Demis Hassabis, co-founder of DeepMind Applied sciences Photographer: Jose Sarmento Matos/Bloomberg by way of Getty Photographs
Bloomberg | Bloomberg | Getty Photographs
Relying on the place you sit, Google both has probably the most enviable place in synthetic intelligence or is bleeding high expertise to main AI labs and different startups on the frontline of innovation.
It is a distinction that is been on full show over the previous two weeks, starting with the corporate reporting 82% income development in its cloud division, adopted by a shakeup on Wednesday in Google’s AI group, as chief scientist Jeff Dean introduced his departure after 27 years.
For Google, residence to the well-known 2017 transformer paper that paved the way in which for the generative AI increase, the latest occasions underscore a central problem going through the $4 trillion firm: the place to speculate. Constructing frontier fashions requires big upfront prices for compute and analysis with no assure of future returns, whereas the cloud enterprise is proving to be extremely environment friendly and is rising a lot sooner than rival choices at Amazon and Microsoft.
Alphabet CEO Sundar Pichai stated on final month’s earnings name that 90% of Fortune 100 firms are utilizing Gemini Enterprise, underscoring the corporate’s means to promote AI companies to cloud prospects.
Tomasz Tunguz, founding father of Concept ventures, stated it is turning into clear that top-of-the-line fashions aren’t required in relation to assembly most enterprise demand.
“I believe we’re at that place with AI, significantly for lots of white-collar work, the place most of the fashions which can be affordable are ok,” Tunguz stated. “The following evolution of fashions are more likely to be useful in domains the place you could have actually fancy computer systems.”
Google’s full-stack method to AI is an enormous motive the inventory is up 16% this 12 months after leaping 65% in 2025, when it outpaced all of its megacap friends.
It has been a bumpier street of late. Alphabet shares fell after the newest earnings report attributable to issues about capital expenditures, and dipped additional on Wednesday following the announcement that Dean is departing and Demis Hassabis is stepping down as CEO of Google DeepMind to turn into chairman of the unit.
Whereas the tone on Wall Avenue has been usually favorable, not everyone seems to be celebrating within Google.
Some researchers have grown pissed off over entry to the computing capability they should pursue formidable tasks whereas watching Google Cloud promote TPUs to exterior prospects, together with Anthropic, in response to folks conversant in the matter who requested to not be named attributable to confidentiality. Tensor processing items, or TPUs, are the the corporate’s homegrown AI chips that compete with Nvidia’s graphics processing items.
Google’s forms is a typical supply of frustration, with layers of approval required to maneuver analysis into merchandise. That may make rising firms like OpenAI, Anthropic and even youthful startups extra interesting, particularly for AI researchers and builders preferring lab work to steadiness sheets.
Dean is leaving together with Google stars Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to start out Discovery Loop. On X, Dean stated the startup, backed by Google, can be a public profit company “whose mission is to automate machine studying, science, and engineering to speed up discoveries and progress.”
Their exit follows the departures of different distinguished researchers, together with Noam Shazeer, one of many authors of the landmark 2017 paper “Consideration Is All You Want,” which offered the inspiration for generative AI. All eight authors have now left Google. Shazeer left for OpenAI in June, lower than two years after Google paid practically $3 billion to carry him again by means of an acquihire. His exit got here shortly earlier than Nobel laureate John Jumper left DeepMind for Anthropic.
‘A part of historical past’
Gil Luria, an analyst at D.A. Davidson, stated there is a clear pattern when trying on the exodus of high expertise.
“They are not keen on commercializing AI,” stated Luria, who recommends holding Alphabet inventory. “They’re keen on being a part of historical past, and they also have a look at Anthropic, OpenAI or one other startup as being the place the place they’ll pursue historical past.”
At Google, Dean was one of many only a few high-profile voices keen to criticize the Trump administration, and earlier this 12 months he was vocal in opposing the Pentagon’s choice to designate Anthropic as a supply-chain danger, warning that the transfer might harm the broader U.S. AI business.
Extra importantly, from a technical perspective, he helped construct the computing infrastructure and neural community methods that established the corporate as an early chief in trendy AI.
Hassabis, who co-founded DeepMind in 2010 and bought it to Google 4 years later, is turning into chairman of the division and can assume the newly created position of chief scientist at Alphabet, specializing in longer-term analysis and the societal implications of synthetic basic intelligence, or AGI. He additionally plans to commit extra time to Isomorphic Labs, the AI drug-discovery firm that grew out of DeepMind.
