At the Agentic AI Summit held recently at the University of California, Berkeley, Google DeepMind Chief Strategy Officer Jasjeet Sekhon once again pulled market attention back to the frontier narrative of “Recursive Self-Improvement” (RSI). According to information from the event, Sekhon not only characterized RSI as the core investment thesis behind the industry’s capital expenditures, but also directly conceded that current AI revenue is “not yet sufficient” to support the scale of these investments, warning that the industry could face a “revenue vacuum” risk.
On the spending front, the scale of the RSI bet is extraordinary. Google’s total capital expenditure (capex) on AI data centers, chips, and infrastructure this year is described as being on the order of $200 billion, with plans to increase further next year. Meanwhile, Wall Street shareholders’ concerns over cash burn and the pace of revenue realization are also mounting: if revenue growth from AI software or cloud services fails to cover capex in the near term, the room for market imagination will compress.
Against this backdrop, Sekhon described the industry’s current capital spending as “the biggest scientific bet in the history of human civilization,” comparable in scale to the U.S. government’s Apollo moon landing, the Manhattan Project, and internet development spending. His key rationale: the market broadly believes that once AI gains the ability to further improve itself, marginal costs will decline significantly, reshaping business models and even industry structures.
RSI, as defined in the summit discussions, is not merely “AI-assisted research.” The more critical concept is that AI can independently redesign its own underlying architecture, or even build an entirely new generation of models from scratch, without frequent human scientist intervention in the closed loop. Once that tipping point is reached, the pace of AI capability improvement could leap from “iterations that take humans decades” to “weeks or even days.” It is for this reason that Sekhon emphasized in discussing industry competition: don’t bet against RSI materializing—if RSI does arrive, the company that masters it first could enjoy near-“supernormal productivity” advantages, creating a “winner-take-all, losers-exit” competitive dynamic.
Despite the grand narrative, Sekhon also put financial reality on the table. He admitted that AI’s current revenue is insufficient to support the ongoing scale of capital expenditure, and the industry faces the possibility of falling into an “AI air pocket”: money is spent, data centers are built, chips are installed, and power is connected—but “the thing that can truly monetize this investment” is slow to arrive. This logic echoes market concerns that “capex is running ahead of revenue.”
To address the debate over whether “RSI is actually happening,” the summit discussions introduced samples from multiple companies, painting a picture that “precursors of recursive self-improvement” are emerging, though still far from a true closed loop.
For example, on the Google side, DeepMind’s AlphaEvolve system was cited for using AI to optimize algorithms: it reportedly accelerated a key kernel in Gemini training by 23%, reducing training time by 1%; meanwhile, some modified circuit designs were written into next-generation TPU chips, making it a routine tool in the infrastructure. The description emphasizes that “AI can complete engineering and optimization tasks faster,” but it does not yet equate to “AI fully and independently restructuring its own architecture.”
In Anthropic’s case, the summit referenced its April experiment: nine Claude agents were placed in a constrained alignment research task, allowed to propose hypotheses, run experiments, and share findings. According to the disclosure, just two human researchers recovered 23% of the performance gap in one week; the nine agents collectively spent 800 hours and roughly $18,000 to recover 97%. However, the same material also presented a counterexample: when researchers transferred the “most effective” method to Claude’s real production training environment, it did not yield statistically significant improvements.
OpenAI was likewise used as a sample to illustrate that “self-improvement is emerging in pockets.” According to related information, in July OpenAI disclosed that GPT-5.6 Sol, via Codex, autonomously rewrote GPU kernels in the production environment, reducing end-to-end inference costs by 20%; it also ran several hundred architecture experiments on a speculative decoding model, improving token generation efficiency by more than 15%. Additionally, OpenAI published an “RSI Index” specifically designed to evaluate self-improvement capabilities. But the discussion also noted: to date, GPT-5.6 Sol has not yet reached the company’s “High” self-improvement threshold—it can solve some real research problems, but cannot yet reliably design and execute complete post-training pipelines.
These comparisons collectively point to a relatively cautious conclusion: AI-assisted AI R&D is indeed happening, but “true RSI”—where models independently rebuild complete architectures, train, and deploy stronger successors—may still require a transition period.
