The most under-priced input in artificial intelligence is not compute or capital. It is the walking distance between the people building the models. That was the argument this site made ten days ago from King’s Cross, where AI firms took 661,100 sq ft of London office space during 2026 to 10 July on Knight Frank’s count, and where PwC recorded a 61% rise in UK job adverts for specialist AI roles. The claim was that clustering buys something specific — not prestige, not a postcode, but faster coordination. Fewer days between a question and an answer.
Then Google shifted the centre of gravity of its AI research out of London and towards Mountain View, and the reason given was essentially the same one.
That is an uncomfortable thing to read about your own argument, and it is worth sitting with rather than explaining away.
Article Brief
Key Takeaways
6 Points36s Read
The reporting comes from Forbes on 6 August, Bloomberg the same day, and Fortune on 10 August, with a follow-up on 13 August. Google disputes the framing, so treat the internal detail as well-sourced reporting rather than company account.
The structural facts are not in dispute. Demis Hassabis has stepped back from day-to-day operations to become chairman of Google DeepMind and Alphabet’s Chief Scientist. Koray Kavukcuoglu, now senior vice president, runs AI research and operations day to day — from Mountain View, roughly 6,000 miles from the lab’s founding office. Sebastian Borgeaud, who headed a key AI coding effort, relocated from the UK to California. And Jeff Dean, after 27 years and a run of work that shaped much of modern Google, left to start companies of his own.
Engineers quoted by Forbes described the reshuffle as a slow-motion pull of power from London towards Mountain View, and some used the phrase “America-fication” for it. Google’s response was direct: it said it was not aware of any rivalry between the two offices, that London remains an important hub for its AI talent and ambitions, and that Google DeepMind is a global team with a global footprint. Fortune reported the company’s position that DeepMind retains separate leadership and research autonomy under Kavukcuoglu.
The backdrop is what makes the reorganisation legible. Gemini 3.5 Pro missed a June target, then a mid-July target, and was still unreleased going into August. In June, Noam Shazeer — a Gemini co-lead and one of the authors of the transformer paper — left for OpenAI. John Jumper, who shared a Nobel for AlphaFold, went to Anthropic, and AlphaFold veterans Jonas Adler and Alexander Pritzel followed him there. A lab does not lose that roster in a quarter and reorganise by coincidence.
Of everything in the shake-up, the coding effort is the most legible signal.
Coding is where this generation of models is being judged. It is the capability with the shortest path from a benchmark to a paying customer, the one enterprise buyers evaluate first, and the one where the gap between labs is narrow enough that a two-month slip is expensive. It is also where Anthropic and OpenAI have both concentrated, which means the comparison is made in public and made often.
A company relocating the head of its lowest-stakes research group tells you nothing. A company relocating the head of the effort attached to its most contested product line is telling you where it thinks it is losing time. Google has been shipping AI at a fast and slightly uneven cadence — we wrote in July about a release where the most interesting of three simultaneous Gemini variants was the one no ordinary customer could access. That is what a research organisation looks like when the output is strong and the productisation is not yet keeping up.
It helps to remember how this structure came about. Google bought DeepMind in 2014 and left it in London with an unusual degree of independence — its own leadership, its own research culture, its own name. In 2023 it was merged with Google Brain, which sat in Mountain View, to form Google DeepMind. The research combined. The researchers stayed on separate continents, and the question of where the centre of gravity sat was never fully settled.
For most of a decade that ambiguity was affordable. Research on a multi-year horizon does not care much about an eight-hour time difference. What changed is the clock. When the release cycle compressed from years to months, a structure built for patient research began paying a tax it had never really been charged before, and the tax is levied in exactly the currency a schedule cares about: elapsed days.
Read that way, the move is less a verdict on London than an overdue reconciliation. Google has been running a two-continent research organisation on a timetable that suits a co-located one.
Here is the part that took a few days to work out.
The King’s Cross argument was about coordination among peers. If the person who has already solved your problem is a ten-minute walk away, you find them in an afternoon instead of a fortnight. That effect is real, it compounds, and nothing in the last two weeks contradicts it.
What Google appears to be addressing is coordination up the chain. The bottleneck there is not “can two researchers find each other.” It is “how long does it take for a call to get made, and how many time zones does the decision cross on the way.” When a lab is shipping on a compressed cadence and its flagship has slipped three times, the expensive latency is not peer-to-peer. It is the gap between the team and the person who can say yes.
