AI spend’s endgame: machines that improve themselves
▼ Summary
– DeepMind’s chief strategy officer Jasjeet Sekhon says the trillion-dollar AI buildout is a bet on recursive self-improvement (RSI), where AI systems rewrite and upgrade themselves without human involvement.
– Sekhon admits AI revenues “don’t sustain the capital expenditures,” framing the spending as a promise rather than a current return, and compares the effort to Apollo or the Manhattan Project.
– Alphabet’s capital spending hit $44.9bn in one quarter, with 2026 guidance up to $205bn, but the company posted its first-ever negative quarterly free cash flow of about $5.9bn.
– Sekhon warns of a possible “AI air pocket” where spending happens but revenue never arrives, and RSI remains a research hope with doubts about safety, control, and feasibility on a 2027-2028 timeline.
– The modest version of RSI already exists—models can generate code to improve themselves—but the leap to full autonomous self-enhancement is large, making the bet explicit and risky for investors.
The trillion-dollar question hovering over the AI buildout has finally received a blunt answer from a top insider. The massive capital expenditure, according to a senior Google DeepMind executive, is essentially a wager that machines will eventually learn to build better versions of themselves.
Jasjeet Sekhon, DeepMind’s chief strategy officer, laid out this thesis at a summit hosted by UC Berkeley. The concept, known as recursive self-improvement (RSI) , is quickly becoming the central pillar of the industry’s investment narrative, as first reported by The Information. RSI describes a future where AI systems can rewrite their own code and upgrade their capabilities without requiring human intervention, creating a cascade of increasingly powerful successors.
What makes this admission so striking is its sheer transparency. Sekhon openly acknowledged that current AI revenues “don’t sustain the capital expenditures we’re making so far.” In plain terms, the industry is pouring cash into a promise rather than a profit. He argued, however, that dismissing the potential of RSI would be a mistake, pointing to what he called “the makings of RSI” already present in today’s systems. His analogy was elegant: steam engines, after all, were used to build the next generation of steam engines.
A new north star for the industry
The power of this framing lies in what it replaces. For years, the sector justified its spending spree by chasing artificial general intelligence (AGI) . Sekhon is effectively trading one distant horizon for another. In this telling, RSI becomes the new AGI: the ultimate payoff that transforms today’s sprawling data centres from a massive cost centre into the most valuable machines humanity has ever constructed.
The financial scale is undeniable. Alphabet alone spent $44.9bn on capital projects in a single quarter, roughly double what it spent a year prior, and has raised its 2026 guidance to as much as $205bn. The company has also signaled another “significant” increase for 2027. Amazon, Microsoft, and Meta are singing from the same hymn sheet. Sekhon compared the collective effort to something exceeding the scope of the Apollo programme or the Manhattan Project.
There is evidence that parts of the strategy are working. Google Cloud saw its revenue surge by 82% in the quarter, backed by a order backlog exceeding $500bn. Yet the cost side of the ledger is staggering. Alphabet reported its first-ever negative quarterly free cash flow, landing roughly $5.9bn in the red. The gap between what is being spent and what is coming back in is widening at an alarming rate.
A gamble that might not pay off
Sekhon did not shy away from naming the primary risk. He warned of a potential “AI air pocket,” a scenario where the heavy spending continues but the anticipated revenues simply never materialize. This is the quiet anxiety lurking beneath every hyperscaler earnings call, finally vocalized by the very person tasked with justifying the outlay.
It is crucial to remember that RSI is not a product that can be shipped. It remains a research aspiration fraught with genuine uncertainties, including safety protocols, control mechanisms, and technical feasibility. Whether it can even be achieved on the aggressive timelines executives are hinting at, roughly 2027 to 2028, is an open question. Competitors are already taking jabs at DeepMind, questioning whether it possesses the self-improvement expertise to outpace OpenAI or Anthropic in this race.
A narrower version of the claim is already demonstrably true. Modern models can generate code and, in limited capacities, assist in their own refinement. However, the leap Sekhon is peddling, from these incremental steps to full, autonomous self-enhancement, is monumental.
What he has accomplished, perhaps intentionally, is making the underlying trade explicit. The industry is spending Apollo-era sums today, banking on a capability that does not yet exist and might not for years. His candour is certainly refreshing. But for anyone holding a stake in these companies, it is also, to put it mildly, deeply unsettling.
(Source: The Next Web)
