Anthropic’s IPO prospectus, made public on 26 October 2026, commits hundreds of billions of dollars to computing infrastructure. That bill says more about investing in AI than the size of the listing itself. AI is not one trade. It is at least seven, and they behave very differently.
The largest commitment, $161.2 billion, is with Broadcom, a chip designer. Google, Amazon and SpaceX each account for more than $100 billion and Microsoft for about $45 billion, with some contracts running to 2036.
Much of the coverage so far has focussed on size. Reports suggest the offering could raise around $100 billion, which would make it the largest IPO on record. For brokers and platforms, that means another surge of clients asking for AI exposure.
The spending commitments by Anthropic matter more than the headline. One company’s operating costs become revenue for chipmakers, cloud providers, data-centre builders and power generators. This means that two clients who both say they hold AI can be holding very different risks.
AI is a supply chain, not a sector
In The Architecture of Disruption, published on 29 June 2026, my colleague 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. Fabrication sets the pace of compute. Networking decides whether compute can run at cluster scale. Power decides how quickly finished systems can be switched on. The application layer decides whether any of it earns a return.
Anthropic’s bill touches every one of them. The Broadcom agreement sits in compute. The cloud contracts pass down to memory, networking, physical infrastructure and power as the hyperscalers build capacity. Anthropic itself sits in applications, the layer that has to pay for the rest.
The layers are connected, but the market has not rewarded them together. Since 29 June, analysts have raised earnings forecasts most for fabrication, compute, physical infrastructure and memory. Power was the only layer revised lower. Share prices moved almost in reverse. Fabrication, with the strongest upgrade, fell 25.6% to 21 September. Applications, with one of the weakest, rose 33.3%.
The market is paying for revenue that has arrived, not revenue that has been ordered. Applications fell in the first half of the year as investors questioned whether the layer that has to pay for the buildout could do so. Microsoft’s 30 million paid Copilot seats went some way to answering that.
Profitability splits the stack in two. The layers closest to the silicon are growing and becoming more profitable, with compute and memory margins above 55%. Further down the physical chain, revenue lags orders. These layers have full order books, but they are being asked to invest before the cash arrives.
An AI position is also a credit position
The buildout is now financed in the bond market as much as from cash flow. In 2025, the five largest US hyperscalers issued $121 billion of investment-grade bonds. Their annual average between 2020 and 2024 was $28 billion. By October 2025, AI-related debt had become the largest segment of the US investment-grade market. However, AI hyperscalers are also using off-balance sheet arrangements such as special purpose vehicles and long-term leases and purchase agreements funded by private credit and insurers..The Bank for International Settlements describes this as “shadow borrowing” and has highhilighted that these methods are potentially creating new shock transmission channels.
As a result, a client who holds AI equities and is also exposed to rates or credit risk. Therefore the real level of AI risk exposure may be far greater than their equity allocation suggests.
The binding constraints are physical
In Beneath the Gold Rush, published on 15 October 2026, I argue 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 (IEA) expects data-centre electricity demand to more than double to around 945 terawatt-hours by 2030. Grid connection lead times in some established data-centre hubs stretch to ten years.
These constraints do not affect every layer equally. A shortage raises prices for whoever controls the bottleneck and raises costs for everyone who depends on it. The same physical constraint can therefore support one layer’s margins while compressing another’s.
Spend is not the same as return
The case for the whole stack rests on one assumption: that AI produces enough economic value to justify the capital. That is still a claim to be tested.
OECD modelling puts the potential productivity gain at 0.4 to 1.3 percentage points a year, provided workers reskill and institutions adapt. Until the gain appears in national accounts, AI spending on its own is a cost line.
This is why the application layer matters disproportionately. If its revenue keeps growing, it validates the capital beneath it. If it stalls, the most capital-intensive layers are likely to reprice first. Anthropic is a test case: it reported $16.3 billion of revenue in the first half of 2026, against commitments that run for a decade.
What the layers mean for brokers
For firms offering AI exposure, the layer view has three practical consequences.
Products labelled AI can carry very different risks. A basket weighted towards applications behaves differently from one weighted towards memory or power, as the gap of almost 60 percentage points between the best and worst layers since June shows.
Positions that look diversified may not be. A client holding a chipmaker, a utility and a hyperscaler bond is exposed to the same macroeconomic catalysts. Risk and margin models that treat these as separate exposures can understate concentration.
As a result, client communication carries more weight. When demand is driven by a headline listing, explaining what a theme actually contains is critical.
Which layer a holding actually sits in
Treating AI as one trade hides most of the risk. A more useful approach starts with three questions, for the firms offering the exposure as much as for the people holding it. Which layer does a holding actually sit in? Is that layer’s growth showing up in margins and cash flow, or only in orders? Which physical or financing constraint would hurt it most?
Anthropic’s listing will attract attention for its size. The more lasting lesson is in its cost base. Every dollar of AI revenue at the top of the stack is a dollar of spending somewhere below it. Firms and investors who understand where that money lands, and what it depends on, are better placed than those who only sell or buy the label.
About Renée
Dr Renée Friedman is Global Head of Research at EXANTE, where she leads analysis on geopolitics, macroeconomics, and public market capital trends. A former Managing Editor at the Economist Intelligence Unit and fund manager in the City of London, she has advised institutions from the UK Parliament to the UNDP. She holds a PhD in Economics from London Business School.
Subscribe to her team’s research here: exante.eu/subscription
This article draws on findings from the EXANTE Beneath the Goldrush report available for download here.
Disclaimer and Risk Warning: The information contained herein is provided for informational purposes only and should not be regarded as an offer or solicitation of an offer to buy or sell any investments or related services that may be referenced here. Investing involves a high level of risk. Past performance is not a reliable indicator of future results.
