Anthropic 630k London Salaries|Are AI Job Cuts and AI Hiring Booms the Same Trade?

A City Divided: The Week AI Cut 8,000 Jobs and Advertised £630,000 Ones

Anthropic is paying some engineers in London up to £630,000 a year before equity. That number landed this week in the same news cycle as Standard Chartered announcing plans to eliminate almost 8,000 back-office roles, with its chief executive explicitly linking the cuts to AI investment. Both headlines are about the same technology. They describe opposite outcomes for the people it touches.

Standard Chartered is not alone. Over 50,000 tech jobs have been cut globally so far this year, with firms including Meta, Oracle and Atlassian connecting the redundancies directly to AI deployment. In banking, Morgan Stanley research published last year estimated that AI could put more than 200,000 European banking positions at risk by 2030 — around 10 per cent of industry roles. That figure is now being cited in boardrooms as permission, not as a warning.

Against that backdrop, Anthropic is expanding into a new London office capable of housing 800 staff. More than 40 roles are currently advertised in the capital, including six research engineering positions at up to £630,000 a year. OpenAI has 32 London vacancies live simultaneously, with several software and infrastructure roles approaching £450,000. Google DeepMind continues hiring across frontier research, safety and communications. The cluster has formed around King's Cross — London's answer to Silicon Valley — and the bidding war for talent is now generating compensation packages that rival what the most senior traders in the City earn.

The week's numbers, then, are not contradictory on the surface. They describe two ends of the same market: capital flowing away from roles AI can replicate and concentrating into roles that build the replication engine. What the surface reading does not account for is who moved first — and what that ordering reveals about which bet is already crowded.

Why £630,000 Does Not Mean AI Is Creating Jobs — It Means the People Who Build It Are Becoming Scarce

The salaries at Anthropic, OpenAI and DeepMind are not evidence of broad hiring. There are, by most estimates inside the industry, only a few hundred researchers globally capable of building and scaling frontier AI systems — training large language models, advancing reinforcement learning, managing the infrastructure at the scale these labs require. The number is genuinely small. A Deliveroo software engineer this week told City AM they had barely coded manually in almost a year; instead, routine engineering work is now supervised rather than performed. That shift is hollowing out the middle of the profession while compressing extraordinary value into the top.

That compression is what makes the Standard Chartered move a timing signal, not just a cost story. The bank's chief executive described cutting lower-value human capital and replacing it with financial and investment capital — a framing his own staff pushed back on publicly. But behind the language, the positioning move is clear: the bank is now long AI infrastructure spend and short the operational headcount it previously used to justify that spend. That rotation happened before the AI tools have demonstrably replaced the functions being cut, which is precisely what makes it a positioning shift rather than a productivity outcome.

The counter-evidence arrives here as a complication rather than a comfort. OpenAI chief executive Sam Altman said this week he is "delighted to be wrong" about the speed of white-collar job displacement. He had expected AI to have eliminated far more entry-level office roles by now. He admitted it has not happened as fast as he anticipated. His own firm experimented with AI-generated internal communications before reverting some interactions to humans. That admission, from the person with arguably the most direct visibility into AI capability timelines, suggests the cutting is running ahead of the capability that justifies it — which means the market may be pricing a displacement that is not yet operational.

The Pinsent Masons case adds a third data point from a different direction. A junior solicitor at the firm used an AI tool to research a legal question and draft a letter to the High Court, which contained a legal provision that does not exist. The judge called it a hallucination that had clearly not been checked. Critically, the AI tool itself had warned the solicitor against using the output without verifying it against authoritative sources — and the warning was ignored. The firm has referred itself to the Solicitors Regulation Authority. What the case surfaces is not that AI is unreliable in general, but that the reliability gap is concentrated precisely in the verification layer — the work that requires expert judgement applied to AI output. That layer is not being replaced. It is becoming the scarce input.

The Verification Layer: Which London Capital Is Priced Correctly?

The unresolved question from those three data points is whether the London AI labour market has already priced the scarcity correctly, or whether the £630,000 figure represents the early phase of a bidding war that has not yet cleared.

The historical parallel is the 2000 to 2001 period in London's internet hiring market. By late 1999, web developers and e-commerce architects were commanding salaries that rivalled investment banking, drawing capital away from established institutions. The premium held for roughly 18 months before the supply of trained candidates expanded faster than demand, compressing compensation sharply. The parallel is imperfect — the barrier to training frontier AI researchers is considerably higher than it was for early web skills, and the commercial applications are already generating revenue at scale rather than projecting future clicks. But the dynamic of institutions bidding against each other before candidate supply adjusts is recognisable.

The condition that sustains the current premium is a continued inability to automate the verification and judgement layer — precisely what the Pinsent Masons case made visible. If AI tools continue producing outputs that require expert scrutiny to be legally and commercially usable, the people who provide that scrutiny at the frontier of model development retain their scarcity value. Anthropic's London expansion includes not just research engineering but 12 research positions, seven sales roles, four safeguards positions, and vacancies in legal and policy — roles that are explicitly about managing what the technology cannot yet manage itself.

The condition that breaks it is faster-than-expected progress on autonomous verification. If the next generation of models can check their own legal citations, replicate the quality-control layer internally, and reduce the need for senior human review, the candidate pool required narrows further and the salary premium concentrates into an even smaller cohort — or collapses for the broader hiring wave currently in progress.

For the domestic investor, the verification benchmark is Anthropic's London fill rate on those 40-plus open roles over the next quarter. If the firm cannot hire at the advertised packages — because even £630,000 is insufficient to move the people it needs — that scarcity is real and structurally durable. If the roles fill quickly and new rounds open at lower salary bands, the first-mover bidding phase is ending. Standard Chartered's AI-for-headcount trade looks more exposed in the second scenario than the first — the capability may have arrived before the headcount reduction was justified, and the cut cannot easily be reversed.

The question the week leaves open is not whether AI displaces work. It is whether the institutions cutting now are doing so because AI can already do the work, or because the narrative has given them permission to act before the capability is confirmed. Those are different risks. One resolves in their favour. The other surfaces in the next round of results.