Oracle Expands Bloom Energy Deal to 2.8 GW, Days After $400 Million Stock Warrant

Four business days separated two Oracle disclosures that together explain how hyperscalers actually finance AI now. On Thursday, April 9, Bloom Energy issued Oracle a warrant to buy 3.53 million Bloom shares at $113.28 each. On Monday, April 13, the two companies announced that Oracle had expanded its commitment from roughly 500 megawatts of on-site fuel cell capacity to a 2.8 gigawatt ceiling, with 1.2 gigawatts already under firm contract. By Monday's close, Bloom shares were at $176.67. By after-hours, they were near $203. Oracle's warrant, which had been worth nothing five months earlier, was in the money by $316 million before it bought a single electron.

The financial press read this as a clean-tech story. Oracle bought more fuel cells, Bloom stock soared, move on.

That reading is wrong. The warrant is the whole point. The gigawatts are the cover story. What Oracle actually did on April 13 was use Bloom's equity market, not its own bond market, to finance gigawatts of AI data center power. You are watching the first public-market invoice for a financing structure the AI infrastructure business has been quietly building for six months. And it matters because Oracle cannot easily do this with debt anymore.

What the four-day gap reveals

Start with the sequence, because it is the argument.

The warrant was negotiated on October 28, 2025, under a master agreement filed via Bloom's Form 8-K two days later. The strike was locked to that day's closing price of $113.28, with a cap of 3.53 million shares and a final expiration of October 9, 2026. The instrument would not actually issue until five and a half months later, on April 9, 2026, carrying customary anti-dilution adjustments and registration rights. Bloom's own April 13 press release opens by confirming that April 9 issuance date.

Between October 28 and April 13, Bloom shares climbed 56%, from $113.28 to $176.67. The warrant went from an at-the-money option worth zero intrinsic value to a position worth roughly $224 million. That was before the 2.8 gigawatt announcement added another $92 million in paper gains by Monday night, taking CNBC's total to $316 million.

Then, four business days after the warrant finally struck, the expanded commitment landed on the wires. Not four months. Not four weeks. Four days. The timing is surgical. It is the cleanest possible signal that Oracle's public gigawatt commitment was always the thing that would validate the warrant's value, and that both parties timed the announcement to land once the instrument was officially in Oracle's hands.

Structured-finance readers will recognize the shape. This is a two-stage option trade dressed as a procurement contract. Stage one: supplier writes a strike-at-spot call to the customer, with the understanding that the customer's future volume commitments will re-rate the supplier's equity. Stage two: customer delivers the commitment, equity re-rates, option pays. The structure is legal. It is standard in biotech collaborations. At this scale, in physical-power procurement, it is new.

Oracle is the most debt-cornered hyperscaler

To understand why Oracle needed a structure like this, read the Oracle balance sheet.

As of early 2026, Oracle carried roughly $100 billion in on-balance-sheet debt, with some analyses putting the number closer to $124 billion after the February 2026 financing round. Add another $248 billion in off-book data center leases, most of them locked in for 15 to 19 years. Fiscal 2026 capex guidance came in at $50 billion. Last year's was $21 billion. The math doesn't lie. Free cash flow had turned sharply negative. Oracle disclosed plans to raise another $45 to $50 billion during the year through a balanced mix of debt and equity, including a $25 billion bond priced in February and an at-the-market equity program of up to $20 billion.

Sitting against all of that: a $553 billion remaining performance obligations backlog, up roughly 325% year over year, with an estimated $300 billion of it tied to a multi-year OpenAI compute contract.

The market has noticed. Oracle's credit default swap spreads doubled from roughly 40 to 55 basis points to 100 to 110 through late 2025. A securities class action filed in early 2026 alleges that management made misleading statements about AI capex returns. By late 2025, Oracle's investment-grade bonds had begun trading like junk, with Barclays modeling a cash shortfall by late 2026. Oracle's stock was down roughly 20% for the year by April 13, even after a 12.7% single-day rally that morning.

Oracle's problem is not that it cannot raise capital. It is that every incremental dollar of Oracle corporate debt sits on top of a $553 billion backlog whose largest counterparty is a private company with its own capital constraints. Another bond issuance makes the credit picture worse. Another equity issuance dilutes Larry Ellison. Neither is free.

A Bloom warrant costs Oracle nothing. Oracle paid no cash on October 28, and no cash on April 9 or April 13. The 1.2 gigawatts under firm purchase orders will cost Oracle capex, but that capex was already committed to data centers Oracle was already building. What the warrant delivered, effectively, was a $316 million unrealized mark-to-market gain that Oracle can narrate to credit analysts as evidence that its AI strategy is working. The expansion announcement delivered credible, deliverable gigawatts Oracle could not procure from any utility in the United States on a timetable that matches its OpenAI contracts.

Bloom's cap table is Oracle's new balance sheet

Here is the part the procurement framing misses entirely. Oracle did not pay Bloom for fuel cells using Oracle money. It paid Bloom with access to Bloom's own capital markets.

