The AI economy is experiencing a historic surge in capital expenditure, model development and market value. Central banks and multilaterals now warn that elevated technology valuations—fueled by hopes of dramatic productivity gains—could correct sharply if expectations outpace delivery. The Bank of England recently cautioned that the “risk of a sharp market correction has increased,” while the IMF has compared the current AI boom to dot-com dynamics that may end in a bust even if systemic risk to banks remains contained [1][2]. This article weighs the evidence for and against an AI bubble, identifies the tell-tale signs, explores plausible timelines and outcomes, and sets out practical responses. It also integrates a central contention shaping today’s capex spree—that speculation about Artificial General Intelligence (AGI) is a major driver of infrastructure investment—and evaluates what happens if AGI fails to appear within a few years.
Why this matters now. AI-related spending is scaling rapidly. Gartner forecasts total worldwide AI spending of nearly $1.5 trillion in 2025, with a substantial share flowing into cloud, chips and data-centre infrastructure required to train and serve increasingly large models [3]. McKinsey estimates that keeping pace with compute demand through 2030 could require around $6.7 trillion of global data-centre investment, most of it AI-oriented capacity [4]. Energy systems will feel the strain: the IEA projects data-centre electricity consumption will more than double by 2030 to roughly 945 TWh, with AI the most significant driver [5]. Policymakers are starting to integrate compute and energy into national AI strategies, highlighting trade-offs between competitiveness, sustainability and resilience [6].
Are we in a bubble? The case for caution. A classic bubble features valuations detached from fundamentals, herd behaviour and over-build ahead of adoption. Several present-day signals rhyme with that template. First, hyperscaler capex has sprinted ahead of monetisation clarity. Meta, for example, guided 2025 capital expenditure to $66–72 billion, explicitly to scale AI infrastructure; peers have issued similar guidance, implying a multi-hundred-billion annual run-rate for AI-related capex across the largest platforms [7]. Second, the revenue mix from AI remains uncertain for some suppliers; even as leaders post extraordinary growth, analysts note spending caution and uneven data-centre contributions through earnings cycles, a potential sign of digestion after a rapid buildout [8]. Third, private and sovereign capital is moving into large, long-dated infrastructure exposures—such as a recent $40bn data-centre takeover—concentrating risk if demand proves slower than hoped [9]. Fourth, consultants estimate the current building spree would require unprecedented levels of annual AI revenue by decade-end to justify the asset base—raising the bar for commercial pay-off [10].
The counter-case: tangible progress and deep pockets. AI is already delivering value in pockets and the largest investors are among the world’s most cash-generative enterprises. McKinsey’s 2025 global survey finds adoption continuing to climb and early bottom-line impact where organisations redesign workflows and deploy at scale under strong governance [11]. Gartner also expects end-user spending on specialised generative models to grow rapidly in 2025, signalling demand for targeted, domain-specific deployments [12]. And the capex is building durable capabilities—high-density campuses, power interconnects, advanced cooling—that can be repurposed across use-cases and cycles. Even individual sites underscore the scale and optionality: Meta’s new AI-ready campus in Texas, designed to scale toward a 1-GW footprint, is intended to match use with renewable energy and water-efficiency commitments [13].
What’s driving the surge: the AGI expectation premium. A central, often unspoken, thesis behind the infrastructure boom is the belief that AGI could arrive within the foreseeable planning horizon. If models approach human-level competence across tasks, the addressable market for AI services, agentic automation and scientific discovery could expand by orders of magnitude—justifying today’s fixed-asset sprint. That thesis is reflected both in bullish lab rhetoric and in cautious takes from prominent researchers. Andrej Karpathy, formerly of OpenAI, recently suggested truly capable agents may be ~a decade away, citing gaps in planning, multimodality and continual learning [15]. By contrast, Sam Altman has argued that we may achieve AGI sooner than most expect and has written that his team is confident they “know how to build AGI,” even as OpenAI pivots harder toward enterprise monetisation to fund that vision [17]. On the sceptical side, Meta’s Yann LeCun has repeatedly said that current LLM paradigms are unlikely to yield general intelligence quickly and that AGI will take years, if not decades, requiring new architectures and learning approaches [16]. Independent analyses also lay out reasons why AGI may still be decades away, emphasising unresolved scientific bottlenecks [18].
