.png "Apple iPhone Frictionless Bill Splitting Feature 2026")
In much of that demand growth, you can see yourself. The data centers behind your searches, chatbots and videos are colliding with the oldest constraint in computing: the physical limits of the electrical grid. This piece explains what's actually straining the grid, why power not chips is now the binding constraint on AI, and what it means for electricity bills, a nuclear renaissance, and the next generation of AI hardware.
KEY TAKEAWAYS
- PJM's July 2026 auction (for the 2028/29 delivery year) cleared at the FERC-approved cap of $325/MW-day and fell 6,831 MW short of its reliability target the third consecutive miss. Uncapped, the price would have been $554.72, a 70% jump.
- The price arc is the story: capacity prices went from $28.92/MW-day (2024/25) to $325+ (2028/29) roughly 11x in three auctions. Total auction cost rose from $2.2 billion to $16.4 billion.
- Demand is outrunning supply: 2026 reporting cites projections of global data center electricity use above 1,000 TWh by the end of 2026 roughly Japan's entire annual consumption. (The IEA's own 2024 projection was 945 TWh by 2030; the upward revision is part of the debate.)
- Projects are stalling: roughly half of planned U.S. data center builds were projected delayed or canceled as of April 2026; interconnection waits run 4–10 years; transformer lead times are 36–48 months.
- Everyone is buying nuclear: all four major hyperscalers have signed a dozen-plus nuclear agreements worth nearly 10 GW, per Forbes (July 2026).
- Efficiency is improving but chasing a rising curve: NVIDIA's Vera Rubin claims 10x performance per watt over its predecessor and draws about twice the system power.
- Your bill is in the mix: PJM estimated 1.5%–5% bill increases for some ratepayers in the 2026/27 delivery year.
Source: PJM auction reports, via Utility Dive and Oil Price. A "capacity auction" is how the grid operator pays power plants in advance for being available; when prices hit the cap year after year, something structural has changed.
WHY THIS MATTERS
For users: capacity costs flow into electric bills. PJM estimated its 2026/27 auction would add 1.5% to 5% for some ratepayers depending on state. The AI you use runs on someone's grid increasingly, yours.
For the industry: "speed to power" has replaced latency and fiber as the primary site-selection criterion in 2026 reporting. Available megawatts now matter more than millisecond latency.
For markets: Big Tech has become, in Forbes' words, a de facto power company making procurement decisions that determine which reactors get financed and which plants stay open.
A record with a warning label
PJM's July 2026 auction, announced July 15, is the cleanest snapshot yet of the crunch. It cleared at the FERC-approved price ceiling of $325 per megawatt-day, and fell 6,831 MW short of the capacity needed for reliability the third consecutive auction missing the target. Without the cap, the clearing price would have been $554.72 a 70% jump and one distribution zone (COMED LDA) would have cleared at $776.69.
The cap itself is a story. PJM established the price "collar" in coordination with the governors of all 13 PJM states and the Federal Energy Regulatory Commission, after last December's report said the next auction would run without one. In other words: the market was signaling prices high enough that regulators stepped in.
And the demand side names its driver plainly. In the prior year's auction, forecast peak load grew about 5,500 MW "mainly from data centers," per PJM. Julia Hoos, head of USA East at Aurora Energy Research, put the supply-side problem bluntly: "It's unsurprising that almost no new capacity showed up because projects in PJM are still struggling to get built."
What's actually straining the grid
In simple terms: the chip race ran on one assumption that compute, not energy, was the limit. The grid is disproving it. Data centers are adding 5–7 GW of new load to U.S. grids every year, while new generation comes online at only 2–3 GW per year a persistent 2-to-1 gap that analyses expect to continue through 2032.
Under the hood, the intensity of AI workloads is the difference-maker. Rack density has surged from 10–15 kW in conventional data centers to 40–70+ kW for AI clusters, with the newest AI training clusters exceeding 100 kW per rack and individual sites approaching 1 GW, per Gartner. Industry analyses put a single AI task at consuming up to 1,000x the electricity of a traditional web search. Gartner's June 2026 forecast adds scale: global data center power demand grew roughly 26% in 2026, AI-optimized servers now account for about 31% of total data center consumption, and AI servers are projected to overtake conventional servers in electricity use during 2027. Goldman Sachs Research forecasts U.S. data center power demand jumping from 31 GW in 2025 to 66 GW by 2027.
