“What happens when the hype cycle meets the hard limits of engineering?” This question gets a bright-colored response in Oracle’s recent experience. Only three months ago, it was the phoenix-like rise of the re-born old-tech giant, riding the latest wave of investment mania for artificial intelligence infrastructure. On one day in September, its shares rose by an incredible 43%, beating JPMorgan’s market value at one point, pushing Larry Ellison above Elon Musk as the richest man in the world. This is the same stock that is down by an incredible 46% from that level, troubled for the very reasons that took it there.
This sharp reversal has been based on a shift in market sentiment, where investors who were previously enamored with any news related to AI are suddenly looking under the hood and expecting results. Oracle’s bold increase in data center AI capex spend, upped to $12 billion for Q2 compared with analyst estimates of only $8 billion, has been matched by an increase in full-year guidance to $50 billion compared with previous guidance of $35 billion. Of course, such large outlays on capex have also squeezed the year’s free cash flow, which came in sharply negative—a staggering $10 billion—a problem further exacerbated by the fact that a major investor withdrew from a $10 billion data center deal, further casting doubt on the company’s ability to finance such obligations without drastically increasing leverage.
The debts provide the background for such worries. The long-term debts at Oracle have escalated from an average of US$60 billion in 2020 to over US$100 billion, and an additional US$38 billion is expected to be raised through debts that are meant to develop infrastructure. The debt/equity ratio at Oracle currently stands at nearly 500% compared to Amazon’s 23%, Microsoft’s 7%, Meta’s 16%, and Google’s 3%. The credit default swaps at Oracle are also at their highest levels since 2009.
If anything, the scope of Oracle’s AI infrastructure development effort is impressive from an engineering standpoint. Hyperscale AI data centers involve massive GPU clusters, dense racks, and innovative cooling technologies like direct-to-chip liquid cooling or immersion cooling to tackle increasing temperatures. Its multicloud offering, combining with AWS, Azure, and Google Cloud, provides portability but involves complex management across different environments. There are supply chain issues related to GPUs, high-bandwidth networking infrastructure, and slower lead times for connects to the power grid.
The economic risks, on the other hand, are also quite daunting. McKinsey research indicates that there will be a projected investment of $5.2 trillion in AI-related data centers by the year 2030, with hyperscalers and businesses competing for available capacity. Over-investment could result in stranded infrastructure in the event of lower adoption rates or reduced demand for computation. Under-investment could result in losing market edge. Oracle has a backlog of non-cancelable contracts of $523 billion from clients Meta, NVIDIA, OpenAI, and other companies, which can be considered potential revenue.
This injection of market discipline got reflected because there are some broader worries expressed by industry leaders, says Sundar Pichai, CEO of Google, warning of “some irrationality” in cycles of investment in AI, adding a no company can escape a bubble when it bursts. Sam Altman, Open AI, expressed auyuva mapping of “insane valuations” given to very small AI startups, and Dario Amodei, Anthropic, warns: “if you just make a timing error… bad things could happen.” These warnings actually contrast sharply with market funding, as there may appear to be a disconnect between actually developing AI infrastructure and funding patterns given to AI infrastructure at this time.
Technical risk is not limited to construction timelines. The workloads for AI keep increasing at breakneck speeds. The data centers currently being constructed for large language models may become obsolete in the event of a transition toward models that have more compute-efficient architectures or edge inference models that are nearer to metropolitan areas. The customer concentration risk is significant. Oracle Communications has a $300 billion deal with OpenAI that has a five-year deadline. But customer concentration risk is high for any firm with this type of deal. If any large customer stumbles or changes direction, this could lead to reduced usage levels and a chance that ROI may not be as estimated.
Investor sentiment has responded by increasing the short interest levels for any institution that has a stake in tech stocks with AI connections, meaning firms whose prices may be decoupled from present market performance. The recent market downturn has led to a reduction by hundreds of billions for firms exposed to AI. It is clear that “easy gains” for any AI firm. Since Oracle has a huge deal with OpenAI, its engineering puzzle is now indistinguishable from its financial puzzle: that is, whether this firm’s infrastructure is capable of delivering ROI that is consistent with the unprecedented level of investment that has been made.
