Integrated Circuit, Film-Layout of a Printed Circuit Board. (Photo by Mediacolors/Construction Photography/Avalon/Getty Images)

Getty Images

The largest capital expenditure cycle in the history of technology is now running at roughly $400 billion a year from just four companies, and the leading analyst tracking this wave says nobody, including the companies writing the checks, knows whether the underlying product has found its market. That is the core provocation in Benedict Evans’s latest annual presentation, "AI Eats the World", released in November 2025 and updated this month. The deck is required reading for anyone deploying capital in enterprise software, infrastructure, or any sector AI is supposed to disrupt.

Evans, a former Andreessen Horowitz general partner whose annual presentations have been circulated inside Sand Hill Road and every major tech company for over a decade, frames the current moment as a classic platform-shift problem with one new and genuinely unsettling variable. In every previous shift, from mainframes to PCs to the web to smartphones, the physical constraints of the technology set a ceiling on how good the next year's product could be. With large language models, that ceiling is unknown. Sergey Brin says the race to AGI is underway. Demis Hassabis says AGI requires multiple further breakthroughs. Both can be right, which is precisely what makes rational capital allocation so difficult.

The infrastructure numbers Evans assembles are staggering in isolation and alarming in context. The four hyperscalers, Microsoft, Alphabet, Amazon, and Meta, spent roughly $400 billion on capex in 2025, more than the entire global telecommunications industry spends annually. That figure was itself a revision: planned 2025 spending nearly doubled during the year as each company raised guidance. Microsoft said it expects FY26 growth to exceed FY25. Meta said Capex dollar growth in 2026 will be notably larger. Alphabet said it expects a significant increase. OpenAI separately announced commitments for 30 gigawatts of capacity at $1.4 trillion, with an aspiration of one gigawatt of new construction per week. Evans calculates that ambition, at $20 billion per gigawatt, implies roughly $1 trillion in annual construction, equivalent to two-thirds of the entire current global data center base, every year.

US data center construction spending has already overtaken office construction and is approaching the pace of warehouse building, according to US Census figures Evans cites. Power availability, not capital, is now the binding constraint: a Schneider Electric survey from February 2025 found utility power access ranked as the top constraint to US data center construction, ahead of chip supply, permitting, and land. Microsoft CTO Kevin Scott has said it has been almost impossible to build capacity fast enough since ChatGPT launched.

The demand side is where Evans’s analysis is the deepest. ChatGPT reported 800 million weekly active users as of late 2025, but Evans notes that apparently only 5 percent are paying subscribers, and the platform announces weekly actives rather than daily actives, which invites skepticism about engagement depth. Deloitte surveys conducted in June 2025 show that substantially more Americans and British consumers use AI chatbots occasionally than daily. Across nine separate surveys between August 2024 and September 2025, daily active usage in the US ranges from roughly 5 to 20 percent of the population, with wide variance depending on methodology.

Evans’s framing is not that this usage is disappointing but that it is structurally predictable. Chatbots as currently deployed require the user to proactively seek out use cases, have a flexible enough job to experiment, and then consciously optimize their workflow, yet most people do not have that profile. The implication for product strategy is pointed: for most of the addressable market, AI capability needs to be wrapped in opinionated tooling and workflow software before adoption scales. That is the opportunity Evans sees for the application layer, but it is also the critique of the current market structure, where three years of enormous infrastructure investment has produced no clear product moat, no dominant distribution model, and no settled answer on where value will accrue.

On model differentiation, Evans shows benchmark convergence data from LMArena and Artificial Analysis as of October 2025: on the most general evaluations, the leading Western models from OpenAI, Google, and Anthropic score within a few percentage points of one another, with Chinese and open-source models closing rapidly; the leading model changes weekly. Evans summarizes the competitive situation as no apparent moats and no clarity on product or value capture. That is a direct challenge to the valuations currently assigned to frontier model companies, and to the venture theses that assume infrastructure providers and model developers will capture the majority of AI-generated value.

The investment question Evans leaves open is the same one the market is pricing in real time: whether generative AI is "only" as transformative as mobile or the internet, in which case the capex cycle is rational and the application layer opportunity is enormous, or whether the ceiling on model improvement is high enough to make current infrastructure spending look conservative in retrospect. For founders, the clearest signal in the data is that distribution and product wrapping matter more right now than raw model capability. For investors, the slide that deserves the most attention is the one Evans titles "rational actors": each major player in the ecosystem is behaving logically given its own balance sheet and competitive position, yet the aggregate dynamic has the structure of a bubble. Evans does not call one out in his analysis; he notes only that every bubble is different - and they are generally still bubbles. The full presentation is available at ben-evans.com/presentations.