AI Investments

Strategic Flywheel or Fragile Stack?

Next semester at Harvard Business School, we will teach a new case on the AI infrastructure boom that I wrote with my colleague Aliya Korganbekova titled “Circular AI Deals: Strategic Flywheel or Fragile Stack?”.

The case looks at the web of investments and commercial commitments linking Microsoft, OpenAI, Amazon, Anthropic, NVIDIA, CoreWeave, AMD, Oracle, and others. On the surface, it is a case about artificial intelligence. But the deeper purpose is different. It is about helping students connect accounting, financial statements, strategy, capital allocation, and valuation in a consequential setting where the right answer is not obvious.

That is what makes the case so interesting. The AI boom is not just a story about models, chips, or data centers. It is a story about how capital moves through an ecosystem, how reported numbers shape investor beliefs, how private valuations affect public companies, and how leaders make large commitments before the economics are fully proven.

One of the most revealing facts in the case is that Microsoft and Amazon made investments in frontier AI labs that looked strategically similar but showed up very differently in their financial statements.

Microsoft invested heavily in OpenAI and reported losses tied to that investment. Amazon invested heavily in Anthropic and reported gains tied to that investment.

Both companies were backing frontier AI labs, strengthening their cloud platforms, and positioning themselves for what could become the next great computing platform. Yet the reported outcomes looked very different.

That contrast opens the door to a much bigger question.

Are these circular AI deals creating a strategic flywheel, where capital builds infrastructure, infrastructure improves models, models attract customers, and customers generate cash flow? Or are they creating a fragile stack, where capital, revenue, valuation, and future commitments reinforce one another before the underlying economics are fully visible?

The same facts, two interpretations

The optimistic interpretation is compelling. AI labs need enormous amounts of compute. Hyperscalers have the balance sheets, data centers, and distribution channels to provide it. Chip companies supply the hardware. Neoclouds add capacity when the large platforms cannot move fast enough. Enterprises and consumers adopt better AI products, usage grows, and revenues fund the next generation of infrastructure.

In that version of the story, capital is not being wasted. It is being deployed ahead of a massive platform shift. Microsoft’s relationship with OpenAI, Amazon’s relationship with Anthropic, NVIDIA’s relationship with CoreWeave, and AMD’s partnership with OpenAI all become pieces of a larger industrial buildout. The ecosystem is coordinating investment at a scale that no single company could handle alone.

The skeptical interpretation starts from the same facts but follows the money differently. A hyperscaler invests in an AI lab. The AI lab uses that capital to buy cloud compute. Cloud revenue rises. The hyperscaler builds more data centers. Chip suppliers sell more GPUs. Some of those suppliers invest in the companies buying or renting their chips. Private valuations rise. Those valuations support more fundraising. The new capital funds more compute.

The question is not whether the money is moving. It clearly is. The question is whether each turn of the loop is creating more economic value, or simply validating the previous turn of the loop.

That is the difference between a flywheel and a fragile stack.

Why accounting matters here

Accounting matters in this story not because it gives us all the answers, but because it shapes what investors think they are seeing. Microsoft’s OpenAI exposure showed up as losses. Amazon’s Anthropic exposure showed up as gains. That does not automatically mean Amazon made the better investment or Microsoft made the worse one. It means the structure of the investments, the accounting treatment, and the valuation dynamics produced very different reported outcomes.

The same issue appears when we ask a simple question: who is growing faster, OpenAI or Anthropic? At first, that sounds like a straightforward comparison. Look at revenues. Compare growth rates. Compare annualized run rates. Compare customers, subscriptions, API demand, and enterprise adoption.

But the question quickly gets complicated because revenue recognition differs. If one company recognizes partner related revenue on a gross basis and another recognizes economically similar activity on a net basis, the top line numbers are not directly comparable. The reported growth rate may say as much about the transaction structure as it does about customer demand.

That is not a technical footnote. Revenue numbers flow into valuation. Valuation affects the ability to raise capital. Capital affects the ability to buy compute, hire talent, improve models, and sign the next infrastructure commitment. Those commitments then affect the cloud providers, chip suppliers, and investors tied to the ecosystem.

