An alternative take on the AI bubble theory: examining the sustainability of high valuations through capital allocation trends

Many investors focused on the AI industry may have noticed that, whether in business models or product positioning, numerous Agent products on the market are converging towards uniformity.

On April 9, 2026, OpenAI raised the price of ChatGPT Pro to $100 per month, precisely aligning with the same pricing tier as Anthropic’s Claude Max, which was launched a year earlier. The three-tier pricing structure of $20/$100/$200 is identical. Five days later, on April 14, Bloomberg reported that Anthropic was receiving investment offers valuing the company at $800 billion, nearly matching OpenAI's valuation.

Does something feel off? In theory, two competing companies in their early growth stages should focus on finding their own differentiated positioning or leverage unique advantages for growth. Yet these two companies increasingly resemble the war between Uber Technologies and Didi, though they haven't yet reached the stage of a price war.

The homogenization of products is only a surface-level phenomenon; behind it lies a deeper pricing mechanism that deserves clarity.。

Starting in 2014, Uber Technologies and Didi burned through tens of billions of dollars in subsidies, while their valuations skyrocketed from tens of billions to hundreds of billions during the same period; the food delivery wars and shared bike battles followed the same playbook. But in those scenarios, the valuation was based on the future narrative of 'fighting until only one player remains, then collecting monopoly rents,' with the premise set in the future. In the case of AI this time, the $800 billion valuation isn’t set for future monopoly; everyone knows that these two companies have virtually no chance of merging.

If not for monopolistic positioning, then what is it for?

SoftBank, NVIDIA, Amazon, Middle Eastern sovereign funds, Tiger Fund, and Fidelity are sitting on trillions of dollars that must be deployed into the AI theme due to performance benchmarks, strategic narratives, and investor commitments. However, there are very few independent targets in the AI space capable of absorbing multi-billion-dollar investments—OpenAI, Anthropic, xAI—and xAI has already been fully integrated into SpaceX by February 2026.

There's too much money chasing too few investable targets, which is why allocation pressure is mounting.。

As a result, valuations aren't derived from bottom-up fundamental analysis (under that algorithm, neither company would be worth $800 billion). Trillions of dollars need to be invested, and whatever amount needs to be absorbed dictates how high the valuation goes.Differentiation and who wins or loses aren't the main concerns; as long as the target can absorb the capital, the valuation holds.。

With this foundational fact in mind, the strategic logic at the product level starts to make sense. Basic model capabilities are rapidly becoming commoditized, with open-source large model ecosystems also closing in. Enterprise clients demand feature parity and retain the right to switch providers. High-value use cases driving willingness to pay are limited to coding and deep research, forcing both companies into the same circle, executing nearly identical product strategies.

If you're an investor betting on AI growth, you must ask yourself this: The ultimate payers in nodes such as AI infrastructure, power companies, private credit, and cloud vendors are really just a handful of players. As money flows down this chain, where it goes, where it settles, and which segment will break first are the real risks you need to assess.

01 Everything looks good.

There is a need to align on one key understanding: fundamental analysis, in the current wave of AI, especially within investment frameworks that focus on moving along the AI industrial chain,cannot adequately and objectively reflect a company's risks.。

Oracle is one of the listed companies that has made the most significant turnaround in the AI cycle. In its Q3 FY26 earnings report released on March 10, its cloud infrastructure business grew more than 80% year-over-year, with contract backlog reaching an all-time high. The stock price jumped about 10% on the night the earnings were announced. For a long-standing database company, this marks a complete transformation and represents the second growth curve that the market has waited several years to see.

However, in September 2025, Oracle signed a five-year, $300 billion cloud services contract with OpenAI. On an annual basis, this single contract amounts to approximately $60 billion per year, accounting for more than half of Oracle’s total contract backlog. Meanwhile, OpenAI itself still has negative operating cash flow during the same period and relies on the next round of financing to cover daily expenses.

In order to fulfill contracts with a group of major clients including OpenAI, Oracle announced in February 2026 that it would raise $45 to $50 billion through debt and equity financing.A supplier, in order to serve a batch of clients whose current cash flows are unstable, also needs to leverage itself first.A significant portion of Oracle's revenue story in the coming years will depend on this currently loss-making company fulfilling its obligations step by step.

In most industrial chains, the scale of contracts signed at each level typically does not far exceed the range that can be covered by the current cash flow. Auto manufacturers' orders are usually placed on a quarterly to yearly basis, corresponding to vehicle delivery cycles; multi-year contracts for enterprise software match foreseeable IT budgets of customers; advertising and subscription revenues of internet platforms are mostly settled on a current basis, with contract sizes almost synchronizing with current income. Fundamental analysis does not need to specifically track who the counterparties are at each level or where their credit foundations lie because the relationship between contract size and current operational rhythm naturally holds.

