For months, analysts, investors, and tech observers have asked the same question: what is the trillion-dollar AI infrastructure buildout actually for? The stock market has rewarded companies that pour billions into data centers, but the underlying revenue streams remain oddly thin. Now a senior Google DeepMind executive has offered an unusually blunt answer. The spending, he says, is a bet on machines that improve themselves.

Jasjeet Sekhon, DeepMind's chief strategy officer, made the case at a summit at UC Berkeley. Recursive self-improvement, or RSI, is “becoming a key component of the AI investment thesis,” according to a report from The Information. RSI refers to AI systems that can rewrite and upgrade their own code, generating increasingly capable successors without needing humans in the loop. It is a concept long familiar to science fiction fans, but now it is being invoked to explain real-world capital allocation decisions worth hundreds of billions of dollars.

The candid admission

The most striking part of Sekhon's remarks is not the theory but the candor underneath it. He admitted that AI revenues “don’t sustain the capital expenditures we’re making so far.” This is a startling confession from a senior executive at one of the world's most prominent AI research labs. The money is being spent, in other words, on a promise. He argued that betting against it would be unwise, since there are already “the makings of RSI.” To illustrate the point, he offered a neat analogy: steam engines built the next steam engine.

The analogy captures the essence of recursive self-improvement. Early steam engines were crude and inefficient, yet they were used to manufacture better steam engines, leading to rapid advances in design and performance. In the same way, today's AI systems, though imperfect, can already generate code and assist in designing new AI architectures. If this process becomes autonomous, the technology could improve at an exponential rate, transforming the economics of the entire industry.

The new north star

What makes Sekhon's framing land is what it replaces. For years, the industry justified its massive spending by pointing at artificial general intelligence, or AGI, a still-nebulous goal of creating machines with human-level cognitive abilities. Sekhon is effectively swapping one distant goal for another. RSI, on this telling, is the new AGI: the payoff that turns today's data centers from a cost into the most valuable machines ever built. The difference is that RSI is more concrete and arguably more testable. You can measure whether a model improves its own code. You can benchmark self-generated training data. AGI, by contrast, remains a moving target.

This rhetorical shift matters because it aligns with the strategic realities of the companies involved. Alphabet, DeepMind's parent company, has been under pressure to show that its enormous AI investments will eventually generate acceptable returns. By pointing to RSI, Sekhon gives investors a narrative that is at once ambitious and specific. It also hedges the timeline: if AGI remains decades away, RSI might arrive sooner, perhaps within the next few years.

The scale of the bet

The scale of the financial commitment is staggering. Alphabet spent $44.9 billion on capital projects in a single quarter, roughly double the amount from a year earlier, and lifted its 2026 guidance to as much as $205 billion. It has promised a “significant” increase again in 2027. Amazon, Microsoft, and Meta are all making similar pledges. Collectively, the leading hyperscalers are spending more than the Apollo program and the Manhattan Project combined, adjusted for inflation. Sekhon likened the effort to something bigger than those historical benchmarks.

Some of it is clearly working. Google Cloud revenue jumped 82% in the quarter, with a backlog above $500 billion. The demand for AI compute is enormous, and companies are willing to pay premium prices for access to cutting-edge chips and models. Yet the bill is equally enormous. Alphabet posted its first-ever negative quarterly free cash flow, about $5.9 billion in the red. Spending and revenue are moving at very different speeds, and the gap is widening.

An air pocket risk

Sekhon named the risk himself. There could be an “AI air pocket,” he warned, where the expenditure happens but the revenue never arrives. That is the quiet fear under every hyperscaler earnings call, said out loud by the person whose job is to justify the outlay. The term calls to mind the aviation phenomenon where an aircraft suddenly loses altitude, and in business it perfectly describes a situation where expectations exceed reality. Investors have been willing to fund the buildout on faith, but faith can evaporate quickly.

The risk is compounded by the fact that RSI is not a shipping product. It is a research hope with real doubts attached, including safety, control, and technical feasibility. There are open questions about whether RSI is even doable on the timeline executives imply, roughly 2027 to 2028. Some experts argue that self-improvement at the level Sekhon describes would require breakthroughs in several other fields, such as automated theorem proving, curriculum learning, and reinforcement learning from synthetic data. Others point out that the current generation of large language models, for all their impressive capabilities, still lack the robustness and reliability needed to be trusted with their own code.

Rivalry and skepticism

Rivals are already needling DeepMind over whether it has the self-improvement know-how to get there before OpenAI or Anthropic. Both of those companies have made significant advances in AI reasoning and agentic systems, and both have large research teams dedicated to alignment and safety. The competitive pressure is intense, and the race to RSI could become as consuming as the race to AGI once was.

There is also a philosophical dimension to the debate. If RSI becomes a reality, it would represent a fundamental shift in the relationship between humans and machines. An AI that can improve itself without human intervention would free itself from the constraints of its creators. That is the dream of some technologists, but it is also the source of deep existential concern. The idea that an AI system might rewrite its own goals or values, even inadvertently, is a scenario that safety researchers have warned about for decades.

A modest version already exists

There is a modest version of the claim that is already true. Models can now generate code and, in narrow ways, help improve their own. For example, a model can write test cases for its own output, or propose optimizations to its own inference pipeline. Some research labs have demonstrated that AI can automatically tune hyperparameters, select better training data, or even design smaller, more efficient versions of itself. This is not full recursive self-improvement, but it is a step in that direction.

The leap Sekhon is selling is from this narrow, constrained self-assistance to full, autonomous self-enhancement. That is a very large leap. It would require an AI to understand its own architecture, identify weaknesses, and invent new algorithms that are superior to anything devised by human researchers. It would need to run millions of experiments, validate its own results, and then implement changes in production systems. This would demand both immense computational resources and a degree of reliability that current AI systems simply do not possess.

The trade-off is now explicit

What Sekhon has really done is make the trade explicit. The industry is spending Apollo-sized sums today against a capability that does not yet exist, and may not for years. His honesty is refreshing in a sector known for hype and obfuscation. It is also, if you are an investor, slightly terrifying.

The implied timeline is critical. Alphabet has penciled in 2027 as a year when capital expenditures could exceed $205 billion. If RSI fails to materialize by then, or if it arrives in a form that is too limited to justify such spending, the financial consequences would be severe. The stock prices of tech giants would likely suffer, and the broader market would feel the ripple effects. Even if RSI arrives, it would bring new risks: job displacement, security vulnerabilities, and the possibility of an AI that evolves beyond human control.

What's at stake

The stakes extend beyond corporate balance sheets. The infrastructure being built today, the massive data centers, the advanced semiconductor fabs, the transoceanic cable networks, all of it represents a bet that intelligent machines will be the engine of the next economic era. If that bet pays off, the rewards could be almost unimaginable. If it does not, the world will be left with a vast stock of expensive hardware that loses value fast, especially if the power consumption becomes unsustainable.

There is also an environmental dimension. Data centers already account for a significant share of global electricity demand, and that share is growing. The more ambitious the AI buildout, the larger the carbon footprint. RSI, if achieved, might help optimize energy use, but it could also lead to even more compute intensive training runs. The net effect is unclear.

For now, the industry seems content to run on the assumption that RSI is more than a mirage. Executives like Sekhon are increasingly willing to say so, perhaps because they believe that intellectual honesty will buy them credibility when the next earnings call goes badly. Or perhaps they genuinely believe that the singularity is just around the corner. Either way, the trillion-dollar question has finally received a direct answer: the machines will build the machines, and that is the whole point.

Source: TNW | Google News