Stripe Sees AI Firms Scale 3x Faster Amid Compute Theft Fraud

Agents Reshape the Internet Economy

Emily Glassberg Sands, head of data and AI at Stripe—which processes 2% of global GDP—describes a fundamental shift: the internet's core actor is evolving from humans to AI agents. "The internet has this new kind of actor on it. Over time, this actor, these agents will become the predominant actors on the internet." Humans still drive many interactions, but increasingly through AI interfaces, delegated agents, or software-to-software exchanges. This breaks assumptions baked into products, from discovery and purchase to developer tools and payments infrastructure.

Stripe is adapting by making its stack "agent-ready," including fraud detection, billing, and identity layers. Builders must follow suit: rethink product discovery for agents, enable agent-driven building (e.g., LLM traffic to Stripe docs up 10x YoY), and prepare economic rails for agent-to-agent commerce. Sands emphasizes every stack layer needs evolution, not marginal AI tweaks.

Host Dan Shipper probes specifics, like fraud in an agent world where models might misuse cards. Sands clarifies agents introduce novel risks beyond traditional payment theft.

Compute Theft Emerges as AI's Core Fraud Vector

Traditional SaaS fraud targeted low-value free tiers, but AI's high marginal costs (every prompt incurs real compute expense) make abuse existential. Fraudsters steal compute as the "new CAC," exploiting free trials and credits that fuel growth.

Key vectors:

Multi-account abuse: Fraudsters cycle aliases for repeated credits; 7% of AI signups on Stripe.
Free trial abuse: Exploded 4x in six months. One large AI firm saw 4% conversion from trials costing $25 each in LLM spend—$625 per paying customer, mostly abusers. Stripe blocks 250k fraudulent trials weekly for one user.
Non-payment abuse: Rack up thousands in overages on 30-day invoicing, then ghost.

"Fraud used to be a transaction thing. Now it is a customer thing. It is a full-funnel thing," Sands explains. Stripe expanded Radar from checkout to signup, overages, and beyond via API calls at lifecycle moments: signup (block credits), payment, overages (throttle/top-up). AI arms race favors defenders with data scale—Stripe's network sees cross-processor patterns agents detect hourly.

Advice for AI builders: Integrate Radar at signup before granting credits; check dashboards for anomalies; block virtual cards judiciously (15% legit AI transactions). Shipper tests live: Every's fraud at 0.02% early warnings, 2% total—low, but upfunnel blind spots lurk.

"Giving someone credits... letting them rack up a bunch of tokens... is a major fraud vector," Sands warns, noting some drop trials (throttles growth) or blanket-block virtual cards (hurts legit users).

AI Companies Outpace SaaS History on Revenue Velocity

Stripe data reveals AI's blistering growth: Top 100 AI firms hit $30M ARR in 18 months—3x faster than 2018's top SaaS cohort. Milestones like $1M/$5M ARR scale "orders of magnitude" quicker. Horizontal model providers and vertical "wrappers" (domain-specific apps) drive adoption, but payments reveal full picture: who's buying what, retention/churn.

Sands attributes speed to AI's pull across economy, but flags monetization flux. Seat-based SaaS suited zero-marginal-cost humans; AI demands usage-based billing mirroring costs/customer value: tokens, API calls, workflows, outcomes. Hybrids dominate—subscriptions + overages, prepaid credits, real-time topups.

Lovable exemplifies: Started simple subscriptions via Stripe Billing for quick launch, iterated to Lovable Cloud/AI with metered outcomes.

"These AI companies are just growing from a revenue perspective faster than any previous cohort we've seen," Sands states. Model firms see tokens; gateways see flows; Stripe sees revenue reality.

Billing Evolves to Outcome-Based Precision

AI pricing reflects inference costs: "Way more hybrid monetization models... subscriptions with usage overages or prepaid credits that burn down." Builders meter granularly—every event rated with metadata—for precision aligning value and expense.

Outcome-based billing replaces seats: Charge per workflow completed or result delivered, not inputs. This handles variable costs absent in legacy SaaS.

Sands notes free compute as growth lever (like paid media), but fraud erodes it. Non-payment after massive overages leaves firms "holding the bag."

Agentic Commerce Demands New Primitives

Agents span "assisted buying" (human-in-loop) to autonomous purchases. Early signals: Agents buy low-stakes items like Halloween costumes. Stripe builds for full spectrum.

Key innovations:

Shared payment tokens: Agents pass buyer credentials safely to merchants, embedding Radar scores for off-Stripe processing.
Stripe Link: Consumer wallet for delegated agent buys.

Developer experience shifts: Machines use docs/infra (Stripe's LLM traffic 10x YoY). Fraud defenses as "public good"—Radar now covers all payment methods, crypto, non-Stripe via API.

"Stripe how are we getting agent ready? ... helping businesses get agent ready?" Sands frames Stripe's dual role.

Key Takeaways

Integrate fraud detection (e.g., Stripe Radar) at signup, payments, and overages before granting AI credits—full-funnel, not just checkout.
Monitor for compute theft: 7% multi-account abuse, 4x free trial surge; block 250k/week if scaled.
Expect AI revenue velocity: $30M ARR in 18 months (3x 2018 SaaS); prioritize usage/outcome billing over seats.
Hybrid monetize: Subscriptions + metered tokens/workflows/outcomes with real-time topups to match costs.
Prepare for agents as buyers: Build shared tokens, Link wallets; fraud mitigation is public good via comprehensive data.
Check your fraud dashboard now—0% visible may hide upfunnel abuse costing $625/payer.
Use AI defensively: Agents scan network anomalies hourly; scale beats fraudster creativity.
Iterate pricing fast: Start simple (subscriptions), add metering as products evolve (e.g., Lovable).

Notable quotes:

Emily Glassberg Sands: "Over time this actor these agents will become the predominant actors on the internet."
Emily Glassberg Sands: "Compute theft is the new payment fraud... free compute is kind of the new CAC."
Emily Glassberg Sands: "These AI companies are just growing from a revenue perspective faster than any previous cohort we've seen."
Emily Glassberg Sands: "Fraud used to be a transaction thing. Now it is a customer thing. It is a fullfunnel thing."
Dan Shipper (probing): "What even counts as fraud now in the sense of it's possible that my agent could go steal someone's credit card?"