On the more direct financial logic, market attention is gradually shifting from “whether the technology can be achieved” to “when the capital will pay off.” According to the summit’s extrapolation, if inference and training costs decline in the next phase due to the closed loop, then capex has a chance to shift from “fixed asset investment” to a “compounding, rolling logic”: each generation of AI produces the next faster, and the next drives the next. At that point, compute consumption would no longer depreciate strictly linearly but would accelerate in line with capability iteration speed.
However, the risk side was also repeatedly emphasized in the same context. At the summit, Google DeepMind research discussions and Berkeley computer science professor Dawn Song (who recently joined Meta’s “Superintelligence” division) focused on extreme AI risks, particularly the “imbalance between attack and defense speed.” According to the discussion, an attacker only needs to succeed once, while a defender must cover all possible paths, so in the near term AI is more likely to benefit attackers.
On cybersecurity, the material noted: attackers could use smarter AI agents to poison open-source codebases, scan for and exploit vulnerabilities left by human programmers. The discussion specifically flagged that critical infrastructure—power grids, hospital systems, financial networks—may have insufficient resilience against sophisticated attacks, increasing vulnerability.
On biosecurity, the discussion was more severe. The information indicated that as technology advances, people might be able to design viruses or proteins simply by conversing with models in natural language, making long-term risks more favorable to attackers. To that end, the material suggested that future society will need stricter licensing, monitoring, and tracking mechanisms for “biology-related materials,” and revealed that Google is attempting to extend SynthID—its watermarking technology for identifying AI-generated content—into the biological domain, helping DNA synthesis companies screen for potentially risky AI-generated biological sequences.
On the timeline front, more specific projections emerged in the summit’s panel discussions: DeepMind researcher Oriol Vinyals and OpenAI co-founder Wojciech Zaremba were cited as believing RSI could be achieved between 2027 and 2028; Sekhon leaned toward RSI “likely emerging within the next few years.” While these windows carry uncertainty, they are enough to make capital markets more sensitive to the time gap between “infrastructure in place” and “revenue realization.”
Meanwhile, another variable equally relevant to capital realization is quietly reshaping the industry’s pace: the high-intensity churn of talent competition is reshaping R&D organizational costs and collaboration efficiency. Previous reports have noted that top researchers in the AI industry frequently move between companies, with compensation and compute resources serving as major attractions—but this can also bring project disruptions, double costs, and even litigation risks between firms. This phenomenon suggests, to some extent, that even as the RSI narrative accelerates, the industry still faces the management challenge of keeping key talent “stably at the same table.”
In terms of impact analysis, Sekhon’s remarks carry at least three layers of market implications.
First, for the AI infrastructure and compute supply chain, this means the direction of capital expenditure remains more firmly set. As long as the market believes RSI is the “prerequisite for capital payback,” data centers, chips, and power will likely continue to see high-intensity construction, and order visibility across the related supply chain will remain relatively high.
Second, for the financial cadence of listed companies, the “revenue vacuum” concern will more readily become a pricing variable for investors. If the industry cannot convert AI capex into sustainable software or cloud service cash flows in the near term, market sensitivity to cash flow and income statements will further increase.
Third, for R&D strategy, partial closed-loop capabilities are rapidly permeating, but “true RSI” still needs to bridge the gap between technology and validation. The samples above show AI can already deliver gains in optimizing algorithms, rewriting kernels, and improving engineering efficiency; but the transferability from experimental environments to production training environments, and the reliability of moving from “making local improvements” to “fully restructuring architecture,” remain key gaps.
If RSI is a “winner-take-all” narrative, then the core contradiction in today’s market is this: giants are betting on potentially near-exponential returns in the future, while cash outlays are a certain, linear cost in the present. Until this contradiction is resolved, any signal about “commercialization arriving faster/slower” could amplify market volatility.
For investors, the takeaway from this round of discussion is not simply whether RSI will ultimately be achieved technologically, but when the chain of “capex—revenue realization—risk governance” forms a more stable closure. In other words, the bet continues, and the market is asking: in the next phase, which arrives first—RSI, or another validation of the air pocket?