TL;DR
The AI data center buildout is routing around political defeat by occupying former industrial land, federal sites, military bases, and active oil fields where the four constraints killing standard development don't apply.
Gas turbine prices are up 195% and delivery slots are sold out through 2031, but practitioners are split on whether behind-the-meter solar-plus-storage reaches reliability thresholds before the turbine lock-in becomes a half-decade gas dependency.
RGGI carbon allowance prices moved 110% in 90 days as Virginia prepares to rejoin the cap-and-trade system, repricing every gas-fired megawatt-hour in PJM starting July 1.
The gap between EV sales share and fleet share remains the most misunderstood variable in energy demand forecasting, with Norway at 97% of sales but only one-third of its fleet electric.
Quiet week on the forwarding signal. No anomalies cleared the full discipline gate for a standard deep-dive this cycle, but the week's posts trace a coherent line through a single structural question: what happens when AI power demand collides with every constraint simultaneously, from site availability to turbine supply to carbon pricing to fleet turnover timelines.
Coverage This Week
The gas turbine bottleneck is the real AI constraint — Gas turbine prices up 195%, global orders at a 25-year high, and US data centers outbidding every country on earth. The post frames the fossil power surge not as a comeback but as a desperation signal: you pay double for the slow, dirty option when the fast, clean one can't deploy on your timeline. The practitioner debate was the richest of the cycle, with eight qualified comments splitting between operators who see solar-plus-storage as the overlooked alternative and those who argue dispatchable power is the only option that meets the timeline. The fork is clear: does behind-the-meter storage reach 99.999% uptime before locked turbine delivery slots open in 2028-2029, or does the interconnection bottleneck force a half-decade gas dependency? Read →
The AI buildout stopped using standard data center sites — Every standard site got harder to build in 2026. So the buildout rerouted onto federal land, military bases, Superfund parcels, and active oil fields. DOE identified 16 federal sites. The Air Force opened 3,100 acres across five bases. EPA rewrote Superfund guidance to fast-track redevelopment. California's biggest oil company filed permits for a 600,000-square-foot data center inside the Elk Hills oil field, powered by a 550 MW gas plant already running below capacity. The four constraints killing sites everywhere else don't apply where heavy industry already operates. Read →
RGGI carbon allowances moved 110% in 90 days — From $25 to $52.50 per ton. Virginia rejoins the cap-and-trade system on July 1 with the world's largest concentration of AI data centers. The March 2026 auction exhausted all 7.9 million tons of cost containment reserve allowances before Virginia even formally returned. Every gas-fired megawatt-hour in PJM just got materially more expensive. Dominion filed to charge residential customers $13 per month under Rider RGGI, triple the prior rate. Read →
The federal heat pump credit died on December 31 — The 25C credit covered up to $2,000 per heat pump installation. Physics hasn't changed: heat pumps still use less energy than gas furnaces, still cut emissions on a cleaning grid. But the affordability claim just took a hit. The post separates durable from fragile in the electrification thesis. Read →
China's solar subsidy window closed on schedule — 372 GW deployed in 2025. BNEF projects 341 GW in 2026. That drop was scheduled the day the five-year-plan pricing reset got announced. The post reframes the decline as a policy design feature, not a market signal, and asks what happens when the world's largest installer transitions from subsidy-driven to market-driven deployment. Read →
Norway sells 97% electric but two-thirds of its cars still run on gas — The distinction between sales share and fleet share is where demand forecasts go wrong by a decade. Fleet turnover runs at 5-6% per year. Norway's road oil use fell 12% from 2021 to 2024 after years at number one. The post argues a majority-electric American road is a 2045 story, not a 2035 story. Read →
Asia's solar generation overtook gas for the first time — 1,727 terawatt-hours of solar in the 12 months to April against 1,711 from gas. The post argues the real signal is the flat blue line: Asian gas generation barely moved in five years while solar quadrupled. Read →
This Week’s Signals
The AI buildout is rerouting onto former industrial land — federal sites, military bases, Superfund parcels, and active oil fields — because the four constraints killing standard data center development don't apply there.
