MLOps engineer: the role, the stack and how to break in

An MLOps engineer builds and runs the systems that models are trained, shipped and served on: GPU clusters, deployment pipelines, serving stacks and monitoring. In 2026 the title is rare but the work is not. On Tokn, as of 14 October 2026, only 3 of about 800 AI job ads say "MLOps" in the title, while 54 ads at 20 companies hire for the same work under other names.

Below: the job since LLMs, the stack ads ask for and how to get in. It is one of the paths in our guide to careers in AI.

In short

• Search for the work, not the label: inference, ML platform, training infrastructure, AI infrastructure and AI reliability are the titles that hire MLOps engineers in Europe.
• Ads ask for Kubernetes, GPUs and Python first, then CUDA, NCCL, Slurm and vLLM. MLflow appeared in 2 of 53 ads we read in full, Kubeflow in none.
• DevOps, SRE and backend engineers are the most direct hires, and several lab ads said so outright. Junior roles are rare: 31 of 36 ads with a level were senior or lead.

What does an MLOps engineer do?

An MLOps engineer makes machine learning work in production: they build the path from training to a served model, then keep it fast, cheap and up for the researchers and ML engineers who use it.

Google Cloud's MLOps reference architecture (2024) defines MLOps as "an ML engineering culture and practice that aims at unifying ML system development (Dev) and ML system operation (Ops)". The idea goes back to a 2015 Google paper, Hidden Technical Debt in Machine Learning Systems: model code is a small part of a real ML system, and the data, configuration and serving around it carry large ongoing maintenance costs.

A day in the job

From ads advertised in October:

• "Detecting, draining and repairing unhealthy GPU nodes" (H Company, ML Compute).
• Building CI/CD "that carry a model from the registry to a served endpoint", and taking part in on-call (Doctolib, Senior ML Ops).
• Releasing the serving stack through "automated performance gates and progressive rollout" (Mistral AI, Inference Foundation).
• Designing evaluations that "gate CI on the result" (Supabase, AI Platform Engineer).

Perplexity's AI Inference Engineer ad put the range in one line: "You can read a research paper on Monday, write a kernel on Wednesday, and debug a production incident on Friday."

How LLMs changed the MLOps engineer job

The job moved from retraining pipelines to GPUs, serving cost and evaluation: the hard problem is now serving huge models to millions of users without wasting accelerators.

The vLLM launch post of June 2023 marks the shift: existing serving systems wasted 60–80% of KV-cache memory, and LMSYS halved the GPUs its traffic needed after switching to vLLM. In August 2026, Gartner forecast that inference will overtake training in AI-optimised cloud infrastructure spending in 2026: $23.3bn against $19bn, or over half of the total.

Evaluation is the other change. "Evaluation is how we decide which models, checkpoints and recipes ship," read Mistral AI's Eval Platform ad. For LLM products, evals are now the release gate that tests are in DevOps. Some call this LLMOps; no ad on Tokn used the word in a title.

Five names for the MLOps engineer job

Search for the work, not the label. On Tokn, as of 14 October 2026, 54 ads at 20 companies describe ML infrastructure work under five families of titles. Only 3 say MLOps: Alice & Bob's Senior MLOps Engineer and two Doctolib roles, all in Paris.

Inference is the largest family. Another 21 ads, for data scientists, ML engineers or AI architects, list MLOps as a skill inside another job.

MLOps engineer skills: the stack real ads ask for

Kubernetes, GPUs and Python come first, then a GPU layer (CUDA, NCCL, Slurm, vLLM). The toolchain most courses teach, MLflow and Kubeflow, barely appears.

Method: we read the full text of 86 postings from employers' public application systems and counted ads naming each term anywhere, nice-to-haves included, so shares are minimums of what ads name, not usage. "General infrastructure" means SRE, DevOps and platform ads with no ML word in the title.

CUDA and NCCL are what really separates the two columns. Inference ads lean on Rust (10 of 13), CUDA, TensorRT and vLLM or SGLang; training infrastructure ads add Slurm and NCCL. For the board-wide view, see the most in-demand AI skills.

MLOps vs DevOps, SRE and ML engineer

MLOps applies DevOps practices to a different artefact: a model plus its data, which runs on GPUs, degrades when the world changes and ships through evals. ML engineers build the model; MLOps engineers build and run the road it travels on.