If AGI disappoints near-term: implications for the asset base. Should broadly general agents fail to materialise within several years, today’s infrastructure looks less obviously accretive in the short run. Demand could still grow—through specialised models, vertical copilots and retrieval-augmented systems—but the revenue slope might be shallower than capex curves assume. That mismatch risks lower utilisation, delayed payback periods and a valuation reset for the most AGI-levered names. Because the build also strains power grids and water systems, political scrutiny could intensify if societal returns lag expectations; several governments are already reviewing compute policy and energy impacts to balance competitiveness with resilience [6][5].
Key signs to watch over the next 12–24 months. First, guidance versus delivery: track whether leading vendors can convert GPU-hours into high-margin, recurring software and services revenue. Second, earnings breadth: watch for diffusion of AI revenue beyond a handful of mega-caps into the wider enterprise software, chips, power and real-estate ecosystems. Third, unit-economics: look for falling inference costs, better model-quality-per-dollar and higher server utilisation that validate returns on capex. Fourth, energy and permitting: grid interconnect lead times, power contracts and policy reforms will determine how quickly new capacity can earn a return. Finally, macro signals: central-bank and IMF commentary points to valuation sensitivity if productivity evidence underwhelms [1][2].
What’s likely to happen—and when. Near-term (next 12 months), valuations could remain elevated but more discriminating, rewarding firms that show clear enterprise traction and sustainable gross margins. As 2026 approaches, outcomes bifurcate. In a benign scenario, inference costs fall, enterprise copilots show measurable productivity and vendor lock-in stabilises revenue quality, allowing the capex wave to look prescient. In a cooler scenario, digestion sets in: utilisation lags, model switching increases, and infrastructure ROIC disappoints, prompting a 20–30% multiple compression among the most AI-levered names. In a harsher variant, an earnings air-pocket exposes overbuild; a sharper drawdown ensues but remains non-systemic—consistent with the IMF’s baseline [2]. The exact path will hinge on whether vendors can translate experimentation into durable, scaled adoption by late 2026.
For, against—and the AGI speculation question. For: the case that we are not in a classic bubble rests on real progress and strong balance sheets. Enterprises report early value where they rewire processes for gen-AI, not just bolt it on; specialised models and retrieval-augmented systems are already improving service, fraud, coding and knowledge work in measurable ways [11][12]. Infrastructure is not single-purpose; it can serve cloud, analytics and traditional workloads across cycles. Large pools of private and sovereign capital can extend timelines, smoothing volatility in utilisation [9]. Against: the scale of the build presumes steep revenue ramps and big unit-cost declines. If compute-hungry models encounter diminishing returns or slower enterprise absorption, the mismatch between expectations and earnings could trigger a valuation reset. Power constraints, permitting frictions and grid decarbonisation challenges raise execution risk and could delay pay-back [5].
“AGI speculation is a principal driver—and a fragility”. Speculation about AGI is one of the biggest reasons for the infrastructure spree; if AGI fails to be invented within a few years, doubt could cloud AI’s commercial future. Weighing the for and against: For: even if AGI slips, the intermediate advances pursued along the way—domain-specific models, reasoning tools, agentic copilots, AI-accelerated science—can justify much of the build by creating profitable, compounding businesses. History shows over-investment sometimes lays the foundation for the next wave (as fibre overbuild later powered broadband and cloud). Against: the attention, pricing and board-level narratives anchoring today’s capex assume step-changes in capability and adoption; if generality stalls, CFOs may revisit ROI hurdles, stretching amortisation and depressing returns. Research leaders such as LeCun argue that new architectures are required and that current systems will not scale smoothly to general intelligence, which increases timeline uncertainty [16]. We also feel, the extension of existing AI technology is not enough to reach AGI. Karpathy’s decade-ish horizon for capable agents suggests that the payoff for the most aggressive builds may arrive later than equity markets currently discount [15], while Altman’s confidence anchors the bull case that near-AGI capability (and monetisation) can emerge sooner [17]. Independent assessments underline that AGI may still be decades away, which—if true—raises the risk of an expectations reset [18].
