What Counts as "AI Infrastructure"? Almost Everything, Now

Andy Jassy said the quiet part out loud. In May, defending Amazon's planned $200 billion in "AI infrastructure" spending, the CEO listed what the figure covers:

AWS is to lay out cash for land, power, buildings, chips, servers, and networking gear in advance of when we can monetize it, typically six to 24 months before we start billing customers, depending on the component. — Andy Jassy

. Land. Power. Buildings. Those are the costs of putting up a data center for any purpose, AI or otherwise. They are not incremental spending driven by a new technology. They are the cost of a building ^[1]. That sentence from Jassy is the cleanest admission of a pattern that runs across the entire "AI capex" story. The label has expanded, company by company, until it covers spending that would occur with or without AI. Amazon is far from alone. Morgan Stanley projects $3 trillion in global data center construction by 2028 and attributes the full amount to "enterprise spending on AI agents," though the figure encompasses buildings, power systems, cooling, and land, not just AI-specific compute ^[2]. McKinsey's $7 trillion-by-2030 estimate works the same way: it covers all data center construction capable of housing computing units, regardless of whether those units run AI workloads ^[3]. Gartner projects total public cloud spending exceeding $1 trillion by 2027 and ties it to "agentic AI workloads," but public cloud spending includes storage, databases, and SaaS hosting that have existed for a decade ^[4]. Then there is SpaceX. Bank of America initiated coverage of the rocket-and-satellite company with a Buy rating, calling it "a foundational AI and infrastructure provider rather than solely a rocket launcher" ^[5]. SpaceX's own IPO filing claims a $28.5 trillion total addressable market, of which $26.5 trillion, or 93 percent, is tied to AI ^[6]. A company that launches rockets and operates a satellite internet constellation has classified nearly all its future value as AI. The relabeling is not pure invention: SpaceX has real GPU clusters, including 220,000 Nvidia GPUs in Tennessee leased to Anthropic and 110,000 leased to Google ^[6]. But the claim that 93 percent of a rocket company's market is AI is a valuation argument dressed as a market analysis, and the IPO filing gives SpaceX a direct financial incentive to make it ^[6]. Nvidia, the company most precisely associated with AI, is stretching the label in the other direction. Its new Vera Rubin platform targets the CPU market, a $200 billion opportunity Nvidia "has never addressed before" ^[7]. CFO Colette Kress framed the expansion plainly:

VeraCPU opens a brand new $200 billion (total addressable market) TAM for Nvidia, a market we have never addressed before. — Colette Kress

. General-purpose data center CPUs are now AI infrastructure. Jensen Huang has gone further, redefining Nvidia itself from a chip company into an "AI factory that converts electricity to tokens" ^[4]. The reframe is rhetorically powerful and commercially useful: it pulls Nvidia's entire product roadmap, including networking gear, data center operations, and a new revenue-sharing financing arm, under the AI umbrella ^[7]. At the smaller end, Credo Technology makes active electrical cables. It holds roughly 80 percent market share in its niche. Brown Capital Management, a fund that owns the stock, calls it "positioned at the center of the buildout for AI data centers" ^[8]. The cables are general data center components. They are not specific to AI workloads. But because hyperscaler data center construction drives demand for them, the company is now classified as an AI play ^[8]. Meta's recent stock movements tell the same story from the demand side. Shares rose 3 percent to $603.58 on July 6, driven not by any AI product breakthrough but by hardware dominance in smart glasses and a potential cloud-computing business ^[9]. That cloud business, internally called Meta Compute, would sell surplus AI compute capacity. Zuckerberg has acknowledged the company may have "overbuilt" its data centers ^[10]. The market rewarded that admission with a 10 percent stock rise, valuing the infrastructure-pivot story over AI product success. Meta's broader $600 billion three-year plan bundles data center construction, custom silicon, and AI product development into a single "AI infrastructure" figure, while a 17 percent stock drop in June showed investors are not sure the bundled spending generates high-margin revenue ^[11]. The spending itself is real, and some of it is genuinely AI-specific. Google Cloud grew 63 percent year over year to $20 billion with a $460 billion backlog, anchored in part by Anthropic's $200 billion five-year commitment on purpose-built AI processors ^[12]. Broadcom's custom AI chip sales surged 65 percent to $20 billion with projections of $60 to $90 billion by 2027 ^[13]. Jefferies reports a 12 gigawatt global data center capacity deficit, with $2 trillion in cloud service backlogs, attributing the shortage to supply chain bottlenecks rather than speculative overbuilding ^[14]. The infrastructure demand is not fictional. But the same quarter that saw Broadcom's custom AI chip sales surge also saw pure AI software companies BigBear.ai and C3.ai post revenue declines and widening losses ^[13]. The companies that would most directly validate AI-specific application demand are struggling. The money is flowing to hardware and chips serving hyperscalers' mixed workloads, not to AI software proving the technology's value. TS Lombard analysts put global AI capex at $800 billion for 2026, with U.S. hyperscalers accounting for $700 billion, and warn the valuations may be unsustainable by historical analogy:

significant expansion — Alphabet Inc.

^[15]. The comparison is to Britain's Railway Mania, where infrastructure investment created long-term economic benefits but severe investor losses ^[15]. The hyperscaler capex figures that anchor this debate sit at roughly $570 to $700 billion for 2026, with Amazon at $200 billion, Alphabet at $180 to $190 billion, and Microsoft at $190 billion ^[16][17]. AWS revenue rose 28 percent to $37.6 billion, and Google Cloud hit $20 billion ^[18]. That revenue is real. But it is cloud revenue, driven by general capacity expansion as much as by AI workloads. The $200 billion Amazon figure, by Jassy's own description, includes land, power, and buildings ^[1]. The $3 trillion Morgan Stanley figure includes all construction costs ^[2]. The $1 trillion Gartner cloud figure includes all cloud services ^[4]. This is why the bubble debate has stalled. The hawks, led by TS Lombard's railway analogy, say the spending will not deliver proportional returns ^[15]. The bulls point to Google Cloud's $460 billion backlog and the 12 gigawatt supply deficit as proof that demand outruns supply ^[12][14]. Both are arguing over a number that bundles land, power, buildings, general-purpose CPUs, electrical cables, cloud storage, satellite internet, rocket launches, and smart glasses into one line called "AI infrastructure." Neither side has disaggregated it. The question worth asking before the next earnings season is simpler than the debate suggests: of the $600 billion to $1 trillion in headline "AI capex," how much is spending that would not exist without AI? No company or analyst has answered that with a number. Until someone does, both the bubble warnings and the spending forecasts are built on a category that has lost its meaning.