The artificial intelligence ecosystem is a multi-trillion-dollar gamble not only for the companies that are building it, but also for nations seeking new competitive advantages and network power in a world of geo-economic fragmentation. The U.S. is in a good position to win the AI race with its deep financial markets and strong suite of innovating companies. But not all firms will emerge as the strongest.
The AI build-out is developing along two distinct structures with four necessary components or layers: chips; cloud and storage infrastructures; model design, development and training; and model deployment for use by consumers and firms. Competition is occurring in all four of these areas, but there are also rival business models in how these parts are put together.
Tower vs. Lego Stack
We see two different vertical structures in the U.S.: the AI Tower and the AI Lego Stack, with some hybrids around each. An example of an AI Tower is Alphabet Inc., the parent of Google, which has all four components inside the company. Examples of AI Lego Stacks include NVIDIA Corp. and OpenAI Group. Through financial linkages and partnerships, they have assembled their own vertical by snapping together several companies. The firms within these Lego-like stacks, and in some cases substacks, are different in terms of market capitalization, revenue generation, cash flow, and profitability.
AI capital expenditure is estimated to require as much as $5.8 trillion between 2025 and 2030. Investors have recently grown concerned about the ability of these AI stacks to generate enough revenues and profits to justify the scale of investments. The focus is on two vulnerabilities. First, the growing interconnectedness within the AI ecosystem, which results in higher correlation and greater counterparty risk. Think for example at the circularity of recent deals, with OpenAI having reportedly committed to a $1 trillion deal-spree with companies such as NVIDIA, CoreWeave Inc., Advanced Micro Devices Inc., Broadcom Inc. and Oracle Corp.
Second, there is increased use of leverage. While the initial phase of AI infrastructure buildout has relied on cash flow generated by hyperscalers such as Amazon.com Inc., Google, Microsoft Corp. and Meta Platforms Inc., debt is increasingly being used to finance AI data centers.
In AI Towers, cash flow is available, there is less leverage, revenues are generated across different business lines – Google’s search engine and cloud businesses are examples — and there is greater ability to embed AI into existing business lines with existing paying clients.
By contrast, companies assembled into Lego-like stacks rely more on debt (in some case of lower quality), interdependence is greater, as are correlation and counterparty risks. More than 60% of CoreWeave’s revenue, for example, comes from a single partner: Microsoft.
AI capital expenditure is estimated to require as much as $5.8 trillion between 2025 and 2030. Investors have recently grown concerned about the ability of these AI stacks to generate enough revenues and profits to justify the scale of investments. The focus is on two vulnerabilities. First, the growing interconnectedness within the AI ecosystem, which results in higher correlation and greater counterparty risk. Think for example at the circularity of recent deals, with OpenAI having reportedly committed to a $1 trillion deal-spree with companies such as NVIDIA, CoreWeave Inc., Advanced Micro Devices Inc., Broadcom Inc. and Oracle Corp.
Second, there is increased use of leverage. While the initial phase of AI infrastructure buildout has relied on cash flow generated by hyperscalers such as Amazon.com Inc., Google, Microsoft Corp. and Meta Platforms Inc., debt is increasingly being used to finance AI data centers.
In AI Towers, cash flow is available, there is less leverage, revenues are generated across different business lines – Google’s search engine and cloud businesses are examples — and there is greater ability to embed AI into existing business lines with existing paying clients.
By contrast, companies assembled into Lego-like stacks rely more on debt (in some case of lower quality), interdependence is greater, as are correlation and counterparty risks. More than 60% of CoreWeave’s revenue, for example, comes from a single partner: Microsoft.
Speedbumps and Innovation Risks
There are a number of risks that could disrupt the AI ecosystem. Let’s look at a few.
- Expectations of energy availability needed to support AI data center capacity currently embedded in debt structures may need to be significantly scaled down, leading to a repricing of credit risk.
- Broad adoption of new technologies such as small language models — a breakthrough in computing on local devices instead of the cloud — could undercut revenue assumptions in all four of the existing AI buckets.
- Chip depreciation could be more rapid than the 4 to 6 years as is often estimated if innovation cycles speed up.
- Technological advancements could lead to cheaper, less energy-intensive, and more customized chips that could reduce demand for graphics processing units or GPUs, impacting assumptions about collateral value embedded in current pricing of debt structures.
- Just how competition results in winners has implications for profitability and debt sustainability throughout all companies in the interlocking AI stack.
Exuberance Fades
Investors are increasingly concerned about the financial risks embedded in the AI buildout race. For example, the share prices of cloud computing companies CoreWeave and Oracle have fallen off their highs over the past few months, while the cost of insuring their debt against default has climbed.
Which AI stack—the AI Tower or AI Lego—is better positioned to navigate growing financial risks and the cone of unpredictable outcomes of rapid AI technological innovation? Who will be the winner?
There is no obvious answer. Right now, it looks like a contest between agility and financial strength. The AI tower can be patient, has significant financial resources, is less interconnected, so it is less susceptible to fragilities in the AI ecosystem. But it is also less nimble, and more exposed to sudden accelerations in technological innovation.
By contrast, the AI Lego stack is assembled from independent companies each with their own ambition, governance, and decision-makers and is therefore more agile – a crucial feature in a world of rapid technological changes. But it has generally weaker balance sheets across the individual companies of the stack so it is more reliant on riskier debt, more interdependent, more correlated, and more exposed to a domino effect if one member of the stack weakens.
Which AI stack—the AI Tower or AI Lego—is better positioned to navigate growing financial risks and the cone of unpredictable outcomes of rapid AI technological innovation? Who will be the winner?
There is no obvious answer. Right now, it looks like a contest between agility and financial strength. The AI tower can be patient, has significant financial resources, is less interconnected, so it is less susceptible to fragilities in the AI ecosystem. But it is also less nimble, and more exposed to sudden accelerations in technological innovation.
By contrast, the AI Lego stack is assembled from independent companies each with their own ambition, governance, and decision-makers and is therefore more agile – a crucial feature in a world of rapid technological changes. But it has generally weaker balance sheets across the individual companies of the stack so it is more reliant on riskier debt, more interdependent, more correlated, and more exposed to a domino effect if one member of the stack weakens.
