$BE, $CRWV, $NBIS And $OKLO: What The Second Quarter Actually Revealed About The AI Power Boom

01 What Actually Happened

On August 11, after the close, CoreWeave reported second quarter results. On the morning of August 12, Nebius did the same. By the middle of the August 12 session CoreWeave was up 20.68% and Nebius up 17.02%, and almost every listed company that touches data centre power was moving with them: Vertiv up 5.78%, GE Vernova up 3.60%, Quanta up 3.48%, Vistra up 3.02%, Constellation up 2.31%. Bloom Energy, which had reported two weeks earlier on July 28, was up 11.76% on the same day.

The reflex reading is that the AI build-out is accelerating and everything attached to it goes up together. The numbers in the three releases say something more specific and more useful than that, because the three companies are exposed to identical demand through completely different business models, and the quarter exposed exactly how different those models are.

Read the chart from the bottom up rather than the top down. Oklo is down 33.3% this year and NuScale down 30.1%. These are the companies whose entire proposition is that artificial intelligence will need enormous quantities of firm, carbon-free power in the 2030s. That proposition has not become less true in the last twelve months. What has changed is that the market has stopped paying in advance for it.

Above them sit Constellation, down 19.4%, and Vistra, down 7.5%, which own large existing generation fleets. They have the electricity today. And above those, in a different world entirely, sit the companies that ship equipment, connect substations, sell cooling, and rent out graphics processing units by the hour.

The line separating the two groups is not technology and it is not carbon. It is whether there is an invoice.

02 Bloom Energy: The One That Turned The Demand Into Profit

Bloom Energy makes solid oxide fuel cells. They sit on the customer’s site, run on natural gas or hydrogen, and produce electricity without a combustion process and without waiting for a grid interconnection. For a data centre operator whose constraint is not capital but the queue for a substation, that last property is the entire product.

In the second quarter of 2026 the company reported revenue of $1,065.4 million, its first quarter above a billion dollars, up 165.5% from $401.2 million. Product revenue was $935.4 million, up 215.4%.

The composition matters more than the headline. When a hardware company grows this fast, margin normally compresses: it buys components in a hurry, pays overtime, and discounts to win the reference account. Bloom did the opposite. Gross margin was 33.4%, up 668 basis points from 26.7%. On a non-GAAP basis it was 34.3%, up 604 basis points. Operating income was $182.2 million against an operating loss of $3.5 million a year earlier, a swing of $185.7 million. Net profit attributable to common stockholders was $196.3 million against a loss of $42.6 million.

Cash followed the accounting, which is the part that separates this from most growth stories in the sector. Operating activities generated $226.4 million of cash in the quarter, against $213.1 million consumed a year earlier, a swing of $439.5 million. Diluted earnings per share were $0.62 against a loss of $0.18; on a non-GAAP basis $0.78 against $0.10. The company raised full year 2026 revenue guidance to a range of $3.9 billion to $4.2 billion, which is roughly 100% growth at the midpoint, with non-GAAP gross margin around 34%.

The sequential progression is worth setting out, because it shows a curve rather than a single good quarter. Revenue went from $751.1 million in the first quarter of 2026 to $1,065.4 million in the second. Operating income went from $72.2 million to $182.2 million. Adjusted EBITDA went from $143.0 million to $253.4 million, against $41.2 million in the second quarter of 2025.

Chief executive KR Sridhar framed the demand side in a sentence that is worth quoting precisely because it is a claim about adoption rather than about technology: “all the major US hyperscalers and over a dozen US neoclouds, AI labs, and colocation data center operators have validated and approved our power solutions for their AI factories.”

03 CoreWeave: Revenue Doubled, And So Did The Interest Bill

CoreWeave rents graphics processing units. It buys the hardware, houses it, powers it, cools it, wraps software around it, and charges by consumption. Second quarter revenue was $2,575 million against $1,212 million a year earlier, growth of 112.5%. Revenue backlog stood at approximately $104 billion at June 30, 2026, and the company disclosed more than $25 billion of net new customer commitments added in early in the third quarter, which sits outside that figure.

Now the other side. Operating expenses were $2,624 million against revenue of $2,575 million, so the company recorded an operating loss of $49 million against operating income of $19 million a year earlier. Below that line, net interest expense was $640 million for the quarter, against $267 million a year earlier. The net loss was $626 million against $290 million, and basic and diluted loss per share was $1.14 against $0.60.

