As tech giants pour unprecedented sums into AI infrastructure, the underlying logic of this high-stakes bet is being clearly articulated: a wager on AI achieving "recursive self-improvement."
On August 3, Google DeepMind's Chief Strategy Officer, Jasjeet Sekhon, publicly stated at the Agentic AI Summit held at the University of California, Berkeley, over the weekend that recursive self-improvement (RSI) is the "core investment thesis" behind the industry's capital expenditure.
Recursive self-improvement refers to AI's ability to autonomously create superior versions of itself, which is considered by the industry as the next major milestone after achieving Artificial General Intelligence (AGI).
Sekhon also acknowledged that current AI revenue "is not yet sufficient" to support the current scale of capital spending, warning that the industry faces the risk of entering a market vacuum period.
Alphabet is projected to spend around $200 billion on AI data centers and related equipment this year, with plans to increase that figure further next year.
Meanwhile, DeepMind researcher Oriol Vinyals and OpenAI co-founder Wojciech Zaremba, who shared the stage, both indicated that recursive self-improvement could be realized between 2027 and 2028.
Logic Behind the Massive Bet: Recursive Self-Improvement Replaces AGI
In the AI industry's narrative, recursive self-improvement is quickly replacing AGI as the most sought-after concept.
Sekhon characterized the current capital expenditure in the AI industry as "the biggest scientific gamble in the history of human civilization," surpassing in scale the U.S. government's Apollo program, the Manhattan Project, and the development of the internet.
True recursive self-improvement means AI models can independently redesign their own overall architecture and develop entirely new models, creating a continuous cycle of self-enhancement.
Sekhon admitted that current technology "is far from reaching this stage," but pointed out that AI companies are now using models to assist in designing components of other models, which can be seen as a "precursor" to recursive self-improvement. He compared this phenomenon to the historical use of steam engines to build the next generation of steam engines.
Despite this, Sekhon stated: "It seems unwise to bet against recursive self-improvement happening."
However, he predicted that recursive self-improvement will "most likely emerge within the next few years."
DeepMind's Oriol Vinyals and OpenAI co-founder Wojciech Zaremba offered a more specific timeline in a panel discussion at the same summit: 2027 or 2028.
Revenue Gap Reality: A Concern for Alphabet Shareholders
Despite the grand strategic narrative, Sekhon's description of the current financial reality was remarkably blunt.
He explicitly stated that AI-generated revenue "currently cannot support the capital expenditure we are undertaking," meaning the industry risks falling into an "AI air pocket"—where capital spending is significant but corresponding revenue fails to materialize.
This statement aligns closely with the current concerns of Alphabet shareholders. While major AI investors, including Alphabet, are seeing accelerated growth in software or cloud service sales, shareholders have shown some unease about the company's continuously rising cash burn.
In contrast, Amazon's AWS recorded 37% revenue growth this quarter, with quarterly revenue reaching $422 billion. Its operating margin improved from 37% in the previous quarter to 39%, demonstrating a clear path for accelerating demand for AI cloud services.
To some extent, this provides support for the commercial logic behind the industry's massive investments, but it also highlights the structural gap between revenue and expenditure that Alphabet has acknowledged.
Risk Landscape: From Cyberattacks to Biological Threats
At the summit, Sekhon, along with University of California, Berkeley computer science professor Dawn Song, who recently joined Meta's "superintelligence" unit, spent considerable time discussing the security risks posed by AI, particularly its potential as a tool for cyberattacks and biological attacks.
Song pointed out that AI will "benefit attackers more" in the near term due to the inherent asymmetry between offense and defense: "Attackers only need one successful exploit, while defenders must defend against all attacks."
Sekhon agreed, warning that as attackers pollute open-source code repositories and exploit low-quality code written by humans, "the coming period will be quite difficult," with vulnerable systems like the U.S. power grid and hospitals being particularly exposed.
However, Sekhon pointed to a more severe threat in the biosecurity domain. He stated: "We are very close to a world where a person can simply talk to a model in natural language and design a virus or protein. This is an extremely dangerous world where the attacker still has the advantage over the long term."
He said that addressing biological risks requires stricter licensing, monitoring, and tracking mechanisms for "biologically relevant materials," and revealed that Alphabet is extending its AI-generated content watermarking technology, SynthID, to the biology field to help DNA synthesis companies screen for potentially risky AI-generated biological sequences.
Comments