Klarna Built a European Fintech Giant. Then Handed the Keys to Amazon.

Klarna is the European fintech story of the last twenty years. The company was founded in Stockholm in 2005, holds a full Swedish banking licence under Finansinspektionen as Klarna Bank AB, and was valued at around 14 billion dollars when it listed on the New York Stock Exchange in September 2025. Sebastian Siemiatkowski, the founder and chief executive, is one of Europe's most visible technology figures.

Klarna is also the European fintech that talks most loudly about replacing humans with AI. In February 2024 the company announced an OpenAI powered customer service assistant that it claimed handled two thirds of customer chats in its first month, doing the workload equivalent of 700 full time agents. Siemiatkowski subsequently briefed the press that Klarna was moving away from SaaS in favour of large language models. The company later walked the framing back; the assistant exists, but Salesforce was not replaced.

Where does all this run? Almost entirely on Amazon Web Services.

Methodology. Eight infrastructure layers, scored zero to ten on European ness, weighted by sovereignty impact. Every claim cites a public source. As a fully licensed bank, Klarna falls under EU banking supervision; the choice of cloud provider does not change banking regulation, but it does change which jurisdiction can compel data disclosure. Klarna's IPO risk factor disclosures explicitly flag operational dependence on cloud providers, payment networks and banks.

Compute and Orchestration

Klarna operates more than 1,000 AWS accounts. Microservices run on Amazon EKS, the managed Kubernetes service, sitting on EC2. AWS announced Klarna as its preferred cloud provider in December 2019, and Klarna engineering posts have referenced EKS adoption continuously since then. Host: AWS, primary cloud. Layer score: 0 of 10.

Storage and Databases

The Klarna engineering blog confirms AWS hosted data services as part of the core banking platform on AWS. Neo4j is used as a self managed graph database to consolidate data formerly held in SaaS applications. Specific managed services (Aurora, DynamoDB, RDS) are referenced indirectly in posts about AWS Step Functions and inventory tooling but never fully enumerated. Host: AWS. Layer score: 0 of 10.

Data Warehouse and Analytics

Klarna's cloud data platform uses Firebolt as the analytics engine on top of an AWS S3 data lake. Firebolt is US incorporated, hosted on AWS. There is no public evidence of Snowflake or Redshift as a system of record. Host: AWS plus Firebolt SaaS. Layer score: 0 of 10.

AI and Machine Learning

The OpenAI customer service assistant launched on 27 February 2024 and the company says it handled 2.3 million conversations in its first month. Internal ML for credit underwriting runs on Klarna's own AWS based stack: more than 1,700 features, 64 data sources, real time per transaction decisioning. Klarna reports its US Gini coefficient on credit decisions improved from 0.36 in 2019 to 0.72 in 2024. The assistant runs on the OpenAI API. The underwriting models run on AWS. Host: OpenAI (US) for the assistant; AWS (US) for credit risk. Layer score: 0 of 10.

Observability and Monitoring

No primary public source confirms a specific observability vendor. The Klarna engineering blog discusses cloud inventory tooling without naming a stack. Layer score: not publicly evidenced.

Identity and Access

Klarna's customer identity sits inside the bank's regulated KYC and onboarding flows, which run on Klarna infrastructure (AWS). Workforce identity sits on internal systems plus the post 2024 consolidation onto SaaS tools like Deel for HR. The named workforce identity provider is not publicly disclosed. Host: AWS for customer side; mixed for workforce. Layer score: 2 of 10.

CDN and DNS

No primary public source confirms a specific CDN or DNS vendor. Layer score: not publicly evidenced.

Payments

This layer is genuinely Klarna's own. The company operates its own buy now pay later network and switch. The new Klarna Card (2025) runs on Visa debit rails. Merchant acceptance covers Visa, Mastercard, Maestro, and Discover. Host: Klarna's own infrastructure (running on AWS, but the switch logic is Klarna IP), plus global card network rails. Layer score: 6 of 10. Klarna owns the switch and is itself the payments provider for other merchants, which is a strong sovereignty position. The card network rails are global and US anchored, and the underlying compute is AWS, which pulls the score down.

Layer	Weight	Score
Compute and orchestration	20	0
Storage and databases	18	0
Data warehouse and analytics	14	0
AI and machine learning	12	0
Observability and monitoring	not scored	unverified
Identity and access	8	2
CDN and DNS	not scored	unverified
Payments	10	6

Weighted total across the scored layers: 11 out of 100.

Badge: American with European Branding.

Klarna's score is lower than Spotify's despite owning more of its payment stack. The reason is the AI layer. Where Spotify could plausibly run TensorFlow Extended pipelines on European infrastructure if it chose to, Klarna's customer service assistant is locked to the OpenAI API by design. That makes the AI layer a structural dependency, not a procurement decision.

• Compute and orchestration. Managed Kubernetes from Scaleway, OVHcloud, STACKIT, or IONOS Cloud all run in EU data centres operated by EU entities. For a regulated bank workload, STACKIT (operated by the Schwarz Group in Germany) has been specifically positioned for financial services.
• Storage and databases. Managed Postgres and MySQL from Aiven (Helsinki) on top of EU infrastructure. Neo4j can run anywhere Klarna chooses to deploy it.
• Data warehouse and analytics. ClickHouse on European infrastructure or a self managed Apache Pinot deployment provide the OLAP layer.
• AI and ML. Mistral AI (Paris) operates the closest European peer to OpenAI for chat assistant workloads. Aleph Alpha (Heidelberg) is positioned for enterprise and regulated deployments, including air gapped installations. For credit risk specifically, internal models on European compute remove the OpenAI dependency entirely.
• Identity. ZITADEL (Switzerland) is a credible workforce IdP. Hanko (Germany) offers passkey first flows for customer authentication.
• CDN. BunnyCDN (Slovenia) and CDN77 (Czech Republic).
• Payments. Klarna is itself the payments provider for other merchants. No change needed on the rails side.

See the European AI APIs hub and the deeper Alternatives to OpenAI API page.

Klarna's AWS commitment runs deep. The bank's core platform was rebuilt on AWS from the ground up over the 2019 to 2022 window. The architecture is not a lift and shift; it is native to EKS, S3, and AWS managed services. Moving away would mean rebuilding a regulated banking platform on a foundation that is, at every layer, less mature than AWS's at this point in 2026.

The OpenAI dependency is even harder to unwind. Klarna's customer service assistant is not a thin LLM wrapper. It is built around a specific generation of OpenAI's models, with prompt engineering, retrieval, and tool use tuned to that surface. A switch to Mistral or Aleph Alpha is technically possible but would require re tuning the whole pipeline.

The IPO risk factors flag EU AI Act high risk classification for credit scoring. We did not find a Klarna risk factor naming the CLOUD Act explicitly. The exposure is structural regardless.

Siemiatkowski has been candid about both the ambition and the walk back. In a 2024 interview with Diginomica he conceded that the SaaS replacement framing had been overstated. The Salesforce relationship was complicated by his earlier comments, and Klarna had to clarify publicly that it had consolidated SaaS, not replaced it with an LLM.