Read those two paragraphs together and the shape of the business becomes clear. CoreWeave’s interest expense in a single quarter is now roughly a quarter of its revenue, and it is growing faster than revenue: revenue rose 112.5% while net interest expense rose 140%. Adjusted EBITDA was $1,510 million at a 59% margin, which sounds transformative until one notices that adjusted operating income, which subtracts depreciation, was $128 million against $200 million a year earlier, and that the adjusted operating margin fell from 16% to 5% while revenue doubled.

That gap between adjusted EBITDA and adjusted operating income is not an accounting quirk. It is the cost of the graphics processors wearing out. In a business where the principal asset has a useful life measured in a handful of years and is replaced by a faster generation on a roughly annual cadence, depreciation is a real economic expense and excluding it does not make it go away.

How the buildout is funded

The financing disclosures in the same release make the mechanism explicit. During the quarter CoreWeave completed a $3.1 billion term loan, described as the first publicly syndicated delayed draw facility backed by high performance computing infrastructure. It received a $1 billion strategic investment from Jane Street. And it raised more than $10 billion of unsecured debt and convertible bonds, including its first Eurobond issue.

On the operating side, active power expanded by nearly 500 megawatts during the quarter to reach 1.5 gigawatts, and total contracted power reached approximately 3.7 gigawatts. The company also completed the industry’s first bring-up and validation of NVIDIA’s Vera Rubin NVL72 platform, and was selected for inclusion in the Nasdaq-100.

04 Nebius: Growth Of 454%, And $5.66 Billion Of Capital Expenditure In One Quarter

Nebius Group is an Amsterdam-headquartered AI cloud company listed on Nasdaq, led by founder and chief executive Arkady Volozh. Second quarter revenue was $582.3 million against $105.1 million, growth of 454%. For the six months, revenue was $981.3 million against $156.0 million, growth of 529%. Adjusted EBITDA turned positive at $236.2 million against a loss of $21.0 million.

The operating leverage is genuine. Cost of revenues fell from 29% of revenue to 23%. Product development fell from 41% to 33%. Sales, general and administrative fell from 65% to 30%. Depreciation and amortisation fell from 72% to 45%. Total operating costs fell from 206% of revenue to 130%. Every one of those ratios moved in the right direction, and by a lot.

Two lines complicate the picture. Share-based compensation was $102.5 million in the quarter against $14.7 million a year earlier, growth of 597%, and now represents 14% of total operating costs against 7%. And the reported net result from continuing operations was a loss of $190.4 million against income of $502.5 million in the second quarter of 2025, a comparison distorted by the prior-year gain on the Toloka transaction; on an adjusted basis the net loss narrowed to $33.2 million from $91.5 million.

The two cash flow lines that matter

Net cash provided by operating activities from continuing operations was $2,246.1 million in the quarter, against $167.7 million used a year earlier. For the six months it was $4,504.1 million provided against $352.0 million used. A company with $582 million of quarterly revenue generating $2.25 billion of operating cash in the same quarter is not generating it from operations in the ordinary sense. It is collecting cash in advance from customers who want capacity reserved.

And that cash goes straight back out. Purchases of property, equipment and intangible assets were $5,657.4 million in the quarter, up 1,008%, and $8,130.3 million across the six months, up 671%. The company spent roughly ten times its quarterly revenue on fixed assets in the same three months.

There is nothing improper about that pattern and it is how capacity gets built. What it means for a reader is that the reported profitability of a business at this stage is close to meaningless as a guide to its economics. The depreciation on $8.1 billion of assets bought in six months has barely begun to run through the income statement. The 45% depreciation ratio that looks like an improvement today is a ratio calculated on a much older, much smaller asset base.

Shares issued and outstanding at June 30, 2026 were 271,855,218, comprising 238,400,165 Class A and 33,455,053 Class B shares, excluding 50,185,726 Class A shares held in treasury.

05 The Divergence The Market Is Actually Pricing

Put the three quarters side by side and the difference is not growth. All three grew. The difference is what each one had to consume to grow.

Bloom sells a box. The customer pays for the box, Bloom books a margin on it, and the cash arrives. The capital intensity sits on the customer’s balance sheet, not on Bloom’s. That is why a 165% revenue increase produced a $185.7 million swing in operating income and a $439.5 million swing in operating cash flow.

CoreWeave and Nebius own the asset. They buy the processors, they carry the debt or the customer prepayments, they absorb the depreciation, and they earn a spread on utilisation over a hardware generation that lasts a few years. That model can work extremely well and it can also be brutal, because the spread has to cover both the cost of capital and the obsolescence of the asset, and the second of those is not under management’s control.