Anthropic’s IPO prospectus, made public on 26 October 2026, commits hundreds of billions of dollars to computing infrastructure. That bill says more about investing in AI than the size of the listing itself. AI is not one trade. It is at least seven, and they behave very differently.
The largest commitment, $161.2 billion, is with Broadcom, a chip designer. Google, Amazon and SpaceX each account for more than $100 billion and Microsoft for about $45 billion, with some contracts running to 2036.
Much of the coverage so far has focussed on size. Reports suggest the offering could raise around $100 billion, which would make it the largest IPO on record. For brokers and platforms, that means another surge of clients asking for AI exposure.
The spending commitments by Anthropic matter more than the headline. One company’s operating costs become revenue for chipmakers, cloud providers, data-centre builders and power generators. This means that two clients who both say they hold AI can be holding very different risks.
AI is a supply chain, not a sector
In The Architecture of Disruption, published on 29 June 2026, my colleague 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. Fabrication sets the pace of compute. Networking decides whether compute can run at cluster scale. Power decides how quickly finished systems can be switched on. The application layer decides whether any of it earns a return.
Anthropic’s bill touches every one of them. The Broadcom agreement sits in compute. The cloud contracts pass down to memory, networking, physical infrastructure and power as the hyperscalers build capacity. Anthropic itself sits in applications, the layer that has to pay for the rest.
The layers are connected, but the market has not rewarded them together. Since 29 June, analysts have raised earnings forecasts most for fabrication, compute, physical infrastructure and memory. Power was the only layer revised lower. Share prices moved almost in reverse. Fabrication, with the strongest upgrade, fell 25.6% to 21 September. Applications, with one of the weakest, rose 33.3%.
The market is paying for revenue that has arrived, not revenue that has been ordered. Applications fell in the first half of the year as investors questioned whether the layer that has to pay for the buildout could do so. Microsoft’s 30 million paid Copilot seats went some way to answering that.
Profitability splits the stack in two. The layers closest to the silicon are growing and becoming more profitable, with compute and memory margins above 55%. Further down the physical chain, revenue lags orders. These layers have full order books, but they are being asked to invest before the cash arrives.
An AI position is also a credit position
The buildout is now financed in the bond market as much as from cash flow. In 2025, the five largest US hyperscalers issued $121 billion of investment-grade bonds. Their annual average between 2020 and 2024 was $28 billion. By October 2025, AI-related debt had become the largest segment of the US investment-grade market. However, AI hyperscalers are also using off-balance sheet arrangements such as special purpose vehicles and long-term leases and purchase agreements funded by private credit and insurers..The Bank for International Settlements describes this as “shadow borrowing” and has highhilighted that these methods are potentially creating new shock transmission channels.
As a result, a client who holds AI equities and is also exposed to rates or credit risk. Therefore the real level of AI risk exposure may be far greater than their equity allocation suggests.
The binding constraints are physical
In Beneath the Gold Rush, published on 15 October 2026, I argue 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 (IEA) expects data-centre electricity demand to more than double to around 945 terawatt-hours by 2030. Grid connection lead times in some established data-centre hubs stretch to ten years.
These constraints do not affect every layer equally. A shortage raises prices for whoever controls the bottleneck and raises costs for everyone who depends on it. The same physical constraint can therefore support one layer’s margins while compressing another’s.
Spend is not the same as return
The case for the whole stack rests on one assumption: that AI produces enough economic value to justify the capital. That is still a claim to be tested.
OECD modelling puts the potential productivity gain at 0.4 to 1.3 percentage points a year, provided workers reskill and institutions adapt. Until the gain appears in national accounts, AI spending on its own is a cost line.
This is why the application layer matters disproportionately. If its revenue keeps growing, it validates the capital beneath it. If it stalls, the most capital-intensive layers are likely to reprice first. Anthropic is a test case: it reported $16.3 billion of revenue in the first half of 2026, against commitments that run for a decade.
What the layers mean for brokers
For firms offering AI exposure, the layer view has three practical consequences.
Products labelled AI can carry very different risks. A basket weighted towards applications behaves differently from one weighted towards memory or power, as the gap of almost 60 percentage points between the best and worst layers since June shows.
Positions that look diversified may not be. A client holding a chipmaker, a utility and a hyperscaler bond is exposed to the same macroeconomic catalysts. Risk and margin models that treat these as separate exposures can understate concentration.
As a result, client communication carries more weight. When demand is driven by a headline listing, explaining what a theme actually contains is critical.
Which layer a holding actually sits in
Treating AI as one trade hides most of the risk. A more useful approach starts with three questions, for the firms offering the exposure as much as for the people holding it. Which layer does a holding actually sit in? Is that layer’s growth showing up in margins and cash flow, or only in orders? Which physical or financing constraint would hurt it most?
Anthropic’s listing will attract attention for its size. The more lasting lesson is in its cost base. Every dollar of AI revenue at the top of the stack is a dollar of spending somewhere below it. Firms and investors who understand where that money lands, and what it depends on, are better placed than those who only sell or buy the label.
About Renée
Dr Renée Friedman is Global Head of Research at EXANTE, where she leads analysis on geopolitics, macroeconomics, and public market capital trends. A former Managing Editor at the Economist Intelligence Unit and fund manager in the City of London, she has advised institutions from the UK Parliament to the UNDP. She holds a PhD in Economics from London Business School.
Subscribe to her team’s research here: exante.eu/subscription
This article draws on findings from the EXANTE Beneath the Goldrush report available for download here.
Disclaimer and Risk Warning: The information contained herein is provided for informational purposes only and should not be regarded as an offer or solicitation of an offer to buy or sell any investments or related services that may be referenced here. Investing involves a high level of risk. Past performance is not a reliable indicator of future results.
