Nobody Bought a Reactor for Speed · BlueShip Research Working Paper No. 19

Credit. The framing of this paper, the test that counts deals against construction sites and the arithmetic of half of one percent, comes from the video essay that seeded it. The verification, the extensions and the prediction ledger are mine.

1. The ledger

I audited every announced nuclear agreement between Microsoft, Google, Amazon and Meta against primary sources. Composition matters more than the total. About a third of the headline megawatts existed before the deals were signed, so the buyer changed and the grid did not. Uprates and two restarts supply the only near-term new electrons. Roughly half the ledger is new build, and 50 MW of that is under construction: one reactor.

Measure	What the record says
Buyers audited	Microsoft, Google, Amazon, Meta
Deals counted	13 rows; other lumpings give 11 to 14
What counts as a deal	Anything from a PPA running 20 years on an operating plant to a memorandum to explore
Ledger at face value	~14.7 GW
Ledger excluding the weakest structures	~13.4 GW
Carnegie Endowment, independent count, 2 Jun 2026	up to 13 GW
Already existing output, reallocated to one buyer	5,187 MW, ~35% of the headline
Under construction, Jul 2026	50 MW, Hermes 2, and nothing else: 0.34%
Excluded	Fusion, from all fission arithmetic. Oracle has no nuclear deal; OpenAI no direct one I could find as of July 2026

#	Buyer	Counterparty / plant	MW	Structure	Announced	Stated first power	Physical work, Jul 2026
1	Microsoft	Constellation, Crane (ex-TMI-1), PA	835	PPA running 20 years, funding restart	Sep 20, 2024	2027	Yes, heavy. 550+ FTEs, transformers ordered, DOE $1B loan closed Nov 2025, FERC waiver Jun 2026, NRC restart authorization still pending
2	Google	Kairos Power, master PPA	500 by 2035	Framework for a new SMR build	Oct 14, 2024	~2030	First unit only (row 3)
3	Google (via TVA)	Kairos Hermes 2, Oak Ridge TN	50	First US utility PPA for a Gen IV reactor	Aug 18, 2025	2030	Yes. Ground broken Apr 17, 2026
4	Google	NextEra, Duane Arnold, IA	615	PPA running 25 years, funding restart	Oct 27, 2025	early 2029, possibly Q4 2028	Partial. Major equipment ordered; NRC license amendments in review; not full construction
5	Google	Elementl Power, 3 undisclosed sites	1,800	Development capital + offtake option, no reactor tech chosen	May 7, 2025	~2035	None. No sites announced
6	Amazon	Talen, Susquehanna, PA	1,920	PPA on existing output through 2042 (restructured in front of the meter Jun 11, 2025 after FERC rejected the ISA behind the meter, 2-1, on Nov 1, 2024)	Jun 11, 2025 (orig. Mar 2024)	n/a, operating since 1983/85	Plant running. Zero new generation
7	Amazon	X-energy	~5,000 ambition by 2039	Equity, anchor of ~$500M round	Oct 16, 2024	early 2030s	None
8	Amazon	Energy Northwest / X-energy, Cascade, Richland WA	320 initial, 960 full	Amazon funds development; EN owns/operates	Oct 16, 2024	early 2030s	None. Construction permit application not yet submitted
9	Amazon	Dominion, North Anna VA	~300	MoU to explore an SMR. Weakest structure in the ledger	Oct 17, 2024	none stated	None
10	Meta	Constellation, Clinton IL	1,121 (incl. 30 uprate)	PPA running 20 years on existing plant from Jun 2027	Jun 3, 2025	n/a, operating since 1987	Plant running. Only the 30 MW uprate is new
11	Meta	TerraPower, up to 8 Natrium plants	2,800 baseload (~4 GW with storage)	Funds first 2 units + rights to 6	Jan 9, 2026	as early as 2032	None on Meta units. Kemmerer WY, under construction, is PacifiCorp's, not Meta's
12	Meta	Oklo, Pike County OH	1,200	Development funding	Jan 9, 2026	first phase as early as 2030	None. Site characterization slated to begin 2026
13	Meta	Vistra, Perry + Davis-Besse + Beaver Valley	2,609 (2,176 existing from Perry and Davis-Besse + 433 uprates across all three, 140 at Beaver Valley)	PPAs running 20 years, purchases from late 2026, full 2,609 by 2034	Jan 9, 2026	n/a existing; uprates through 2034	Plants running. Uprates are the only new MW

