AI Demand Hit Four Ledgers as US Model Tests Stayed Private

Four filings published after the US market close put concrete numbers around AI and digital-finance demand. Palantir reported rapid software growth, onsemi said AI data centres were its fastest-growing business, Backblaze showed storage growth and a large backlog, and Paymentus processed more than 213 million payment transactions. The White House meanwhile said its voluntary frontier-model test framework was complete but did not publish it.

This brief covers material published or substantively announced between 3 August 2026 at 09:02 Iran time and 4 August 2026 at 09:02 Iran time (05:32 UTC to 05:32 UTC). It separates reported results from management attribution, and new disclosures from older events that appeared inside those results.

• Palantir's revenue rose 93% year on year, but the filing does not isolate revenue from its AI platform.
• onsemi reported stronger revenue and cash flow; its AI data-centre growth claim is management commentary, not a separately audited segment.
• Backblaze's storage backlog jumped after a CoreWeave agreement signed in June, so the contract is not new August news.
• Paymentus provides a useful finance-side control case: its release attributes growth to billers and transactions, not to AI.
• Washington's completed model-testing framework remains operationally opaque outside the companies and officials involved in its development.

The evidence shows demand moving through several ledgers. It does not establish one clean “AI revenue” number or constitute investment advice.

The 24-hour signal

The operating decision is to keep those ledgers separate. Our production AI cost-stack guide explains why software, compute, storage and control costs should not be compressed into one model-price figure. The same discipline applies to demand: a contract, a shipment, a transaction and a benchmark are different units.

Palantir put a hard number on the software layer

Palantir's 3 August earnings release filed with the SEC reported second-quarter revenue of $1.935 billion, up 93% from a year earlier and 19% sequentially. US commercial revenue was $764 million, up 149% year on year, while US government revenue was $809 million, up 90%. The company closed 220 deals worth at least $1 million, including 73 worth at least $10 million.

Total contract value booked in the quarter was $3.373 billion. US commercial remaining deal value reached $6.238 billion, up 124% year on year. Palantir also reported $1.220 billion of adjusted free cash flow and raised full-year revenue guidance to $8.150 billion–$8.158 billion.

Those figures show strong software demand and cash conversion, but no standalone result for the Artificial Intelligence Platform. The underlying 8-K points to the exhibit, which defines contract measures using agreements that may include options or termination provisions. Neither measure is recognised revenue.

Chief executive Alex Karp attributed the performance to demand around AI and sovereignty. That is management's explanation of the result, not an independent measurement of how much revenue AI produced. Operationally, the most useful next evidence would be stable product-level retention, usage or revenue measures that connect adoption claims to recognised economics.

onsemi showed the power layer inside the AI bill

onsemi's 3 August results filing reported second-quarter revenue of $1.604 billion, up 9% year on year and 6% sequentially. GAAP gross margin was 38.4%, GAAP operating margin was 16.1%, and GAAP diluted earnings were $0.56 per share. Free cash flow reached $425.4 million, more than four times the prior-year figure reported in the release.

Power Solutions Group revenue increased 19% year on year to $829 million. Chief executive Hassane El-Khoury said AI-driven demand was a particular source of strength and described AI data centres as the company's fastest-growing business, with revenue expected to more than double in 2026. The company also guided third-quarter revenue to $1.65 billion–$1.75 billion.

The distinction between evidence and attribution matters here. The quarterly 10-Q provides segment and financial detail, but it does not report “AI data centre” as a separate audited revenue line. Investors and infrastructure planners can observe stronger power-product revenue, margin and cash flow. They cannot derive the exact AI share from the filing.

For buyers, model capacity depends on power conversion and delivery as well as accelerators. A plan tracking only GPUs can miss constraints in semiconductor lead times, rack designs or facility commissioning.

Backblaze separated storage growth from contract optics

Backblaze's 3 August earnings exhibit reported revenue of $42.7 million, up 18% year on year. B2 Cloud Storage revenue rose 34% to $26.6 million, while Computer Backup revenue declined 2% to $16.1 million. Annual recurring revenue reached $177.3 million, and B2 net revenue retention was 113%.

The eye-catching figure was remaining performance obligations of $396 million, up $319.5 million from the prior quarter. The company said the increase was primarily driven by its CoreWeave agreement. But the event-time check changes the interpretation: Backblaze announced the five-year storage agreement on 23 June, and the contract filing says it became effective on 16 June. The deal itself is not an August 3 announcement.

Backblaze estimated roughly $335 million could be payable, subject to actual storage used, and issued warrants connected with the arrangement. The August results show how a prior AI-infrastructure contract entered backlog and guidance—not that $335 million became current revenue.

