Oracle Is Spending Almost Every Dollar It Makes on AI. The Ledger Remembers What the Ego Forgets.

Oracle just told the world, without saying it on a slide, that it is spending almost all of its annual revenue on AI infrastructure. The number is not on the earnings deck. It is in the arithmetic between management guidance and the analyst consensus. For fiscal 2025, Oracle reported around 59.9 billion dollars in revenue and roughly 19.6 billion dollars in capital expenditure. That is a 33 percent capex-to-revenue ratio. For fiscal 2026, the consensus is moving toward 40 to 45 billion dollars in capital expenditure. If revenue grows to 65 billion, the ratio crosses 70 percent. On a cash-flow basis, Oracle is spending more than it generates before asking a single Wall Street banker for a loan. The headline from Crypto Briefing says investors are not thrilled. That is an understatement. They are pricing Oracle as a leveraged GPU fund, not as a software annuity. The ledger remembers what the ego forgets. I have watched this movie before. In 2017 I spent months auditing ERC-20 contracts in Remix, looking for integer overflow and other ways the code could disconnect from the promise. I found two mid-cap projects with critical vulnerabilities before they launched. That experience taught me a simple rule: code security correlates directly with market viability. A contract that cannot pay what it promises is not a contract. It is a narrative with a timestamp. Oracle's capex guidance is the same kind of contract. The code is the 10-K. The narrative is the investor day. The footnotes are the obfuscation. This article is not a defense of Oracle. It is an autopsy of the market's reaction. To understand whether the fear is rational, I need to break down the balance sheet, the customer contracts, the GPU depreciation schedule, the power bottleneck, and the hidden competition from NVIDIA itself. Start with the balance sheet. Oracle's core business is still a high-margin database and enterprise software franchise. That franchise generates steady cash flow. But the AI infrastructure strategy is converting that cash flow into physical assets at a speed that no software company has tried before. Microsoft, Amazon, and Google can spend billions because their cloud businesses already generate enough cash to absorb the pain. Oracle does not have that luxury. Its revenue base is smaller, its margin profile is thinner after accounting for cloud infrastructure, and its debt load is not trivial. The market does not care about the AI story. It cares about the free cash flow statement. If Oracle spends 40 billion in capex and generates only 25 billion in operating cash flow, the gap must be funded by debt, equity, or customer prepayments. Each of those sources has a cost. Debt increases interest expense and raises the risk of a downgrade. Equity dilutes existing shareholders. Prepayments reduce the cash value of future revenue. The market is not stupid enough to price all three as positive developments. The second issue is customer concentration. Oracle has signed OpenAI, xAI, and Meta as large AI infrastructure customers. Those names sound strong. But a brand name is not a commitment. A commitment has volume, duration, and a price floor. The public reporting does not disclose the price floor. If AI compute prices fall as supply floods the market, the customer will demand a repricing. If Oracle says no, the customer will walk to CoreWeave, AWS, or Azure. The customer commitments that the Crypto Briefing article mentions are not a guarantee. They are a stablecoin peg. And a stablecoin peg only works until the market tests it. I shorted UST in 2022 after watching the liquidity imbalance in the Curve pool. I saw more sellers than the algorithm could absorb. The peg looked stable until it was not. The same dynamic is visible in the GPU cloud market today. Supply is being built on demand forecasts that have never been tested at these prices. If the market price for an H100-hour falls below Oracle's cost, the customer has no reason to honor the old contract. They will ask for a new one. The ledger still remembers the old number. The human in the negotiation will not. The third problem is depreciation. A GPU cluster is not like a corporate campus. It is a technology asset with a useful life determined by NVIDIA's product release calendar. Oracle buys H100 and H200 clusters today. Twenty-four months later, NVIDIA ships a new architecture that is three times faster. The old cluster still has three years on the accounting depreciation schedule. But its economic value drops immediately. The accounting ledger says the asset is worth something. The secondary market says otherwise. Oracle is not alone in this problem. Every hyperscaler has it. But Oracle has the least margin of safety. If AI training demand slows or shifts to inference, the need for massive training clusters could decline faster than the depreciation schedule. The result is stranded assets and an ugly mark-to-market moment for future earnings. The fourth problem is power. AI infrastructure is no longer a chip problem. It is a power and cooling problem. A single modern GPU cluster can consume more electricity than a small city. Oracle has been signing long-term power agreements and building data centers near available grids. That is the correct move. But power contracts are not flexible. They are take-or-pay obligations. If utilization falls, Oracle still pays the utility bill. The capex number only tells you the hardware cost. The real fixed cost is the land, the power, the cooling, and the network. Those costs do not disappear when demand slows. Alpha hides in the friction of chaos. The friction here is the difference between the reported capex number and the real cash commitment. Oracle can slow GPU purchases. It cannot slow power contracts. Investors who only look at capex miss the physical contracts underneath. The fifth issue is the business model itself. Oracle is selling capacity for compute contracts. That means it is buying capacity before selling it. This is a classic merchant model. A merchant that builds a refinery before locking in customers takes on construction risk. If the refinery is built on time and under budget, the returns are exceptional. If it is delayed or over budget, the operating leverage works in reverse. Oracle is effectively building a network of AI refineries. The customer commitments are the throughput agreements. The percentage of revenue being spent on capex is the refinery's construction cost. The market is saying the refinery is being built too fast and too expensively for the current demand curve. I built a dashboard in 2024 to track GBTC and IBIT flows after the ETF approvals. I saw institutional accumulation before the Q4 rally. The lesson was simple: flows move price, but not always on the same day. Oracle's stock price is a flow of narratives. The current narrative is fear. But the underlying cash-flow data is still too incomplete to confirm the fear. That does not mean the fear is wrong. It means