The post maps a structural rerouting in AI infrastructure siting. Every standard data center site got harder to build in 2026: New York, Texas, Pennsylvania, and Utah all tightened permits in the last month. The buildout responded by moving onto land most developers wouldn't have considered two years ago. DOE identified 16 federal sites for rapid data center construction. The Air Force opened 3,100 acres across five bases. EPA rewrote Superfund and Brownfield guidance to fast-track redevelopment. The sharpest example: California's biggest oil company filed permits for a 600,000-square-foot data center inside the Elk Hills oil field in Kern County, powered by a 550 MW gas plant already running below capacity, with closed-loop cooling and carbon injection into depleted wells next door. The thesis is that the four constraints killing data centers everywhere else (land, power, water, community opposition) don't apply where heavy industry already operates.
A senior technology leader in energy infrastructure endorsed the thesis but pushed the argument further: "Communities don't stop at the property line anymore. Their concerns reach city halls, state capitols, regulatory agencies, and even the halls of Congress. I'm not convinced there's any truly 'safe' land left or any site completely immune from community opposition." The argument is that developers who treat industrial sites as opposition-free zones are underestimating the next constraint. The competitive advantage won't be finding places with less resistance. It will be building repeatable frameworks that treat community alignment as a critical project workstream alongside land, power, and permitting.
The CEO of a consulting firm endorsed the Elk Hills example specifically: the 550 MW plant already running below capacity, closed-loop cooling, and carbon injection next door amount to "a purpose-built industrial stack that happens to fit a data center's needs almost exactly. The mechanical infrastructure already exists; someone just has to adapt it."
Downstream from this: the exposure sits with environmental remediation and site-preparation contractors. If data centers route onto Superfund parcels and depleted oil fields at scale, the remediation-to-construction handoff becomes the new critical path. Current industrial landowners (DOE, DoD, oil producers) hold an asset class that didn't exist two years ago: data-center-ready industrial land with pre-provisioned power and no residential opposition.
> Does EPA's fast-track Superfund redevelopment guidance survive its first contested environmental liability transfer on a data center project, or does legacy contamination risk freeze the brownfield pipeline before it scales?
The gas turbine bottleneck thesis drew the richest practitioner debate in this batch — eight qualified comments splitting heavily toward challenges.
The post argues the AI bottleneck has moved from chips to gas turbines. Prices are up 195%. Global orders hit 130 GW last year, a 25-year high. US data centers building captive plants have outbid every country on earth except the US itself. Data centers drive half of all US electricity demand growth through 2030, and that power is needed in 18 months, not eight years. Solar-plus-storage is cheaper, but you can't will an interconnection queue to move faster. So hyperscalers order turbines at whatever the cost, with delivery slots booked years out. The IEA says US turbine demand is now limiting availability for the rest of the world.
The CEO of a consulting firm endorsed the power-constrained framing directly: "The interconnection queue problem you mentioned is the part that doesn't get enough attention - turbines are a workaround for a grid access problem, not a solution to it. Until interconnection timelines compress, every large compute deployment is essentially an energy procurement problem first."
The challengers outnumbered endorsements. The CEO of an energy sector firm pushed back on the turbine bet itself: gas turbines carry $8 million to $22 million per day in OPEX, require large infrastructure, and are sold out until 2031. A senior technology leader in energy sector countered on different grounds: "If you can build gas behind the meter, why don't you build solar + storage behind the meter?" The question cuts to the thesis's weakest joint. A second energy sector CEO challenged the durability of the turbine demand signal, arguing that global AI cost competition (citing cheaper Chinese LLMs competing at a fraction of the cost) makes the turbine bet risky. A third energy sector CEO argued the US is running backward while China scales renewables massively.