Synthesia's ML platform ad warned that "this is not a pure ML Engineer role". If you want to build models, read about becoming a machine learning engineer or what an LLM engineer does.

How to become an MLOps engineer from DevOps, SRE or backend

The most direct route is from DevOps, SRE, platform or backend engineering: add GPUs, serving and evals to the infrastructure you know. Ads advertised in October said so plainly:

• OpenAI, ChatGPT Infrastructure: "We welcome candidates from backend engineering, distributed systems, platform engineering, and reliability backgrounds."
• OpenAI, Model Deployment: candidates "do not need specialized experience with GPUs or low-level hardware."
• Anthropic, AI Reliability Engineering: "we're looking for reliability-minded software engineers and SREs."
• Mistral AI, ML Infrastructure Engineer / DevOps: "2+ years of experience in a DevOps or Site Reliability Engineering role".

From ML engineering or data science, move toward serving, deployment and evals rather than modelling.

What to learn and build

1. Kubernetes, Terraform and one cloud.
2. Serve an open model with vLLM and measure latency, throughput and cost.
3. Run a multi-GPU job on Slurm or Kubernetes and monitor it.
4. Build an eval suite that blocks a release in CI.
5. Write it up. ElevenLabs' inference ad said it does "not require any formal certifications or degrees" and looks at "past projects, designs, or GitHub contributions".

Resources, as of October 2026: MLOps Zoomcamp is free and self-paced, with no live cohort planned for 2026 and no LLM module. Made With ML is free and built on Ray (now part of Anyscale). For serving, Baseten's free, vendor-published Inference Engineering; for system design, Chip Huyen's Designing Machine Learning Systems (2022). Then read how to get a job in AI.

Seniority, where the jobs are and pay

These are senior jobs, mostly at AI labs in London and Paris. In the examples below, frontier-lab bands run two to four times senior SRE pay elsewhere.

Seniority and experience

Of 36 levelled ads on Tokn, 31 were senior or lead. Where ads state a minimum (16 individual-contributor roles), the median is around 4–5 years, from 2 to 10. Only two were tagged junior, and one of them still asked for 2+ years of DevOps or SRE: the usual entry is a DevOps, SRE or platform job first.

Who hires, and where

AI-native companies hold 46 of the 54 ads. Anthropic, Mistral AI and Together AI had 7 each, OpenAI 5, H Company and Wayve 4 (see working at Mistral AI). By city: London 20, Paris 16 (counting Ubisoft near Paris), Amsterdam 5, Dublin 3. Seven were fully remote, often within a list of countries; more in remote AI jobs in Europe.

Amsterdam is small but clean: Tokn's infrastructure category there holds only infrastructure engineering roles, mostly at Together AI.

Pay examples

Tokn read a salary on 12 of the 54 ads, mostly at labs that publish bands. These are examples, not averages; 4 of the 6 GBP ads are Anthropic's.

For a wider benchmark, IT Jobs Watch counted 407 UK permanent ads citing MLOps in the six months to 14 October 2026, about twice as many as a year earlier, with a median of £80,500 (10th–90th percentile £56,000–£115,000). That covers all UK ads citing MLOps, not only AI companies. More in AI salaries in Europe.

Find MLOps engineer jobs on Tokn

Start with Tokn's infrastructure and MLOps category, which also includes security and data platform roles; some inference roles sit under research, so also search "inference" or "GPU". Labs posted new infrastructure roles in October, so set up a job alert to see the next ones first.

FAQ

Is "MLOps engineer" still a job title?

Rarely. On Tokn, as of 14 October 2026, 3 of about 800 AI job ads use it, while 54 describe the same work under other names, inference first. MLOps survives mostly as a skill listed in other jobs.

What is an inference engineer?

An inference engineer makes model serving fast, cheap and reliable: batching, caching, quantisation, kernels and engines such as vLLM, SGLang or TensorRT. It was the largest ML infrastructure family on Tokn, with 13 ads.

Do I need machine learning knowledge to become an MLOps engineer?

You need to know how models are trained and served, not how to research them. OpenAI's Model Deployment ad did not require GPU experience, and H Company's ML Compute ad said "You don't need ML research experience."

Is MLOps a good career, or is it dying?

The label shrank; the work grew. Inference was the biggest ML infrastructure family on Tokn, UK ads citing MLOps roughly doubled in a year according to IT Jobs Watch, and Gartner forecasts that inference spending overtakes training in 2026.