Why this is not a judgement about quality

None of this makes Bloom a better company than CoreWeave or Nebius. They are not in the same business and they are not competing for the same dollar. It does mean that the three cannot be valued on the same multiple of anything, and that the habit of trading them as a single basket, which is exactly what happened during the August 12 session, is a habit rather than an analysis.

The one comparison that does hold across all three is the direction of the demand signal. A hardware company raising full year guidance to double, a compute provider reporting a $104 billion backlog, and a cloud provider spending ten times quarterly revenue on capacity in a single quarter are three independent confirmations of the same thing. Whatever else is uncertain, the demand is not being imagined.

06 Oklo, NuScale And The Names That Were Paid In Advance

The clearest way to understand what the market rewarded this quarter is to look at what it stopped rewarding.

Oklo is down 33.3% year to date and 38.98% over twelve months. NuScale Power is down 30.1% year to date and 74.04% over twelve months. Both design small modular reactors. Both have credible engineering, real regulatory engagement and a plausible role in supplying firm, carbon-free power to data centres in the next decade. Neither has commercial revenue. Oklo’s reported operating margin is a number so large in the negative that it is arithmetically meaningless: a rounding of costs against a revenue base close to zero.

Twelve months ago these were among the strongest performers in the complex, on precisely the argument that has since been validated. The demand thesis did not fail. The willingness to pay for it years before delivery did.

The middle of the table is the interesting part

Constellation Energy is down 19.4% year to date and Vistra down 7.5%, despite owning exactly the generation that data centres are competing for. Both were among the most crowded expressions of the AI power trade in 2025. The reason they have lagged is instructive: owning existing generation gives exposure to power prices, not to growth. A merchant generator with a fixed fleet cannot sell more electricity because demand rose; it can only sell the same electricity at a higher price, and even that is capped by contract structures, regulatory intervention and the pace at which new supply arrives.

Meanwhile Quanta Services, up 64.4% year to date, builds the transmission and substation work that connects a data centre to the grid. GE Vernova, up 60.4%, sells the turbines and grid equipment. Vertiv, up 84.0%, sells the power distribution and thermal management inside the building. All three convert the same demand into an order book, a delivery and an invoice within a normal business cycle.

07 Where The Contract Actually Sits

The single most useful question to ask about any company in this complex is narrow and answerable: between the electricity and the graphics processor, where does this company’s contract sit, and how long is it?

Behind the meter

Bloom sits behind the meter, on the customer’s site. It is not competing with the grid; it is competing with the queue. The scarce resource in United States data centre development is not generating capacity in the abstract, it is an interconnection agreement and a delivery slot for a transformer. A product that bypasses both has a value that is not a function of the price of electricity at all. That is why Bloom’s margin expanded while it grew: it is selling time to market, and time to market is the thing in shortest supply.

The compute layer

CoreWeave and Nebius sit at the other end, selling compute by consumption. Their contracts are long, their backlogs are large, and their capital is committed years before the revenue is recognised. CoreWeave’s disclosure of 1.5 gigawatts of active power and 3.7 gigawatts contracted is the clearest public statement anyone in this sector has made of the physical scale involved: 3.7 gigawatts is comparable to the output of several large power stations, contracted by a single company that was not public two years ago.

The grid layer

Between the two sit the equipment and construction names, whose contracts are the shortest and most conventional: an order, a delivery, a payment. This is why they have performed well without being volatile in the way the compute names are. They carry execution risk, not obsolescence risk and not financing risk.

What this means for reading the next quarter

For the equipment names, watch the order book and gross margin. For the behind-the-meter names, watch whether product revenue keeps growing faster than total revenue, because that is what drove Bloom’s margin expansion. For the compute names, watch three lines specifically: interest expense against revenue, adjusted operating income against adjusted EBITDA, and capital expenditure against operating cash flow. And for the pre-revenue nuclear names, the only thing that changes the picture is a signed power purchase agreement with a delivery date, at a price, with a counterparty named.

08 What To Watch From Here

These are checkpoints drawn from the filings, not predictions, and none of them carries a probability.

09 The Comparison The Trade Does Not Want To Make

Anyone who has followed a capital-intensive build-out before will recognise the shape of this quarter, and the comparison most often raised by sceptics is the fibre optic build of 1999 to 2001. Their argument deserves an accurate hearing, because it is not a frivolous one, and it also deserves to be tested against where the analogy breaks.

