Affirm launches in UK on Amazon and unveils ARC, an attention-based underwriting model that outperformed tree-based models by 2x

Affirm launches in UK on Amazon, unveils ARC underwriting model

Affirm is now live on Amazon in the UK, extending a US partnership that Levchin describes as long-established enough that the two engineering teams know each other well. The practical pitch to Amazon, a company with ample engineers and AI tooling, is specialization: Affirm has 15 years of underwriting data, a capital markets program built to fund the loans at scale, and the kind of regulatory credibility that a general-purpose retailer would have to spend years building from scratch. Shopify and Costco are also live in the UK ahead of the Amazon launch.

“The ARC based model outperformed the next planned improvement by a factor of two. I don't remember the last time I've seen a factor of two implementation improvement... Today's news: we are available in The UK on Amazon. That's a massive moment.”

ARC: the underwriting model shift

The larger technical announcement is ARC, a new family of attention-based underwriting models that Affirm has been developing for roughly three years. Affirm built its reputation on gradient-boosted tree models, which are deterministic, auditable, and easier to explain to regulators. The move toward transformer-style attention architecture reflects a belief that attention mechanisms can surface complex behavioral patterns across users in ways tree models structurally cannot.

Levchin says ARC outperformed the next planned improvement to Affirm's existing tree-based models by a factor of 2x. For context, Affirm ships minor model updates quarterly and major architectural updates roughly annually, each measured against the last. A 2x improvement in that cadence is, by Levchin's account, something he hasn't seen before.

The practical goal is inclusion: say yes to more borrowers while holding credit performance flat. Whether ARC can deliver both at scale in a new market like the UK will be the real test.

Developer productivity

Around January of this year, Affirm split off a team of roughly 12 engineers whose sole mandate is to keep the company's development environment at the cutting edge without forcing disruptive wholesale switches every time a better model or harness appears. The team maintains a curated menu of tools, including Cursor and Claude, and manages transitions so engineers don't spend months individually evaluating options.

The results Levchin cites are concrete. AI-assisted code as a share of total output increased by 10x after the team was formed. Cost per pull request, fully loaded with salaries and AWS spend, is down 30% since the team launched. The framing is that prescriptive tooling with managed optionality beats the "go explore" approach that leaves individual engineers spinning for months.

Agentic payments

On whether the agent economy needs new financial infrastructure, Levchin is skeptical that a smaller, less profitable BNPL player can displace Affirm by wrapping itself in an agentic pitch. His view is that agents handling routine purchases will rely on the same underlying plumbing, and the companies with the deepest underwriting expertise and capital markets relationships will accrete that volume. He also argues that shopping is partly entertainment, which means human-in-the-loop will persist for a meaningful slice of consumer purchases regardless of how capable agents become.

The 2x ARC improvement and the 30% cost-per-PR reduction are the two numbers worth holding onto. Both point to a company compounding on technical infrastructure in ways that are hard to replicate quickly, which is exactly the kind of durable edge that matters when a new market like the UK opens up.