The AI Spend Layer

TL;DR

AI spend is observable in real time, hard to fake, correlated with future income and survival, and not yet captured by any incumbent bank or card network. The vertical-banking playbook the best fintech investors already know (SoFi, Figure, the Bill.com pattern) extends cleanly into this category.

The primary investable position is consumer-side: a banking product (credit, debit, or hybrid charge) for the student-and-builder cohort, underwritten on AI-tool spend and engagement. Two adjacent surfaces — the corporate-card vertical and the cross-card data layer — are worth tracking but harder to win on the same timeline.

Scope: this document maps the AI-spend category across the value chain — current entrants, where positions exist, who the buyer is, and where the structural gaps remain. It is not a recommendation on a specific company.

The Macro Insight

Three observations make this insight non-consensus.

First, AI-spend depth predicts speed of building. Developers using Copilot complete tasks 55% faster than non-users. The gap widens at the frontier: the people running multiple LLM sessions in parallel, layering Cursor + Claude Code + Replicate + n8n into a single workflow, ship prototypes in hours that used to take quarters. The depth of someone's AI-tool stack is a real-time proxy for how fast they can build, ship, and learn — which is the real input to future income, more than degree or employer.

Second, the signal is concentrated in transaction data. Time-on-tool is hard to observe from outside. Subscription depth is not. Most paid AI tools — and almost all of the frontier ones — sit behind a paywall: $20/mo ChatGPT Plus, $20/mo Claude Pro, $20/mo Cursor, $200/mo Claude Max, $200/mo ChatGPT Pro, $200/mo Cluely. Free tiers, school-issued seats, employer-paid subscriptions, and open-source models exist alongside the paywalled stack; spend cadence is what differentiates the long-tail builder from the curious user. A 19-year-old paying $400/mo across five AI tools out of summer-job money is a stronger forward signal than a 24-year-old with a $90K starting salary and no AI subscriptions. No card today surfaces that comparison.

Third, nobody is underwriting on it. Brex and Ramp see corporate expense data but do not segment AI. SoFi sees student-loan repayment but not tool usage. Greenlight sees teen spend but not AI-tool category. Cherry sees elective-healthcare repeat-spend but not tech. The category is empty.

The thesis in one sentence: AI spend is the new FICO for the under-30 builder generation, and the cheapest way to capture it is by issuing them a banking product — a credit card, charge card, or debit account built for the cohort.

Signal Noise — and What It Does Not Break

AI spend is messy. Students use school-paid seats, employer-issued tools, parents' cards, free tiers, and open-source models. Heavy spend can mean a serious builder — or bad subscription discipline. Low spend can mean a non-builder — or a capable user running open-source models locally. Anyone who has run consumer credit knows the raw signal will be noisy. Three reasons it does not break the thesis.

1. Pattern matters more than amount. Affirm and Cherry did not succeed by relying on spend totals; they succeeded on spend cadence and category mix. A consistent $40/mo spread across Cursor + Claude + a Replicate API is a different signal from a one-time $200 ChatGPT Plus annual. The card sees both — the underwriting model differentiates them.
2. The cohort selects in by carrying the card. The leakage cases (school-paid, parent-paid, OSS-only) describe non-cardholders. A 19-year-old who applies for, qualifies for, and uses an AI-spend card has already revealed willingness to put discretionary income behind tools they could otherwise free-tier. That self-selection IS the signal.
3. The data-layer consideration is built to absorb the noise. Cross-checking AI-spend with employer-paid SaaS, education partnerships, and AI-vendor API logs is how the messy ground truth becomes a clean dataset. The cards supply the spend; the data layer supplies the cross-validation.

The pilot underwriting-lift test in Falsifiability (point 1) is the load-bearing check. If noise really does swamp signal, the lift test reveals it before significant capital is deployed.

Why Now

Three shifts converged in 2025–2026 and make this fundable for the first time.

1. AI subscriptions hit consumer mass. OpenAI passed 400M weekly active users by February 2025 and roughly 900M by February 2026 — a 2.25× expansion in twelve months. Cluely and similar AI-study-and-work assistants have made the under-25 cohort the heaviest AI-tool stackers per capita. The customer base for a consumer AI fintech now exists at scale and is concentrated in identifiable populations.