the fear is early. Early fear is still fear. The order book can stay silent longer than the narrative can stay patient. Now I need to address the strategy from the other side. Oracle is not trying to be the next AWS. It is executing a side-flank. It is building a vertically integrated cloud plus database plus AI infrastructure offering for customers who do not want to put all their compute on AWS or Azure. OpenAI and xAI need second sources. They need alternatives to Microsoft and Google. Oracle is a credible alternative because its network supports high-bandwidth, low-latency GPU clusters and because its enterprise software relationships provide an on-ramp for traditional companies to buy AI compute. This strategy is real. It has won real contracts. The question is not whether the strategy works. It is whether the price Oracle is paying for it will destroy the company before the strategy pays off. The competition is intensifying. Microsoft, Amazon, and Google are all raising their AI capex budgets. NVIDIA is also building its own cloud and investing in third-party GPU clouds like CoreWeave. That is a hidden competitor. Oracle wants to be a merchant of GPU compute. But NVIDIA controls the supply of the merchant's inventory. If NVIDIA decides to prioritize its own cloud or its own partners, Oracle's delivery schedule suffers. If NVIDIA launches a new GPU every year, Oracle's installed base suffers. Oracle cannot diversify away from NVIDIA without a massive ASIC project, and it has not publicly committed to one. This is what the analysts call single-vendor risk. The market is pricing it, but the Crypto Briefing article does not mention it. The blockchain world has already walked down this road. I have said for years that the dedicated data availability layer is overhyped. Ninety-nine percent of rollups do not generate enough data to justify a dedicated DA layer. They build the infrastructure first and hope the demand arrives later. Oracle is doing the same thing with AI data centers. It is building a dedicated infrastructure layer for a workload curve that might be less steep than the capex curve. Not because AI demand is fake, but because the current buying cycle is crowded. Everyone is building at the same time. The capacity will arrive at the same time. The price will fall at the same time. That is not a thesis. It is mechanical. The same logic applies to the governance of these contracts. Code is law is a myth. In DAO governance, the underlying smart contracts may be immutable, but the upgrade rights sit with a few multi-sig signers. The signers can change the rules when the market turns. Oracle's AI contracts are not public smart contracts. They are private agreements governed by procurement lawyers and board votes. The customer commitments are not immutable. They are renegotiable. When the downturn comes, the multi-sig will choose survival over the promise. That is the human variable the spreadsheets miss. There is also a crypto-native lens that the analysts on television ignore. The same capital that once chased DeFi yields and NFT floors is now chasing AI infrastructure through corporate bonds and cloud contracts. This is not irrational. It is the same behavior with a different wrapper. The yield has a cost. The denominator is now a data center. The question is whether the data center can generate a return above the cost of the capital that built it. In the 2020 DeFi summer, I ran leveraged yield farming on Aave. When the flash loan attack hit, the models said the positions were safe. The ledger said otherwise. I froze the positions and withdrew. I kept ninety percent of the capital. The same principle applies to Oracle. The models say AI revenue will eventually cover the capex. The ledger says the cash is leaving the balance sheet faster than revenue is arriving. The exit strategy is more important than the thesis. For Oracle, the exit strategy would be a pause in capex. But a pause in capex would also break the customer commitments. That is a trap. The contrarian case deserves to be stated cleanly. The market hatred of Oracle is not necessarily wrong, but it might be early. Oracle is being priced as if it will become a utility. Utilities are not exciting, but they are not zeros. If AI training demand stays hot into 2026, Oracle will have capacity when AWS is telling customers to wait. Oracle can also attach AI infrastructure to its existing database and enterprise software relationships. A bank that runs Oracle databases can buy Oracle AI clusters without introducing a new vendor. That is a real sales motion. CoreWeave and Nebius do not have that distribution. The retail screen sees a massive capex number and screams. The smart money sees a call option on AI capacity, and the call option is being priced as if it were already worthless. The blind spot in the bear case is the assumption that all AI infrastructure is the same commodity. It is not. A GPU cluster attached to a financial services customer with strict data residency rules is different from a GPU cluster in a merchant cloud. Oracle has a path to high-margin enterprise AI workloads, but that path is hidden inside the cloud revenue line. The company is not telling investors how much of the new AI revenue comes from enterprise workloads versus pure training-for-hire. That distinction determines whether the gross margin is 70 percent or 20 percent. I do not know the split. The public 10-K does not disclose it. That is the information gain the market is missing. The market is trading Oracle as if every AI dollar is a commodity dollar. If even a quarter of the AI infrastructure revenue is enterprise-attached, then the bear case is overdone. But if the large customer commitments are priced at zero margin, then the investor fear is rational. The silence in the order book is louder than the noise in the headlines. The takeaway is not a price target. The takeaway is a set of three numbers that will settle the debate. First, next quarter's operating cash flow. If Oracle cannot grow operating cash flow while spending forty billion in capex, the funding gap is real. Second, the ratio of customer prepayments to capex. If prepayments cover a large part of the construction cost, then the refineries are already paid for. If not, Oracle is betting its balance sheet on future demand. Third, the average duration of the NVIDIA supply and power agreements. Long contracts protect against price increases but also lock in fixed costs. Short contracts raise flexibility but create delivery risk. Watch those three numbers. Do not watch the stock price. The question is not whether Oracle can build data centers. It can. The question is whether the data centers can generate cash flow before the bond market stops answering the phone. The market is already telling Oracle that trust has a price. The ledger remembers what the ego forgets. The next earnings report is not a narrative event. It is a ledger entry. Read it like a smart contract: check the inputs, check the outputs, and verify the reserves before you sign the transaction.