A founder advising grid infrastructure operators added context without taking a side: China's eastern data centers run on 70% coal because that's where the grid is, but half their western data centers sit at 20-30% utilization. "Abundant power doesn't equal competitive AI. The bottleneck moved again."
The downstream question is whether behind-the-meter solar-plus-storage reaches 99.999% uptime reliability before locked turbine delivery slots open in 2028-2029, or whether the current lock-in creates a five-to-seven-year gas dependency that outlasts the interconnection bottleneck it was designed to bypass.
> Does behind-the-meter solar-plus-storage reach 99.999% uptime reliability before locked turbine delivery slots open in 2028-2029, or does the interconnection bottleneck force a half-decade gas dependency that the turbine bet was supposed to be a bridge past?
RGGI carbon allowance prices moved 110% in 90 days as Virginia prepares to rejoin the cap-and-trade system on July 1.
The post tracks a 110% move in RGGI carbon allowances, from $25 to $52.50 per ton in 90 days. RGGI is the cap-and-trade system pricing carbon emissions from power plants across 11 northeastern states. It normally moves single digits per year. Virginia rejoins on July 1 with the world's largest concentration of AI data centers entering a tightening cap. The March 2026 auction exhausted all 7.9 million tons of cost containment reserve allowances before Virginia even formally returned. The cap declines toward a 92% reduction by 2037. Every gas-fired megawatt-hour in PJM just got materially more expensive. Dominion filed to charge residential customers $13 per month under Rider RGGI, triple the prior rate.
A senior technology leader in energy sector noted that the pricing signal changes data center energy strategy: "If carbon costs start moving into power contracts and utility riders, then data center energy strategy has to account for more than generation availability. It has to account for regulatory exposure and long-term cost pass-through."
Save activity ran at 3.11x the 90-day average rate with zero send activity. Readers kept this for reference but did not pass it along. That reference-only pattern typically appears on content with regulatory or compliance implications: readers save it for their own planning rather than circulating it.
The EV sales-versus-fleet distinction is where demand forecasts go wrong by a decade.
The post makes a straightforward structural argument: Norway sells 97% electric but two-thirds of its cars still run on gas. Fleet turnover runs at 5-6% per year. Even when every new car sold is electric, the combustion fleet doesn't vanish. It keeps driving for a decade, sometimes two. Norway's road oil use fell 12% from 2021 to 2024 after years at number one. The US is at 10% of sales. Most decks quietly assume a majority-electric American road by 2035. The fleet math says 2045.
A founder advising power generation operators challenged the linear retirement assumption: as EV adoption reaches a tipping point, the resale value of used internal combustion vehicles will likely collapse, and fuel infrastructure (gas stations) will become less profitable. "This economic pressure could dramatically accelerate the scrappage rate, making the fleet curve bend much faster than a decade-long projection suggests." The counterargument matters because it identifies a nonlinear accelerant the post's analysis explicitly excludes.
A founder advising energy sector operators pushed back from the opposite direction, arguing that China's EV sales have declined as subsidies become unviable given debt-to-GDP exceeding 300%. "The EV party and 'everything electric' craze is crashing into reality."
A senior technology leader in energy sector endorsed the fleet-curve framing: "Markets can change quickly, but physical assets turn over slowly. The infrastructure transition often follows the fleet curve, not the sales curve."
A second senior technology leader in energy sector offered a different measurement entirely: "The most important number is km driven by EV and ICE. Many households have both types of car and use the EV more because it is cheaper."
Field Notes
Asia's solar generation overtook gas for the first time. 1,727 TWh of solar in the 12 months to April against 1,711 from gas (Ember, via Carbon Brief). The post argues the real signal is the flat line: Asian gas generation barely moved in five years while solar quadrupled. Read →
Offshore wind went silent this cycle. Zero posts on the topic against a 46-post baseline in the database. Not a signal by itself, but the absence is worth tracking if it extends.
Reply if any of this is playing out at your company, or contradicting what you're seeing on the ground. Every reply goes directly to our analyst desk and feeds our intelligence.