DeepMind: the surge in AI capital expenditure is a wager on self-enhancing AI.

      Many have been inquiring about the purpose of the trillion-dollar AI investment. A high-ranking executive from Google DeepMind has provided a surprisingly forthright response. He states that this expenditure represents a wager on autonomous machines that can enhance themselves.

      Jasjeet Sekhon, the chief strategy officer at DeepMind, presented this perspective during a summit at UC Berkeley. According to him, recursive self-improvement (RSI) is becoming a fundamental aspect of the AI investment rationale, as reported by The Information. RSI refers to the concept of AI systems that can rewrite and upgrade themselves, producing increasingly advanced successors without human intervention.

      What stands out is the frankness of Sekhon’s remarks. He acknowledged that the current AI revenues do not cover the capital costs being incurred. Essentially, the funds are being invested based on a future promise. He contended that it would be unwise to bet against this potential, given the emergence of RSI already in progress. He presented a compelling analogy: steam engines were responsible for creating the next generation of steam engines.

      The new guiding vision

      The impact of this framing lies in its contrast to past justifications. For years, the industry pointed to artificial general intelligence (AGI) as the rationale for its spending. Sekhon appears to be replacing one distant aspiration with another. Under this perspective, RSI becomes the new AGI – the anticipated outcome that could transform today’s data centers from mere expenses into the most valuable machines ever constructed.

      The scale of investment is significant. Alphabet invested $44.9 billion in capital projects in just one quarter, which is about double the amount from the previous year, and has raised its projections for 2026 to as much as $205 billion. Additionally, it has indicated a "significant" further increase for 2027. Other major players like Amazon, Microsoft, and Meta are echoing similar sentiments. Sekhon compared this endeavor to monumental projects like Apollo and the Manhattan Project.

      Evidence of progress is visible. Google Cloud saw an 82% increase in revenue for the quarter, with a backlog exceeding $500 billion. However, the costs are substantial, and Alphabet recorded its first negative quarterly free cash flow, amounting to approximately $5.9 billion in deficit. There is a notable disparity between expenditures and revenue growth.

      A gamble with uncertain returns

      Sekhon identified the risk involved himself. He cautioned of a potential "AI air pocket," whereby spending occurs but revenue fails to materialize. This concern looms over every major tech company's earnings report and was articulated by someone whose role is to rationalize the expenditure.

      Moreover, RSI is not a tangible product; it remains a hopeful research objective accompanied by genuine uncertainties regarding safety, control, and technical feasibility. There are questions about whether it can be achieved within the timeline suggested by executives, estimated around 2027 to 2028. Competitors have already begun to challenge DeepMind regarding its capability to achieve self-improvement ahead of OpenAI or Anthropic.

      There is a more limited version of this claim that is already evident. Current models are capable of generating code and can, to a certain extent, aid in their own improvement. The significant leap Sekhon is promoting is from this capability to full autonomous self-enhancement, and it is a considerable challenge.

      In essence, Sekhon has made the trade-off clear. The industry is currently investing Apollo-like amounts into a capability that does not yet exist and may take years to develop. His transparency is refreshing but may also be somewhat daunting for investors.