This misalignment is not unique to Oracle, nor is it Oracle’s own strategic choice, nor an accidental result of certain accounting treatments—it is a widespread phenomenon across the entire AI industrial chain.

What to do when money runs out? At the peak of the internet bubble in 1999, Lucent had annual revenue close to $38 billion but committed up to $8.1 billion in supplier financing to its customers, approximately a quarter of its revenue. Nortel and Cisco did the same: lending money to cash-strapped emerging telecom operators to buy their equipment. According to McKinsey's statistics from that time, nine major telecom equipment suppliers collectively lent $25.6 billion to their customers.

In September 2025, NVIDIA announced a $100 billion investment commitment to OpenAI, most of which would be used to purchase NVIDIA's own AI chips. Four months later, Jensen Huang publicly clarified that "this was never a commitment" and would be assessed round by round based on deployment progress. This is strikingly similar to what Lucent did back in the day.Back then, supplier financing also advanced quarterly as an intention, and the market gave a full valuation until customer defaults revealed major problems.。

The bubble burst in 2001 when customers defaulted en masse, causing Lucent’s loan portfolio bad debt ratio to soar from single digits to over 40%. Dozens of telecom service providers across the industry collectively defaulted and went bankrupt between 2001 and 2002. Lucent’s revenue collapsed from a peak of nearly $38 billion to $8 billion in 2006, eventually selling to Alcatel for $3 per share. Nortel’s stock price plummeted from $86.75 to $0.18 before heading into bankruptcy.

The script currently unfolding in the AI supply chain shows some similarities, with multiple key nodes locking in future growth through contract sizes far exceeding their current cash flows. For example, CoreWeave, a computing power contractor, had projected annual revenue of $5.1 billion in 2025, but its contract backlog surged to approximately $88 billion by April 2026, with Meta and OpenAI together accounting for about two-thirds of this total.

At the very top of the chain are two leading AI ventures.OpenAI alone has accumulated contractual commitments reaching $1.15 trillion, involving seven suppliers: Broadcom, Oracle, Microsoft, NVIDIA, AMD, Amazon, and CoreWeave, while its annualized revenue at present stands at only $24 billion.。

Fundamental analysis cannot clearly discern the gap between contract size and current cash flow because the contract size appears on financial statements merely as 'contract backlog,' which is not an audited asset nor included in the current income statement. In case of issues, the ultimate recourse is limited to the upper layer fulfilling payments as per the contract. Ultimately, aside from large corporations like Google and Microsoft capable of self-funding, others must rely on the primary market to alleviate funding pressures.

As such, each node in the supply chain can present seemingly legitimate financial figures individually, but they all hinge on one shared implicit assumption:Funds will continue to flow into the AI theme at the current pace, and downstream customers will continue making payments as per contracts.This is where fundamental analysis falls short in the AI infrastructure chain, as it fails to see the hidden dependencies between layers linked by contracts and overlooks how a single node's 'good business' actually rests on others’ precarious contractual obligations.

02 Money is flowing

In the entire AI industry chain, CoreWeave is a very interesting company.

According to the company's 2025 annual report, total revenue for the year was $5.1 billion, of which 67% came from Microsoft. Based on this set of figures alone, CoreWeave presents itself as a B2B business reliant on large, stable clients — one being the world’s second-largest tech giant by market cap, with impeccable credit backing.

However, within the $88 billion contract backlog, historical contracts with Microsoft only account for a small portion; the majority consists of several major contracts signed after 2025. After Meta signed a $14.2 billion contract in September 2025, it added another $21 billion in April 2026, doubling its total commitment to CoreWeave to $35.2 billion within half a year, accounting for about 40% of the contract backlog. OpenAI has cumulatively committed $22.4 billion, representing approximately 25%.

This means that nearly two-thirds of the company's growth in the coming years depends on the continued payments from these two tech giants.Meta appears more stable than OpenAI, at least having independent advertising cash flow as a safety net. However, Meta’s approach to building AI data centers takes a different route: it only puts up 20% equity, while the bulk of the funding comes from a joint venture led by Blue Owl, keeping hundreds of billions in debt off Meta’s own balance sheet. One client is the loss-making OpenAI, and the other is Meta, which relies on external joint ventures for expansion. Neither client fully controls their ability to make these payments.