2. Real-time alt-data underwriting became standard. Affirm built BNPL on permissioned data via Spark and Databricks. Cherry built medical financing on alternative income-cadence signals, achieving 80–90% approval at a fraction of CareCredit's MDR. Karat Financial proved you can underwrite a previously-uncreditable population (creators) on platform earnings rather than W-2. The infrastructure layer is now mature enough for an early-stage entrant to ship a defensible model. Underwriting on AI-spend pattern is technically possible in a way it was not in 2023.

3. Card issuance compressed from years to weeks. Sponsor-bank-backed card programs from Lithic, Stripe Issuing, Marqeta, Highnote, and Treasury Prime have made the BIN partner layer effectively a commodity, with go-to-market timelines on the order of weeks rather than quarters. The total cost of launching a vertical card pilot fell from the $5–10M range typical of de novo card programs in the 2018–2020 era to roughly $200–500K to ship a sponsor-bank -backed pilot today (estimate, based on public Lithic / Stripe Issuing pricing plus a benchmark cost-stack for a 5K-cardholder pilot).

These three shifts mean a category that was structurally impossible to enter in 2023 is now economically attractive in 2026 — and will be obvious to every other fund by 2028. The 12–18 month window to take the first conviction position is open right now.

Who, What, When, Where, Why — The User Underneath the Math

The macro argument above is logical. This section is empirical: the people the math represents.

Who

Two illustrative profiles drawn from cohort observation — not yet from formal survey work.

Maya, 20, CS junior at Stanford. Pays ~$260/mo across ChatGPT Plus, Cursor, Claude Pro, Cluely, and a Replicate API for class projects. Charges everything to her mom's Sapphire Reserve. Has no credit history of her own. Wouldn't qualify for a Chase Sapphire on her own income (campus job, $400/mo). No card she could use today rewards her actual spend pattern.

Aman, 24, second year out of UC Berkeley. Works at a Series A AI startup; spends ~$400/mo on personal AI tools to ship a B2B side project nights and weekends. Has a thin credit file because student loans went to direct creditors. Wants a card that reads his AI engagement as a credit signal — not yet another rewards card optimized for restaurants and travel.

What — The Pain Point

• No card recognizes AI spend as a category. Builders charge their tools to whatever they can — a parental Sapphire, a debit card, a school-issued procurement card.
• No credit is being built on the most predictive part of their financial behavior. Maya's $260/mo of AI tool engagement doesn't accrue to her credit file or her financial relationship.
• No product underwrites on this signal. A 19-year-old paying $200/mo for AI tools out of summer-job money is a fundamentally different credit risk than a 19-year-old with no AI engagement, and no card today sees that.

When — When the Need Surfaces

• First semester of college: first time managing their own subscriptions, AI tool stack ramps up.
• First post-grad job: builds a side stack outside employer-paid tools, needs a card that recognizes the spend.
• Sponsorship / fundraising transitions: builders moving from school to independent shipping need a card that maps to a non-W-2 income story.

Where — Channel and Context

• App-first onboarding. The cohort applies on a phone, in the UX rhythm of Apple Cash or Cash App.
• Distribution through communities the cohort already lives in: hackathons, AI builder programs (Build Club, YC Startup School), university CS departments, AI-tool waitlists.
• The card UX surfaces spend insights as the product — "you spent $237 on AI tools this month; here is what that compares to" — turning the financial product into a builder dashboard.

Why — Why This Product, Not Another Card

• Builds credit on the signal that actually matters. AI engagement → credit history → eventual access to the broader financial stack.
• Reframes AI spend from cost to investment. Spend insights + benchmarks + cashback on AI tools position the card as part of the building workflow, not adjacent to it.
• Cohort-native. Brand, AOV, eligibility, and rewards mix designed for under-30 builders — not retrofitted from a generalist card.

Why Now

• The cohort exists at scale for the first time. ChatGPT WAU 400M → 900M in twelve months. The buyer is observable today; it wasn't in 2023.
• Alt-data underwriting is now standard. Affirm, Karat, and SoFi (via its Nova Credit partnership) have built the playbook. The risk infrastructure exists.
• Card issuance is a commodity. Lithic / Stripe Issuing collapse the cost of a pilot to ~$200–500K.
• No incumbent is moving. Brex (acquired by Capital One), Ramp (horizontal), Chase / Amex (wrong cohort eligibility). The window to take the first conviction position is open and narrow.

Current Entrants and the Empty Intersection

A scan of the adjacent spaces — who is actually building today, with named players. Three adjacencies surround the category; the intersection is empty.

