On the Record

About 6.7 GW of nuclear is now under contract to hyperscalers — Microsoft at Crane, Meta at Clinton, Amazon at Susquehanna, Google at Duane Arnold, Meta again at Perry and Davis-Besse. It is the most quoted number in the AI-energy conversation.

One of those plants is actually delivering. Susquehanna was already running. Crane and Duane Arnold are restarts and both are still shut down; Duane Arnold is not expected on the grid until the first quarter of 2029. Contracting takes a signature. Delivering takes restart licences, fuel, staff, and an interconnection, all on nuclear timelines.

A GB300 NVL72 rack draws up to 142 kW today. Vera Rubin is reported at roughly 180–220 kW for the second half of 2026. NVIDIA's own roadmap puts Rubin Ultra at around 600 kW on 800 V DC in 2027. A megawatt rack is one generation past anything published.

Power density is the variable that quietly rewrites everything downstream of it — cooling, switchgear, floor loading, interconnection, and the entire economic case for putting generation on the same site as the load. It is the number I watch most closely and the one I almost never see discussed outside a data-centre engineering meeting.

If I am wrong about all fourteen predictions at once, this is how. The binding constraint on AI turns out to be neither chips nor capital nor models, but interconnection queues and transformer lead times — the physical layer nobody writes essays about because there is no curve to draw.

It is also the least glamorous thing to be right about, which is roughly how you can tell it matters. Every forecast in this document, including mine, assumes the electrons show up on schedule.

China holds roughly 95–97% of global solar wafer capacity and 93% of polysilicon. Wafer is the hardest stage — the most capital-intensive, the most energy-intensive, and the one with the least Western activity. It is also the one that reshoring announcements quietly skip.

Module assembly is cheap and fast to stand up, so that is where the ribbon-cuttings are. Cells are harder. Wafer is where you find out whether industrial policy actually worked, and we are not close.

Constellation received FERC approval in June 2026 and NRC approval of its fuel receipt and possession amendment on 1 July 2026, clearing the way for restart in the second half of 2027, with the reactivated operating license targeted for May 2027. The $1B DOE loan closed in November 2025 and roughly ninety operator candidates are in the licensing pipeline. The binding constraint here is not reactor hardware — it is regulatory sequencing and staffing a plant that has been dark since 2019.

I'm at 55% rather than 75% because “delivers power to the grid” is a harder line than “on schedule,” and because every schedule in this industry has a hidden month in it. But this is the single most important row in the register: it is the first megawatt of the entire AI-nuclear story to actually show up. Everything else in that trade is paper.

TSMC's own guidance puts Fab 2 at N3 volume production in the second half of 2027, and on its Q4 2025 call it said it was pulling that schedule forward. Taiwanese press reports — not TSMC guidance — add that equipment installation is running early and that US capacity is fully booked; I treat those as unconfirmed, and TSMC's fourth Arizona fab only began construction in early 2026. I am betting that TSMC hits its own published date, which sounds trivial and is not — leading-edge node transitions on new soil with a new workforce slip as a matter of routine.

Note what this row deliberately excludes. Intel is already running 18A in volume at Fab 52 in Arizona, so “leading-edge logic in volume on US soil” is a question that has already been answered. The open question is specifically whether TSMC can replicate Taiwan's ramp discipline in Arizona, which is the one that matters for supply-chain resilience.

GPT-4 launched in March 2023 at $30 per million input tokens. Today the cheapest credibly frontier-adjacent model publishes $0.22 per million for off-peak cache-miss input, rising to $0.44 at peak. That is roughly a 140x decline in three and a half years, and the mechanism — better hardware utilization, distillation, brutal competition among a half-dozen labs — has not run out.

The wording here is doing a lot of work, and deliberately. Cached input runs far below the standard rate — OpenAI lists $0.005 per million on its smallest tier — and batch tiers cut it further, so on a loose reading this resolved a year ago. Legacy small models already sit at or below $0.10. The bet is specifically that a model people would call frontier — top-3 in public head-to-head ranking at the moment of measurement, not a cheap tier a generation behind — lands under a dime at standard list price. The risk is that “frontier” keeps running away from the price curve, which is exactly what it has done so far.

UBTech's FY2025 results put full-size humanoid sales at 1,079 units in 2025 alone, across BYD, Geely, FAW-VW, Dongfeng and Foxconn — but no single buyer is close to a thousand. AgiBot announced its 15,000th unit in June 2026, which is a manufacturing milestone rather than a deployment number; Omdia's independent count of its 2025 shipments was 5,168. Tesla's Optimus, per Musk on the January 2026 earnings call, is not in material use in Tesla's own factories.

