Bloom Energy reports second-quarter results after Tuesday’s close, with a management call at 5 p.m. ET. Six weeks ago that would have been a victory lap. The stock had run roughly 247% in the first half of 2026 on a single, very good idea: AI data centers need power now, grid interconnections take years, and Bloom’s solid oxide fuel cells arrive on a truck and switch on in months.
Then June 25 happened. The stock peaked at $351.28 and has spent the month since giving back roughly 47% of itself — about 40% in the last four weeks alone. Nothing in the business broke. What broke was the market’s willingness to take management’s numbers at face value.
Three things did it. On July 8, short-seller Hunterbrook Media published a report alleging Bloom sources scandium oxide — the dopant that makes its ceramic electrolyte efficient enough to sell — from Chinese suppliers routed through intermediary countries, contradicting the company’s public position that it has no China supply chain. Bloom fired back in a July 9 8-K calling the claims “false and misleading” and saying its supply chain supports 25 GW of annual production. A follow-up report, and then a second rejection of the natural gas pipeline feeding Oracle’s Project Jupiter campus in New Mexico — New Mexico’s public lands commissioner turned down Energy Transfer’s 17-mile line again on July 14, citing emissions and water — kept the pressure on. That campus was designed around up to 2.5 gigawatts of Bloom cells.
The $492 million question
The number the bears keep returning to is the gap between the “$20 billion contracted backlog” Bloom cites in presentations and the roughly $492 million in remaining performance obligations disclosed in its SEC filings.
Both can be true, and mostly are. RPO is a GAAP construct: only revenue the company is contractually obligated to recognize under currently enforceable terms at prices already agreed. Bloom’s Energy Servers ship with 10-to-15-year service contracts, and project-financed deals don’t convert into audited RPO until financial close. The $20 billion is a commercial-momentum figure. The $492 million is the part an auditor will sign.
That distinction is fair. It is also exactly why the composition question matters more than the headline: how much of the $20 billion sits behind binding take-or-pay terms versus milestone-dependent structures that slip when a pipeline permit gets denied twice.
Our take: This is the first AI-infrastructure earnings report where the interesting number isn’t revenue — it’s conversion. Every AI power story of the last year has been an announcement: gigawatts reserved, master agreements signed, campuses planned. Bloom is where announcements have to become invoices first, because it actually ships hardware. Whatever Tuesday reveals about how backlog turns into recognized revenue is a preview of what the rest of the buildout will look like when the same question reaches everyone else.
The second-half math
Bloom raised full-year guidance after a record Q1 — revenue up 130.4% to $751.1 million — to a range of $3.4 billion to $3.8 billion. Wall Street’s consensus for Q2 is roughly $766.9 million and $0.39 a share. Hit that, and the back half still has to deliver around $2.08 billion to reach the midpoint. That is roughly 36% above the implied first-half run rate, and it requires the Oracle, Nebius and Brookfield-backed pipelines to accelerate at once.
The bull case hasn’t vanished. JPMorgan raised its target to $346 from $267 this month; Baird, UBS and Susquehanna have all reiterated positive ratings. Industrial Development Funding and Oaktree committed $1.7 billion on July 16 to fund Bloom cells across Nebius’s AI cloud buildout — capital allocators voting with balance sheets, not press releases. And the physics still favors the product: roughly 54% electrical efficiency at point of use against 35–40% for open-cycle gas turbines, no grid interconnection required, no rotating mass to limit how fast it tracks a load that can swing from 20% to 150% in milliseconds.
Options are pricing a 27–32% move in either direction off Tuesday’s print, against a historical average nearer 12%. That is not noise. That is a market that no longer knows which story it’s holding.
What to watch
- Guidance. Any trim to the $3.4–$3.8 billion range concedes the second-half ramp isn’t happening on schedule. Reaffirming it without named-account delivery timing won’t settle anything.
- Project Jupiter. Does the 2.8 GW Oracle master agreement protect revenue if physical deployment slides into 2027?
- A scandium tonnage number. Bloom has given a capacity claim (25 GW/year) but never a quantity. One figure investors can cross-check against independent supply estimates ends this argument in a sentence.
- Backlog structure. Take-or-pay share of the $20 billion. That single disclosure is worth more than the quarter.
The broader read-through is the one we flagged when AI spending stopped working as a buy signal and when Moody’s put a credit lens on Big Tech’s bill: the market has moved from rewarding announced capacity to auditing delivered capacity. Bloom is simply first in line, and it isn’t alone — FuelCell Energy has a Siemens co-development deal and PowerCell just signed 300 megawatts with a data center operator. The category Bloom created is getting crowded at precisely the moment its own numbers are being read with a magnifying glass.
Tuesday, 5 p.m. ET. The gigawatts are contracted. The invoices are the story.