Why it can't be fixed quickly is the third layer. Grid interconnection queues now run 4–10 years ERCOT's large-load queue alone reportedly approaches 226 GW. High-voltage transformer lead times stretch to 36–48 months. Of the 12 GW of U.S. AI data center capacity announced for 2026, only about 5 GW was under active construction, per one industry analysis; as of April 2026, roughly half of all planned U.S. builds were projected to be delayed or canceled. The financial damage of a slip is immediate: one industry analysis calculated that a six-month delay on a 60 MW facility nearly halves its internal rate of return, from 17.1% to 8.8%.
The green conflict
Here's the uncomfortable part. Most cloud companies have net-zero mandates, and data centers must run 24/7 without a second of downtime a profile that intermittent wind and solar can't match without enormous storage. So the interim has been fossil. In PJM's 2027/28 auction, gas-fired generation accounted for 43% of cleared capacity and coal for 20%, per PJM's auction report. The bridge between AI's power hunger and its climate targets is, for now, partly running on gas and coal.
Critics make the mirror image: that nuclear money is diverting capital and political attention from faster-to-deploy solar and wind. Both sides are describing the same squeeze — firm, cheap, clean baseload doesn't exist in sufficient quantity yet, and everyone is racing to be first in line.
The nuclear pivot
That race has a paper trail. In late July 2026, Forbes reported that every major AI company Microsoft, Google, Amazon and Meta has now signed at least one nuclear deal, committing across a dozen-plus agreements to nearly 10 GW of capacity, "enough to power roughly 7 million homes."
- Microsoft signed a 20-year, roughly $16 billion power purchase agreement for the entire output of Three Mile Island Unit 1 renamed the Crane Clean Energy Center an 835 MW reactor being restarted by Constellation Energy with an expected 2027 return to service. The U.S. Department of Energy closed a $1 billion loan for the restart in November 2025.
- Amazon secured 1.92 GW from the Susquehanna nuclear plant in Pennsylvania (via Talen Energy) and is investing $700 million in X-energy's Xe-100 small modular reactor program, alongside advanced-nuclear work with Energy Northwest.
- Google signed the first corporate SMR agreement with Kairos Power (October 2024), targeting roughly 500 MW with a first reactor around 2030, reinforced by a 50 MW TVA partnership and site-development deals (including Elementl Power), plus a separate contract for 1,800 MW of new capacity, per Forbes.
- Meta has the largest commitment: up to 6.6 GW across TerraPower's Natrium fleet (up to 4 GW, targeting 2032–2035), Oklo's 1.2 GW Aurora campus in Pike County, Ohio, and existing-fleet PPAs with Vistra and Constellation including a 20-year deal for 1,121 MW from Clinton in Illinois that begins in 2027.
The reality check matters. The first nuclear power to actually arrive for a hyperscaler is the 835 MW Three Mile Island restart in 2027. Almost everything else the SMRs that most of these deals depend on lands in the 2030s. No small modular reactor has been completed commercially in the U.S. yet, and the cautionary tale is Vogtle: the last large U.S. reactors finished seven years late and $36.8 billion over budget more than the most expensive power plant ever built, by its own measure.
The efficiency pivot and its limit
The semiconductor industry's response has been real, not rhetorical. At CES 2026, NVIDIA unveiled its Rubin platform; in February, CNBC's first look reported the Vera Rubin system would deliver 10 times more performance per watt than its Grace Blackwell predecessor, out of 1.3 million components from 80+ suppliers, shipping in the second half of 2026. At GTC 2026, the specs got more specific and more telling: Rubin GPUs run at 1,800 to 2,300 watts each (versus roughly 1,000 W for Blackwell), and 100% liquid cooling is mandatory.
Read that combination carefully. The per-watt numbers are genuinely improving, but the absolute power envelope is climbing faster. Performance-per-watt is the industry's most powerful tool against the crunch yet it's chasing a demand curve that's roughly doubling in the same window. Efficiency buys time. It does not end the bottleneck.
What it means for you
Three practical implications. First, bills: capacity charges are already showing up in rate cases across PJM's 13 states, and the auction trajectory suggests the pressure persists. Second, the AI you consume may become regionally differentiated services and training clusters increasingly follow power, not population. Third, watch the corporate layer: as one case study put it, the emerging nuclear contract is less a power deal and more a financial instrument a hedge, and eventually a competitive moat, for the companies that can secure firm power first.