The case they make runs as follows. A genuine technological shift creates genuine demand. Capacity is built ahead of that demand because whoever has capacity first captures the customer. The build is financed with debt and with vendor arrangements rather than with operating profit. Backlog and contracted capacity are quoted as evidence of future revenue. Then demand grows more slowly than capacity, utilisation falls, pricing collapses, and the companies that borrowed to build are left servicing debt against an asset that has become abundant. On that reading, a $104 billion backlog and 3.7 gigawatts of contracted power are not reassurance but the measure of the commitment that has to be honoured either way, and $640 million of quarterly interest is the running cost of having made it.

Three things in the current data cut against that parallel, and they should be stated with the same care.

The first is that the asset depreciates rather than lasting. Fibre laid in 2000 was still usable in 2010, which is precisely why the glut persisted for a decade. A graphics processor is functionally obsolete for frontier training within a few years. Overcapacity in this cycle self-corrects far faster, and the cost of that correction shows up as depreciation rather than as a decade of dead pricing.

The second is that the buyers are different. The dark fibre build was substantially financed by carriers with weak balance sheets buying from vendors who lent them the money. The counterparties in CoreWeave’s backlog, on the disclosure available, include hyperscalers and large enterprises with substantial cash generation of their own. That does not make the contracts riskless, but it changes who absorbs a downturn.

The third is the constraint itself. In 1999 the binding constraint was capital, and capital was abundant. In 2026 the binding constraint is electricity and interconnection, and neither is abundant. A shortage of the scarce input is a very different environment from a glut of the built output, and it is the reason a company like Bloom, which sells a way around the queue, can expand margin while growing revenue 165%.

Where the sceptical case retains force is on financing structure rather than on demand. Nebius collected $2.25 billion of operating cash in a quarter in which it recognised $582 million of revenue, and spent $5.66 billion. CoreWeave raised more than $10 billion of unsecured debt and converts, plus a $3.1 billion term loan, in three months. Those are the numbers to watch, not the demand headlines, because the demand is visible and the financing is the part that has to keep working.

10 Bottom Line

The second quarter of 2026 did not tell us whether artificial intelligence will need enormous quantities of electricity. Three independent sets of accounts confirmed that it already does: a fuel cell manufacturer doubling revenue and raising guidance, a compute provider carrying a $104 billion backlog against 3.7 gigawatts of contracted power, and a cloud company spending $5.66 billion on fixed assets in three months. That question is settled for the purposes of reading a balance sheet.

What the quarter did tell us is that the same demand produces radically different financial outcomes depending on where a company stands in the chain, and that the market has become considerably more discriminating about the difference. Bloom Energy converted the demand into $196.3 million of net profit and $226.4 million of operating cash in a single quarter. CoreWeave converted it into $2,575 million of revenue and a $626 million net loss, with $640 million of that gap explained by interest. Nebius converted it into 454% revenue growth, $2.25 billion of customer cash collected in advance, and $5.66 billion spent immediately on the capacity to serve it.

All three are rational responses to the same opportunity. Only one of them currently pays for itself out of operations, and it is the one that does not own the data centre.

What is verified. Bloom Energy: Q2 2026 revenue $1,065.4 million, up 165.5%; product revenue $935.4 million, up 215.4%; gross margin 33.4%; operating income $182.2 million; net profit to common $196.3 million; diluted EPS $0.62; operating cash flow $226.4 million; adjusted EBITDA $253.4 million; FY2026 guidance raised to $3.9-4.2 billion. CoreWeave: revenue $2,575 million against $1,212 million; operating loss $49 million; net interest expense $640 million; net loss $626 million; diluted loss per share $1.14; adjusted EBITDA $1,510 million at 59% margin; adjusted operating income $128 million against $200 million; backlog approximately $104 billion at June 30, 2026, plus more than $25 billion of commitments added in early Q3; active power 1.5 GW, contracted power approximately 3.7 GW. Nebius: revenue $582.3 million against $105.1 million; adjusted EBITDA $236.2 million; total operating costs $758.2 million, 130% of revenue; share-based compensation $102.5 million; operating cash flow $2,246.1 million; capital expenditure $5,657.4 million; shares outstanding 271,855,218 at June 30, 2026.

What is not verified. Bloom’s characterisation of hyperscaler adoption, which is a commercial claim and not a disclosed contract list. CoreWeave’s and Nebius’s forward guidance, which was given on their respective calls rather than in the releases read for this piece. Nebius’s operating income, which the release does not state as a line and which is shown here as implied arithmetic from disclosed revenue and cost figures. Bloom’s and CoreWeave’s capital expenditure, not disclosed in their releases. And every share price percentage quoted here, which is a Finviz reading of August 12, 2026 and will be wrong by the time this is read.