Is Oracle’s AI Cloud a Hidden Money‑Sink? Uncovering the Real Profit Margins

Last month Oracle executives announced a bold forecast: over the next five fiscal years the company expects to generate $381 billion in revenue by renting dedicated cloud servers to OpenAI and other AI developers.

The market cheered, celebrating the former database‑centric giant’s transformation into an AI powerhouse.

However, an internal document obtained by The Information shows that Oracle’s AI cloud business margin is painfully low – only a few dozen percentage points, even less than Walmart’s toothpaste sales.

Financial reports indicate that in the three months ending August, Oracle earned about $900 million from renting Nvidia GPU servers but only $125 million in gross profit – roughly 14 cents of profit for every dollar of revenue. While AI cloud revenue has tripled over the past year, margins have stayed between 10 % and 20 %, averaging around 16 %. Some batches even lose money on older chips and lose more on newer ones. After accounting for labor, electricity, and depreciation, the effective net margin drops another 7 percentage points.

By comparison, traditional cloud services enjoy margins around 60 %; Oracle’s AI cloud operates at less than a quarter of that level, effectively building a “super‑data‑center” for OpenAI at a near‑break‑even cost.

01

“Buying chips and renting them out isn’t a good business.” Nvidia GPU prices are sky‑high, and the operating costs far exceed those of conventional cloud servers. Oracle’s quarterly spending on chip leases, power, and networking compresses profit margins, while deep discounts offered to marquee customers like OpenAI further erode profitability.

According to the Associated Press, Oracle earned about $10 billion in revenue from cloud server rentals in the fiscal year ending May, with 20 % coming from GPU servers. This share rose to 27 % in the most recent quarter, and internal forecasts suggest GPU cloud revenue could match Oracle’s traditional database and ERP businesses by 2028.

However, this growth comes at the expense of profit. Deploying the new Nvidia Blackwell GPUs caused a $1 billion loss in a single quarter because new data centers sit idle for weeks or months before customers begin using and paying for them, reducing asset utilization while power and OPEX costs remain.

Internal data shows GPU server utilization ranging from 60 % to 90 % depending on chip model. Nvidia CEO Jensen Huang claimed newer chips would instantly replace older demand, yet Oracle’s financials show the opposite: 2020’s Ampere chips still generate steady income, while the latest Blackwell chips depress margins.

Oracle’s biggest risk is customer concentration. In the three months ending August, $317 billion in cloud contracts were signed almost entirely with OpenAI. This creates a paradox: without OpenAI, Oracle’s cloud business cannot grow; with OpenAI, its margins are squeezed to the limit.

Guggenheim Securities analyst John DiFucci notes that Oracle’s internal debate pits high margins against low‑margin, high‑volume contracts, and the company has chosen the latter, betting on revenue scale to offset thin profits.

While Oracle’s stock rose nearly 40 % in three months, analysts warn this is a revenue‑driven boom, not a profit‑driven one.

02

Systemic risk is forming as a single company intertwines chips, cloud, energy, and capital. OpenAI ties together Nvidia, AMD, Oracle, CoreWeave, and dozens of Fortune 500 firms, controlling the compute backbone for most leading AI applications.

OpenAI’s recent deal with AMD grants it up to 10 % of AMD’s stock at a $0.01 exercise price, effectively allowing it to acquire shares for free upon meeting shipment and market‑cap targets.

The “purchase‑hold‑re‑purchase” financial loop is now closed, and the announcement sent AMD’s share price soaring while also leveraging Nvidia.

Estimates suggest OpenAI’s deals with Nvidia and AMD could total $5 trillion and $3 trillion respectively, with $3 trillion in long‑term compute contracts with Oracle and $22 billion with CoreWeave. The combined “Stargate” super‑data‑center project could require 20 GW of compute power – roughly the output of 20 nuclear plants – at an estimated cost of $500 billion per GW, approaching $1 trillion in total.

Analysts warn that OpenAI’s $40 billion financing, while appearing to be led by SoftBank, is actually a complex structured bond issuance that packages future API receivables for sale to pension funds, sovereign wealth funds, and hedge funds, creating a potential high‑yield “time‑bomb” if cash flows falter.

Currently, 65 % of Fortune 500 companies heavily rely on OpenAI’s models or APIs; a disruption could cause economic losses exceeding $100 billion.

03

AI infrastructure’s importance drives low margins; low margins force companies to scale aggressively to sustain valuations, creating a feedback loop across the industry.

OpenAI sits at the center, consuming compute and simultaneously generating demand narratives that inflate GPU values and attract capital. Suppliers like Oracle quietly bear the cost of this expansion.

If any external shock hits – energy price spikes, supply delays, capital market risk aversion, or tighter regulation – the “story‑driven boom” could turn into a “structurally amplified pull‑back.”

Thus, OpenAI has evolved from a “too‑big‑to‑fail” firm into a systemic risk for the AI era, where its stability impacts the broader intelligent‑infrastructure and macro‑credit cycles.