Among industries heavily favored by US stock investors, there is another category of companies that are relatively distant from AI itself but still benefit significantly: power equipment manufacturers and power plant construction contractors. GE Vernova, specializing in gas turbines and grid equipment, was spun off from the GE Group and listed in April 2024. Its stock price rose from around $140 at IPO to nearly $1,000 by April 2026, appreciating almost sevenfold in two years. Argan, which handles EPC contracting for gas-fired power plants, saw its share price rise from $130 to over $600 in the past year, close to a fivefold increase.

HoweverTheir gains over the past year don't align with the fundamentals of traditional power equipment and engineering contracting businesses. The core issue lies in a severe mismatch between production capacity and contract scale.GE Vernova’s gas turbine contract backlog surged from tens of gigawatts in 2024 to 83 GW by the end of 2025, with a target of 100 GW by the end of 2026. However, the company’s annual production capacity will not reach 20 GW until mid-2026, meaning it will take at least four to five years to work through the current order backlog. Argan’s project backlog grew from $1.4 billion to $2.9 billion, triple its annual revenue.

These contract backlogs essentially represent expected revenues over multiple future years, not current assets, and do not appear on income statements. If tech giants slow down their AI data center expansions, delay orders, or renegotiate terms, the execution pace of these backlogs could decelerate instantly. Even if contracts include penalty clauses, actual enforcement often involves significantly reduced renegotiation rather than full compensation.

Looking at the first two examples together, the nodes on the chain generally fall into three categories.

The first category includes tech giants like Microsoft, Google, Meta, Amazon, and Oracle. Their core businesses were already independently established before the advent of AI, and their cash flows from advertising, search, e-commerce, databases, and enterprise software do not rely on AI to survive.For this type of node, going long requires a test of endurance: strip out the AI premium from the current valuation and assess what the company is truly worth—this is the real margin of safety.。

OpenAI and Anthropic represent the second category. These companies rely on capital allocation pressures for funding, with their credit foundation hanging on whether the next round of financing will succeed and whether their valuation trajectory will continue to rise. Direct investment by secondary market participants is still unrealistic at this stage; instead, they serve more as 'leading indicators' to monitor, such as the slope of the next private equity valuation, which can be an early warning sign for potential issues across the entire chain.

The third category consists of mid- to downstream nodes that rely on contracts for survival. This group is the most complex, including compute contractors, data centers, power companies, grid equipment suppliers, and private credit funds. The weighting should depend on the composition of counterparties in these contracts: those primarily dealing with large firms (e.g., Constellation’s electricity contract with Microsoft) can be considered defensive holdings.

However, if the proportion of speculative contracts with entities like OpenAI or Anthropic is high among counterparties, or if major firms themselves are using SPV structures to maintain infrastructure expansion (e.g., nearly two-thirds of CoreWeave’s contract reserves coming from Meta and OpenAI), vigilance needs to be heightened.

03Conclusion

Another signal that those following this chain should watch closely is liquidity at the far end of the chain.

AI represents the largest asset allocation trend at present, with sovereign wealth funds and major capital continuing to deploy over the long term. However, money in the primary market does not flow in at a constant rate—it tends to focus on several key events:Whether the next round of private equity valuations will continue to rise, whether employees are actively selling shares in internal transfers, and whether large private credit funds have started facing redemptions.Blue Owl’s three consecutive anomalies over the past six months already indicate that even the deepest pools of capital can start to feel strained at certain points.

Blue Owl is one of the largest private credit management companies in the US, with assets under management exceeding $300 billion by the end of 2025. Its funds have taken on a large number of structured bonds related to AI infrastructure. From February to April 2026, three consecutive events occurred at Blue Owl: In mid-February, OBDC II suspended its quarterly share repurchase; on February 18, it sold $1.4 billion in direct loan assets to boost liquidity; and on April 2, quarterly redemption requests for a non-traded private credit fund reached 4.999%, just below the mandatory redemption threshold of 5%.

The event on April 2 could be attributed to the impact of geopolitical tensions involving Iran, which placed certain pressure on liquidity, but the two incidents in February already showed some troubling signs. Betting on AI itself is not an issue—it’s the biggest trend today—but the pressure on funding at the top of the chain remains, and the overall growth narrative of the chain has yet to fully play out.

But with every bullish move, one must ask:On this part that I’m standing, whose ledger does the final payment flow end up on?Answering this question clearly puts both the confidence for going long and the position of risk firmly in hand. $Microsoft (MSFT.US)$$Oracle (ORCL.US)$$NVIDIA (NVDA.US)$$Amazon (AMZN.US)$$CoreWeave (CRWV.US)$

Disclaimer: This article is intended for learning and communication purposes only and does not constitute investment advice.

154K Views