The most unsettling twist in today’s AI race is that the United States may not be slowed by a shortage of brilliant researchers or cutting-edge chips, but by something far less glamorous: electricity. Your briefing frames this as an “electrical bottleneck” that could downgrade American technological leadership if the grid cannot power the next wave of AI infrastructure.
That sounds almost too mundane to be geopolitically dangerous. And yet the numbers are anything but mundane. The briefing highlights projections that, by 2035, US data centers alone could demand around 106 GW (BloombergNEF) or even 123 GW (Deloitte). This acceleration is not incremental; it’s a shock that has apparently outpaced forecasts so rapidly that BloombergNEF reportedly revised its estimates upward by 36% in just seven months. The hardware reality is brutal: the most advanced AI systems in the world cannot run without a reliable, scalable power backbone.
The piece rightly insists that this is no longer theoretical. Microsoft’s experience becomes a kind of diagnostic case. The CEO reportedly acknowledged a “compute glut” — a surplus of compute capacity that cannot be deployed because energy and ready-to-use facilities are missing. The company’s reported strategy shift — spending $11.1 billion in Q1 2026 to lease existing capacity rather than waiting for new builds — looks less like a tactical adjustment and more like a quiet admission that the grid is now a core limiter of AI scale.
So where exactly does the system fail? The briefing outlines three pressure points that interact in slightly nightmarish ways.
First, reliability risk. NERC warnings suggest the data center boom is unlike anything the system has faced before. The Virginia incident in July 2024 is revealing: a sudden disconnection of 1.5 GW of load nearly destabilized the grid. It’s a reminder that modern risk is no longer just about losing generation; losing enormous demand too fast can also break stability assumptions.
Second, the bureaucratic choke point: interconnection queues. In PJM — the region hosting the famous “Data Center Alley” — the briefing says the time from request to operation can exceed eight years. Eight years is an eternity in the AI cycle. It is also the kind of delay that incentivizes companies to bypass the public system altogether.
Third, physical constraints in the supply chain. Big transformers have lead times three to four times longer than in 2020, and large gas turbines have seen delivery times stretch to roughly 4.5 years, with costs per kW reportedly up 71% over four years. This is where the story becomes especially sticky: even perfect regulatory reform cannot conjure equipment that does not exist or cannot be delivered.
Unsurprisingly, policy has started to treat the issue as national strategy rather than utility management. The briefing points to Executive Order 14318, issued on July 23, 2025, designed to accelerate federal permitting for data center infrastructure. More strikingly, it explicitly pushes “dispatchable baseload” energy including natural gas, coal, and nuclear. The subtext is hard to ignore: in a competition defined by speed, Washington appears willing to trade some near-term decarbonization purity for immediate capacity and strategic momentum.
Industry, predictably, is moving faster than the grid. The “behind-the-meter” trend described in your briefing is essentially an energy end-run. OpenAI’s Stargate project in Texas reportedly includes ten gas turbines totaling 361 MW. Meta’s Prometheus project in Ohio is linked to an additional 200 MW of on-site gas generation. And then there is xAI’s Memphis saga, which the briefing presents as a cautionary tale of what happens when speed and local environmental governance collide, with allegations of unpermitted turbine operations and major NOx pollution implications.
This is the moment where the political economy of AI begins to fray. The briefing calls attention to an emerging “AI tax” on ordinary consumers, where outdated tariff structures socialize grid upgrade costs while hyperscalers reap the upside. In PJM, it estimates that 63% of recent price increases may be driven by data center demand, potentially pushing residential bills up to 25%. The Georgia Power example is similarly stark: a proposed $15 billion spend to serve data centers, with staff estimates suggesting a possible $20/month increase for residential customers.
Environmental externalities stack on top of these distributive tensions. The briefing cites a Cornell study in Nature Sustainability projecting that by 2030 the AI industry could add 24–44 million metric tons of CO? annually and consume 731–1,125 million cubic meters of water per year in an unmitigated scenario. The mitigation roadmap it summarizes — smarter siting, faster grid decarbonization, and efficiency upgrades — reads like a sensible triad. But its feasibility depends on whether the market and regulators can move faster than the hype cycle.
The geopolitical comparison with China sharpens the anxiety. The briefing states that in 2024 China added 429 GW of new power capacity versus 51 GW for the US. Even if one debates the broader structural context, the rhetorical implication is clear: centralized capacity expansion can become a strategic advantage in an AI era that is hungry for baseload reliability.
What I find most compelling — and most disquieting — is the briefing’s core reframing. It argues that energy is now a determinant of AI leadership, not a background condition. If that is correct, then the policy priorities it lists are not optional refinements but guardrails against strategic erosion: clearing interconnection logjams, creating fair cost allocation through a specific data center tariff class, securing domestic supply chains for critical grid components, and accelerating advanced nuclear and SMR pathways (with industrial-scale deployment not expected before 2030).
The emerging picture is messy, pragmatic, and politically combustible. AI supremacy is being negotiated not only in labs and venture rounds, but also in transformer factories, regulatory dockets, and ratepayer hearings. The future of “compute” may hinge on the unromantic world of “capacity.” And the United States may soon discover that the energy transition and the AI transition are not parallel stories at all, but the same story told by two different bureaucracies that can no longer afford to ignore each other.
for more info see this report by FT https://ig.ft.com/ai-power/?spoor-id=3b153ecb-9db0-4b97-8773-40b67d827264