Bloom Energy posted on X on 24 September that Oracle "remains committed to Project Jupiter and its contract with Bloom to deliver 2.4 GW of fuel cell capacity" — twice the 1.2 GW Bloom said in April had been contracted, and the fourth gigawatt figure on one relationship. Spread over the nine quarters to the campus's 2028 target, the smallest of the four claims 27% of the quarterly output Bloom's factory is guided to reach by the end of 2026, and the largest 62%. No filing either company has made says which is the order.
Oracle, STACK and Bloom all say the contract stands
Oracle sent a force majeure notice on 24 September to Blue Owl Capital, owner of STACK Infrastructure, developer of Project Jupiter in Doña Ana County, New Mexico. Bloomberg reported it first: the notice preserves Oracle's protections if the campus misses its 2028 opening. Nobody outside the parties has seen the document. Same campus as the $18 billion of loans quoted at 89 to 91 cents eight days earlier — a piece about whose debt that is, silent on the size of Bloom's order on top of it.
Oracle vice-president Michael Egbert told Engineering News-Record that "Project Jupiter remains on our planned schedule" and that such notices "do not, by themselves, establish a project delay." STACK called itself, Blue Owl and Oracle "fully aligned", with no change to financial commitments. Bloom said Oracle stays committed.
Two of the four are ceilings, one is a contracted order, one is a post
Bloom's 13 April release says Oracle "intends to procure up to 2.8 gigawatts" under a master services agreement, of which "an initial 1.2 GW of capacity has been contracted" — across Oracle projects in the US, with Jupiter never named. Oracle's 27 April redesign, which swapped gas turbines for one Bloom microgrid, put installed capacity at the site at up to 2.45 GW. The April air-permit application to New Mexico specifies 2,462 MW.
Bloom's post is the only one to call a number "its contract" flat, with no "up to" and no "initial." ENR asked Bloom about construction status, permitting and commissioning; Bloom declined and pointed back at the post.
Nine quarters, and the answer swings by a factor of 2.3
Bloom's 10-K holds the one denominator that survives all four. Fremont's annual production capacity run rate is about 1 GW today and 2 GW by the end of 2026 — 500 MW a quarter — with nothing beyond the 2 GW committed.
October 2026 to the end of 2028 is nine quarters. Divide by nine:
- 1,200 MW ÷ 9 = 133 MW a quarter, 27% of 500
- 2,400 MW ÷ 9 = 267 MW a quarter, 53%
- 2,462 MW ÷ 9 = 274 MW a quarter, 55%
- 2,800 MW ÷ 9 = 311 MW a quarter, 62%
Dividing flat is the kindest reading — Oracle expects first capacity in the first half of 2027, which front-loads the draw — and even so the ceiling fits. The spread is the finding. Whether one site takes a quarter of the plant or two-thirds of it for nine straight quarters is a 2.3x difference in the largest claim on Bloom's only factory. At today's 250 MW a quarter, that 267 MW is 107% of everything Fremont builds.
The filings carry $442.4 million and no gigawatts at all
"Jupiter", "New Mexico" and "2.8 GW" appear in zero Bloom filings, ever. The 13 April 8-K, filed the day of the release, discloses the Oracle warrant and no volume. One order figure is filed: $442.4 million of unsatisfied performance obligations at 30 June, mostly product and installation, due within one to two years, plus $51.7 million of service — 11% of the low end of the $3.9–4.2 billion FY2026 guide, under half of June-quarter product revenue alone.
Fairly, the line may not hold the order: Bloom excludes contracts under a year and those billed as invoiced, and says it can power a data centre within 90 days. Granted — which is why the line rising 12% from December while product revenue tripled proves nothing about Jupiter either way, and why nothing filed sizes this order.
February's ~$6.0 billion "product backlog" is a different unit: Bloom's own measure includes anticipated tax credits, and the two sit 13.6x apart. One filed proxy for Oracle progress exists — $17.9 million of the warrant's $324.4 million fair value, 5.5%, had reduced revenue as Oracle systems shipped by 30 June. Bloom does not say what volume that is 5.5% of.
What settles it is the Q3 performance-obligation line
Bloom's Q3 10-Q decides it. Name Jupiter, or move the product performance-obligation line by billions, and 2.4 GW is a purchase order — 53% rather than 27% of the 500 MW a quarter our model assumes, the concentration its August bear case already flagged. Leave the line alone and 2.4 GW is allocation inside a framework, April's 1.2 GW still the only order Bloom has. On Oracle, already published: a two-quarter delay to every megawatt costs 3.6% of base fair value. New Mexico decides the air permit by 23 November; neither company has filed an 8-K.
Bloom's 10-Q to 30 June 2026 holds the performance obligations and the warrant; the FY2025 10-K the Fremont run rates; the deck filed 5 February 2026 the product backlog and its tax-credit definition. The contracted and ceiling gigawatt figures are Bloom's 13 April release and Oracle's 27 April redesign statement; the 2.4 GW wording is Bloom's own X post of 24 September. Engineering News-Record of 28 September carries the permit application, the November deadline and the Egbert and STACK statements. The notice itself is Bloomberg of 24 September, carried by Reuters, CNBC and the FT; none has published it. Filing absences were checked on EDGAR on 30 September. Every division and factory share is ours, as are the model figures of 23 August and 10 September.
