Anthropic's infrastructure plan reveals the hidden layers of AI investment

The anatomy of an AI investment

Anthropic's recent IPO prospectus reveals a massive $161.2 billion commitment to infrastructure, primarily with chip designer Broadcom. This spending highlights that investing in artificial intelligence is not a single trade, but rather a complex web of at least seven distinct layers. These layers include fabrication tools, memory, compute, networking, physical infrastructure, power, and the final applications.

Each layer in this stack depends on the others to function effectively. For example, fabrication tools set the pace for compute capabilities, while networking determines if those systems can run at a large scale. Power availability ultimately decides how quickly these finished systems can be switched on and used by the public.

The application layer sits at the top of this hierarchy and is responsible for generating the actual return on investment. Anthropic itself occupies this space, meaning it must generate enough revenue to pay for the massive costs incurred by the layers beneath it. As Finance Magnates reported, one company's operating costs become the revenue for the chipmakers, cloud providers, and power generators that support them.

Profitability and market reactions

The market has not rewarded every layer of the AI stack equally, leading to a divergence in share prices and earnings forecasts. Analysts have raised forecasts most for fabrication, compute, physical infrastructure, and memory, while power was the only layer revised lower. Interestingly, share prices have moved almost in reverse of these upgrades, with fabrication falling 25.6% while applications rose 33.3%.

This discrepancy suggests that investors are currently prioritizing revenue that has already arrived over revenue that has simply been ordered. The application layer faced skepticism earlier this year regarding its ability to pay for the buildout, but Microsoft's 30 million paid Copilot seats helped provide some clarity. The market is looking for proof that the software can eventually monetize the heavy hardware spending.

Profitability also splits the stack into two distinct categories based on their proximity to the silicon. The layers closest to the hardware, such as compute and memory, are seeing margins above 55%. In contrast, the physical chain layers are seeing revenue lag behind their orders because they are being asked to invest heavily before the final cash arrives from the top of the stack.

Shadow borrowing and credit risks

The financing of the AI buildout has shifted significantly toward the bond market in recent years. During 2025, the five largest US hyperscalers released $121 billion in investment-grade bonds, a sharp increase from their annual average of $28 billion between 2020 and 2024. By October 2025, debt linked to AI had grown to become the primary portion of the US investment-grade market.

However, these hyperscalers are also utilizing off-balance sheet arrangements like special purpose vehicles and long-term leases. These methods are often funded by private credit and insurers, which the Bank for International Settlements describes as "shadow borrowing." This practice is noted for its potential to create new channels for transmitting economic shocks across the financial system.

Because of these financing methods, a client holding AI equities may also be carrying significant exposure to interest rates or credit risk. The real level of risk for an investor might be much higher than what their equity allocation alone suggests. This means that a portfolio looking diversified on the surface could actually be highly concentrated on the same macroeconomic catalysts.

Physical constraints and bottlenecks

The growth of AI is ultimately bound by physical realities rather than just technological progress. Every model relies on chips, energy, and about 40 different critical minerals. The International Energy Agency expects data-center electricity demand to more than double to around 945 terawatt-hours by 2030, creating a massive strain on existing utility systems.

In many established data-center hubs, the lead times for connecting to the power grid can stretch out to ten years. These physical constraints do not affect every layer of the AI stack in the same way. A shortage in a specific area raises prices for whoever controls that bottleneck while simultaneously increasing costs for everyone who depends on it.

These constraints can create a lopsided economic environment where one layer sees its margins supported by scarcity while another sees its margins compressed by high costs. For example, a power shortage might benefit a utility company but hurt a cloud provider trying to expand its capacity. Investors must therefore look past the technology label to see which specific physical constraints will impact their holdings most.

The path to economic value

The fundamental assumption for the entire AI stack is that the technology will produce enough economic value to justify the massive capital expenditure. Current OECD modeling suggests that potential productivity gains could be between 0.4 and 1.3 percentage points a year, provided that workers reskill and institutions adapt. Until these gains appear in national accounts, the spending remains a cost line rather than a proven return.

This is why the application layer is considered disproportionately important for the long-term health of the industry. If the revenue from these applications continues to grow, it validates the capital invested in the layers beneath it. If the application layer stalls, the most capital-intensive layers are likely to be the first to see a repricing of their value.

Anthropic serves as a primary test case for this dynamic, having reported $16.3 billion in revenue for the first half of 2026 against commitments that span a decade. Every dollar of revenue at the top of the stack represents a dollar of spending somewhere below it. Investors who grasp where that spending settles and what it relies on will be better positioned than those who only trade on the general AI label.