Bloom trades at more than 100x forward earnings. That multiple exists because the market has decided Bloom is a scarce bottleneck asset in the AI power crunch, the same way Nvidia's multiple re-rated once the market concluded GPU supply was the binding constraint on model training. Warrants issued against a 100x multiple are extraordinarily cheap financing for the party receiving them. Every dollar Oracle commits to Bloom gigawatts gets paid for, implicitly, in Bloom equity upside. And because Oracle is the reference customer whose commitments drive the re-rating, Oracle is the one collecting the spread.

Think about what this replaces. In a traditional utility power purchase agreement, Oracle would sign a 20-year offtake at regulated rates, pay cash or posted credit for delivered electrons, and absorb the full financing burden of the generation asset on its own cost of capital. Cost of capital for Oracle debt, today, is rising. Cost of capital for Bloom equity, today, is falling. The warrant arbitrages the difference. You could call it synthetic infrastructure financing. Oracle calls it a partnership.

Institutional emotions are on the table. Oracle's credit desk is cornered and defensive. Bloom's management is emboldened and, for the first time in the company's 23-year history, vindicated. And Bloom's long-suffering retail shareholders, who sat through a decade of $10-to-$25 trading ranges, are watching their thesis pay off in real time.

The template that just went public

The broader signal is the one worth paying for. AI power procurement is no longer a utility PPA business. It is becoming a structured-finance business, and the tools are the ones semiconductor buyers used during the 2024-2025 GPU shortage: volume commitments, price protection, delivery guarantees, and equity kickers. Nvidia built its CoreWeave relationship this way. OpenAI's tangle of equity ties to Oracle-adjacent special purpose vehicles works the same way. Brookfield's $5 billion co-development model with Bloom already blurred the line between offtake contract and joint venture. The April 13 warrant is the first time a major hyperscaler has publicly received an equity instrument as part of a gigawatt-scale power deal.

Expect more. GE Vernova, Siemens Energy, Oklo, Fermi America, and Bloom itself should each receive warrant-linked term sheets from hyperscalers whose balance sheets cannot absorb another direct bond issuance. Traditional utility PPAs will look anachronistic for AI workloads by year-end. Regulators, who still think of data center power as a rate-case problem, are about to discover that hyperscalers have quietly opted out of the regulated utility model while they were writing the 2030 capacity docket.

Who wins, who loses

Bloom wins unambiguously. The Oracle ceiling alone would consume about 1.4 years of Bloom's fully expanded 2 gigawatt annual manufacturing capacity at peak utilization, on top of the $5 billion Brookfield partnership and the $2.65 billion AEP deal at the Cheyenne AI Factory in Wyoming. Bloom's 2026 revenue guidance of $3.1 to $3.3 billion suddenly looks conservative. Simon Edwards, whom Bloom named CFO effective April 13, is taking the job at precisely the right moment.

Oracle wins conditionally. The warrant gain is real on paper. The time-to-power advantage is real on paper. But Oracle is still the most debt-constrained hyperscaler, its biggest customer remains a private company with its own uncertainties, and its CDS spreads are still screaming. Bloom gigawatts help at the margin. They do not refinance the $100 billion debt pile.

Utilities lose slowly and quietly. Every gigawatt Oracle procures from Bloom is a gigawatt it does not ask PJM, ERCOT, or CAISO to deliver. The interconnection queue that now stretches seven to ten years in key markets just lost its most motivated customer.

The test that matters

Here is the forward-looking test. If Oracle exercises the warrant before October 9, and if Bloom's 1.2 gigawatts of initial capacity are delivered on the stated "into next year" timeline, the structure becomes the template. Every hyperscaler with a bruised credit profile will want a Bloom-style warrant against a scarce power supplier. If Bloom slips on manufacturing ramp, or if Oracle's AI capex cycle falters before the warrant is exercised, the instrument becomes a case study in how the 2026 AI infrastructure buildout mispriced its own risk.

Either way, forget the gigawatts. Watch the next warrant. Because the one that printed on April 9 was not an accident. It was a prototype.

Mistral Borrows $830 Million to Build Its Own AI Data Center Near Paris

Mistral AI secured $830 million in debt financing on Monday to build a 44-megawatt data center south of Paris, the French startup's first move into debt markets since its founding in April 2023, accor

The Implicator

Oracle Racks. The Pentagon Overreaches. Meta Buys Broken.

San Francisco | Wednesday, March 11, 2026 Oracle posted its best quarter in 15 years and raised its 2027 forecast to $90 billion. The stock jumped. The backlog hit $553 billion. What nobody celebrate

The Implicator

Trump Tells Tech Companies to Build Own Power Plants for AI Data Centers

President Donald Trump announced a "ratepayer protection pledge" during his State of the Union address on Tuesday, telling major technology companies they must provide their own electricity for data c

The Implicator

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]