Category	MW	Share of ~14.7 GW	What it is
Existing output reallocated to one buyer	5,187	~35%	Talen 1,920; Clinton 1,091 ex-uprate; Vistra 2,176. Zero new electrons
Restarts in progress	1,450	~10%	Crane 835 (2027 target); Duane Arnold 615 (late 2028/29)
Uprates	463	~3%	Vistra 433; Clinton 30. Real new MW, no new plant
New build, contracted or framework	~7,560 upper bound	~51%	Kairos 500, Elementl 1,800, Cascade 960, Dominion ~300 MoU, TerraPower 2,800, Oklo 1,200
Of which under construction	50	0.34%	Hermes 2, and nothing else

2. Every shelf is empty

Nobody in the ledger bought speed, because speed was not for sale. Turbines are sold into the 2030s, the restart shelf holds three plants and does not restock, and new build has never come in near its certified cost. What the buyers took was a position in a queue, on gas as much as on nuclear, and Project Kilby shows the position itself is now the tradable asset.

Shelf	The binding number	Source
Turbines, backlog	GE Vernova backlog 116 GW at Q2 2026: 53 GW firm equipment, 63 GW slot reservations. 2031 delivery slots already selling	GE Vernova
Turbines, capacity	~110 GW of 2025 orders against 60 to 70 GW per year of world manufacturing capacity; prices up 195 percent since 2019. A campus decided today gets a 2031 slot and first power around 2031 to 2033	Wood Mackenzie
Restarts, record	Zero US nuclear restarts have delivered a single megawatt hour (MWh) as of 28 July 2026
Restarts, shelf	Three plants: Palisades ~800 to 805 megawatts (MW), Crane 835 MW, Duane Arnold 601 to 615 MW. About 2 percent of the ~97 GW US fleet, and it does not restock
Restarts, clock	Palisades returned to operations status on 25 August 2025, has missed four named dates since and still has no announced restart date. Working number is 3 to 4+ years from decision to grid for plants never gutted	American Nuclear Society, 8 Jul 2026
New build, cost	Vogtle 3&4 was certified in 2009 at ~$14B and finished at over $31B. Full ladder below, all in unless noted
New build, finance	DOE announced $17.5B of conditional loan commitments on 23 June 2026 for ten new AP1000s. No firm AP1000 order for a specific site existed as of late July 2026 that I could verify	DOE
New build, my estimate	Roughly 300 to 425 MWe (megawatts of electric output) of small modular reactor (SMR) capacity on any grid by the end of 2030, essentially none of it contracted by a hyperscaler	my arithmetic
Fuel	Natrium, Xe-100 and Hermes burn high-assay low-enriched uranium (HALEU). Centrus, the only company licensed to make it commercially in the US, produces 900 kg per year at Piketon against a DOE statement of need above 40 tonnes by 2030. One Natrium first core alone is reported at 15 to 20 tonnes	DOE; trade press
Fuel, the fixes	DOE's $2.7B package buys Centrus expansion to 12 tonnes per year, first new capacity in 2029. TRISO-X's plant, licensed 13 February 2026, starts making fuel in the first half of 2028	DOE; TRISO-X
The queue as an asset	Project Kilby: Chevron's 2.67 GW plant, built on 7HA turbine slots reserved before the panic and sold to Microsoft on a PPA running 20 years

