The market is repricing the promise of intelligence. It has not repriced the physics. Megawatts are still the binding constraint, and the premium is moving to the megawatts that depend least on congested grids and contested permits: behind the meter, close to load, already consented.
Two things happened this week that look, at first glance, like headwinds for the AI-energy story. The Nasdaq fell 2.9% and the semiconductor index dropped into a formal bear market. And New York imposed the first statewide executive moratorium on large data centres.
But read them together and a different picture emerges. The sell-off was about whether the software returns justify the valuations. The moratorium is about whether the grid can carry the load without sending household bills through the roof. Neither questions the underlying equation: MW = intelligence; no power, no tokens. Both, in fact, tighten it. If capital gets choosier and permits get scarcer, the value concentrates in what can actually be built, powered and connected. That means behind-the-meter generation, and it means sites (smaller, edge and enterprise-scale, or already consented) that sit below the political waterline.
The tech wreck, round three. US technology shares gave back the prior week’s gains in full. The Dow eased 0.93% to 52,146, the S&P 500 fell 1.55% to 7,457, and the Nasdaq dropped 2.9%. A proximate trigger on Friday was Beijing-adjacent: Moonshot AI’s Kimi K3, billed as the world’s largest open-weight model, sharpened fears that Chinese competition undercuts the subscription economics of US frontier labs and, by extension, the chip spending that funds them. The sell-off drew equally on broader doubts about AI capital expenditure and semiconductor valuations. The Philadelphia Semiconductor Index ended the week 20.2% below its 22 June high, a bear market by the arithmetic, after its worst week in more than a year. The VIX closed at 18.77, up 24%.
Note what did not happen: nobody suggested AI demand is falling. The anxiety is about who captures the margin, not whether the electrons get consumed.
Albany draws the line. On Tuesday, Governor Kathy Hochul signed Executive Order 62, the first statewide executive moratorium on data-centre development in the US. Read closely, the order is narrower and more interesting than the headlines. It applies to facilities that consume, or can consume, 50 MW or more; it directs the Department of Environmental Conservation to hold incomplete applications for discretionary state permits in abeyance while the state prepares a Generic Environmental Impact Statement covering energy demand, water use, air quality and noise; and it expressly leaves local-government permissions untouched. The trigger sits in the order’s own recitals: nearly 12 GW of data-centre load requests were in the New York ISO interconnection queue as of May, more than 8 GW of it arriving in 2025 alone, and it is now stated New York policy that everyday ratepayers should not fund the grid upgrades those loads require.
The order lands alongside the legislature’s separate Responsible Data Center Development Act, passed on 4 June (Senate 44–16, Assembly 102–39), whose principal one-year permitting pause applies at 20 MW, with further requirements reaching down to 5 MW.
Two details matter for anyone siting capacity. First, regulatory intensity is now explicitly scale-dependent. New York has drawn lines at 50 MW, 20 MW and 5 MW, and the burden rises with size; smaller, enterprise-scale facilities are not exempt from scrutiny, but they sit below the heaviest of it. Second, the order does something quietly remarkable: it contemplates requiring data centres to fund new clean generation and storage dedicated to their own operations, including customer-sited distributed energy resources, to the greatest extent feasible. Albany has, in effect, written behind-the-meter generation into its remedy. Beyond New York, the direction of travel is a patchwork: Maine’s moratorium was vetoed, Hill County, Texas rescinded its pause after a developer sued, and a federal moratorium bill sits in the Senate with uncertain prospects. Regulatory uniformity in the US is gone. Site selection is no longer an energy-first exercise; it is an energy-and-consent exercise, and consented capacity just became scarcer, which is another way of saying more valuable.
How the BTM complex traded. The behind-the-meter and power-infrastructure names told their own story this week, one of divergence rather than collapse.
Bloom Energy, the purest listed expression of on-site, grid-bypassing generation, closed the week around $245: still up roughly 180% year-to-date and more than 800% over twelve months, but clearly caught in the tech downdraft, trading below its short-term moving averages after a parabolic run. The fundamentals that fuelled the run haven’t changed: Q1 revenue of $751m, up 130% year-on-year, and a swing to a $75m net profit. What has changed is the market’s tolerance for paying any price for them.