Causes of bubble-risk dynamics. Abundant cash flows at hyperscalers, cheap(er) capital until recently, and geopolitical drives for digital sovereignty all funnelled money into GPUs, campuses and power. The industrial logic is clear—build now to capture scale economies and moats later. The narrative logic, centred on rapid progress toward AGI, adds a premium to today’s valuations. But the physical economy (power, land, water, interconnects) introduces constraints that software cycles have not historically faced; this is why many governments now explicitly consider compute and energy in AI strategy, highlighting a different risk profile from past software booms [6][5].
What would a correction look like? A mild cooling would see slower capex growth, consolidation among AI-native startups, and a rotation from “vision premium” to cash-flow durability. A deeper correction would likely be earnings-led—missed monetisation milestones or utilisation shortfalls—triggering multiple compression, particularly in names with the highest AGI-dependent narratives. Yet even in a drawdown, the IMF expects limited banking-system contagion given the equity-financed nature of much AI investment [2].
What to do now: solutions for investors, operators and policymakers. Investors should diversify across the stack (chips, power, facilities, software and services), prioritise firms with proven pricing power and falling unit costs, and underwrite more conservative adoption curves. Track whether vendors replace one-off trials with embedded workflows, and whether gross margins improve as inference gets cheaper. Operators should stage capex against contracted demand, secure long-dated power at predictable prices, and design for re-use across workloads to protect ROIC if AGI timelines extend. Commercial teams should pivot from “AI everywhere” to outcome-priced offerings tied to customer KPIs. Policymakers should modernise planning and interconnect processes, expand clean-power build-outs, and set transparent compute-impact reporting so markets can price risks. They should also steward skills and safety standards to ensure AI’s benefits diffuse even if revenue ramps are slower than narratives imply [6][5].
Our view (incorporating your opinion). We judge that markets are in an AI bubble-risk phase: enthusiasm and capital are running ahead of broadly proven monetisation, but the underlying technology is progressing and much of the build will remain useful. Crucially, AGI speculation is both the primary accelerator and a key vulnerability of today’s cycle. If AGI-level capability—or a commercially equivalent inflection—arrives within a few years, the current infrastructure will look visionary and value-accretive. If it does not, confidence could ebb, capex could reset, and a more selective, productivity-first approach would take hold. Our base case over the next 12–24 months is a re-rating toward fundamentals: winners will demonstrate durable enterprise adoption, falling cost-to-serve, and clearer ROI; others will see capital rationed and valuations normalised. Either way, the sensible path is to keep building—but stage it against evidence, not headlines—so that if AGI takes longer, the asset base still clears a robust return hurdle [1][2][4][16][15].
Summary: The AI economy exhibits several bubble-like features: aggressive capex, valuation premia and narratives leaning on rapid capability jumps. Against this, adoption is growing, infrastructure is multi-purpose and leading firms have the capital and operating leverage to play long games. The pivotal uncertainty is AGI timing. If general intelligence (or equivalent breakthroughs) arrives soon, the investment case strengthens; if not, a valuation and capex reset is likely, but the long-term arc of AI diffusion should continue. Pragmatic staging of investment, transparency on power and utilisation, and rigorous ROI discipline are the best antidotes to bubble risk.
[1] Is there an AI bubble? Financial institutions sound a warning — link
[2] Opinions split over AI bubble after billions invested — link
[3] Gartner Says Worldwide AI Spending Will Total Nearly $1.5 Trillion in 2025 — link
[4] The cost of compute: a $7 trillion race to scale data centers — link
[5] IEA: AI set to drive surging electricity demand from data centres — link
[6] OECD: AI compute policy and capacity — link
[7] TechCrunch: Meta to spend up to $72B on AI infrastructure in 2025 — link
[8] Reuters: Nvidia outlook and data centre spending caution — link
[9] Financial Times: $40bn Aligned Data Centers acquisition — link
[10] Wall Street Journal: Spending on AI is at epic levels — link
[11] McKinsey: The State of AI 2025 survey — link
[12] Gartner: 2025 spending on specialised GenAI models — link
[13] Reuters: Meta commits $1.5bn for AI data centre in Texas — link
[14] McKinsey PDF: The State of AI (methodology and adoption detail) — link
[15] Business Insider: Karpathy says capable agents likely a decade away — link
[16] AI Business: Yann LeCun says AGI is years/decades away — link
[17] Business Insider: Sam Altman’s AGI confidence and timeline remarks — link
beFirstComment