Despite its record-high orders totaling 664 billion yuan, Oracle reported a cash flow deficit of 5.4 billion yuan: what is it betting on?

The freshly released FY27 Q1 financial results of Oracle have delivered a mixed performance report card:

On one hand, AI demand has overwhelmed this veteran tech giant's server business. Its single-quarter revenue reached USD 19.3 billion (up 30% year on year), marking a rare "positive sequential Q1 growth" in its history. Its remaining performance obligations (RPO) have even soared to an astonishing USD 664 billion, equivalent to more than 7 times the company's full-year revenue guidance.

But on the other hand, there is a "landmine" hidden in the financial report that has made Wall Street's heart race:

Its single-quarter capital expenditure (CapEx) hit as high as USD 28.5 billion, leaving a huge deficit of USD 5.4 billion in free cash flow (FCF). Meanwhile, the company swiftly completed a USD 20 billion stock financing.

With orders piling up so fast that its cash printing machines are overheating, yet cash is draining so badly that it has to replenish capital via equity financing — what kind of game is Oracle playing exactly?

01. Tear off the mask: This is by no means a software company, but a "heavy-asset arms dealer"

Looking at the growth rate of its business segments, you will find Oracle's growth engine is almost entirely taken over by AI infrastructure, with an extremely obvious differentiation:

Traditional cloud applications (SaaS): Only grew by 10% (NetSuite even recorded a mere 6% growth);

Cloud infrastructure (OCI): Skyrocketed by 121%;

Of which computing power infrastructure (CPU & GPU): Surged by 151%.

This sudden shift in business structure is reflected in the financial results as a notable decline in gross margin, yet the operating margin remains rock-solid at 42%.

The underlying business logic is as follows:

Data center construction is a heavy-asset investment, where early-stage depreciation and power consumption eat into gross margin; but OCI has extremely large scale effects, coupled with extremely restrained operating costs, which directly push the operating profit back up through "operating leverage".

Oracle is thoroughly "de-softwareizing" itself, transforming into a super computing power infrastructure operator wrapped in a tech shell.

02. Break Wall Street's common sense: Use other people's money to run your own race

Given that heavy-asset investment burns huge amounts of money, with the total annual under-construction scale reaching as high as USD 70 billion or even USD 90 billion, and S&P even downgraded its rating for this reason, isn't Oracle afraid of being dragged down?

The management presented two highly controversial yet extremely cunning "scalpels" during the earnings call:

The first scalpel: BYOH (Bring Your Own Hardware) and huge advance payments.

Among the USD 664 billion RPO in this cycle, more than USD 30 billion of new AI contracts added in Q1 are mostly under the model of customer advance payment and customer-provided hardware.

This means: If customers want computing power, they will bring their own NVIDIA chips to Oracle to rent computer rooms and configure networks. By doing this, Oracle cleverly passes the ultimate risk of computing power depreciation back to downstream parties.

The second scalpel: Overturn the "GPU residual value zeroing theory".

The market has long been worried that GPUs only have a service life of two to three years and will depreciate sharply after iteration. But the financial report disclosed data contrary to the market narrative:

GPUs that were renewed for contracts or resold this quarter recorded an average premium of 20%; the vast majority of these cards have been in service for more than 4 years, with a comprehensive utilization rate as high as 97.9%.

Old GPUs have not turned into scrap metal at all, but are being snapped up for lease at a premium. This directly shatters the traditional asset impairment model calculated based on 3-year depreciation.

03. Decisive battle for 10,000-card clusters: The behind-the-scenes driving force behind GPT-6 Astra

If you look at the specific construction blueprint, you will find that Oracle is leaving its competitors far behind in the delivery speed of computing power infrastructure.

Abilene Campus: 131,000 GPUs were delivered in a single quarter, and the acceptance cycle was compressed to just 24 hours. As disclosed, the highly anticipated next-generation model GPT-6 Astra is currently undergoing intensive training here;

The next GW-level campus: Construction has already kicked off, and NVIDIA's next-generation Vera Rubin system will be stationed here in Q2;

Multi-cloud breakthrough: Its database business surged 353% year on year, even competitors AWS and Microsoft have to fully access the Oracle database.

From power supply (independent power supply via Bloom Fuel Cell) to network, Oracle is adopting the extreme construction method of engineering infrastructure to push data centers to the front line in phases like building blocks.

