The guidance from Oracle is more explosive than ever, but history has thrown a bucket of cold water on it

After ChatGPT brought generative AI into the public eye at the end of 2022, changes in the investment sector accelerated: corporate investments in AI hardware and data centers are nearing some of the largest investment waves in U.S. history. The market subsequently presented a series of impressive revenue curves, but the questions became sharper—how likely are these predictions to be realized, and is it worth the capital and time costs?

According to news from the Chasing Wind Trading Desk, Michael J. Mauboussin from Counterpoint Global, a subsidiary of Morgan Stanley Investment Management, provided a straightforward methodology in a report on the 10th: to evaluate such forward-looking judgments, one should “start with an initial belief and update that belief as new results appear,” which is the “Bayesian formula”: “New conclusion = Initial judgment (prior probability) × Adjustment factor from new evidence (likelihood ratio).”

Following this framework, the report placed two of the most closely watched predictions back into historical distributions: OpenAI's revenue from $3.7 billion in 2024 to $145 billion in 2029 (corresponding to a 5-year compound growth rate of 108%), and Oracle's cloud business from $10 billion in fiscal year 2025 to $166 billion in fiscal year 2030 (5-year compound growth rate of 75%). The conclusion is quite blunt: in the sample of U.S. publicly traded companies from 1950 to 2024, no company has achieved such a growth rate starting from this scale.

What’s more troublesome is that AI infrastructure is not as simple as “buying a few more servers.” Building data centers is essentially a large project, and large projects have their own baseline failure rates: budget overruns, delays, and returns falling short of expectations are almost the norm. The report also conveniently explained this round of intensive transactions within competitive strategies: they may not only be to meet demand but also to signal to competitors and attempt to deter potential entrants—but this kind of first-mover bet carries high risks.

First, place OpenAI's prediction into historical distribution: 108% compound growth rate is "blank" in the sample

The reference frame used in the report is very specific: it selects a group of companies from U.S. publicly traded companies between 1950 and 2024, with initial revenues between $2 billion and $5 billion (in 2024 dollar terms), with nearly 18,900 “company-period” observations. The average 5-year revenue compound growth rate for this group is only 7.0%, with a standard deviation of 10.6%.

OpenAI's prediction implies: from $3.7 billion in 2024 to $145 billion in 2029, a 5-year compound growth rate of 108%. The report's statement is firm—in the past three-quarters of a century, no publicly traded company has achieved such speed. Even using a normal approximation to describe this, it is close to a result of 9.5 standard deviations, with a very low probability; moreover, the historical growth rate distribution itself does not follow a normal distribution, with heavier tails, but it still does not change the conclusion of “almost invisible.”

A detail worth pondering: because "never happened" in the sample, the baseline probability becomes 0, and the Bayesian formula cannot be directly applied. The report adopts common heuristic treatments (such as 3/N, Laplace smoothing), resulting in an initial probability still below one-thousandth.

The evidence does indeed "raise the probability," but the report does not optimistically tell you how far it goes

The report also acknowledges that the baseline probability is not a hard rule; the world changes. It provides two pieces of evidence that can push OpenAI's success probability from "close to 0" upwards:

• Diffusion Speed: ChatGPT reached 100 million users in 2 months; in comparison, TikTok took 9 months, Instagram 28 months, Facebook 4.5 years; the internet took 7 years to reach 100 million users, mobile phones 16 years, and telephones 75 years. Even considering population changes, this speed is still historically rare. The report also reminds that users do not equal revenue; many do not pay.
• Short-term Revenue Growth Rate: OpenAI expects revenue of about $13 billion in 2025, with a year-on-year growth rate of about 250%. This is far higher than the average compound growth rate over five years.

However, the report immediately pins down the "optimistic boundary": the larger the company, the smaller the fluctuations in growth rate, making it increasingly difficult to maintain high growth rates. Furthermore, OpenAI has also provided a revenue forecast of $200 billion for 2030—pushing the window back, the five-year compound growth rate from 2025 to 2030 is still 72.7%.

Using a reference class of initial revenues of $10 billion to $15 billion (about 3,700 observations) for comparison, the conclusion remains: no one has achieved this before; even if the initial revenue threshold is relaxed to at least $6.5 billion and the sample expanded to over 16,400 observations, no one has achieved this either.

Growth does not equal value: cash flow gaps and equity incentives will pull the "high growth story" back to financing reality

The report here offers a more realistic reminder: growth itself does not create value. It also places constraints on the definition of "Total Addressable Market (TAM)"—not "how much can be sold," but "how much revenue can be generated if 100% market share is achieved under the premise of creating shareholder value"; the core threshold is whether the return on investment exceeds the cost of capital.

