The last 24 hours put a price on the physical work behind AI demand. Nebius said its largest new cloud contracts averaged more than $1 billion in total value and $20 million to $25 million of annual contract value per megawatt. Customers increasingly paid before the capacity existed. At the same time, RADCOM blamed delayed customer infrastructure for a sharp revenue decline, and a power-semiconductor supplier still reported a loss despite strength in AI and server applications.
In brief, for the window from 12 August 09:03 to 13 August 09:03 in Tehran:
- Nebius reported $582.3 million of second-quarter revenue, $5.7 billion of capital expenditure and a financing mix spanning customer prepayments, secured debt and new shares.
- RADCOM's revenue fell 33.4% as component costs and supply constraints delayed customer deployments, even as it announced three post-quarter contracts.
- Alpha and Omega Semiconductor said AI and server demand supported its advanced-computing business, but quarterly revenue remained below the prior year and GAAP operating loss was $11.0 million.
- Usio processed $2.47 billion across its payment channels, up 27%, while its gross margin and first-half operating cash flow showed why volume alone is not the whole fintech result.
| Disclosure | In-window evidence | Operational reading |
|---|---|---|
| Nebius | $582.3m revenue; $5.7bn quarterly capex | Prepayments reduce, but do not remove, delivery and financing risk |
| RADCOM | $11.8m revenue, down 33.4% | Customer infrastructure can delay software revenue after demand is signed |
| AOS | $170.4m revenue; $11.0m GAAP operating loss | AI mix does not by itself establish profitable semiconductor growth |
| Usio | $2.47bn processed; $23.7m revenue | Payment volume must still convert into margin and operating cash |
These are company disclosures, not a common industry dataset. The useful connection is operational: contracted demand becomes economic output only after financing, power, hardware, deployment and payment rails all work. Nothing below is investment advice.
Nebius turned customer demand into a financing instrument
Nebius's 12 August earnings release reported group revenue of $582.3 million, up 454% from $105.1 million a year earlier. Its core AI cloud business generated $574.9 million, about 98% of the group total. The company reported $236.2 million of group adjusted EBITDA, while the AI cloud business reported $285.7 million and a 49.7% adjusted EBITDA margin.
The matching shareholder letter filed on 12 August supplied the more consequential commercial detail. Nebius said it closed four cloud deals with an average total contract value above $1 billion. Roughly 70% of second-quarter deals included prepayments, which management estimated would cover 50% to 60% of associated capital expenditure. It put the estimated payback period for those deals at one year and ten months, based on forecast costs and capacity that is not all built yet.
That structure changes the meaning of backlog. A prepayment is not merely evidence of interest: it transfers some construction funding from the supplier to the customer and creates a delivery obligation in return. Nebius said it expects more than $9 billion of customer prepayments in 2026 and has more than $40 billion of customer commitments. Those figures are forward-looking company measures, not recognised revenue.
This is the next step beyond the constraint picture in yesterday's CoreWeave and Lumentum briefing. AI infrastructure is no longer described only through revenue growth or backlog. It is being contracted through megawatts, prepayments, hardware generations, delivery dates and financing covenants.
The build still consumed more cash than the quarter produced
Nebius spent approximately $5.7 billion on capital expenditure in the quarter, primarily on GPUs, related hardware and data-centre expansion. Its reported purchases of property, equipment and intangible assets reached $5.657 billion, against $2.246 billion of operating cash provided by continuing operations. For the first half, those purchases were $8.130 billion.
The company ended June with $8.042 billion of cash and cash equivalents, but it also showed $8.499 billion of non-current debt plus $46.7 million current. In July it raised a $775 million secured facility at SOFR plus 2.50%, backed by deployed GPU infrastructure and contracted cash flows. During the second quarter it also sold 12.7 million Class A shares through its at-the-market programme for about $2.8 billion of gross proceeds.
The formal operating and [financial review filed later on 12 August](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d1.htm) provides the necessary counterweight to the adjusted figures. Nebius recorded a $175.9 million GAAP operating loss and a $190.4 million net loss from continuing operations for the quarter. Depreciation and amortisation reached $259.7 million, reflecting the expanding hardware base. Adjusted EBITDA is informative about management's view of core operations, but it does not pay for the GPU fleet or replace the GAAP statement.
Operationally, teams evaluating an AI cloud should separate five records:
- contracted power, which may not yet be energised;
- built and commissioned capacity, which may not yet be allocated;
- customer prepayments, which are cash received with performance obligations;
- recognised revenue, which follows delivery and contract accounting; and
- operating cash after the infrastructure and financing cycle.
Combining those records into one “demand” number hides the delivery risk.
