---
type: "post"
id: "shipping-like-a-startup-regulated-like-a-bank-9968e4"
slug: "shipping-like-a-startup-regulated-like-a-bank-9968e4"
title: "Shipping like a startup, regulated like a bank | AI Engineer NYC 2026"
description: "Revolut’s David Abitbol described serving GLM 5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, at a reported blended cost near $0.30 per million tokens."
retrieval_nugget: "Revolut’s David Abitbol described serving GLM 5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, at a reported blended cost near $0.30 per million tokens. The case matters because he attributes improved engineering throughput not just to the model, but to making AI use cheap enough that engineers stop rationing requests."
published_at: "2026-11-03"
updated_at: "2026-11-03"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["anthropic"]
entities: ["Revolut","David Abitbol","GLM 5.3","Nvidia B200","Anthropic","OpenAI"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "Revolut’s case adds a separate company example: 56 B200s serving GLM for about 2,000 engineers, with reported 70% better lead time and frontier tools relegated to exceptions."
prior_refs: ["memory:beeeb201-cb93-8342-a57a-a3b93cae9053"]
source_url: "https://ai.engineer/nyc/2026/talks/revolut"
source_urls: ["https://ai.engineer/nyc/2026/talks/revolut"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:shipping-like-a-startup-regulated-like-a-bank-9968e4"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/shipping-like-a-startup-regulated-like-a-bank-9968e4/","markdown":"https://newruntime.com/posts/shipping-like-a-startup-regulated-like-a-bank-9968e4.md","json":"https://newruntime.com/posts/shipping-like-a-startup-regulated-like-a-bank-9968e4.json"}
---

# Shipping like a startup, regulated like a bank | AI Engineer NYC 2026

## Retrieval answer

Revolut’s David Abitbol described serving GLM 5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, at a reported blended cost near $0.30 per million tokens. The case matters because he attributes improved engineering throughput not just to the model, but to making AI use cheap enough that engineers stop rationing requests.

## Summary
Revolut’s David Abitbol described serving GLM 5.3 on 56 Nvidia B200 GPUs for roughly 2,000 engineers, at a reported blended cost near $0.30 per million tokens. The case matters because he attributes improved engineering throughput not just to the model, but to making AI use cheap enough that engineers stop rationing requests.

## What happened
At AI Engineer NYC 2026, David Abitbol, who leads AI operations at Revolut, presented the company’s approach to coding models and engineering performance. He said his chart covered a sample of about 2,000 engineers and tracked two measures: coding time in hours and lead time in days, from an engineer picking up a feature to its being merged and ready to deploy. Revolut initially used Gemini for information lookup and documentation. Abitbol said the first Claude Code rollout was less compelling economically: some engineers consumed substantial token budgets without a similarly clear return, leading the company to impose limits. Those limits, he argued, encouraged engineers to reserve AI for difficult coding tasks rather than use it across surrounding work.

Revolut moved open models to the default for this workload while keeping frontier services available for exceptions. Abitbol reported that the company serves GLM 5.3 to roughly 2,000 engineers using 56 B200 GPUs. Its blended cost is around $0.30 per million tokens, and its spending is about one tenth of its earlier Anthropic and OpenAI invoice. These are deployment-specific figures, not a general price for self-hosted inference. In the company’s chart, the move to GLM 5.1 coincided with roughly 60% better coding time and 70% better lead time. Abitbol said the delays being measured include work in inboxes, Jira, Confluence and elsewhere, not just writing code. The presentation does not isolate the model change from other causes.

The team’s results depended on tuning the serving system, not simply loading weights onto GPUs. Abitbol described a target time to first token below three seconds; once it exceeds five seconds, users begin complaining. Around 200 tokens per second was adequate for the interactive work he described, as engineers increasingly ran tasks in parallel. Holding GPU count constant, a chart showed 76% more requests per second and time to first token falling from 14.26 seconds during migration to 1.68 seconds in a stable setup. His advice was to tune existing infrastructure before adding GPUs.