GE Vernova, the heavier industrial expression of the same thesis (turbines, grid equipment, and SMR technology selected for Ontario Power Generation’s Darlington project), held up better, around $1,090, up roughly 67% year-to-date and still trading above its moving averages. The contracted generators, Constellation, Vistra and Talen, each with hyperscaler offtakes attached, barely flinched: cash flows under contract remain the market’s preferred way to own the power shortage.
The pattern is the one this briefing has flagged before: when the market doubts the returns on intelligence, it sells the chips and the model companies first, the servers second, and the electrons last. Steel in the ground with a creditworthy offtake behaves like infrastructure. Promises behave like promises.
The Fed’s family fight. Cooling retail and wholesale inflation eased fears of an imminent hike, but futures markets still price the possibility of a quarter-point increase as soon as September, and Chair Kevin Warsh delivered his first congressional testimony this week with notably less forward guidance than his predecessors. Higher-for-longer, or higher-again, remains the wrong weather for long-duration growth stories and the right weather for contracted, inflation-linked infrastructure cash flows.
The war arrives on the bill. This briefing will not re-argue the case that dependence on imported molecules is a security risk priced in someone else’s chokepoint; events keep making it for us. This week’s instalment: Brent ended the week above $85 a barrel, its highest since mid-June, as the US concluded an eighth consecutive night of strikes on Iran and Reuters reported that Tehran has asked the Houthis to stand ready to close the Red Sea oil route. On Saturday the war reached Gulf energy infrastructure directly: Kuwait Petroleum reported severe damage and injuries at a vital oil-sector site after repeated Iranian attacks, and Kuwaiti power generation and desalination plants were also struck. The bill, meanwhile, has physically arrived in Britain: Ofgem’s price cap for July–September rose 13%, taking a typical dual-fuel direct-debit household to £1,862 a year, with the regulator attributing the increase to wholesale gas prices driven by the Middle East conflict. A household in Hull is paying the Hormuz risk premium. And watch what the producers themselves are doing about it: Iraq, which moved some 3.4 million barrels a day through Basra before the war, is now trucking oil across Syria to the Mediterranean port of Baniyas, expanding this month to crude at roughly 50,000 barrels a day, with Reuters reporters describing tanker queues stretching 30 kilometres along war-damaged roads. It is perhaps the most expensive way to move a barrel ever devised, it amounts to about 1.5% of Iraq’s pre-war southern flows, and Baghdad has said the route stays even if Hormuz normalises. When a chokepoint binds, rational actors pay almost any unit cost for the bypass. The megawatt that never transits a strait remains the cheapest and safest on offer.
Burnham’s North Sea instinct. Britain’s new Prime Minister takes office Monday and is expected within days to signal support for new drilling at Jackdaw and Rosebank, the Shell/Equinor Adura fields off Scotland, alongside an expansion of tie-backs near existing infrastructure. Public consultations on both fields launched Thursday, so formal approval must wait; the direction of travel will not. Burnham’s instinct is sound on energy security grounds, and the industry-union coalition lobbying for it makes a fair point: domestic production can carry lower production and transport emissions than some imported alternatives. But the molecules still price at the marginal barrel, still cross commodity markets, still leave the household in Hull exposed to exactly the Hormuz risk premium described above. Rosebank’s first oil is years away, and the operator’s own claim is that it could at peak supply roughly 10% of UK continental shelf output: a meaningful share of a declining base, not insulation from a global price. The outgoing Energy Secretary’s counterpoint - offering eligible legacy low-carbon generators voluntary fixed-price contracts to weaken gas’s grip on electricity prices - is structurally durable, even if it may now be politically overtaken. The EDGE answer, as ever, is the electron that never enters a commodity market at all.
The moratorium contagion question. New York is first, not last. Moratoriums have been proposed in at least a dozen states. Watch which states position themselves as the opposite trade (open for business, power available) and watch whether hyperscalers start paying up for the one thing no legislature can conjure at short notice: sites with consent already in hand.
