David Abitbol, who oversees AI operations at the European digital bank Revolut, described how the company runs coding AI for about 2,000 engineers. He spoke on October 13, 2026, at "AI Engineer NYC 2026." Revolut self-hosts Z.ai's open-weight "GLM" model, from China, on roughly 56 NVIDIA B200 GPUs, and says it holds the blended average price for input and output to $0.30 per million tokens. Abitbol also said costs fell to about one-tenth of what the company had been paying Anthropic and OpenAI. The notable point is not just the model swap: GPU operations tuning and a change in billing structure changed how the AI gets used. However, total cost, including GPU procurement and operations staff, was not disclosed, so the conditions behind the "one-tenth" comparison need careful reading.
From Gemini to GLM-5.3: Why Revolut Moved to Self-Hosting
Abitbol spoke at the session "Shipping like a startup, regulated like a bank" at AI Engineer NYC 2026, held October 12–14. In the roughly 15-minute talk, he explained self-hosting open-weight models on dedicated GPUs. Revolut is a major digital bank with 80 million customers (not 800,000) operating in more than 40 markets.
Revolut first introduced AI internally in January 2025, starting with Google's Gemini. This let the company work with internal, non-public data using AI. In February it began trialing Anthropic's Claude Code, but at the time the benefits did not justify the cost, so it set strict per-engineer budget caps. Claude Code was released as a limited research preview on February 24, 2025, which matches the timing in the talk.
With pay-as-you-go APIs, costs rise as usage grows. As engineers approached their caps, they had to save tokens for important work. Seeking a lower-cost way to give AI to more people, Revolut began self-hosting Z.ai's (formerly Zhipu AI) "GLM-5.1" in April 2026. By the time of the talk it had moved to "GLM-5.3." Open-weight models are now the default, and frontier models such as Claude Code, Codex and Gemini are used only as exceptions.
The timeline is as follows.
| Date | Event |
|---|---|
| January 2025 | Gemini introduced internally |
| February 2025 | Claude Code trial begins |
| April 2026 | Self-hosting of GLM-5.1 begins |
| August 2026 | Time to first token cut to about 1.6 seconds |
| October 13, 2026 | GLM-5.3 operations presented at AI Engineer NYC 2026 |
Hugging Face records show the GLM-5.1 repository was created on April 3, 2026, under the MIT license. The GLM-5.3 repository was created on August 25. According to the talk, about 56 B200s now serve GLM-5.3 to 2,000 engineers. However, the models, period and usage volumes behind the earlier Anthropic and OpenAI spending were not disclosed.
Cutting Time to First Token from 14.26 to 1.68 Seconds Without Adding GPUs
The key performance metric for Revolut is time to first token (TTFT): the time from sending a request until the first token comes back. Abitbol said the target is under 3 seconds, and complaints flood the company's Slack once it exceeds 5. Generation speed is a lower priority. About 200 tokens per second is enough for ordinary interactive use, and after 5 p.m., when fewer people are using the system, speeds of around 800 tokens per second are available.
When operations began in April 2026, at a load of about 3 requests per second, responses took more than 10 seconds to start. By August, the system handled more than 7 requests per second with time to first token of about 1.6 seconds. The company says it did not add GPUs in that period. According to the organizer's session summary, the slides showed requests rising 76% while TTFT fell from 14.26 to 1.68 seconds, a drop of about 88%.
The request figures, however, do not match between the spoken explanation and the slides. A rise from 3 to over 7 requests per second would be an increase of more than 133%. It is unclear whether the spoken numbers were approximate or whether the 76% figure uses different aggregation conditions. The organizer also cautions that the machine-generated transcript may contain errors. Even so, both accounts agree that the same number of GPUs now handles more requests with shorter waits.
The main tuning items were these five:
- Context length limit: Even where models can handle up to 1 million tokens, the company judged 256K tokens sufficient for real use cases, which cuts unnecessary memory consumption.
- FP4 quantization: On Blackwell-generation GPUs it uses FP4, which Abitbol said cuts required GPU resources by about 40% compared with FP8. He claimed accuracy does not drop but showed no evaluation data.
- Tensor parallelism tuning: The GLM model itself fits within TP4 (split across 4 GPUs), but when KV cache capacity runs short, the company chooses TP8.
