DeepMind CEO: AI Bottleneck is Experimental Computing Power, Not Training Large Models
Demis Hassabis, CEO of Google DeepMind, pointed out that the real bottleneck in AI is not the one-time training scores, but the continuous consumption of experimental computing power during the research phase. Researchers must repeatedly validate every new algorithm, architecture, or training method at a sufficiently large scale, and many seemingly feasible ideas fail when scaled up.
He emphasized that this means the demand for AI experiments is continuous, rather than just large training tasks once or twice a year. Meanwhile, English sources indicate that GPU delivery cycles still last 36 to 52 weeks, and global AI data center power consumption has reached 29.6 gigawatts, with suppliers often only able to queue for more even if they are willing to pay higher prices.
Hassabis has previously mentioned that shortages of chips, memory, and power are simultaneously suppressing AI deployment and research advancement. As the gap between research and commercialization is redefined by computing power, those who can complete experimental validation faster are more likely to seize the next breakthrough in models.
Source: Public Information
ABAB AI Insight
This information reveals a layer of misunderstanding in the AI race: while outsiders focus on "who trained the larger model," the real determinant of speed is who can make the research cycle run more densely. Training large models is a few significant events, while experimental validation is a daily consumption battle, and the latter is the core of research throughput.
This means the main battlefield of AI competition is shifting from "capital-intensive training" to "experimental density." A laboratory's true advantage lies not just in having stronger models, but in whether it can quickly send more hypotheses to be tested at real scale. In other words, innovation is not a matter of inspiration, but of computing power scheduling; it is not about who is smarter, but who can ensure more ideas are not stalled by resource constraints.
Hassabis's judgment also indicates that AI infrastructure has shifted from being "supporting resources" to "research prerequisites." In the past, algorithm teams could first experiment on small machines and then gradually scale up; now, the complexity of models and system coupling is so high that many methods are meaningless without validation at real scale. This will push experimental computing power, HBM, networks, and electricity to become common constraints in research, forming a deeper bottleneck than a simple GPU shortage.
In the long term, this structure will amplify the advantages of leading laboratories and large platforms. Only they can simultaneously bear high-frequency trial and error, resource queuing, and infrastructure investment, leaving behind not the "best paper writers," but those who can continuously turn research into verifiable systems. This is why the AI industry is increasingly resembling an infrastructure-led competition rather than a simple model competition.