For any capacity contract, record signing date, commencement, usage variability, warrant effects, remaining obligations and recognised revenue separately. Backlog can improve visibility while leaving timing and concentration risk.

Paymentus supplied the finance-side control case

Paymentus's 3 August results filed with the SEC reported second-quarter revenue of $360.7 million, up 28.8% year on year. It processed 213.4 million transactions, up 21.4%, while adjusted EBITDA rose 54% to $48.8 million. GAAP net income increased to $25.6 million from $14.7 million.

The company attributed revenue growth to more billers and transactions. It did not attribute the quarter's performance to AI or disclose AI-derived revenue. That makes the filing more valuable in this brief, not less: it is a control case against relabelling every digital-finance improvement as an AI result.

Paymentus operates an AI-branded product, but its launch predates this news window and does not establish causation for the reported quarter. The defensible in-window facts are transaction throughput, revenue, profitability and guidance. Full-year revenue guidance was raised to $1.443 billion–$1.458 billion. Future releases would need adoption, usage, cost or outcome measures before an operator could connect a specific AI product to those economics.

Washington completed a test framework without publishing it

Axios reported on 3 August that a White House official said the federal framework for voluntary frontier-model testing was complete. The official did not disclose its contents, who had seen it or when it would be used. The report said Anthropic, OpenAI and Google had provided feedback on a draft and that staff planned a meeting with company representatives on Tuesday.

The legal background is Executive Order 14409, signed 2 June. It directs the Department of Energy to create classified cyber benchmarks and a voluntary framework through which developers may give government access to covered models up to 30 days before release to trusted partners. It expressly rejects mandatory licensing or pre-clearance.

Completion is a milestone, but confidentiality leaves operating questions unanswered: which models qualify, what constitutes a material change, how failures are remediated, what evidence developers retain, and what the public can learn. Participation is not a public safety certificate.

This resembles the boundary in our model change-control baseline: a test result is useful only when it is attached to an identified version, environment, threshold and accountable release decision. A private process may support national-security review while still being insufficient for a customer's own acceptance case.

Read the four commercial ledgers separately

Taken together, the filings describe several places where demand and cash can surface. They should be reconciled rather than blended:

• Recognised revenue records what accounting rules permit in the current period.
• Contract value and remaining obligations indicate future commitments but preserve conditions, options and timing.
• Segment or product revenue locates growth only when the company reports a sufficiently specific line.
• Transaction volume measures throughput, not necessarily unit economics or causal technology.
• Operating and free cash flow show conversion after working-capital and capital-allocation effects.
• Management attribution explains a result but remains a claim until supported by disaggregated measures.
• Benchmark participation describes a control process, not revenue, product performance or certification.
• Guidance is forward-looking and should remain distinct from completed-period results.

Procurement teams can use the same map. Ask which unit supports each claim, the period covered, whether it is audited and what changed from baseline. That stops backlog, launches or private benchmarks becoming proof of live outcomes.

The evidence does not prove an AI-wide boom

This window is unusually rich in company disclosures but narrow in scope. Four US-listed businesses do not represent the full AI or finance economy. Revenue can rise because of pricing, acquisitions, mix, contract timing or broader digitisation. Management can accurately identify AI-driven customer interest without providing enough detail to calculate its financial contribution.

The filings also expose different risk shapes. Palantir's contract metrics do not guarantee recognition. onsemi's data-centre commentary is not a segment disclosure. Backblaze's backlog is concentrated around a previously announced agreement whose value depends on utilisation. Paymentus's throughput says little about AI adoption. The White House framework has no published benchmark evidence to inspect.

Those limits define each claim's boundary. Demand appeared across multiple operational layers; measurement quality declines when they are collapsed into one narrative.

What to watch next

• Palantir product-level adoption or retention measures that make AI attribution reproducible.
• onsemi disclosure that separates AI data-centre revenue, orders or design wins from the wider power segment.
• Backblaze recognition, cash use and customer-concentration effects from the CoreWeave arrangement.
• Paymentus evidence tying a named AI capability to biller adoption, transaction outcomes or cost changes.
• Publication of the US framework, an unclassified summary, benchmark scope or an implementation timetable.
• Any revised filing, regulator response or subsequent event that changes the numbers reported here.

The next useful signal is not another broad claim that AI demand is large. It is a better bridge from a named product or control to a dated unit of revenue, usage, cash, capacity or tested performance.

Primary source trail

• Palantir second-quarter results, SEC exhibit, 3 August 2026
• onsemi second-quarter results, SEC exhibit, 3 August 2026
• Backblaze second-quarter results, SEC exhibit, 3 August 2026
• Paymentus second-quarter results, SEC exhibit, 3 August 2026
• Executive Order 14409 on frontier-model security testing, 2 June 2026
• Axios report on the completed White House framework, 3 August 2026