- Aggressive caching: For coding workloads, he said the cache hit rate from reusing inputs must exceed 90%.
- Inference optimization: The company is considering speculative decoding and separating prefill (input processing) from decode (token generation).
The combination of FP4 quantization and tensor parallelism is especially important. According to Hugging Face, GLM-5.3 has about 753.3 billion total parameters, nearly the same as GLM-5.1's roughly 753.9 billion. At FP4, each parameter takes 0.5 bytes, so the weights alone need about 377GB. At FP8 it would be about 753GB, roughly doubling the memory required.
A B200 has 180GB of GPU memory, so TP4 provides 720GB in total. After placing the FP4 weights, about 343GB remains, but that must also hold the KV cache and runtime working space. The KV cache stores intermediate computation results for input tokens, and it consumes capacity as context grows longer or concurrent requests increase.
With TP8, eight GPUs provide 1,440GB, leaving about 1,063GB after the weights. With 56 B200s, TP4 gives 14 groups and TP8 gives 7, so TP8 halves the number of model replicas that can run in parallel but greatly increases each group's KV cache capacity. By simple arithmetic, the cluster-wide free capacity for KV cache rises from about 4.8TB with TP4 to about 7.4TB with TP8. This estimate excludes quantization metadata and runtime overhead and does not represent Revolut's actual configuration.
Abitbol's explanation that "the model fits in TP4, but TP8 can be advantageous for the KV cache" reflects this difference in memory capacity. Limiting context length to 256K also reduces KV cache consumption. However, Revolut has not said whether it uses an official FP4 release or quantized the model itself, nor what its actual parallel configuration is. It is also unclear whether the April-to-August optimizations targeted GLM-5.1 or GLM-5.3.
Behind the 60% Coding-Time and 70% Lead-Time Improvements
Abitbol also presented internal data from the past year covering about 2,000 engineers. The analysis measured "coding time," spent on implementation, and "lead time," from starting development until code is merged and deployable.
After Gemini was introduced, coding time improved by 20% and lead time by 10%, he said, citing less time spent searching Stack Overflow and technical documents. When self-hosting of GLM-5.1 began in April 2026, he said coding time improved by 60% and lead time by 70%.
Abitbol noted that lead time improved more than coding time because the bottleneck in development is not necessarily programming itself. AI can also be used for email, Jira and Confluence work, helping streamline tasks around development.
The billing change likely played a large role here. Under pay-as-you-go billing, engineers tended to hold back on AI use to stay within budget. Once self-hosting made the bill fixed, there was no longer a need, in Abitbol's words, to "save tokens for quarter-end or hard tasks." As a result, AI became easier to use for everyday development work, not just code generation.
These gains cannot, however, be attributed solely to the move to GLM. Over the same period, the model changed along with budget limits and usage policies, and engineers' growing proficiency may also have contributed. It is also not stated whether the improvement rates are cumulative from before adoption or compared with the previous stage. The control group and sample selection were not disclosed either. All that can be confirmed from the talk is that the GLM-5.1 rollout coincided with large improvements in development metrics.
Security Measures for Running a Chinese AI Model at a Bank
On the question of a heavily regulated bank adopting a Chinese AI model, Abitbol said how data is handled is what matters. Revolut does not use an inference API located in China; it downloads the published model weights and runs them on US-based services. The setup therefore does not send internal data directly to a Chinese service. He did not disclose the country where the GPUs are located or the actual service provider.
The company does not run its coding AI as a fully autonomous agent and keeps a human in the loop at all times. Specifically, it limits permitted operations in the sandbox, allowing grep for code search but not the build tool gradle, for example. Abitbol said these constraints on the AI's autonomous actions can be explained to regulators. The sandbox is also meant to prevent internal misuse, and work on prompt-injection defenses continues.
On data location, he said storage in the US, EU, UK or Israel is fine and the necessary GPUs can be sourced. Under EU personal data protection rules, however, transfers abroad are not uniformly permitted. The European Commission's adequacy decisions include the UK and Israel, while for the US there are conditions, such as commercial organizations participating in the EU-US Data Privacy Framework.