In the case of OpenAI, the report directly lays out the constraints:

• The free cash flow for 2025 is reportedly -$9 billion, and it is expected to be -$17 billion in 2026. In this situation, maintaining "high-speed expansion + heavy investment" will almost inevitably require continuous financing from external investors.
• A significant portion of employee compensation is equity incentives (SBC): it is estimated that SBC will exceed revenue by 45% in 2025, translating to about $1.5 million per employee per year, which is 7 times the intensity of SBC issuance before IPOs of large tech companies This information does not directly negate revenue forecasts, but it brings to the forefront a commonly overlooked issue: even if revenue growth materializes, capital structure, financing conditions, and dilution costs may determine "what shareholders actually receive."

Oracle Cloud's $166 billion target: Signing contracts is an advantage, but delivery and financing are hard constraints

Oracle's narrative comes from another type of evidence: the company announced multiple multi-billion dollar cloud infrastructure contracts in 2025, significantly boosting "Remaining Performance Obligations" (future revenue corresponding to signed customer agreements). Management predicts that cloud business revenue will grow from $10 billion in fiscal year 2025 to $166 billion in fiscal year 2030, corresponding to a 75% compound annual growth rate over five years. This cloud business is expected to account for about 17% of Oracle's total revenue of $57.4 billion in fiscal year 2025.

The report first applies baseline probabilities: over the past 75 years, no company with initial revenue exceeding $10 billion has achieved such growth over five years; even lowering the initial revenue threshold to over $5.6 billion still yields no results.

It also provides a reference closer to the scale of Oracle Cloud: initial revenue between $8 billion and $12 billion, with approximately 4,400 observations, showing an average five-year compound growth rate of 5.7% and a standard deviation of 9.6%. The report also reminds that this comparison is made between "business units of the company" and "the company as a whole," not entirely on the same basis.

Oracle's difference lies in the fact that the RPO scale can indeed allow for adjustments to baseline probabilities, but the report emphasizes that adjustments cannot only consider orders; they must also weigh—financing needs for growth, counterparty risks, and potential delays in infrastructure implementation.

AI data centers are typical "large projects," and the baseline success rate for large projects does not favor you

The "main battlefield" for AI investment lies in hardware and data centers. The report mentions that OpenAI and Oracle are both partners in the "Stargate Project," which is expected to invest up to $500 billion in AI infrastructure by 2029.

The key point is: AI data centers are different from traditional data centers; the hardware is more expensive, electricity demand is significantly higher, and they rely more on cooling systems. The bottlenecks are very real—power access and dedicated hardware supply.

The report references a database of 16,000 large projects collected by Bent Flyvbjerg, and the results are almost "discouraging":

• 47.9% of projects are completed within budget;
• Only 8.5% are completed on time and within budget;
• Only 0.5% are completed on time, within budget, and achieve expected returns.

The insights it provides are straightforward: do not take "implementation as planned" as the default option. It is necessary to focus on key bottlenecks such as electricity, chips, and equipment; at the same time, modular design is relatively easier to succeed, but in an environment where AI demand is rapidly growing and competitors are vying for the lead, "think slowly and act quickly" is not easy to execute.

Intensive Transactions and Expansion Declarations May Be a "Preemptive Deterrence" Competitive Experiment

The report states that OpenAI announced about 15 transactions related to infrastructure construction around 2025. Meanwhile, large-scale cloud providers such as Alphabet, Amazon, and Microsoft have raised their capital expenditure expectations, and companies like Anthropic and CoreWeave have also made significant investment commitments.

The author places this wave of enthusiasm in historical context: The telecom investment boom from the late 1990s to the early 2000s ultimately left behind overcapacity and bankruptcy cases. Today, there is certainly a side that says "demand is far from reaching its ceiling" — the data cited in the report indicates that by the second half of 2025, the global AI diffusion rate (the proportion of people who have used GenAI products) is only 16%.

What is truly interesting is its speculation on motivation: This wave of actions may partly stem from a strategic signal — using large-scale capacity commitments to lock in the market, deter competitors, and potential entrants. The report cites Porter's concept of "preemptive strategy" while also clearly stating the risks: this involves committing massive resources before the market outcome is clear; if it fails to deter competitors, it may trigger a more intense war of attrition. A more realistic differentiation is in financing capabilities: startup AI companies need continuous external funding, while giants like Amazon, Google, and Meta have more abundant cash flow and different tolerances. By 2025, capital will still be available, but the report clearly states: this situation will change.

What This Report Really Wants You to Do: Break the Story into Probabilities and Adjust with Data

The report repeatedly emphasizes not "being bearish on AI," but rather changing the judgment process into an updatable probability issue: First, set thresholds for enthusiasm using baseline probabilities, and then gradually adjust with diffusion speed, real income, engineering progress, and financing conditions. It also emphasizes that it does not provide investment advice — but it offers a starting point that is harder to deceive oneself: when predictions fall into areas that have never appeared in historical samples, optimism itself requires evidence, and continuous evidence at that