. FERC’s new rules are another. 2, 3

"The key credit question is whether state rules provide timely, certain recovery and adequate customer security, so the utility remains an administrative intermediary rather than taking meaningful balance-sheet or stranded-cost risk."

— Maria Scheller, ICF International 3

Why builders should care

For AI builders, the lesson is not that every company should build its own power plant. Most should not.

The lesson is that frontier-scale compute is becoming inseparable from energy strategy, site selection, grid politics, and long-cycle capital planning. If the power plan is weak, the compute plan is weak.

For investors, that changes the diligence checklist. It is not enough to ask whether a lab has GPUs and model talent. You need to ask whether it has secure power, what kind, how quickly it can come online, whether it depends on temporary stopgaps, and who carries the downside if the power plan slips.

For policymakers, Colossus is both a warning and a preview. It shows that permissive siting alone will not solve the AI infrastructure problem. Nor will generic appeals to renewables, because the near-term constraint is often dispatchable, local, and buildable power. The U.S. and European evidence both point in the same direction: data centers cluster where firm power and transmission already exist, not where the cheapest clean-energy theory says they should. 19, 20, 21

That is the real Colossus effect. It is not just that Musk is building a giant AI campus. It is that the campus is helping force a rethink of AI energy policy around physical capacity, not abstract ambition.

"The rapid deployment of artificial intelligence hardware is hitting a physical bottleneck: electricity."

— 1 Minute Signal coverage of Business Insider 4

What to watch next

Three developments will show whether this stays a one-off controversy or becomes the template:

  1. More AI companies sign demand-response and self-supply agreements like OpenAI’s. 14, 15
  2. Regulators harden the idea that large computational loads must internalize more of their grid costs. 2, 3, 17
  3. Utilities start designing dedicated power and interconnection paths for AI campuses, as seen in the U.S. and Taiwan. 13, 22

If those patterns keep spreading, Colossus will look less like a single project and more like the operating model for the next phase of AI infrastructure.

Share this

Tags

Sources

[1] AI data centers just got a government-mandated fast lane to the grid | TechCrunch

[2] FERC lays out AI data center, grid operator reforms

[3] PJM’s revised data center backstop procurement plan reduces utility risks: ICF | Utility Dive

[4] The #AI boom has a problem: billions of dollars in #GPUs, but not enough electricity to power them. | 1 Minute Signal

[5] Former Intel CEO: Why This is the Best Time to Build Hardware | 1 Minute Signal

[6] xAI Colossus: the Memphis AI supercomputer explained

[7] Information Center | SpaceXSI Memphis

[8] Our Commitment to Memphis | SpaceXSI Memphis

[9] Musk has the world's most powerful supercomputer, but it is also the most energy hungry - Digital Journal

[10] How Big Is xAI’s Compute? Colossus, by the Records

[11] What I Saw Outside Elon Musk’s Massive Colossus 2 Data Center | WSJ | 1 Minute Signal

[12] SpaceX won’t remove all of xAI’s unpermitted turbines for another year

[13] PJM unveils plan to tackle AI-driven power demand surge | Reuters

[14] Georgia Power to serve major OpenAI project under new customer protection agreement - Daily Energy Insider

[15] OpenAI takes data center power-flexibility pledge national | TechTarget

[16] Speed to Power: Stakeholder Input and DOE Actions | Department of Energy

[17] NERC Asks FERC to Fast-Track Data Center Registration Rules | Davis Wright Tremaine

[18] Processing Power: The Effect of Data Centers on Wholesale Electricity Markets - Research Dept. Working Paper No. 2606 - Dallas Fed

[19] AI Data Centers Amplify Grid and Energy Price Risks | HEC Paris

[20] Data centres and Energy: the environmental cost of a faulty price signal | Banque de France

[21] The AI Question Even the Experts Can't Answer | The Spillover | 1 Minute Signal

[22] Taipower planning for AI power demand - Taipei Times

Written by: 1 Minute Signal Editorial Team