Consumer Student / Gen Z Banking

• Fizz — credit-building debit card explicitly for college students. Builds credit history through everyday spend; no AI-spend lens.
• SoFi — student refi + bank stack. Partnered with Nova Credit to underwrite on consumer-permissioned transaction data. Closest infrastructure precedent for alt-data underwriting at scale on this cohort, but the consumer card itself is not AI-spend-aware.
• Greenlight — debit card + financial literacy for under-18. Same playbook for a younger cohort.
• Karat Financial — credit card underwritten on creator-platform earnings. The closest non-W-2 alt-data underwriting precedent.

AI Subscription / Spend Management — Corporate Only

• Cledara — issues virtual cards per AI subscription with hard caps. Closest product to "AI-spend banking" today, but an enterprise spend-control tool, not a consumer card.
• Zylo — single system of record for SaaS + AI spend tracking, including consumption-based and seat-based pricing.
• CloudEagle.ai, SpendHound, Torii — adjacent SaaS / AI spend management platforms, all enterprise-side.

AI Builder Community & Education

• Build Club — APAC's largest AI builder community, 70K+ members across 60 cities, $1.75M pre-seed led by Blackbird (Annie Liao, founder). Aiming for 1M AI builders by 2026. Captures the cohort but does not serve it financially.
• AI Builder Club — separate platform; learn AI coding / agents / LLM apps.
• Hackathon networks — community-side distribution channels that own the cohort relationship but not the financial layer.

The Empty Intersection

No named entrant occupies the consumer AI-spend banking card position. The adjacencies surround it on three sides:

• Consumer banking for the cohort exists (Fizz, SoFi, Greenlight) but is not AI-spend aware.
• AI-spend management exists (Cledara, Zylo) but is corporate-only.
• The builder community exists (Build Club, AI Builder Club) but does not own the financial relationship.

The intersection — a consumer banking product that underwrites the cohort on AI engagement, for individual student-and-builder users — has no named player. That is the category whitespace this thesis is pointing at.

Value Chain — Where Positions Could Exist

The category breaks into discrete layers. An investment thesis at this stage maps the layers, not pre-commits to a single one.

A complete entrant builds at the cardholder-facing brand + underwriting layers and partners across the rest. Standalone bets at the data layer or the AI-vendor-partnership layer become viable only after the cardholder-facing layer has scale.

Three Positions Across the Category

The three positions sit on different layers of the value chain mapped above. Each has different unit economics and a different fit against the existing fintech-VC portfolio landscape. The consumer-side position covers the empty intersection identified in the market scan; the corporate-card and data-layer positions are adjacent surfaces worth tracking on different timelines.

Comparable Landscape

The thesis lives inside a defined comp pattern: identify a distinctive group of people with predictable value, build a card or financial product for them, capture the alt-data, underwrite on the dataset others cannot see.

The category does not yet contain a winner. Karat, Cherry, and Greenlight are the closest pattern matches; none of them are AI-vertical.

Why a Startup Wins (and not Ramp / Brex / SoFi / Chase / Amex)

The obvious incumbent objection: why doesn't a generalist card just add an "AI subscriptions" category and ship this as a feature? Three reasons the position belongs to a vertical entrant.

1. Card-of-record + cohort brand. A 19-year-old paying for Cursor, Claude, and Cluely is not opening a Chase Sapphire Reserve or an Amex Platinum — the eligibility, AOV, and brand are wrong. SoFi and Greenlight built dominant under-30 fintech franchises not by extending an incumbent product but by becoming the cohort's first-issued card. The card-of-record economics — primary spend, primary repayment data, primary cross-sell surface — only accrue to the issuer the cohort chooses first. Categorizing a category as a feature does not give an incumbent that data.
2. AI-vendor partnerships are first-mover defended. Cashback, co-marketing, and discount partnerships with OpenAI, Anthropic, Cursor, and Replit have natural exclusivity windows. The first vertical card to lock 3–5 of these owns the buyer-acquisition channel for the segment for at least 18–24 months. Incumbents with diversified spend cannot negotiate exclusivity on a category that is well under 1% of their volume.
3. Data network effects compound only for the segment-native card. A card that sees nothing but AI-builder spend can underwrite better than one that sees AI as a rounding-error category. The underwriting model improves with each marginal cardholder in the segment — a compounding loop the incumbents cannot enter without rebuilding their core risk infrastructure around a population that is rounding error to their P&L.

The base-rate behavior of incumbents in vertical fintech categories — Brex did not pre-empt Cherry, Chase did not pre-empt SoFi, Amex did not pre-empt Karat — is the cleanest empirical answer to the objection. Vertical cards win when the segment-native player ships first and locks the data + the vendor-side distribution before the generalists notice.

