Hey there! Welcome to Platform Weekly. Your weekly read of platform engineering fine print. Every week, we zoom in on another area of the Platform Engineering universe, from best practises and news to updates and lessons from the community.
Plus... take our 5-minute Platform Maturity Quiz to see how your team stacks up against real enterprises, from investment to AI and agent infrastructure, and get free tips to improve or maintain your score.
Revolut saved 90% on AI. Here’s what it took.
Revolut has been everywhere these last couple of weeks. The fintech’s roughly 2,000 engineers have swapped pay-per-token APIs for Z.ai’s GLM 5.3, which Revolut runs itself on 56 Nvidia B200 GPUs. David Abitbol, who leads AI operations at Revolut, told a conference in New York in October 2026 that tokens now cost a blended 30 cents per million, and the bill is a tenth of what Revolut used to pay Anthropic and OpenAI. Oh, and over the same period, coding time improved by about 60% and lead time by about 70%.
Not bad at all. No surprise everybody is talking about it.
It also tracks with what I’m hearing from platform teams. As bills from the big providers get worse, more and more people are asking whether open source models are the way out. Token and infrastructure costs were already the second biggest bottleneck to scaling AI in our State of AI in Platform Engineering Vol. 2, at 23%.
The part of the story that’s missing
We’ve been banging the open source drum on Platform Weekly for a while now. Multi-trillion-dollar companies are giving away the software that sits on top of their chips, and Sam Barlien wrote a full report on what that hands platform teams.
But here’s what the Revolut headline leaves out. And 30 cents is the token price. The talk didn’t put a number on the GPUs, the people tuning them or the tooling to watch it all. Tokens get cheaper, but those costs are still there. There is a reason why in the State of AI in Platform Engineering report, only 11% of platform engineering organizations were self-hosting open source models. 52% go straight to a model provider’s API, and another 26% use aggregated hosting. Revolut is in the 11%. Most teams aren’t.
So why isn’t everyone doing a Revolut? Because it’s not as simple as “it’s cheaper, it’s open, let’s go.” The Revolut team is pretty honest about that. Out of the box inference performance was awful. It took months of tuning to fix that. Capping the context window, quantizing harder and splitting the model across more GPUs let the same 56 chips handle 76% more requests while cutting the wait for a first token by 88%.
Free isn’t free. The license fee doesn’t automatically disappear when you go open. It moves onto your headcount. That means the opportunity is there, but only if you are able to seize it.
What to look at before you chase the 90%
Who pays the tuning tax. Someone has to own performance work, and keep owning it every time a model or driver updates. If that someone is “whoever has time this sprint”, those early savings that you’re seeing are going to evaporate sneakily very fast.
A stable interface in front of the model. A tale as old as time. In this case, Revolut’s self-hosted models answer in the same format as OpenAI’s API, which has become something of an industry-standard plug. So the coding tools its engineers use plug straight in. That means the platform team can swap models as it suits them. CHOICE!
The governance sandwich. The top slice is the platform. Revolut’s coding agents work inside a sandbox where risky commands need sign-off (grep is fine, Gradle isn’t), which is how it convinces regulators a person is still in charge. The bottom slice is self-hosting. Prompts stay, and data stay on GPUs Revolut picks.
Sovereignty through choice. Revolut only runs inference in countries with equivalent data protection rules, on GPU providers that keep zero data. For European teams especially, sovereignty comes from being able to switch. We don’t have our own models (Mostly), or our own cloud (Sorry LIDL), so sovereignty through choice is key while we build it out.
Observability from day one. Revolut leans hard on OpenTelemetry so the team can piece together any incident after the fact. You’ll want the same. Most teams are struggling with visibility. That is a problem. If you’re pushing into open source models without visibility, you’re f*cked.
A realistic savings target. Your mileage WILL vary. IMEC, the Belgian semiconductor research center, recently benchmarked self-hosting open-weight coding models and put the total cost of ownership saving at 24 to 25% for GLM-class workloads. Still great! Just not 90%. In our view, the gap comes down to scale, load, and how much platform you already have.
Why this is still very good news
None of this is a reason to stay away. The 90% is real. It’s just a platform outcome, not a model outcome. We’d argue Revolut got there because it controls the whole setup. It writes its own agent rules, knows its traffic well enough to plan around it, and has engineers whose job is tuning.
And cost isn’t even the whole prize. Owning the model makes spend predictable, keeps sensitive data where it belongs, and lets you post-train on your own data. No API discount gives you that.
The good news is you don’t have to build all of it from scratch the way Revolut did. Some of the biggest names in AI, NVIDIA and AMD among them, are open sourcing the software that sits on top of their chips, and infrastructure providers are packaging it into pre-integrated, composable stacks. That won’t make the tuning tax or the governance sandwich disappear. But it can take a lot of the assembly off your plate. If you want to understand that landscape, from what’s actually being released to how the stacks compare and how to judge whether one fits your team, the full report is the place to start.
BUT do your homework first. Pull last month’s AI token invoice. Then estimate what it would cost to run that same workload yourself, GPUs and engineer-hours included. If the invoice is the bigger number, you’ve got a business case. If self-hosting is, ask why. Sometimes the scale or load just isn’t there yet.
But most of the time… the gap is missing platform. No one owns tuning, there’s no stable interface, no sandbox, no observability. That part is your backlog, and it’s the work your platform needs to do before any model, open or otherwise, pays off;)
And as always... stay crunchy 🥐
Quick bites
Highlight of the week
Missed it live? The future of AI-native is open source with Kevin Cochrane from Vultr is available to watch.
Upcoming webinar
Still fixing drifted infrastructure as code (IaC) by hand? On November 12, Artem Artyunov of env zero shows how agentic workflows get failed environments to near-zero time to resolution, guardrails included. Register here.
From the community
Missed a webinar? Catch up on the Platform Engineering YouTube channel.