So the industry has cleared the pilot bar convincingly and has not cleared the fleet bar at all. I'm at 45% because the gap between an impressive pilot and a thousand-unit fleet that pays for itself is precisely where robotics has disappointed for forty years. This row says deployed and doing revenue work, not ordered — order announcements in this sector are marketing.

GB300 NVL72 racks draw up to 142 kW today. Vera Rubin is reported at roughly 180–220 kW for the second half of 2026. NVIDIA's GTC 2025 roadmap put Rubin Ultra NVL576 at roughly 600 kW on 800 V DC for the second half of 2027 — a figure from the roadmap presentation rather than a published spec sheet.

A megawatt rack is therefore not on the published roadmap — it is one generation past it. I'm betting the roadmap keeps its slope through 2029. This is the row I care about most professionally, because rack density is the variable that quietly rewrites everything downstream: cooling, switchgear, floor loading, interconnection, and the entire economic case for putting generation on the same site as the load.

About 6.7 GW of existing-fleet and restart nuclear is now under contract to hyperscalers: Microsoft/Crane at 835 MW, Meta/Clinton at 1,121 MW, Amazon/Susquehanna at 1,920 MW, Google/Duane Arnold at 615 MW, and Meta/Vistra at 2,176 MW across Perry and Davis-Besse. Amazon's Comanche Peak agreement adds up to 1,200 MW more. I state the sum rather than a round number because no published aggregate survives checking.

Restarted megawatts delivered today: zero. Susquehanna is already delivering under the Amazon contract and the Vistra volumes begin in late 2026, so power does flow under some of these PPAs — but every megawatt of it comes from a plant that was already running. Crane and Duane Arnold, the two restarts, are both still shut down, and Duane Arnold is not expected on the grid until the first quarter of 2029.

That gap is the entire point of this row. Contracting is a press release. Delivering requires restart licenses, uprate approvals, fuel, staff and interconnection, all on nuclear timelines. And note what the numbers do to my own confidence here: with Susquehanna flowing and Vistra starting, the 4 GW bar may well clear on plants that never shut down, long before a single restart contributes. If this row resolves true that way, it resolves true for reasons that have nothing to do with the thesis underneath it — so I am saying that now rather than claiming the win later.

Solar was 53% of US capacity additions in 2023, 66% in 2024, and 54% in 2025 — the largest single source each year, five consecutive years by SEIA's count. EIA projects roughly 51% of utility-scale additions in 2026 against record total additions of about 86 GW.

Three of the seven years are banked and 2026 is tracking. The bet is really about 2027, 2028 and 2029, and the risk is not solar's cost curve — it is a policy shock or a gas-turbine resurgence for data-center baseload. Note what actually drove the 2024-to-2025 dip from 66% to 54%: storage grew faster and solar shrank. SEIA and Wood Mackenzie put 2025 installations at 43.2 GW, down 14% year on year. Solar kept the crown on a falling volume, which is a weaker position than the share number alone suggests.

The efficiency question is closer to settled than most people realize. Trina certified a full-size, industrial-format tandem module at 29.2% and 907 W with TÜV SÜD in June 2026, and LONGi's certified tandem cell record is now 35.5%. The record is not the problem.

Volume is the problem. The largest publicly disclosed commercial tandem order is 1.2 MW — GCL's win for a Huaneng research institute in March 2026, the first publicly tendered tandem order from one of China's five major state generators. Trina has separately announced a commercial tandem order deployed in New Zealand without disclosing volume, so I cannot rule out something larger. GCL's Kunshan line is nameplated at 1 GW; my own read is that realistic output is 100 MW-class, which is an industry estimate rather than a published figure. LONGi's tandem line is a 100 MW pilot, and Trina's investor communications put commercial tandem shipments at scale in 2028–2029.

So I've set the bar where the actual uncertainty lives: 500 MW cumulative shipped at above 28%. That requires someone to solve encapsulation and field durability at scale, not to win another certification. Perovskite stability under real-world stress remains the open question, and it is a materials-science problem, not a manufacturing one.

The US has built enormous module assembly capacity — 65 to 70 GW nameplate — on top of almost no domestic cell production. SEIA's last published operating figure is 3 GW of cell capacity. Against the 31 to 37 GW of modules the US actually manufactured over the past year, that is roughly 8 to 9%.