Both statements can be true, and they do not resolve to the same policy. Talent density gets you a deeper pool and faster informal problem-solving. Proximity to the decision-maker gets you a shorter path from a finished idea to a shipped one. A lab under schedule pressure will optimise for the second even at some cost to the first, because the second is the one visible on a release calendar.
Put plainly: the org chart is a better map of where the work happens than the property market is. That is the part the original piece under-weighted.
It does not mean the London cluster is emptying out. The Knight Frank take-up is signed leases, not sentiment, and it is driven substantially by firms that are not Google. One lab rearranging its own reporting lines tells you about that lab’s internal physics, not about the market for London AI floorspace. Reading this as “the UK cluster is over” is reading a company story as a country story — and Google itself says London remains an important hub for its AI talent and ambitions.
It also does not mean the departures were about geography. Shazeer went to OpenAI, Jumper and the AlphaFold group to Anthropic, Dean to found his own companies. None of those is a bet on a different city so much as a bet on a different set of constraints. The pull in this market is autonomy and equity, and it operates largely independently of where the desk is.
And it does not mean centralisation wins as a general rule. Google is a specific case: enormous scale, a research function assembled by acquisition, a decade of semi-independent identity in another country. The coordination cost it is paying is partly a legacy cost. A twenty-person lab founded in one building does not have that problem to solve.
Arguments like this get decided by results rather than by reasoning, so it is worth writing down in advance what would count as evidence.
If the centralisation thesis is right, Gemini 3.5 Pro ships, and the model after it lands closer to its internal date than this one did. That is the entire justification. If shortening the reporting line does not shorten the release cycle, it bought nothing.
If the cluster thesis holds, London hiring continues at roughly its current pace across the rest of the market, UK-founded AI companies keep raising at competitive valuations, and second-half office take-up does not fall off a cliff. The cluster does not need Google to prove its case. It needs everyone else to keep showing up.
If both are right — the outcome worth betting on — you get a market where research talent concentrates geographically and decision authority concentrates organisationally, and the two are simply different axes. Labs will keep hiring into dense clusters and keep pulling the deciding function towards headquarters, and the tension between those becomes a permanent feature rather than a phase.
The near-term test is the Gemini release cadence. Watch that before watching the property data.
Most organisations reading this are not running a frontier lab, and the useful lesson is not about London or Mountain View.
It is that “we need our people together” is two separate claims wearing one coat. One is about the density of expertise available to a working team — a hiring-market question, and clusters answer it well. The other is about how many hops sit between a team’s conclusion and a decision — an org-design question, and no amount of shared floorspace fixes it. Companies routinely spend on the first while the second is what is actually costing them weeks. It is the same category error that runs through much of AI infrastructure spending: the electricity and data-centre build-out is a genuine constraint with a genuine bill attached, and it is also the constraint companies reach for because it is the one you can buy your way out of. Decision latency has no purchase order.
If your people are in a room together and things still move slowly, the room was never the problem.
The King’s Cross article got the mechanism right and the boundary wrong. Talent density is real, it is measurable, and it is why the leases are being signed. But it was written as though coordination were a single quantity that proximity improves. It is not. It splits, and the half that governs shipping speed runs along reporting lines rather than streets.
The revised claim: talent density is necessary and not sufficient. It determines what a lab is capable of building. It does not determine how fast that lab can decide to build it. Google appears to have concluded it had the first and lacked the second, and acted accordingly.
Whether that was the right read has a publication date attached to it. Gemini 3.5 Pro has missed three of them so far.
Article BriefKey Takeaways4 Points24s Read01The Split-AMD fell 3.71 percent and Intel 4.02 percent on 19…
Article BriefKey Takeaways4 Points24s Read01The Launch-Tesla has confirmed a Cybercab launch event with a rider…
Article BriefKey Takeaways4 Points24s Read01The Jury Is Advisory-Eight jurors were seated, but their verdict only…
Article BriefKey Takeaways4 Points24s Read01The Number-ByteDance and Tencent have each received about 10,000 Nvidia H200…
Article BriefKey Takeaways4 Points24s Read01The Move-Samsung raised advanced contract chipmaking prices by 5 to 15…
Ether finally stopped falling in June, and hardly anyone noticed. The slide that had run…