According to Abitbol, the inference service is by default configured not to retain data, and monitoring via OpenTelemetry has been expanded for incident investigation. When image processing is needed, requests are routed from a local router such as OpenCode to another model such as Gemini. Ordinary text is handled by GLM, and a Flash version of GLM is also used for vision. Z.ai describes "GLM-5.3-Flash" as the first natively multimodal model in the GLM-5 series, but the exact model name of the Flash version mentioned in the talk cannot be confirmed from the transcript. The inference service is offered through an OpenAI-compatible API supporting the Responses API and can be used from OpenCode and the Codex app.
Licensing also deserves attention for commercial use of GLM. GLM-5.1 was MIT-licensed, but GLM-5.3 is under a custom "GLM-5.3 License." It permits use, modification and redistribution, but requires a Z.ai security review before commercial use if the combined revenue of a company providing model services (Model as a Service) and its affiliates exceeds $10 billion (about ¥1.579 trillion) over any consecutive 12 months.
A model-service business means offering inference or fine-tuning via an API or similar in a form where third parties effectively control inputs, parameters or training data. End-user products that embed the model into specific features, and services that merely relay requests to models hosted by others, are excluded from the definition. The talk did not address whether internal use for engineers, as at Revolut, counts as a model-service business.
$0.30 per Million Tokens and a One-Tenth Bill: What the Numbers Do and Don't Show
Abitbol gave two cost figures: a blended average price of $0.30 per million tokens for input and output, and a bill about one-tenth of what had been paid to Anthropic and OpenAI. $0.30 is about ¥47. The conversion uses roughly ¥157.90 per dollar, derived from the European Central Bank reference rates published November 2, 2026 (€1 = ¥177.10, €1 = $1.1216).
For comparison, here are the official API prices of major models as of November 3, 2026, in US dollars per million tokens. They may differ from prices at the time of the talk.
| Model / service | Input | Cached input | Output | Notes |
|---|---|---|---|---|
| Revolut self-hosted GLM | — | — | — | $0.30 blended average |
| Z.ai GLM-5.3 | $1.40 | $0.26 | $4.40 | Z.ai official pricing |
| Claude Opus 5.5 | $4.00 | $0.20 | $20.00 | Anthropic official pricing |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | OpenAI official pricing |
Revolut's $0.30 average is far below Z.ai's standard GLM-5.3 prices of $1.40 for input and $4.40 for output. But the self-hosted $0.30 is a blended average of input and output, so it cannot be compared directly with the separate input and output API prices. Z.ai's cached-input price of $0.26 is close to the self-hosted average. Because coding workloads reuse a high share of common input, the effective API price can vary greatly depending on cache usage.
It is also unclear which Anthropic and OpenAI models Revolut used before, so the Claude Opus 5.5 and GPT-6.1 Sol listed in the table were not necessarily the comparison targets. The input/output mix and cache utilization were not disclosed either, so it is impossible to calculate how much cheaper self-hosting is than official APIs.
A further important distinction is between billed amounts and total cost of ownership (TCO). With pay-as-you-go APIs, bills rise with token consumption. With self-hosting, compute such as GPUs is reserved on a fixed basis, so costs do not change much as usage grows within capacity. Revolut says it increased throughput on the same number of GPUs, and this improvement in utilization may have contributed to the cost reduction.
However, it is unknown whether the roughly 56 B200s were purchased or rented, for example from a cloud provider. GPU procurement, power, cooling, infrastructure management and operations staff costs are not disclosed. The period and token volume used for the "one-tenth" comparison are also not stated, so a drop in billed amounts does not necessarily mean a total cost reduction of the same proportion.
What matters in this case is not simply that expensive commercial AI models were replaced with a cheaper Chinese one. Revolut optimized a fixed pool of about 56 B200s and built an environment that serves AI to around 2,000 engineers with short response times. By easing the usage limits that pay-as-you-go billing imposed, it also made AI easier to use for work beyond coding.
Whether other companies can replicate these results will depend on usage scale, GPU procurement terms and operating capabilities. Revolut's case shows not only the cost-cutting potential of self-hosting but also a shift from pay-as-you-go billing, which discourages AI use, to operations that use compute efficiently. To evaluate the economic advantage accurately, total cost including GPUs and staff would need to be compared with API prices under the same usage conditions.