Recommended Actions

If the thesis holds, three concrete next moves. None blocks the others.

1. Open one consumer-card seed bet by Q4 2026. Target the AI-builder cohort (the consumer-side position). Lead a $2–5M seed at $10–25M post or pre-empt a Series A round at $40–80M. Comps for valuation: Karat's $26M Series A (2021), Greenlight's seed at ~$10M post (2014). The right partner relationship to chase: founder team with prior consumer-fintech + a co-founder embedded in the AI-builder community.
2. Watch the corporate side; don't lead. The corporate-card consideration will produce 3–5 named entrants in 2026. Wait for a traction proof point — interchange run-rate by month 12, 25+ paying corporate customers — before leading a Series A.
3. Map the data-layer consideration as a seed-watch portfolio. The infrastructure bet is too early for a lead position today but worth holding 3–5 seed conversations open through 2027.

Geographic sequencing: US-first for both card bets. HK / Singapore / Tokyo expansion in year 2 — the cross-border arc that MoneyForward executed cleanly.

Capital reserve thesis: if both card bets reach scale, the data-layer consideration becomes the natural acquisition target. Hold reserves for the infrastructure consolidation in 2028–2029.

Regulatory & Compliance — Live Risk Surface

Underwriting consumer credit on alternative data for under-30 borrowers sits in one of the most actively-regulated corners of consumer finance. Three surfaces matter; each is operationally manageable but not free.

1. Fair lending (ECOA / Reg B). Any alt-data underwriting model must be tested for disparate impact on protected classes. Karat Financial cleared this for creators by validating that platform-earnings signal does not proxy for race, gender, or national origin within a tolerance band. The same disparate-impact testing applies here — and the AI-tool engagement signal is, on its face, more correlated with class and education access than with protected attributes, which raises the testing bar. Operational implication: build with a disparate-impact testing harness from day one; budget $200–500K/year for compliance counsel + model audit; expect regulator pre-conversation before national rollout.
2. Data permissioning (GLBA + state privacy law). Transaction data feeding the underwriting model must be permissioned at issuance. The data-layer consideration must be structured so that resale to banks, employers, or AI vendors does not violate the cardholder's GLBA expectation. Plaid's permissioning model is the live template; California (CCPA / CPRA) and the post-2025 state-level patchwork raise the floor on consent UX.
3. CFPB posture toward student credit. The student segment is the most politically-sensitive end of the cohort. Heightened disclosure requirements and a likely APR-cap conversation with state regulators should be assumed for any product marketed to current students. The right strategic posture is to lead the disclosure standard rather than be assigned one — Karat used a comparable posture with the creator economy and it shortened, rather than lengthened, the regulator conversation.

None of these closes the category. Together they raise the floor on operating cost. A vertical card entrant should budget ~$1.5–2.5M/year in compliance + counsel by Series A — substantial, but inside the unit economics modeled in the consumer-side position. The Falsifiability section below treats adverse regulatory tightening (rather than baseline compliance cost) as the thesis-breaking risk.

Falsifiability

This thesis is wrong if any of the following are true by mid-2027.

1. Pilot data shows no underwriting lift. Define lift precisely: lower realized loss rate at a constant approval rate, OR higher approval rate at a constant loss rate, OR measurably better 12-month repayment behavior at constant credit-bureau control, OR faster forward income progression among approved cardholders. The thesis fails if the first 200–500 AI-spend cardholders in a live pilot show none of these versus a FICO-matched control after 12 months of card-on-file data. If the moat is invisible at pilot scale, the data-moat claim is dead.
2. Incumbents extend faster than expected. Brex, Ramp, or Mercury ship AI-spend categorization + benchmark dashboards by Q2 2027, foreclosing the corporate-card consideration.
3. Card economics compress. Interchange compression in the US drops blended take rate below 1.5%, killing the unit economics for the consumer-side position at the under-30 segment's AOV.
4. AI providers refuse partnership. Fewer than 3 major AI vendors sign cashback or co-marketing deals with any vertical card entrant within 12 months. Distribution kills the consumer-side position at the gate.
5. Consumer regulatory tightening. A material shift in US consumer-credit regulation around alt-data underwriting (CFPB, state-level) raises compliance cost above the AOV economics will support. Especially watch the student segment.

Each of these is testable on a defined timeline. If any two trigger together, the thesis should be revised before further capital is committed.