Accounting is not the whole story. But in the AI boom, it is one of the gears inside the machine.

The real test is the outside dollar

The most important question in any circular system is simple: where is the outside dollar?

The outside dollar comes from a customer who is not part of the financing loop. It is the enterprise paying for AI because it lowers cost, improves software development, strengthens customer service, accelerates research, improves decision making, or creates new revenue. It is the consumer paying because the product is useful enough to become part of daily life.

That outside dollar is what turns the loop into a flywheel. If customers outside the ecosystem pay for AI products because those products create measurable value, then the capital flowing among hyperscalers, AI labs, chip suppliers, and infrastructure providers can be justified. The partnerships help coordinate investment ahead of demand, and the system strengthens as adoption expands.

But if too much of the revenue depends on companies inside the ecosystem buying from one another, then the signal becomes weaker. Revenue may grow before economic value is proven. Valuations may rise before cash flow appears. Capacity may expand before utilization is certain. The system can look powerful while capital is abundant and belief is high.

That is why the outside dollar is the first test. Who is the final customer? What problem are they solving? What economic value are they receiving? Are they renewing because the product is embedded in workflow, or experimenting because AI is strategically fashionable?

Does growth convert?

AI has a conversion problem that many earlier software businesses did not face in the same way. In traditional software, once the product is built, incremental usage can be highly profitable. In AI, each query consumes compute. Each model improvement may require more chips, more energy, more engineering, more data centers, and more infrastructure.

That does not mean AI cannot become highly profitable. It means growth by itself is not enough. The key question is whether usage converts into attractive economics. If customers receive large productivity benefits and are willing to pay accordingly, the economics can work. If inference costs fall, utilization improves, and model performance rises faster than cost, margins can expand.

But if usage grows faster than monetization, or if competition pushes prices down while compute costs remain high, revenue growth may disappoint investors. The best AI companies will not simply be the ones with the most users or the largest models. They will be the ones that create the most customer value per unit of compute consumed.

That ratio may become one of the defining metrics of the AI era.

Who bears the downside?

Every boom allocates risk somewhere. The AI boom is no different.

The hyperscaler may build data centers before demand is fully visible. The AI lab may commit to future cloud purchases. The chip supplier may support customers that buy its chips. The neocloud may borrow heavily to purchase GPUs. The public shareholder may absorb dilution. The private investor may depend on higher marks in the next funding round.

This is why deal structure matters. Equity investments, convertible notes, warrants, cloud credits, revenue sharing agreements, purchase commitments, exclusivity provisions, and backstop guarantees are not just technical details. They determine who benefits if the upside arrives and who absorbs the pain if it does not.

The headline announcement rarely answers that question. The footnotes often do. A deal that looks like revenue may partly function as financing. A partnership that looks like customer demand may also provide channel support. A valuation gain that improves earnings may not improve operating cash flow. A commitment that signals strategic alignment may also create concentration risk.

The teaching point

This is why the case is so useful for students. It does not ask them to decide whether AI is real. It asks them to develop judgment in a setting where AI is clearly important, the investments are enormous, the accounting is complex, and the strategic stakes are high.

That is the kind of situation leaders actually face. They rarely get clean problems with complete information. They get partial signals, competing interpretations, market pressure, financial constraints, and decisions that must be made before uncertainty disappears.

The goal is to teach students to ask better questions. What do the financial statements reveal? What do they obscure? Which commitments create strategic advantage? Which commitments create fragility? When does capital allocation build the future, and when does it borrow too much from it?

A final thought

The AI boom is not simply about who has the best model, the largest data center, or the biggest chip order. It is about whether an ecosystem can convert extraordinary technological progress into durable economics.

The Microsoft and Amazon contrast is a useful starting point because it shows how complicated the story has become. One company reported losses tied to its AI lab investment. Another reported gains. That difference does not answer the strategic question, but it tells us where to look.

The real issue is whether AI companies can convert compute into customer value, customer value into revenue, revenue into earnings,.

In every platform shift, the future is built before it is fully proven. That is why courage matters. But when growth starts moving in circles, judgment matters even more.