Project	$/kW	Basis
Barakah, contract	~$3,650	$20.4B / 5.6 GW, 2009 contract
DOE Liftoff NOAK target	$4,700 overnight	DOE Liftoff (Sep 2024); MIT CANES NOAK $4,625, 2024 USD
Summer 2&3 completion, marginal	$6,500–10,900	circulating completion estimates $15–25B / ~2.3 GW, excluding ~$9B sunk
$80B Westinghouse program, implied	~$7,000	$80B / 10 AP1000s, my arithmetic
MIT FOAK for the next 2 at Vogtle	$8,300–10,375 overnight	MIT CANES ANP-201, 2024 USD; greenfield $9,300–11,625
Vogtle, all in	~$13,900–15,700	$31–35B / 2,234 MW
Hinkley Point C, current money	~$19–20k	£48.7B / 3.26 GW (EDF, Feb 2026); my FX conversion, approximate
CFPP, cancelled	~$20,100	$9.3B / 462 MW

The binding constraint sits further upstream, in fuel. Most of the fleet contracted for new build burns HALEU, licensed American capacity falls short of stated need by more than an order of magnitude, and the announced fixes are late by construction's own clock. The queue behind the queue is longer than the queue.

3. Where the market is pricing it

Descriptive only. Constellation has handed back the whole re-rating the Microsoft deal started. GE Vernova, priced on the turbine queue rather than on nuclear, rose in a falling sector on a multiple that pays today, in cash deposits, for delivery in the 2030s.

Ticker	Role in the chain	Close 7/28/26	YTD 2026 (approx.)	vs high over 52 weeks	Note
CEG	Fleet + restart	$259.82	-24.4%	-37%	PE 22.2
VST	Fleet	$148.64	-3.9%	-32%	fwd PE 16.3
TLN	Fleet	$326.05	-13.0%	-28%	fwd PE ~10–12; vendor data conflict flagged
GEV	Turbine queue	$943.38	+51%	-21%	fwd PE 44.3
OKLO	2030s queue	$39.58	-43.5%	-80%	no revenue
SMR	2030s queue	$8.22	-41.3%	-86%	$18.7M TTM revenue
NNE	2030s queue	$16.06	-33.2%	-74%	no revenue
CCJ	Fuel	$86.96	-3.9%	-36%	fwd PE 65.2
LEU	Fuel / HALEU	$169.65	-29%	-63%	PE 59.6

Measure	Value
CEG close, 20 September 2024, the day the Microsoft deal added 22 percent	$254.98
CEG close, 28 July 2026	$259.82
Elapsed	22 months, the entire subsequent re-rating round-tripped
GEV, year to date 2026	+51 percent in a falling sector, at 44x forward earnings
Sector, 28 July 2026	Heavy down day across the complex; no catalyst identified

4. Corrections, and what this paper does not claim

Item	What the record says
"About 10 GW" is wrong on the low side	The likely source is a tracker showing 13 deals and 9.8 GW that lists MW as "TBD" for rows it still counts and omits Duane Arnold (615) and Elementl (1,800)
"Palisades got its operating license back in July 2025" is wrong as stated	The license never lapsed; it runs to 24 March 2031. 24 July 2025 was the transfer of operating authority. 25 August 2025 was the return to operations status
The arithmetic of half of one percent	Survives strengthening. 50/9,800 = 0.51 percent on the video's denominator; 0.34 percent on my larger one
Not verified: HALEU tonnage for the Natrium first core	Trade press only
Not verified: Palisades fuel load	Could not confirm whether it has begun, nor learn the identity of its fuel fabricator
Not verified: the sector down session of 28 July 2026	No catalyst identified
Conflict: Crane capacity	835 MW in Constellation's materials, 837 MWe in World Nuclear News. I use 835 and flag it
Not claimed	The market table is descriptive, not a recommendation. Probabilities on the ledger below are not assigned here

5. Outside readings

Three later readings, counting with different tools, arrive at the same empty shelf.

