Can AI justify the current capex boom?

We’ve reached the end of our AI series, having mapped the productivity upside, the limits of current models and the rise of agentic systems. This closing article tackles the question behind the trillion-dollar build-out: is the AI capex boom rational or wildly premature? The answer depends on history, adoption and the speed at which AI becomes more than hype and finally becomes infrastructure.

The historical context: capex precedes productivity

Every general purpose technology that has transformed the economy required huge, often initially baffling, upfront investments that often outpaced short-term returns.

The pattern is clear. Infrastructure spend is the prerequisite for productivity acceleration. The return on investment becomes visible after the build-out, not before. This raises the central question for investors: can the AI capex boom be justified?

When companies like Microsoft, Amazon, and Google collectively commit to spending hundreds of billions of dollars per year on AI infrastructure, the numbers look concerning. But a closer look at the nature of this spending reveals a critical divergence in risk.

Demand pull and supply push

To understand the rationality of this capex cycle, we must distinguish between the Hyperscalers (Microsoft, Google, AWS) and the model builders (OpenAI, Anthropic). The market often lumps their spending together, but their economic realities are opposites.

• Hyperscaler Capex is "demand pull": They are building infrastructure because customers have already signed contracts to use it.
• OpenAI Capex is "supply push": They are spending huge amounts to create a capability, hoping that demand will materialise to cover the costs.

The hyperscalers: building into a backlog

For Microsoft, Google, and Amazon, the current AI build-out is not a speculative gamble but an industrial necessity driven by contractual demand. Unlike a "Field of Dreams" strategy where infrastructure is built in the hope that users will arrive, these companies are facing supply challenges where their cloud divisions are already capacity constrained. This reality was highlighted when Microsoft announced a record US$35 billion in quarterly capital expenditures alongside a crucial increase in its Remaining Performance Obligation (RPO).

This is the "Cloud 2.0" thesis. The demand signal is flashing red. Every startup, enterprise, and government agency training a model needs GPU clusters today.

• Amazon (AWS) explicitly states that their AI capex has "immediate ROI" because they are selling compute directly to customers who are waiting in line for it.
• Google Cloud has seen its revenue run rate top US$50 billion, driven by what management describes as "surging demand" that outstrips its ability to build data centres.

For hyperscalers, this hardware is fungible. Even if the generative AI "revolution" slows down, the chips and data centres can be repurposed for traditional cloud workloads, which continue to grow.

In contrast to the hyperscalers, OpenAI sits at the speculative frontier of the boom. Its trillion dollar compute commitments are not backed by a predictable order book but by a conviction that future demand will materialise. It is a necessary gamble for a company trying to push the frontier rather than follow it, but the economics still boil down to a leap of faith. Fingers crossed! 

The four structural risks

While the hyperscaler bet is supported by current contracts, the sustainability of those contracts depends on the long-term economic reality of AI deployment. Four specific risks threaten to decouple the cost of infrastructure from the value it generates:

1. Model efficiency driving down value

There is a paradoxical risk that AI becomes too efficient. Rapid advances in Small Language Models (SLMs) are delivering high performance at a fraction of the compute cost. If the cost of intelligence collapses towards zero, the revenue per user may shrink faster than the market expands. We could end up in a scenario where AI is ubiquitous but demonetised.

2. Enterprise adoption is too slow

The gap between technological capability and organisational reality is wider than anticipated. While a model can be coded in seconds, deploying that code in a regulated bank can take months. Corporate inertia and legacy data issues are acting as a brake on adoption. If enterprises remain stuck in purgatory for another 2–3 years, the demand curve will flatten, leaving them with idle silicon.

3. Value capture doesn't replicate prior eras

In the PC and Cloud eras, IT players captured a significant share of the productivity surplus. With AI, there is a risk that the benefits accrue primarily to the customer rather than the vendor. If AI is viewed as a commodity utility rather than a differentiated software suite, the "capture rate" for tech giants may be lower than historical norms.

4. The limits of the knowledge base

Our US$5.25 trillion productivity thesis assumes that AI will elevate a vast swath of the global workforce. But what if the "benefit radius" is smaller? If current LLMs only truly unlock productivity for elite coders while remaining hallucination-prone toys for the broader administrative workforce, the Total Addressable Market (TAM) is drastically smaller than the capex implies.

The upside of agentic AI

Despite these risks, the ceiling for this technology remains uncapped, specifically due to the emergence of Agentic AI. This is the pivot that turns the "speculative" bet into a "rational" one.

We are moving from "chatbots" (which wait for instructions) to "Agents" (which execute workflows). This shifts the economic model from selling "seats" to selling "outcomes." In an agentic future, a company doesn't buy software to help a human write an invoice; it buys an Agent that handles invoicing. This decoupling of revenue from human seats allows AI value capture to scale infinitely, breaking the "Knowledge Base" constraints mentioned above. If Agents work, the friction of enterprise adoption dissolves because the ROI becomes immediate and tangible: labour replacement and augmentation at scale. This is the "Blue Sky" scenario that justifies every dollar of the current spend.

Final Thought

The current capex boom is a tale of two timelines where the hyperscalers are rational actors fulfilling the concrete demands of the present, while the model builders are visionaries gambling on the agentic promise of the future. The bridge between them is precarious and built on expensive silicon and hopeful assumptions, yet if the agentic transition holds, today's absurd spending will look like a bargain in hindsight.

Article 1 - The glacial start of the PC era

Article 2 – How centralised capex heralded the speed of AI

Article 3 – Compressing decades of diffusion into single digit years

Article 4 – How much productivity is truly at stake. An economic framework

Article 5 – Can AI justify the current capex boom