Around 300 British employers name this platform, and the reason is increasingly not the platform itself. It is the on-ramp: teams arrive because they want Gemini models under an enterprise agreement, or because they want TPUs, and the surrounding tooling comes with the territory.

That matters for hiring, because it changes what the role is. A candidate whose experience is model deployment on this platform and a candidate whose experience is building on Google’s frontier models through it are answering different job adverts, and you should know which you are writing.

The supply problem you are walking into

Be realistic before you set a budget. In the British market, Google Cloud skills are materially scarcer than the equivalent for Amazon or Microsoft — the enterprise install base here is smaller, so fewer engineers have accumulated years on it.

The practical consequences are worth planning for rather than discovering. Expect a longer search for equivalent seniority. Expect to consider strong engineers from another cloud and allow a genuine conversion period, which is reasonable because the concepts transfer and the vocabulary does not. And expect candidates to know their scarcity, which shows up in the rate rather than in the interview.

If you are choosing a platform now rather than inheriting one, the hiring pool is a legitimate input to that decision and it is routinely left out of it.

What transfers from another cloud, and what does not

Ask a candidate from elsewhere what they found different, because the good ones have a crisp answer.

The genuinely different things are the identity and project model — projects and folders rather than accounts, and service accounts behaving unlike roles elsewhere — along with networking defaults, the quota system, which bites hard on accelerators, and the data stack, where BigQuery is so central that a candidate who has not used it is missing the thing most work revolves around.

What transfers cleanly is everything conceptual: containers, pipelines, registries, monitoring, the shape of a training job. An engineer who has genuinely run ML infrastructure on another cloud is usually productive here inside a month, and screening them out for lacking the specific vocabulary is how teams spend two quarters failing to fill a role.

Accelerator quota is the thing that delays projects

This is the most common operational surprise and it is worth asking about explicitly. Accelerators are quota-controlled by region and type, quota is not granted instantly, and a training plan that assumes availability can stall for a fortnight while somebody negotiates an increase.

An engineer who has worked at any scale here will raise it without prompting, will know that regional availability differs, and will have a habit of checking before committing to a schedule. Someone who has only trained small models on shared infrastructure will not have met it, which is fine if your workloads are small and a real risk if they are not.

If TPUs are part of your reason for being here, ask specifically about that experience rather than accelerators in general — the programming model differs enough that GPU experience does not fully carry, and our JAX engineer guide covers the framework that path usually implies.

Lock-in here is tighter than it looks, and that is a choice to make deliberately

The platform is more coherent than its main competitor — fewer overlapping products, a cleaner path from data to model — and coherence is achieved by integration, which is another word for coupling.

Work built on BigQuery as the feature source, the platform’s own pipeline definitions and its managed endpoints is genuinely productive and genuinely difficult to move. That can be entirely the right trade: most organisations do not migrate cloud, and pretending otherwise costs real productivity. But it should be a decision somebody wrote down, not a drift.

A useful interview question is what they would keep portable if they were starting again. Thoughtful answers protect the training code and the container while accepting coupling in the orchestration, because the orchestration is cheap to rewrite and the model code is not.

Screening for the Gemini half of the role

Where the reason for the platform is access to frontier models, much of the job is application engineering rather than infrastructure, and the questions change accordingly. Ask about grounding against the organisation’s own data, about safety filter behaviour and what happens when a legitimate request is blocked, about quota and rate limits under real traffic, and about cost per thousand requests.

That is a different candidate from the platform engineer described above — closer to the profile in our AI engineer guide, with this platform as the delivery mechanism rather than the subject. Conflating the two in one advert is the most common error here and produces shortlists where nobody matches.

The BigQuery question, which decides who is suitable

Ask how they have used the warehouse alongside the platform, because on this cloud the two are close to inseparable and a candidate who has avoided it has avoided the centre of the stack.

What you are listening for is whether they treat it as the feature source rather than as somewhere data happens to live: pulling training sets directly, understanding what a badly written query costs when it scans a partitioned table, and knowing that the boundary between analytics and machine learning is thinner here than elsewhere. An engineer who has only exported CSVs out of it and worked on those files has been using this platform as generic compute, which is a legitimate way to work and a poor reason to pay a premium for platform experience.

Rates

Indicatively £80,000 to £120,000 permanent, with day rates from around £600 — a modest premium over the equivalent elsewhere, driven by scarcity rather than by difficulty.

Given that, the pragmatic strategy for most teams is to hire on cloud-agnostic ML infrastructure ability and treat the specific platform as trainable, while insisting on it for one senior person who sets the patterns. That converts a scarce search into an ordinary one, and it is the same shape of answer as the framework problem: one person with the specific experience, then good engineers converted into what they establish.