A white paper.
Enterprise AI spending tripled to $37 billion in 2025 from $11.5 billion the year before, the fastest category expansion in enterprise software history (Menlo Ventures, State of GenAI in the Enterprise 2025). More than half of that — $19 billion — went into the application layer: coding tools at $7.3B, general-purpose copilots at $8.4B, vertical AI at $3.5B. AI subscriptions climbed from 8.8% to 26.4% of all enterprise SaaS purchases in fourteen months. And MIT NANDA's State of AI in Business report finds roughly 95% of these AI pilots fail to deliver the ROI buyers expected. Every CFO is asking the same question. Nobody is answering it.
Brex and Ramp see expense data but are not AI-vertical. Vendr and Spendflo see SaaS spend but do not finance. There is no source of truth for what a company's real AI ROI looks like.
But this is the surface. The deeper problem we care about is older. Learning should be a function of outcomes, not of access. Education today is sold backward. You pay four hundred thousand dollars for four years at Stanford before you know what those four years will produce, and the institution you pay has no skin in what you do after.
As a result, education has bent away from learning and toward marketing. The top professors at these institutions are not the best teachers; they are the best researchers, because publication and prestige set the brand, and the brand sets the price. We pay for pedigree. Pedagogy is a side effect.
The question underneath everything is: what would it take to make outcome-aligned human capital the default at scale, not the exception?
Apoorva Mehta's Abundance argues AI can solve the capital-market allocation problem. Capital A applies the same logic one layer up: using AI to solve the human capital allocation problem.
A corporate credit card for AI-native companies, issued via Lithic (with Marqeta as the Series A upgrade path) and a sponsor bank.
Three things land on day one:
By month twelve, the second product layer is credit lines underwritten on AI ROI signals.
The pattern is established. Vertical credit cards win when they identify a distinctive group with predictable value.
AI-spending companies are the next distinctive group with value. Coherent, growing, predictable, AI-native. Nobody has built the card for them yet.
The card is the entry point. The data is what compounds.
AI spend is not just a transaction. It is a forward-looking signal. The companies that adopt AI tools deeply and use them daily are the same companies that hire AI-savvy operators, ship faster, and have the kind of talent density that predicts survival. Real AI spend is a leading indicator of which companies will exist in five years.
That makes the dataset valuable to at least three buyers:
The dataset compounds with every transaction. Six months in, with 500 companies on the card, the underwriting models, benchmark reports, and risk signals Capital A produces are not replicable by any incumbent. That is the real long-term defensibility.
The card is how we get into the market. The data is why we stay.
The card is the wedge. The bank is the long arc. The school is the mission.
Each layer pays for the next. This is why we are building the bank first.
Three things converged in 2025 and 2026.
First, AI tool sprawl hit the "no one knows what we're spending" pain threshold. VC-backed CFOs expect to roughly double median AI tool budgets from $20K to $50K in 2026, with 47% of AI deals reaching production — almost twice the conversion rate of traditional SaaS. 56% of companies miss their AI cost forecasts by 11–25%; one in four miss by more than 50%.
Second, value capture is shifting from the infrastructure layer toward the application and data layers. Gartner forecasts worldwide AI spending will reach $2.5 trillion in 2026, with AI agent software alone projected at $206 billion (up from $86 billion in 2025). The applications are where the dollars are flowing, and the spend data is where the next layer of value will be built — a16z's own framing on this is that competitive advantage is shifting toward the application and data layers as infrastructure homogenizes.
Third, card-issuance infrastructure that took years now takes weeks via Lithic. Real-time underwriting against alternative data is normal (Affirm, Cherry, Belong all prove this). The MIT 95% number is the demand-side wedge. The pieces line up cleanly for the first time.
The core founding team is forming around three pillars: the hackathon-network builder (Chinat, fifty-plus hackathons run, MentorMates founder, Stanford Founders Demo Day Connect builder), a fintech operator with card-issuance and bank-stack experience, and a research-and-data partner.
We are studying and reaching into the orbit of operators who have done this before. The most relevant playbook is Tom Blomfield (Monzo, GoCardless, now YC Group Partner) — the fastest documented path from prepaid card to full UK banking license, with community-driven distribution. His arc is the template.
Lead-investor commitment is in place. The round is open to partners who recognize the long-arc thesis under the short-arc wedge.
Get in touch: calendly.com/chinat/personal · linkedin.com/in/chinat-yu
The card is the wedge. The bank is the long arc. The school is the mission.