ByteByteGo measured inference performance on Wise’s AI stack, pitting NVIDIA’s RTX PRO 6000 Blackwell against an H100 in Akamai Cloud. At 100 concurrent requests the Blackwell delivered 24,240 tokens per second per server, roughly 1.6× the H100’s 1,863 TPS. Such throughput continually matters for agentic systems that must answer thousands of queries per second without queuing.

Wise’s platform keeps performance reliable across its massive service mesh. Over 1,000 microservices, 700+ Java repos and dozens of front‑ends start from a shared chassis artifact that bundles security, observability and Kafka integration. Because the chassis is a versioned dependency, patches flow downstream automatically, turning manual migrations into simple version bumps. Gradle plugins also standardize GitHub Actions, letting a single update propagate through 700+ repositories.

Deployments use Spinnaker’s canary workflow: 5% of traffic receives a new version, metrics are collected for 30 minutes, and any anomaly triggers an automatic rollback. In 2024 that guardrail stopped hundreds of faulty releases, saving thousands of engineering hours. The discipline underpins Wise’s ability to settle cross‑border payments instantly for millions, moving £36 billion each quarter with settlement rates above 60%.