Five gigawatts is a target, not a live fleet
Nebius raised its target for year-end 2026 contracted power to 5 GW and said it plans to deploy more than 1 GW of capacity per year beginning in 2027. It also described an asset-light model in which partners finance and own data-centre assets while Nebius supplies architecture, software and commercial demand.
Those statements matter, but their verbs are different. “Contracted” refers to land and power commitments under the company's definition. “Deploy” concerns the future pace of bringing systems into service. Neither means that five gigawatts is currently energised, filled with accelerators, commissioned and billing customers.
The shareholder letter also says a majority of the second-quarter deals relate to capacity arriving late in 2026 and contributing mainly to 2027 revenue. The $20 million to $25 million annual-contract-value range per megawatt is therefore a disclosed deal-economics measure, not a rate that can be multiplied by every contracted megawatt today. The same discipline applies to the four-ledger distinction in our AI demand and model-testing analysis: bookings, cash, delivered service and accounting revenue answer different questions.
RADCOM showed the distance from order to deployment
RADCOM's 12 August results supplied a smaller but unusually clear counterexample. Quarterly revenue fell 33.4% to $11.8 million. The company said higher component costs and supply constraints slowed customers' infrastructure build-outs, delaying both deployment of RADCOM's network-assurance software and purchasing decisions.
Management characterised the delays as timing rather than lost demand. It pointed to three contracts signed after quarter-end, including a multi-year AI-driven assurance deployment at CETIN Networks in Slovakia, and maintained full-year revenue guidance of $57 million to $63 million. RADCOM also ended June with $109.7 million in cash, equivalents and short-term deposits and no debt.
The confirmed fact is the reported revenue decline and the stated cause. The claim that all demand is intact is management's interpretation, not yet an observed recovery. The next evidence is deployment: customer infrastructure installed, software accepted, revenue recognised and cash collected. A signed AI contract can still wait behind someone else's servers, networking and procurement cycle.
Power semiconductors gained AI mix without escaping the loss
Alpha and Omega Semiconductor's fiscal fourth-quarter release on 12 August reported revenue of $170.4 million, up 4.0% sequentially but down 3.5% year on year. Chief executive Stephen Chang said strength in advanced computing—particularly AI and server applications—and communications offset weakness in traditional PCs.
The financial boundary remains important. GAAP gross margin was 23.1%, and the company recorded an $11.0 million operating loss. It guided the September quarter to $176 million of revenue, plus or minus $10 million, and a 23.8% GAAP gross margin, plus or minus one percentage point. Management expects the advanced-computing mix and pricing environment to support higher margins later in calendar 2026.
The release does not disclose AI-specific revenue, units or gross profit. It supports a qualitative statement that AI and servers helped the mix; it does not support attributing the whole sequential increase to AI. For buyers, the operating point is that accelerator systems also require power-conversion parts, packaging and capacity, and those suppliers can face different margins from the cloud companies using their components.
Payments grew on a different constraint stack
Usio's 12 August fintech results reported $23.7 million of quarterly revenue, up 19%, on $2.47 billion of payment dollars processed and 17.9 million transactions, both up 27%. ACH transaction volume rose 34%, credit-card revenue rose 28%, and adjusted EBITDA increased to $1.1 million from $0.5 million. Usio raised its full-year revenue-growth expectation to 14%–16% from 10%–12%.
This was a fintech result, not an AI revenue disclosure. It belongs in the brief because payment execution is a downstream constraint for automated commerce and agentic workflows. The limits are visible too: gross margin fell to 24.2% from 25.8%, first-half operating cash flow declined to $0.3 million from $1.1 million, and cash ended June at $6.4 million. Processed value, transaction count, revenue, gross profit and cash conversion should remain separate measures.
What to watch next
The next proof points are mechanical rather than rhetorical:
- Nebius: receipt of the expected prepayments; energised and commissioned capacity; capital expenditure not covered by customers; debt and equity issued; concentration by customer; and the conversion of commitments into recognised revenue.
- RADCOM: completion of delayed customer infrastructure, software acceptance and a visible return of revenue rather than only post-quarter contract announcements.
- AOS: disclosed advanced-computing contribution, factory utilisation, gross margin and whether the September-quarter outlook is reached without another operating loss.
- Usio: gross margin, return rates, sponsor-bank and processor resilience, operating cash conversion and whether faster transaction growth produces durable earnings.
The day's strongest signal is not simply that AI demand is high. It is that customers are now helping finance physical capacity before delivery, while the suppliers around that capacity still show deployment delays, hardware-margin pressure and cash-conversion limits. The relevant unit of progress is the completed chain from committed capital to energised hardware to accepted service to collected cash.
