The money is already moving
The clearest verified number in this story belongs to the company that sells the hardware under most AI projects. Nvidia's cost of revenue, the direct cost of producing what it sells, was $16.621 billion in fiscal 2024, $32.639 billion in fiscal 2025 and $62.475 billion in fiscal 2026.1 That is roughly a fourfold rise in two years, and each year came in at nearly double the one before. These are SEC-filing figures that Via News has checked against the source of truth.1
To picture $62.5 billion: it is about nine times the roughly $570 billion figure Penguin AI cites for wasteful US healthcare administration. Wait, it is the other way round. The $570 billion is about nine times Nvidia's annual cost base, which shows how large the labour budgets AI vendors say they are targeting really are (more on that below).2
The quarterly series tells the same story. Nvidia's cost of revenue was $2.544 billion in the first quarter of fiscal 2024 and $20.458 billion in the first quarter of fiscal 2027, about eight times higher.1 Growth is slowing from that pace, though. The latest first quarter was $20.458 billion against $17.394 billion a year earlier, a rise of roughly 18%.1 We are describing costs, not profits, and the dossier gives no revenue figures, so it cannot tell us how profitable any of this is.
A cash balance that moved sideways
Nvidia's year-end cash was $7.28 billion in fiscal 2024, $8.589 billion in fiscal 2025 and $10.605 billion in fiscal 2026.3 Quarter-end balances swing more: $15.234 billion in the first quarter of fiscal 2026 and $13.237 billion in the first quarter of fiscal 2027.3 Cash fell over that year even as costs rose. The dossier does not say why, so we do not guess.
A seam in the earnings data
The verified earnings per share series reads 11.93 dollars (fiscal 2024), 2.94 dollars (fiscal 2025) and 4.9 dollars (fiscal 2026).4 Read naively, that is a collapse and then a recovery. Because costs rose steadily across those years, we would not read it that way, and the dossier does not explain the drop. Treat year-to-year comparisons of this series with caution until the basis of each figure is clear. We are showing the numbers as filed rather than smoothing them into a story.
The bottleneck is data, not ambition
An MIT Technology Review piece reports that within two years 100% of respondents plan to use agentic AI, with 69% expecting to use it widely.5 The constraint it identifies is trustworthy, structured data. On average only about 45% of company data is accessible to AI, and about half of organizations trust their agents' decisions.5 In firms the survey labels 'data laggards', AI access to company data falls to 30% or less.5
A reliability warning applies. Our checks found that 0% of 11 checked claims from this MIT Technology Review document held up.5 That is a small sample, and we do not know which claims failed. Treat the survey figures as a directional signal rather than settled fact. They are consistent with what the companies below say about governance, but they are not independently confirmed.
Incumbents are buying the governance layer
Manulife and Microsoft announced a five-year agreement, with Manulife adopting Microsoft's Frontier Suite and Agent 365, and expanding Microsoft 365 Copilot to more than 30,000 employees.6 Shamus Weiland said: "Our partnership with Microsoft is a critical enabler of Manulife's continued evolution into a truly AI-driven organization."6
Box announced controls for AI agents working across enterprise content: agent guardrails, oversight of third-party agent activity, prompt injection detection and access policies based on agent classification.7 Nomura Research Institute's Tatsutoshi Murata said he expects Box "to provide the administrative features needed to safely leverage this new era of AI."7
Both announcements are about control and safety, which fits the data-trust gap above. Both come from NewsEOD, a source where 57% of 4,954 checked claims held up.6,7 That is better than a coin flip but not a high bar, and both are company announcements.
Startups size the market by payroll
The most useful financial idea in the startup interviews is the choice of budget being targeted. Covecta's Ben Thomas said: "for them we are not just disrupting their software budget but their labor budget as well." He described the market as tens of thousands of financial institutions globally, served today across the US and UK.8
Penguin AI defines its market as administrative labor spend rather than the healthcare IT software budget.9 Glenn Herzberg said US healthcare administration "runs about a trillion dollars a year" and that published estimates put around $570 billion of it "in work that has no effect on health outcomes."9 A trillion dollars is a quarter of total US health spend, by his account.9 These are executives sizing their own markets, not audited figures.
Maisa AI's David Villalon describes his market as process automation of core tasks at regulated industries, covering back office, operations and finance, where work must be auditable, reproducible and hallucination resistant.10 Sizing a market by labor is a bigger claim than sizing it by software licences, but it also gets tested harder.
Funding and the wider Nvidia web
Latitude, a stablecoin-to-local-currency payments platform founded by ex-Stripe crypto team member Cyril Mathew, raised a $35 million Series A on September 10, 2026.11 That is roughly one eighteen-hundredth of Nvidia's fiscal 2026 cost of revenue.1,11 It is a payments company rather than an agent builder, so we do not treat it as direct evidence of agent funding. CB Insights also published interviews with executives at Shepherd, Arlo, LARX, Veridox and Binary World on September 22 and 24, 2026.12 Those show where its coverage is pointing, not how much money is moving.
Our entity graph links Nvidia to customers including Mount Sinai Health System, Yum! Brands and the State of Alaska Legislative Affairs Agency, and names Advanced Micro Devices as a competitor.13 The graph says nothing about size or weighting, and we make no claim about whether readers hold these companies through index funds.
What to watch
- Whether Nvidia's cost of revenue keeps growing at a slower rate than the roughly 18% seen in the latest first-quarter comparison.1
- Whether the earnings per share series is explained in future filings, so year-to-year comparisons mean something.4
- Whether the 45% AI-accessible data figure improves, and whether a better-checked source confirms it.5
- Whether vendors selling against labor budgets, such as Covecta and Penguin AI, publish customer results rather than market-size claims.8,9