He also described choices that affect capacity. The context window competes with concurrent requests for GPU memory, so Revolut sizes it for actual coding requests rather than the maximum advertised length. Quantization reduces memory footprint; Abitbol favored FP4 on this Blackwell-generation hardware. He said an eight-way tensor-parallel configuration can outperform four-way when the latter leaves too little room for KV cache. Cache reuse matters when requests repeat similar context; he recommended checking cache-hit rates and considering speculative decoding or separating prefill from decode. For image understanding, he suggested routing to another model rather than weakening a text-focused coding endpoint. Compatible APIs let engineers use their preferred coding harness.

The operational design also includes controls around the model. Abitbol described sandbox and approval rules that require intervention for restricted operations, and telemetry for investigating incidents and understanding ordinary use. He framed self-hosting as a different data path from sending inference requests to a model developer’s hosted service; he did not claim that a particular model origin or region automatically meets another company’s regulatory obligations.

## What the archive adds
This is a new deployment case, not a restatement of our earlier notes. Our archive has discussed coding-agent histories as evaluation material in [a guide to turning session histories into evaluations](https://newruntime.com/posts/ibragim-bad-published-a-guide-and-a-companion-e8d8ee/), but that is a different subject and does not establish prior coverage of Revolut’s serving economics.

## Trend signal
This case supports [agent economics moving to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/): Abitbol’s account ties lower inference costs to more delegated work and faster reported lead time, rather than treating token price as the outcome by itself. The evidence is suggestive, not causal proof; the presentation does not isolate the model or infrastructure change from other contributors to the charted results.

## Why now
The October 13 conference talk offers a concrete operational account of the tradeoffs behind self-hosting: cost, latency, concurrency, GPU memory and controls. Its timing makes the case useful now for teams weighing whether increased AI use justifies operating their own inference stack.

## Why it matters
A low per-token figure is not enough to decide between hosted APIs and self-hosting. Revolut’s account suggests that the more meaningful question is what happens to the work when engineers no longer need to ration model calls. If cheaper inference makes people delegate documentation, ticket and coordination work as well as code production, the impact may show up in lead time rather than only in coding speed.

That operating model comes with a substantial burden. Teams need enough demand to use dedicated GPUs, engineers who can tune serving, and controls for what coding agents may execute. The reported cost is blended across Revolut’s deployment, and the serving improvements reflect tuning at a fixed GPU count; neither number is a direct benchmark for another company. For builders, the transferable lesson is to measure cost, latency, request volume and verified workflow outcomes together. Without that measurement, a cheaper endpoint can look successful while shifting expense into infrastructure or weakening review and security.

## Limits and uncertainty
The source is a conference presentation by a Revolut leader, not an independently audited study. It gives no detailed methodology, baseline periods, confidence intervals or control group for the reported coding-time and lead-time improvements. The chart’s results coincide with the move to GLM 5.1 but do not show that the model alone caused them. The cost and throughput figures describe Revolut’s hardware, workload and utilization; the talk does not establish comparable economics for smaller teams. Nor does it quantify total GPU ownership costs, staffing, reliability, or the share of tasks routed to hosted frontier models.

## Skill to build now
Before expanding an inference fleet, benchmark one representative coding workload at realistic concurrency. Track time to first token, sustained generation speed, requests per second, cache-hit rate and cost per completed task; test context sizes and parallelism rather than reserving memory for maximum advertised context by default. Keep a record of which tasks require a hosted model, and pair the evaluation with sandbox permissions and telemetry so cost gains do not obscure operational risk.

## Reader takeaway
Revolut’s reported gains came from treating inference as an operating system for engineering work, not just choosing a cheaper model. The useful question for another team is whether lower marginal cost changes which tasks people delegate—and whether serving, security and verification can support that broader use.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)