WHAT TO WATCH NEXT
- PJM's next base auction (July 2027) for the 2029/30 delivery year the fourth and likely final auction under the current price collar. Whether the cap holds is the single best leading indicator of grid stress.
- 2027: the Three Mile Island restart the first hyperscaler nuclear delivery.
- 2027: Gartner's projected year when AI servers overtake conventional servers in data center electricity consumption.
- The 2030s: SMR deliveries (Kairos, X-energy, TerraPower, Oklo) and whether any actually arrive on schedule.
- Regulatory action: interconnection-queue reform in PJM, ERCOT and NYISO, and the question of who pays for grid upgrades analyses put the required investment in the hundreds of billions.
FAQ
BOTTOM LINE
What's confirmed: the grid is the binding constraint on AI. PJM has missed its reliability target three auctions running, capacity prices are roughly 11x what they were two years ago, about half of planned U.S. data center builds are delayed or canceled, and all four hyperscalers have committed nearly 10 GW of nuclear capacity.
Why it matters: the industry spent years assuming compute was the limit. It turns out the limit is electrons — and the people paying for them are in the same ratepayer pool as everyone else.
What remains uncertain: whether the IEA's fast-rising 2026 demand projections hold, whether 2030s SMR timelines survive contact with U.S. construction reality, and whether efficiency gains can ever outrun a demand curve that roughly doubles in a few years. The next clean data point is PJM's July 2027 auction. Until then, the honest summary is this: the smartest models on Earth are now waiting in the same queue as a new transformer.
SOURCES / ATTRIBUTION
- Oil Price (July 15, 2026) — PJM 2028/29 auction: 6,831 MW short, $325/MW-day FERC cap, $554.72 uncapped, third consecutive miss
- Utility Dive (Jul 2024, Jul 2025, Dec 2025) — PJM auction history: $28.92 → $269.92 → $329.17 → $333.44/MW-day; $2.2B → $16.4B costs; fuel mix; bill-impact estimates; Julia Hoos (Aurora Energy Research) quote
- Forbes (Ken Silverstein, July 26, 2026) — all four hyperscalers' nuclear commitments; ~10 GW / dozen-plus deals; ~7 million homes; Vogtle cost overrun
- CNBC (Katie Tarasov, Feb 25, 2026) — Vera Rubin first look: 10x performance per watt, ~2x system power, H2 2026 shipping
- Tech Insider (GTC 2026 coverage, Aug 25, 2026; April 2026) — Rubin 1,800–2,300 W/GPU, 100% liquid cooling; IEA 1,000 TWh reporting; Uptime/Gartner figures
- SMRIntel (July 6, 2026) — nuclear deal database: Microsoft TMI $16B/835 MW/2027 + DOE $1B loan (Nov 2025); Meta 6.6 GW breakdown; delivery timelines
- The Word 360 (Aug 23, 2026) — Meta–Constellation Clinton 1,121 MW PPA; Google–Kairos/Elementl detail; corporate nuclear as financial instrument
- Coro Advisors (Jul 30, 2026) — Gartner June 2026 forecast (26% demand growth, 31% AI share, 2027 overtake); IEA 415 TWh (2024) → 945 TWh (2030); Goldman 31 → 66 GW (2027)
- InformedClearly (May 7, 2026; Jun 2, 2026) — 5–7 GW vs 2–3 GW gap through 2032; 36–48 month transformer lead times; 12 GW announced / 5 GW under construction; IRR impact of delays; 100+ kW racks
- Enkiai (Apr 8, 2026) — Gartner 40%-of-AI-data-centers-by-2027; 1,000x web-search comparison; speed-to-power site selection
Note on conflicts: 2026 reporting cites IEA projections of >1,000 TWh for 2026 while also citing the IEA's 2024 forecast of 945 TWh by 2030 — both are presented as reported rather than reconciled, since the agency's current projection set could not be independently pulled. Last December's PJM reporting said the July auction would run without a price cap; the July results show a FERC-approved collar was established instead — disclosed rather than silently resolved.
.png)
.png)
.png)
.png)
.png&description=The Digital Divider: Apple's New Seamless Feature Completely Eliminates the Pain of Splitting the Check){kind=link}
0 Comments