The end of computing power is power: Why choose "expensive" fuel cells instead of traditional gas turbines?

In the AI infrastructure sector, the real decisive point often does not lie in chips, but in transformers.

The grid connection queue on the U.S. public power grid often gets stuck for 3 to 5 years. By the time the grid is fully connected, the 10,000-card cluster in hand will have long been outdated. To avoid being held back, Oracle simply adopts the "behind-the-meter off-grid power generation" solution.

But when choosing between two power generation solutions, Oracle made a radical decision that caught the traditional energy industry by surprise:

It placed large-scale bets on Bloom Energy (BE)'s Solid Oxide Fuel Cell (SOFC), instead of the large heavy-duty gas turbines of energy giant GE Vernova (GEV).

On the surface, this is a "anti-economics" decision:

GEV's large combined cycle power plants have extremely high thermal efficiency, and their levelized cost of electricity (LCOE) is much lower; while the LCOE of fuel cells is significantly higher.

But Oracle is calculating a different "time discount account":

Extremely fast delivery time: GEV's large heavy-duty gas turbines take at least 3-5 years to go from order placement, customization to complex power plant civil construction before they can be put into operation; while BE's standardized fuel cell modules only take a few months to one year from factory to deployment, perfectly matching the urgent demand pace of 10,000-card clusters.

Avoid the environmental protection death line: Large gas turbines operate based on high-temperature combustion, which causes severe NOx emissions. In states with strict environmental regulations, the emission permit approval process alone can drag the project to a halt. BE's SOFC operates through electrochemical reactions, with almost zero NOx and particulate matter emissions, so the environmental impact assessment can be passed smoothly.

Lego-like on-demand capacity expansion: Fuel cells are modular. When a batch of cabinets is added to the computer room, dozens of megawatts of power can be supplied on site, avoiding the heavy civil construction burden of investing a huge sum of money in the power plant at the very beginning; in the future, it can even run on hydrogen blend or even 100% hydrogen, fully supporting the "green carbon reduction" narrative.

Prefer to bear higher electricity prices, and never wait a single day for the grid.

Prefer to buy the more expensive one, not the slower one. At the cost of extremely high LCOE, Oracle has forcibly broken through the wall of the traditional power cycle, squeezing the time difference of "computing power monetization" to the extreme.

04. The Sword of Damocles behind the prosperity

The full-year FY27 guidance given by Oracle is extremely confident: revenue is guaranteed to reach at least USD 90 billion, with a 34% growth.

But this is by no means without cost. The market will only examine it with more critical eyes in the future:

Maturity mismatch risk: The majority of the huge RPO will be realized within the next 36 months or even FY28, but the huge CapEx has to be paid out right now.

When will FCF turn positive: Although the management predicted that "the EBITDA conversion rate will be close to 100% after the project ramps up", it has never given an exact timetable for when the cash flow will turn from negative to positive.

Concentration concern: When large customers develop their own mature chips, or the commercialization of AI large model applications is blocked, will these leading customers who have paid huge advance payments suddenly hit the brakes?

Before the Investor Day on October 28, the market will continue to fluctuate drastically between these two sentiments: on one hand, marveling at its endless tens of billions of dollars in orders, on the other hand, keeping a close eye on how much real money it will burn in the next quarter.

Oracle has taken the fastest train in the AI era, but outside the window is a bottomless cash flow cliff. This high-stakes bet is entering its climax.

Disclaimer: The copyright of this article belongs to the investment research team of Beijing Gelonghui Investment Consulting Co., Ltd. (Duan Yuehan: A0160625050003). This report is produced based on the principles of independence, objectivity, impartiality and prudence. All information is sourced from public materials and collected, abstracted and edited legally, reasonably and appropriately.

Stock market investment carries risks, and caution is advised when entering the market. No investment advice in this article shall be deemed as the basis for your investment transactions. You shall make your own investment decisions independently and bear your own risks. Any unit or individual that unauthorizedly publishes, reproduces, disseminates, or publishes/reposts this article on the Internet shall be deemed to have committed an infringement, and our company will pursue legal responsibilities for the infringement.

This article is from the WeChat official account "Gelonghui Finance Hotspot" (ID: glh_finance), author: Gelonghui Editorial Team, and 36Kr is authorized to publish this article.