Koray Kavukcuoglu, DeepMind’s expertise chief and Alphabet’s chief AI architect, will take over day by day administration of the division and growth of the subsequent Gemini mannequin. In accordance with an individual near the DeepMind group, Kavukcuoglu had been taking up a broader set of tasks from Hassabis over the previous 12 months, together with directing mannequin growth and presenting main Gemini releases. Hassabis, in the meantime, has been spending extra time away from the lab, specializing in regulation and the longer-term implications of superior AI.

One of many greatest factors of friction inside Google is compute.
Google is investing greater than virtually any firm on this planet in knowledge facilities, chips and associated infrastructure. However capability stays scarce. Each TPU assigned to coaching a mannequin, serving a Google product, or fulfilling a contract with a cloud buyer displays a selection amongst competing priorities.
Frustrations over entry to compute may be particularly acute when Google pronounces massive infrastructure commitments to competing labs like Anthropic, whose fashions compete straight with Gemini, sources with data of the matter stated.
One of many folks stated Google has projections for demand in numerous areas, together with analysis and mannequin coaching, serving merchandise resembling search and Gemini, and dealing with cloud prospects. These necessities are modeled years prematurely, the individual stated, although capability could shift over shorter intervals if a product grows sooner than anticipated or if priorities change.
Pichai has stated on the previous two earnings calls that Google continues to prioritize DeepMind’s compute wants at the same time as demand from cloud prospects grows. Requested in July about TPU allocation, he stated Google’s “first precedence” is securing the compute wanted to compete on the frontier in AGI growth, calling that work “the inspiration for all the pieces we do.”
Pichai went on to say that Google balances that want towards the capability required to run its shopper merchandise and AI fashions, whereas more and more inserting TPUs straight in third-party knowledge facilities to assist fulfill exterior demand.
Dan Niles, founding father of Niles Funding Administration and a Google shareholder, stated entry to compute is a pure supply of pressure.
“Google has all of those different companies, they usually’ve received to determine who they will give a few of these sources to,” Niles stated. “Any person’s at all times going to be sad in that state of affairs.”
Bringing DeepMind nearer to Cloud
In January, on the World Financial Discussion board in Davos, Hassabis and Kurian appeared onstage collectively to debate enterprise merchandise and use circumstances.
It was a hanging sight contemplating how individually the 2 organizations have traditionally operated and the way far eliminated Hassabis was from a lot of the corporate, in response to an individual conversant in Google Cloud’s operations who requested to not be named as a way to communicate candidly on the matter.
The individual seen the joint look as an indication that Hassabis was turning into extra engaged with enterprise use circumstances of Google’s AI, significantly in areas resembling coding and customer support, and mirrored a broader effort to carry the corporate’s analysis and industrial operations nearer collectively because it competes with OpenAI and Anthropic.
Within the months that adopted, Google bumped into a lot of snags with its fashions, most notably delaying the launch of its latest flagship mannequin, Gemini 3.5 Professional. On the similar time, Kurian’s enterprise has seen its most explosive development on file.
Google Cloud CEO Thomas Kurian speaks on the Google Cloud Subsequent occasion in San Francisco, April 9, 2019.
Michael Quick | Bloomberg | Getty Photographs
Kurian, a former high Oracle govt, has led Google Cloud since 2019, constructing a bustling enterprise gross sales group at an organization recognized for its dominance in shopper web. Google Cloud designs its personal AI chips, operates a world knowledge heart community, and sells fashions, databases, safety software program, and instruments for constructing AI brokers.
The technique offers Google a number of methods to revenue from AI demand. It might probably promote infrastructure to OpenAI, Anthropic, and different labs, whereas providing Gemini to enterprises and integrating AI throughout Search, YouTube, Workspace and its different merchandise.
The rising query, virtually 4 years into the generative AI craze, is whether or not Google must develop the very best AI fashions or if it is higher off letting different firms foot the invoice.
Kavukcuoglu advised CNBC on the firm’s developer convention in Might that Google goals to push the frontier whereas additionally bettering effectivity. He stated the corporate’s Flash mannequin delivers frontier-level capabilities whereas operating 4 occasions sooner and extra effectively than comparable fashions, permitting Google to increase superior AI throughout enterprise and shopper companies.
Like Tunguz, Niles stated probably the most highly effective mannequin is pointless for a lot of industrial purposes.
“The fashions are ok for 90% of what must get accomplished,” he stated. “You do not want a Ferrari for these things. A Ford will work for 90% of the use circumstances.”
However for the scientists and researchers attempting to supply the subsequent transformer-scale breakthrough, being ok is not at all times ok.
WATCH: Demis Hassabis on agentic AI deployment


