Anthropic’s compute bill shows where AI investment really flows

Anthropic's IPO prospectus, made public on October 26, 2026, reveals commitments totaling hundreds of billions of dollars to computing infrastructure. The largest commitment, $161.2 billion, is with Broadcom, a chip designer. Google, Amazon and SpaceX each account for more than $100 billion, while Microsoft has commitments of about $45 billion, with some contracts extending to 2036.

While media coverage has focused on the potential $100 billion IPO size, which would make it the largest on record, the spending commitments reveal more about AI investment flows. These commitments show that AI is not a single trade but at least seven distinct layers that behave very differently.

AI as a Supply Chain

In 'The Architecture of Disruption,' published June 29, 2026, EXANTE's Horacio Coutiño divides the AI buildout into seven investable layers: fabrication tools and process materials, memory, compute, networking, physical infrastructure, power, and applications. Each layer depends on the others, with fabrication setting the pace of compute and networking enabling cluster-scale operations.

Anthropic's bill touches every layer. The Broadcom agreement sits in compute, while cloud contracts extend to memory, networking, physical infrastructure and power as hyperscalers build capacity. Anthropic itself operates in applications, the layer that must pay for the rest.

Since June 29, analysts have raised earnings forecasts most for fabrication, compute, physical infrastructure and memory, while power was the only layer revised lower. Share prices moved almost in reverse, with fabrication falling 25.6% while applications rose 33.3%.

Profitability and Credit Risk

Profitability splits the AI stack in two. Layers closest to silicon are growing and becoming more profitable, with compute and memory margins above 55%. Further down the physical chain, revenue lags orders as these layers are asked to invest before cash arrives.

AI hyperscalers are also using off-balance sheet arrangements such as special purpose vehicles and long-term leases funded by private credit and insurers. The Bank for International Settlements describes this as 'shadow borrowing,' potentially creating new shock transmission channels.

Physical Constraints

In 'Beneath the Gold Rush,' published October 15, 2026, Friedman argues that AI exposure cannot be assessed as a technology allocation alone. Chips, energy and about 40 critical minerals sit beneath every model.

The International Energy Agency expects data-center electricity demand to more than double to around 945 terawatt-hours by 2030, with grid connection lead times in some hubs stretching to ten years. These constraints affect layers differently, raising prices for bottleneck controllers while increasing costs for dependent layers.

Implications for Investors

OECD modeling puts potential productivity gains from AI at 0.4 to 1.3 percentage points annually, provided workers reskill and institutions adapt. Until gains appear in national accounts, AI spending remains a cost line.

For brokers, the layer view has three practical consequences: AI products carry different risks, positions may appear diversified but aren't, and client communication about what a theme actually contains becomes critical. Anthropic reported $16.3 billion in revenue in the first half of 2026 against decade-long commitments.

Treating AI as one trade hides most risk. A more useful approach asks: which layer does a holding sit in? Is that layer's growth showing in margins and cash flow or only orders? Which physical or financing constraint would hurt it most?