The money is real. The permission to use the data is the open question.
Every enterprise-AI pitch we reviewed this month rests on one promise: software agents can take over work that companies now pay people to do. The infrastructure behind that promise is very expensive. Nvidia's cost of revenue, the direct cost of producing what it sells, rose from $16.621 billion in fiscal 2024 to $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026, according to SEC filings.1 That is roughly 3.8 times in two years. At the fiscal 2026 level, the cost of delivering Nvidia's products was about 5.9 times its own year-end cash balance of $10.605 billion.2
Via News's analysis of the data suggests the more useful question for an investor is whether the customers can put that capacity to work. A finding tied to Google Cloud research, as summarized in our narrative for this piece, says organizations grant AI agents access to only 45% of company data, and that barely half trust the decisions the agents make.3 That is the trust gap. The rest of this piece tests it against what the checked sources actually say.
What the verified numbers show about the supplier
Nvidia is the clearest place to see the spending. Its year-end cash was $7.28 billion in fiscal 2024, $8.589 billion in fiscal 2025 and $10.605 billion in fiscal 2026.2 Cash grew far more slowly than cost of revenue over the same years.1 The quarterly figures are lumpier. Cash was $15.234 billion in the first quarter of fiscal 2026 and $13.237 billion in the first quarter of fiscal 2027.2 Cost of revenue in that latest first quarter was $20.458 billion, against $17.394 billion a year earlier.1
These filings show Nvidia's own spending and balances. They do not show whether its customers are using what they buy. To be plain about it, this data cannot prove or disprove a trust gap. If you hold a broad index fund you may already own a slice of Nvidia, but the dossier gives no weighting, so we will not guess at one.
The connections in our data show how widely Nvidia's products are deployed. It develops NeMo Guardrails, the NeMo Agent Toolkit and the Nemotron 3 Super model. Its named customers include Mount Sinai Health System, Yum! Brands and the State of Alaska Legislative Affairs Agency, and it competes with Advanced Micro Devices. That is a supplier positioned to profit from agents and also to sell the guardrails around them.
The startups are selling labor budgets, not software budgets
The vertical companies interviewed by CB Insights all describe their market in labor terms. Covecta's Ben Thomas said its total addressable market is "tens of thousands of financial institutions globally, and for them we are not just disrupting their software budget but their labor budget as well."4 He said Covecta deploys "seasoned banker agents" for mission-critical tasks, and that it currently serves clients in the US and UK.4
Penguin AI's Glenn Herzberg makes the same move in healthcare. He said US healthcare administration "runs about a trillion dollars a year, about a quarter of the total health spend," and that published estimates put about $570 billion of it in work with no effect on health outcomes.5 Note that this is a published estimate quoted by a company's head of marketing, not a verified figure, and the trillion-dollar total is his own. It is a sales pitch about market size.
Maisa AI's David Villalon defines his market as process automation of core tasks at regulated industries, where the work must be auditable, reproducible and hallucination resistant.6 Those three requirements are the trust gap stated as product features. Customers in regulated industries are paying for proof that the agent behaved, not only for the automation.
Capital is still arriving, but the evidence on returns is thin
On September 10, 2026, Latitude, a payments platform founded by a former Stripe crypto team member, Cyril Mathew, raised a $35 million Series A for a stablecoin-to-local-currency payments product.7 Investors are also explaining their picks in public. Emily Man of Primary said her firm was "immediately really excited" about Casap's founders because of their backgrounds, and that the two saw "the amount of internal effort and organizational work that it took to solve those challenges even with a really strong engineering team."8
Two cautions apply. First, CB Insights' interviews are the companies' and investors' own accounts, so they show how the sector wants to be seen, not how it performs. Second, none of the dossier's interviews report customer revenue, retention or price. Those are the numbers that would show whether a labor-budget pitch turns into paying customers.
Incumbents are treating the gap as a governance problem
The large institutions in our data are spending on control. On July 22, 2026, Manulife and Microsoft announced a renewed five-year agreement. Manulife will adopt Microsoft's Frontier Suite, deploy Microsoft Agent 365 and expand Microsoft 365 Copilot to more than 30,000 employees.9 Manulife's Shamus Weiland called the partnership "a critical enabler of Manulife's continued evolution into a truly AI-driven organization," and said the Frontier Suite gives Manulife "the trusted foundation to advance AI."9
Box, on July 21, 2026, announced agent guardrails, third-party agent oversight, prompt injection detection and access policies based on agent classification.10 Tatsutoshi Murata of Nomura Research Institute said his firm expects Box "to provide the administrative features needed to safely leverage this new era of AI."10 Both are vendor announcements, but they point the same way: the buyers are asking for controls before they widen access.
How much to trust the adoption claim
The most quoted adoption figures come from MIT Technology Review's "Scaling AI agents with trustworthy data." It says 100% of respondents plan to use agentic AI within two years, with 69% expecting to use it widely. It also says that in organizations categorized as "data laggards," AI access to company data falls to 30% or less.11 Our own measurement of that source is poor: 0% of 11 checked claims held up.11 The sample is small, but it means those figures should be read as a direction and not as data.
The Manulife and Box press releases came through NewsEOD, whose measured record is 57% of 4,952 checked claims holding up.9,10 Company announcements are useful for what an institution says it will do, not for what it achieves. The fully verified figures in this piece are Nvidia's filings. The rest is company narrative.
Our narrative for this topic also points to insider selling at C3.ai and a slower cadence of headline chip launches, including Alibaba T-Head's Zhenwu V900, scheduled for Q3 2027.3 The dossier holds no verified filings for C3.ai, so we treat the insider-selling point as Via News's reading of the market, not established fact.
What to watch
- Whether Nvidia's cost of revenue keeps rising faster than its cash, as it has across fiscal 2024 to 2026.1,2
- Any customer-side evidence, such as revenue, retention or pricing, from Covecta, Penguin AI or Maisa AI. Their interviews give market size but no results.
- Whether governance products (Microsoft Agent 365, Box's agent controls) show up as paid line items, since that would signal spending on trust rather than only on capability.
- A checked source on the share of company data agents can access. Until one exists, the 45% figure is our narrative's summary of Google Cloud research, not a verified statistic.
