AI Capex Economics and Bubble Risk

The current AI infrastructure buildout is characterised by a stark mismatch: unprecedented capital expenditure with modest, commoditising revenue — a structure that critics argue mirrors the telecom bubble of the late 1990s.

What It Is

Fawkes Capital Management (December 2025) presents a bear case on AI economics. The core argument: Big Tech companies are spending at levels that cannot be justified by near-term returns. Google alone will spend ~$60B more in annualised capex in 2025 than it did pre-ChatGPT, with capex now exceeding net profit — a sharp departure from its historically capital-light model. Across the industry, an estimated $500B in AI capex will be deployed between 2022 and 2026, approximately $1,500 per American or $4,000 per household.

The investment is being driven not by calculated return expectations but by fear: when a technology carries the possibility of existential competitive disruption, it becomes rational to invest defensively even at negative returns. Fawkes describes this as a prisoner’s dilemma — each firm keeps spending until all firms simultaneously stop believing AI threatens their dominant position.

Why It Matters (for the Economy)

If the economics of AI infrastructure are as poor as Fawkes argues, two macro-level risks emerge. First, the misallocation of hundreds of billions of dollars toward AI infrastructure could crowd out more productive capital uses. Second, a correction — triggered either by slowing usage growth or a recession that forces cash preservation — could be deflationary in a way that mirrors the 2000–2001 dot-com bust, where telecom capex sprees ended in collapse and a shallow recession.

The “commoditisation of LLMs” thesis is particularly important for investors: if LLMs converge to “good enough” performance and competition shifts to price, the high-margin period for AI infrastructure providers (especially NVIDIA) is likely limited. Sara Hooker (SSRN, 2026) echoes this in her analysis of scaling laws — the period of rapid improvement from simply adding compute may be ending.

Evidence & Examples

  • Google’s capex now exceeds its net profit; cumulative AI-related capex of ~$85B since late 2022 has generated only ~$21B of annualised revenue — a 5% uplift on a massive capital base (2025.12.19-Why-We-Worry-Part-1.pdf)
  • ~$500B in AI capex 2022–2026 requires ~$80B in annual net income to justify; achieving that would require something like 333 million paying ChatGPT users — roughly the entire US population. Current paying users: ~20 million as of Dec 2025 (2025.12.19-Why-We-Worry-Part-1.pdf). 📅 NOTE: The State of Consumer AI (Raw/, March 2026) reports ChatGPT had ~900M weekly active users (total AI apps crossed 1B WAU) by March 2026 — though WAU ≠ paying users, demand growth is substantially stronger than the Fawkes analysis implied.
  • NVIDIA’s ~75% gross margins are a key vulnerability: Google’s TPU infrastructure now rivals NVIDIA’s latest GPUs at significantly lower cost; Amazon Trainium 3 is also closing the performance gap (2025.12.19-Why-We-Worry-Part-1.pdf)
  • Sara Hooker argues the relationship between training compute and model performance is “highly uncertain and rapidly changing” — the scaling formula that justified capex may no longer hold (ssrn-5877662 (1).pdf)
  • Historical parallel: telecom operators in the late 1990s deployed enormous capex on bandwidth infrastructure, found services rapidly commoditised, and triggered a market correction when growth slowed and pricing power collapsed (2025.12.19-Why-We-Worry-Part-1.pdf)
  • Two conditions Fawkes identifies as bubble triggers: (1) AI adoption growth slows materially, or (2) a recession forces cash preservation over technological “defence spending” (2025.12.19-Why-We-Worry-Part-1.pdf)

Tensions & Open Questions

  • The bear case may be too near-term. The Jones (2026) GPT analysis suggests productivity impacts from general purpose technologies routinely take decades to appear. A mismatch between current capex and current returns doesn’t necessarily imply a bubble — it may simply reflect the normal investment-ahead-of-payoff structure of infrastructure cycles.
  • Who captures the value? Even if aggregate AI value creation is large, it may accrue to downstream users (enterprises, consumers) rather than to infrastructure providers. This distinguishes a “misallocation to infrastructure” story from a “no value created” story.
  • The prisoner’s dilemma framing: Fawkes argues spending continues as long as no one firm believes the existential threat has passed. But if one large player (e.g. Google) achieves sufficient AI capability advantage, the game dynamics could shift rapidly — either triggering consolidation or competitive withdrawal.
  • ⚠️ CONTRADICTION: The Contrary Tech Trends Report 2026 (investor perspective) takes a more bullish view of AI infrastructure economics and vertical applications — worth cross-referencing when that source is more fully extracted (see AI Investment Thesis Capex and Returns).

LLM Commoditization · Scaling Law Uncertainty · AI Investment Thesis Capex and Returns · AI as a General Purpose Technology