Read that number carefully, because a competing one is circulating. PV Tech puts domestic cell coverage at one-third to just over 40% by comparing crystalline cell capacity against crystalline module capacity. I am using operating cell output against actual module output, which is the stricter basis and the one this row resolves on. The ramp itself is real: ES Foundry completed its expansion to 3 GW in July 2026, Qcells Cartersville began cell production in June 2026 and is ramping toward 3.3 GW, Suniva and Silfab add another 2 GW, Canadian Solar's CS PowerTech opened in Indiana in July 2026 with over 6 GW at full buildout, and T1 Energy starts by end-2026. That is roughly 10 GW operating or ramping.

This is a coin flip, and I have specified the denominator as assembly output rather than nameplate capacity, because that distinction is where most published claims about domestic content quietly cheat. Cell manufacturing is genuinely harder to stand up than assembly — it is a chemistry and yield problem, not a bolting-together problem — and more announced cell lines get cancelled than commissioned.

LBNL's 2025 Update puts 2024 at 192 TWh, or 4.7% of US electricity, with a 2030 reference case of 649 TWh — about 11.8% — inside a range of 9.5% to 15.3%, or 521 to 843 TWh. Worth knowing if you are grading me against it: that same update revised its own history downward, on reduced reported GPU shipments for 2023 and 2024. The resolver's baseline moves too.

Thirteen percent is above the reference case and below the high case. I am betting the over, because load is contracted years ahead, capex is already committed, and the historical pattern in this specific series has been that LBNL revises upward. The risk is that efficiency gains outrun deployment, or that interconnection — not capital, not chips — becomes the binding constraint. That second one is underrated and it is the same queue problem that governs row 7.

The IEA's April 2026 update estimates 2025 at roughly 485 TWh — its own modelled figure rather than a measured actual — with overall data centre demand up 17% in the year and AI-focused facilities up about 50%.

Nine hundred is a deliberately slightly-below-central line, which is why this is one of my higher confidences. Worth noting: the IEA's 2030 projection barely moved between April 2025 and April 2026. The prose of the newer report says around 950 TWh while its own Annex A base case table still gives 945 — the identical figure to a year earlier. So the modelled output is essentially unchanged despite a full year of enormous announcements. Either the model is insensitive to news or the announcements are mostly re-sequencing the same buildout. I lean toward the second, and that is a mildly deflationary read on the whole sector.

Zero SMRs operate in the US today. Oklo broke ground at Idaho National Lab in September 2025, is proceeding via the DOE authorization pathway rather than NRC licensing, has cleared two of five DOE gates as of June 2026, and has moved its own language from a late-2027-to-early-2028 window to 2028. TVA's Clinch River BWRX-300 is furthest along the NRC route, with the safety analysis report accepted in July 2025 and no construction permit yet issued.

The definitional trap matters here: Aurora at INL is a DOE-authorized demonstration, so even at startup it may not count as grid-connected commercial operation. I've pushed my date to 2030 and sit just below a coin flip. The regulatory path is genuinely accelerating and nuclear timelines genuinely slip, and I don't think either force clearly dominates.

TOPCon accounted for more than 94% of module shipments among the top ten manufacturers in the first half of 2025, against roughly 706 GW shipped globally. The 88.3% figure that circulates is cell shipments among the top five — a different measure, and one worth keeping separate. Tandem is under half a percent on any of these bases. ITRPV's roadmap puts silicon tandem above 17% only by 2036, which implies well under 5% at 2030.

I am betting against the published roadmap, which is why this is my lowest confidence and one of only two rows where I'm taking a genuinely contrarian position. Tandems must beat TOPCon on delivered dollars per watt, not on efficiency, and Chinese TOPCon overcapacity keeps that floor on the ground through roughly 2032. If I'm right, learning curves cracked an entrenched incumbent faster than the industry's own consensus expected. If I'm wrong, incumbency in commodity manufacturing is even more durable than I already think it is — and that has implications well beyond solar.

China holds roughly 95–97% of global solar wafer capacity and 93% of polysilicon. On modules, IEA-PVPS puts it at 86.4% of production against 83% of capacity — worth separating, because production share and capacity share are not the same claim. Wafer is the hardest stage — the most capital-intensive, the most energy-intensive, and the one with the least Western activity.

Note how easy it would be to make this row unlosable. Betting that China still holds most of global module capacity in 2030 is not a prediction, it is a description. Betting on wafer share dropping below 90% is a real bet: it requires meaningful non-Chinese ingot and wafer capacity to be built and operated, not announced. Everyone reshoring solar starts at the assembly end because it is cheap and fast.