Google DeepMind Executive: AI's Hundreds of Billions in Capital Expenditure Is the Biggest Scientific Gamble in Human History, Centered on "Recursive Self-Improvement"

Author | Bao Yilong

      Editor | Hard AI

As tech giants burn money on an unprecedented scale to build AI infrastructure, the fundamental rationale behind this massive gamble is being made clear—a bet that AI will achieve "Recursive Self-Improvement".

On August 3rd, Google DeepMind's Chief Strategy Officer Jasjeet Sekhon openly stated during the Agentic AI Summit at the University of California, Berkeley:Recursive Self-Improvement (RSI) is the "core investment logic" behind the entire industry's capital spending.

Recursive Self-Improvement refers to AI's ability to autonomously create ever better versions of itself—a milestone seen in the industry as the next significant step, following Artificial General Intelligence (AGI).

Sekhon also acknowledged,that AI revenue at present is "not enough to support" the current scale of capital expenditure, and the industry is at risk of falling into a market vacuum period.

This year, Google is expected to spend about $200 billion on AI data centers and related equipment, with plans to further increase investment next year.

At the same time, DeepMind researcher Oriol Vinyals and OpenAI co-founder Wojciech Zaremba, who appeared together at the summit, both stated,that Recursive Self-Improvement could arrive as soon as 2027 or 2028.

The Logic of Mega Bets: Recursive Self-Improvement Replaces AGI

Within the AI industry narrative, Recursive Self-Improvement is rapidly replacing AGI as the hottest buzzword.

AGI stands for Artificial General Intelligence, referring to an AI system that can achieve or surpass human-level performance in the vast majority of economically valuable tasks—one of the ultimate goals of AI research.

Sekhon describes the current AI industry capital expenditure as "the biggest scientific gamble in human civilization", surpassing even the US government's Apollo Moon program, Manhattan Project, and the investment in developing the internet.

True Recursive Self-Improvement means an AI model could autonomously redesign its own architecture and develop entirely new models, thereby creating a continuous loop of self-improvement.

Sekhon admits that the current technology "is far from this stage," but points out thatAI companies are already using models to assist in designing components of other models, which could be seen as a "premonition" of Recursive Self-Improvement. He likens this to using a steam engine to build the next generation of steam engines in history.

Nevertheless, Sekhon says:

Betting that Recursive Self-Improvement will never be realized seems unwise.

However, Sekhon expects Recursive Self-Improvement to "likely emerge within the coming years".

DeepMind's Oriol Vinyals and OpenAI co-founder Wojciech Zaremba, during the same summit's panel discussion, gave a more specific timeline: 2027 or 2028.

Revenue Gap Reality: Concerns Among Google Shareholders

Despite the grand strategic narrative, Sekhon offered a strikingly candid view of the current financial reality.

He stated clearly that AI-driven revenues currently "cannot support the capital expenditures we are making", meaningthe industry faces the risk of falling into an "AI air pocket"—that is, huge spending being incurred, while corresponding revenues fail to materialize.

This view strongly echoes current Google shareholders' concerns. Although major AI investors such as Google are seeing software or cloud service sales accelerate, shareholders have shown some unease about the company's growing cash consumption.

For comparison, Amazon AWS saw 37% revenue growth this quarter, reaching $42.2 billion in quarterly revenues; the operating margin climbed from 37% last quarter to 39%, showcasing the pace at which demand for AI cloud services is translating into actual results.

This to some extent supports the commercial logic behind the industry’s massive outlays, but also further highlights the structural gap between Google’s admitted expenditures and its revenues.

Risk Landscape: From Cyber Attacks to Biological Threats

At the summit, Sekhon and Dawn Song, Professor of Computer Science at UC Berkeley and recently joined Meta's "Superintelligence" division, spent considerable time discussing the security risks posed by AI, especially its potential as a tool for cyber and biological attacks.

Song pointed out that in the near term, AI will "benefit attackers more", citing an inherent asymmetry between attack and defense:

Attackers need only succeed once, whereas defenders must succeed every time.

Sekhon agreed, warning that as attackers poison open source code repos and exploit low-quality human-written code, "the next period is likely to be quite tough," especially for vulnerable systems such as US power grids and hospitals.

However,Sekhon highlights an even more serious threat: biosafety. He stated:

We are very close to a world where anyone can design a virus or a protein simply by conversing with a model in natural language. This is, in the long term, an extremely dangerous world that favors the attacker.

He said that mitigating bio risks will require stricter licensing, monitoring, and tracking mechanisms for "bio-related materials" and revealed that Google is extending its AI-content watermarking technology SynthID to the field of biology, to help DNA synthesis companies screen for AI-generated biological sequences with potential risks.