Outside reading	Logged	What it adds
Tai, Noffsinger and DeFabrizio, Powering AI: How real is the risk of overbuilding?, McKinsey, July 2026	8 Aug 2026	Same asymmetry: a short-term shortfall is likely and the greater near-term risk for power players is underbuilding. Data centers are roughly 75% of expected US power demand growth over the next decade, built at a rate near 30 gigawatts a year. Against roughly 120 gigawatts of projected demand growth the grid holds about 40 gigawatts of dispatchable headroom and about 100 gigawatts committed to be built, while 50 to 75 gigawatts of coal and steam-gas retires: a nationwide capacity need of 30 to 55 gigawatts by 2030. Their gigawatt-accounting sidebar makes this paper's own point about counting: 500 megawatts of IT load can mean 900 megawatts of interconnection served by 1,000 megawatts of gas nameplate
The data center construction boom: risks and claims trends, Allianz Commercial, 2026	13 Aug 2026	The underwriter's view, and the mismatch this paper predicts is now a claims category. A large data center builds in 18 to 24 months while the grid to feed it takes years; key grid equipment runs ~30% short, with switchgear lead times to 80 weeks and transformers to 50; local opposition delayed or blocked at least 75 US projects worth roughly $130bn in the first quarter of 2026 alone. Who wins, in their summary: those who can deploy fastest, not simply finance it
Bommarito and Bommarito, 50-state policy survey, SSRN, June 2026	14 Aug 2026	The pipeline denominator this ledger lacked: 890 announced US projects, roughly $1.79 trillion of announced investment, about 240 gigawatts of reported capacity among disclosing projects. Set 240 announced gigawatts against a build rate near 30 a year and the gap is the thesis: either announcements deflate through cancellation, already the fate of $130bn of projects in one quarter, or the constraint binds harder than consensus assumes. Announced is not built; China's 20 to 30% utilization is the cautionary case

6. The prediction ledger

Twelve predictions with binding dates and bands from base rates. I assign no probabilities; pricing them is a human decision, kept outside the automated research pipeline behind this site.

#	Prediction	Band
P1	Palisades synchronizes to the grid on or before December 31, 2026	45 to 70%
P2	Palisades delivers contracted power to Wolverine and Hoosier by March 31, 2027	55 to 80%
P3	Crane synchronizes to the grid on or before December 31, 2027	40 to 65%
P4	US and Canadian capacity connected to the grid from units of 345 MWe or less, first synchronized after January 1, 2027, reaches at least 300 MWe by December 31, 2030	30 to 55%
P5	At least one MWh is delivered by December 31, 2030 from a new SMR built under a hyperscaler contract	15 to 40%
P6	GE Vernova reports gas equipment backlog plus slot reservations of at least 125 GW for the end of 2026	70 to 90%
P7	GE Vernova is publicly taking orders for 2032 turbine delivery by its Q2 2027 report	60 to 85%
P8	New capacity actually delivering under the nuclear contracts of the big four, excluding reallocated existing output, totals less than 3,000 MW by December 31, 2030	75 to 95%
P9	US commercial HALEU production licensed by the US Nuclear Regulatory Commission (NRC) reaches at least 5 tonnes in calendar 2029	25 to 50%
P10	TRISO-X's TX-1 plant begins commercial fuel fabrication by June 30, 2028	35 to 60%
P11	A binding final investment decision to complete V.C. Summer 2&3 is announced by December 31, 2028	25 to 50%
P12	A new reactor built under a hyperscaler contract delivers grid power by December 31, 2028. The thesis falsifier	2 to 10%

Verdict. Announced is not built. This ledger is a queue ticket, not a power plant. P12 is the falsifier: if a reactor built under a hyperscaler contract delivers grid power by 31 December 2028, speed was real and this paper was wrong.