The AI trade gets its next data points from earnings rather than macro. Alphabet and Tesla report Wednesday. Evercore expects Alphabet to accelerate cloud growth and potentially lift capex guidance towards $300bn for next year, which would be the clearest possible signal that the demand side of MW = intelligence is intact regardless of what the Nasdaq did this week. GE Vernova also reports Wednesday, a direct read on turbine and grid-equipment order books. Thursday brings Intel and Digital Realty, the latter a bellwether for data-centre leasing and power procurement. ServiceNow and IBM give the enterprise-AI adoption read. Elsewhere: flash July PMIs offer an early third-quarter snapshot, and the Bank of Japan is expected to hold at 1%.
The question behind this week’s question. If training built the gigawatt campuses, inference, now the majority of the workload, decides the map for the next decade. So: can inference go anywhere?
The fair answer is that a substantial part of it can, and that part is enormous.
Start with the scale. Deloitte estimates inference will account for roughly two-thirds of all AI compute in 2026, about double its share of a few years ago, and predicts more than $50bn of inference-optimised chip sales this year. McKinsey projects inference demand growing at roughly 35% a year through 2030. Training is episodic and capital-intensive; inference is recurring and scales with every user and every query. Deloitte’s caveat matters too: most inference is still expected to run in data centres on serious silicon, not on cheap edge hardware. The question is which data centres, and where.
Then the geography splits. Training is latency-indifferent: it goes wherever power is cheap, firm and abundant, which is why it centralised into a handful of remote hubs. Inference divides into two species. Latency-tolerant inference (batch processing, overnight agent runs, document analysis, anything asynchronous) genuinely can go anywhere power is available. That workload will chase stranded and behind-the-meter megawatts to the ends of the earth, and it is a large and growing share.
Latency-sensitive inference is different. For real-time voice and video, industrial control, autonomous systems and agentic workloads making continuous tool calls, regional facilities can cut network latency to single-digit or low-double-digit milliseconds for nearby users, an advantage a centralised campus hundreds of miles away cannot match, however many megawatts it commands. Property advisers including CBRE and Cushman & Wakefield have concluded that inference is pushing demand into regional and secondary markets, and enterprises are repatriating inference from cloud to local colocation. And when Nvidia launches an AI Grid reference architecture spanning regional points of presence, telecoms central offices and edge sites, not only giant centralised campuses, the direction of travel is being drawn in public.
There is a counter-case. More efficient models could reduce compute per task; hyperscalers may slow marginal capex; and behind-the-meter generation is not free of consent, with fuel, emissions, noise and local land-use approvals all still to be won. But lower unit costs have, in every previous computing cycle, expanded usage rather than shrunk it. And the investment question is not whether every proposed megawatt gets built. It is which megawatts remain viable when capital, grids and political consent all become more selective at once.
The week’s three stories all point the same way. A market repricing centralised AI economics, a state drawing regulatory lines by scale, a workload migrating towards distributed facilities near load: every force in play favours efficient and decentralised generation of energy, and the sites with fewer dependencies on congested grids and long interconnection queues.
The cheapest, most secure, cleanest megawatt remains the one never wasted, generated close to where it is used. This week, it also became the megawatt least exposed to a pause.
A Very British Problem — why efficiency-first is the answer to Britain’s heat, cold and bills alike; this week’s price cap makes the case for me.
The Forcing Function — how EDGE infrastructure shortens the fuse between a missile strike in the Gulf and a bill arriving in Grantham.
Last week’s Weekend Edition — the hyperscalers give up waiting for the grid.
If you’ve found this useful, do share it — the Weekend Edition is free to all.
— Jonathan Maxwell Founder & CEO, SDCL · author of The Edge
Jonathan Maxwell is the CEO of Sustainable Development Capital LLP and author of The Edge. He writes about energy, climate, finance, and geopolitics. Views expressed are personal and do not constitute investment advice. To learn more about energy efficiency, visit the website of SEIT plc, or SDCL Group.
CNN Business, Kimi K3 and Friday’s sell-off: https://www.cnn.com/2026/07/17/investing/us-stocks-asia
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.