The claim that matters, and the caveat that comes with it
The most quoted number in this story is a warning. A report titled "Scaling AI agents with trustworthy data" says that in organizations it classes as 'data laggards', AI access to company data falls to 30% or less.1 Across all companies, the average is about 45% of their data, according to Google Cloud research summarised in our narrative record.1 The same report says 100% of respondents plan to use agentic AI within two years, and 69% expect to use it widely.1
Read that as a gap: nearly everyone plans to deploy agents, and the systems feeding them are only partly connected.
Before you rely on it, note the trust problem. Via News has checked 11 claims from that source against their originals, and 0% held up.1 That does not prove the 30% or 45% figures are wrong. It does mean they should be treated as a survey's headline, not a verified fact. We use them because they match what the company executives below say, and we flag them because a reader who cannot check should be told.
Where the money has verifiably gone: the chip supplier's cost base
Most of the verified numbers in our dossier come from Nvidia's SEC filings. Nvidia is not an agent company, but our entity graph links it to the tooling agents run on. It develops the NeMo Agent Toolkit, NeMo Guardrails and NeMo microservices, and Mistral AI is recorded as its customer.2 Mount Sinai Health System is also recorded as a customer.2
Nvidia's cost of revenue is what it spends to deliver what it sells. It was $16.621 billion in fiscal 2024, $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026.3 That is roughly a doubling each year, and fiscal 2026 is nearly four times fiscal 2024. Nvidia's cost of revenue grew by a factor of nearly four in two years. These figures are canonical and checked against the filings.
The quarterly series shows the same pattern. Cost of revenue was $17.394 billion in Q1 2026 and $20.458 billion in Q1 2027.3 Growth continues, but the Q1 gap is far smaller than the annual jumps.
Nvidia's cash tells a less tidy story. Year-end cash was $7.28 billion in fiscal 2024, $8.589 billion in fiscal 2025 and $10.605 billion in fiscal 2026.4 Quarterly cash peaked at $15.234 billion in Q1 2026, then read $11.639 billion in Q2 2026, $11.486 billion in Q3 2026 and $13.237 billion in Q1 2027.4 A cost base that nearly quadrupled did not produce a cash balance that did the same. Our dossier does not say where the rest went, so we will not guess.
The dossier also carries earnings per share of $11.93, $2.94 and $4.90 for fiscal 2024, 2025 and 2026.5 That series falls and then rises, and the dossier does not explain why. We would not read a trend into it without the underlying filing notes.
Do you own Nvidia already? The dossier does not give index weightings, so we cannot say how much of a typical index fund it makes up. Check your fund's holdings list.
The buyers: three pitches, one shared idea
The startups profiled by CB Insights all say they sell labor, not software. Covecta's chief revenue officer, Ben Thomas, put its market as "tens of thousands of financial institutions globally," adding that "we are not just disrupting their software budget but their labor budget as well."6 Covecta says it serves corporate and commercial banks, specialist non-bank lenders, building societies, credit unions and private credit organizations, currently in the US and UK.6 It describes deploying "seasoned banker agents" for mission-critical tasks, workflows and portfolio activities.6
Penguin AI makes the same move in healthcare. Its head of marketing, Glenn Herzberg, says the company defines its market "as administrative labor spend rather than the healthcare IT software budget."7 He adds: "US Healthcare Administration runs about a trillion dollars a year, about a quarter of the total health spend, and the published estimates put around $570 billion of that in work that has no effect on health outcomes."7 Those are the company's own figures and are unverified, but they show the scale of the target: if the $570 billion estimate holds, it is more than half of that trillion.
Maisa AI's CEO, David Villalon, frames a narrower target: "process automation of core business and production tasks at regulated industries."8 Its emphasis is on work that must be auditable, reproducible and resistant to hallucination, the failure mode where a model states something false with confidence.8 That emphasis is itself evidence that data trust is the bottleneck.
These are executives describing their own companies in interviews CB Insights published. They are not audited results. None of the three interviews in our dossier includes revenue, funding or customer counts.
Incumbents are buying governance, not just capability
Two large announcements this July are about control. Manulife and Microsoft announced a five-year agreement in which Manulife adopts Microsoft's Frontier Suite, deploys Microsoft Agent 365 and expands Microsoft 365 Copilot to more than 30,000 employees.9 Manulife's Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."9
Box, for its part, announced agent guardrails, oversight of third-party agent activity, prompt injection detection and access policies based on agent classification.10 Tatsutoshi Murata of Nomura Research Institute said he expects Box to provide the administrative features needed to safely leverage AI agents, and valued its support for multiple AI vendors.10
Both announcements are company press releases carried by NewsEOD. Via News's measured reliability for that source is 57%: of 4,953 claims we checked, 57% held up.9,10 The quotes above are accurately attributed to the named speakers, but their promotional framing should be weighed with that number in mind.
The funding side, and the mixed signal at C3.ai
On September 10, 2026, Latitude, a payments platform founded by ex-Stripe crypto team member Cyril Mathew, raised a $35 million Series A for stablecoin-to-local-currency payments.11 Our dossier does not identify the investors, so we cannot say who is backing it.
Also on September 10, Nasdaq Data Link began carrying CB Insights' private company datasets, including financials, investors and founders dating back to 2010.12 For anyone trying to size these startups, that is the data-access theme in miniature: the material exists, and access is the constraint.
On C3.ai, Via News's narrative analysis notes insider selling at the incumbent and reads it as a sign of mixed investor conviction, even as the wider ecosystem accelerates.13 That is our analysis, not a verified fact. The dossier contains no transaction sizes, dates or seller names, so we cannot quantify it, and one sale does not establish a view.
What to watch
- Nvidia's next filing. Watch whether cost of revenue keeps rising faster than cash. Q1 2027 cost of revenue was $20.458 billion against $13.237 billion in cash.3,4
- Named revenue from the agent startups. None of the three interviews includes customer counts or revenue. Until one does, the "labor budget" claim is a pitch, not a result.
- A better data-readiness measure. The 30% and 45% figures come from a source that failed our checks. A verified replacement would change how much weight this whole story can carry.
- Insider activity at C3.ai. Filed details would show whether the selling is routine or a real signal.
