the AI compute
denominator
problem.

The path from a hyperscaler's capex budget to a deployed GPU cluster passes through a surprisingly narrow set of chokepoints. ASML in the Netherlands holds a legal and technological monopoly on extreme ultraviolet lithography equipment. Without it, TSMC cannot produce the leading edge chips on which all current AI accelerators depend. TSMC's CoWoS packaging line assembles those chips with the high-bandwidth memory that makes them useful. Three memory companies produce that memory. Then the assembled hardware must be powered. The network below shows the relationships. Drag any node to explore.

Between fiscal year 2022 and the guidance each company has given for 2025, the four largest hyperscalers have increased their combined stated capital expenditure by roughly $190 billion. The increases are not gradual. Alphabet's capex more than doubled between 2023 and 2024, then the company guided a further 43% increase for 2025. Microsoft announced $80 billion in AI infrastructure spending for fiscal 2025 at a moment when its previous year total was $55.7 billion. These are the numbers from the filings, from the earnings calls, and from the company press releases.

The figure below shows each hyperscaler's 2023 capex as the background bar (baseline) and 2025 guidance as the foreground bar. The orange line marks $60 billion, the level above which each company has either reached or committed to reach.

Capital commitments are a claim on future production. Whether that production materializes depends on the physical capacity that exists or can be built in time. There are four layers where the math does not work on the stated timeline.

The first layer is lithography. ASML is the only company in the world that produces extreme ultraviolet lithography machines, the equipment that etches transistors at the dimensions required for current AI accelerators. In its 2024 annual report, ASML disclosed shipments of 44 EUV systems for the full year. Its High-NA EUV next generation tools, required for the chip geometries beyond the current generation, were just reaching initial production volumes. ASML cannot be quickly substituted. The machines take years to manufacture and require a supply chain of their own: the light source technology alone involves hundreds of specialized components from suppliers that took decades to develop.

The second layer is packaging. TSMC's CoWoS (chip on wafer on substrate) process is the packaging technology required for NVIDIA's Hopper and Blackwell GPU families. Assembling a GPU means bonding the logic die to high-bandwidth memory stacks on an interposer, then mounting the assembly on a substrate. TSMC is the primary provider of this service at scale. CoWoS capacity cannot be doubled in a quarter. New capacity requires new equipment, new cleanroom space, and a trained workforce. TSMC guided capital expenditure of $38 to $42 billion for 2025, much of it directed at capacity expansion. The packaging expansion is one of the reasons NVIDIA's Blackwell production ramp in late 2024 ran behind initial commitments.

The third layer is high bandwidth memory. HBM is the memory stack that sits adjacent to the GPU logic die and moves data fast enough to feed it. Without HBM, a Blackwell or MI300X GPU is not functional as an AI accelerator. Three companies produce HBM: SK Hynix, Samsung, and Micron. SK Hynix holds the supply relationship for NVIDIA's current H200 and B200 series. Converting standard DRAM production lines to HBM is capital intensive and takes 12 to 18 months. HBM supply was described as tight through 2025 and into 2026 in investor presentations by all three producers.

The fourth layer is power. A data center cannot operate without a grid connection, and large new loads in the United States face interconnection queues that average two to five years. In the PJM region, which includes Northern Virginia (the densest AI infrastructure market in the world), the interconnection queue had more than 280 gigawatts of pending requests as of 2024. Microsoft, Google, Amazon, and Meta have each disclosed power acquisition as a risk factor and an operational constraint in their most recent annual filings.

NVIDIA's FY2025 results (fiscal year ending January 2025) are the most direct measure of how much AI compute was actually delivered. Data center revenue reached $115.2 billion, up 142% from the prior year. Total revenue was $130.5 billion. Gross margin was 75.0%. These are exceptional numbers by any standard. They are also a ceiling, not a floor: NVIDIA can only recognize revenue on hardware that has been manufactured, packaged, and shipped.

The Blackwell architecture, NVIDIA's current generation, launched in late 2024. Initial production ramps ran behind schedule, disclosed in NVIDIA's Q2 and Q3 FY2025 earnings calls, with the company attributing delays to the complexity of the new packaging design. Blackwell requires a new substrate configuration that increased the difficulty of CoWoS packaging at scale. The delays were resolved through 2024, and Blackwell shipments ramped significantly in Q4 FY2025. But the episode illustrates the mechanism: each new chip generation encounters the packaging bottleneck fresh, and resolving it takes time.

The hyperscaler order books for NVIDIA hardware extend well into 2025 and 2026. Alphabet, Microsoft, Meta, and Amazon have all disclosed NVIDIA as a critical supplier. The concentration is notable: four of the five largest companies in the world by market capitalization are dependent on a single chip designer whose production is dependent on a single foundry whose packaging line is dependent on a single memory supplier for the specific HBM variant required. At each link, there is no backup.

Every one of the four hyperscalers discussed AI infrastructure spending on their most recent earnings calls. Every one of the same four companies disclosed supply chain dependency as a risk factor in the annual filings released in the same reporting period. The two sets of disclosures use different language about the same physical reality.

The pattern is consistent across all four companies. On earnings calls: confidence, commitment, dollar figures, competitive urgency. In the annual filings: dependency on a limited number of suppliers, risk of disruption, acknowledgment that the hardware may not be available on the timelines required. Neither set of disclosures is false. They describe the same situation from two different angles, and together they define the denominator problem precisely.

All data in this note comes from public annual filings, earnings call transcripts, or official company disclosures. No third-party financial data vendors were used. Where exact filing page numbers are provided, they refer to the PDF version of the filing as submitted to the SEC or relevant exchange regulator.