Big Tech Is Borrowing Heavily for AI | Future Intelligence #33

Oracle confirmed its target of more than $225 billion in revenue in fiscal 2030 and, for the first time, paired it with a budget: no more than $20 billion in new capital, while some of its computing capacity will be built on third-party silicon. Meta is going to the market to raise money for the same purpose, with an issuance of up to $95 billion; Amazon has opened an $80 billion facility backed by purchased hardware, and Microsoft has returned to the bond market after nine years with a 10-tranche offering. Financing for computing capacity is moving from operating cash flows to balance sheets and capital markets, shortening the period during which companies can get by simply talking about returns. Alongside this, Intel is buying quantum software company Classiq for $2.3 billion, and Google DeepMind’s Genie 4 generates interactive worlds in real time, complete with sound.

Future Intelligence is an exclusive analytical report published once a week and available only to Bulios Black members. Members receive it automatically by email every Sunday morning—in full, including specific scenarios and market implications. Get permanent access to the report by joining Bulios Black.

Oracle sets an AI budget: $225 billion in revenue by 2030 and a cap on new capital

Oracle $ORCL held an investor day on Wednesday, October 28, and reaffirmed a target first announced in October 2025: revenue of more than $225 billion in fiscal 2030. That implies a compound annual growth rate of over 35% from FY26.

That is an increase of more than $158 billion from today. The new element was a cap on new financing, and the share price initially jumped 10.7% to $144.54 in response; by Friday, however, it had given back the entire gain and closed the week at $130.67, practically where it stood before the investor day.

The second half of the story began a day earlier. At the Oracle AI World conference in Las Vegas on October 27, the company announced that it would bring Google TPUs and Cerebras inference computers to its cloud. This makes Oracle Cloud Infrastructure the first hyperscale platform outside Google to offer TPUs—with both expected to be available in calendar year 2027.

Why the market valued the cap, not the target

Alongside its growth metrics, Oracle promised a non-GAAP operating margin above 40% and non-GAAP earnings per share above $21 in FY30. Free cash flow is expected to turn positive in the second half of FY28 and exceed $30 billion in FY29—only after the heaviest wave of data center spending has passed.

But the financing limit had the biggest impact: Oracle committed to raising no more than $20 billion in new capital from the markets through FY30—$10 billion in equity and $10 billion in gross debt, excluding the refinancing of existing obligations. The share-price reaction suggests investors were more concerned about financing risk than profitability risk.

The logic is simple. Without a cap, every new AI capacity contract raised an open question: how many new bonds and how much shareholder dilution would it bring? A quantified limit turns expansion from an open-ended variable into a budgeted item—and only then do the margin targets make sense.

Third-party silicon as a cloud product

Cerebras $CBRS builds chips in what is known as a wafer-scale design: rather than dozens of processors cut from a single silicon wafer, the entire wafer is one chip. This eliminates communication between chips and the movement of data, reducing response-generation time—a key advantage for long reasoning chains and agentic tasks.

This gives Oracle a third leg of computing capacity. Alongside Nvidia Vera Rubin NVL72 systems, which Oracle announced for OCI with plans to scale them into multigigawatt superclusters, the offering now includes eighth-generation TPUs in two versions—8t for training and 8i for inference—as well as wafer-scale hardware for fast inference.

“AI inference demands choice,” said Mahesh Thiagarajan, president of Oracle Cloud Infrastructure, summing up the logic. For Oracle, this is more than marketing: the more accelerators it offers, the less bargaining power any single supplier has—and the more room there is to deliver the promised margin above 40%.

Whose economics are changing

For Google $GOOG, this is a major shift in its business model. Until now, TPUs were reserved for its own data centers and Google Cloud customers; deploying them on an external platform turns them into a standalone source of hardware revenue. Wells Fargo estimates TPU sales of $99 billion in 2027, while TD Cowen forecasts $62.5 billion.

Cerebras, which went public in May, is expected to triple its revenue next year to nearly $3 billion; after the announcement, its shares gained as much as 10%. Oracle did not say how much Cerebras equipment it would take—the volume remains the weakest point in the announcement, as with TPUs, where only a multiyear capacity ramp-up has been mentioned.

At the same time, Oracle is planning FedRAMP Class D certification (the new name for the FedRAMP High level), which would open the door for regulated government workloads to use its multicloud AI database on Exadata through Oracle AI Database@AWS, @Azure, and @Google Cloud. It is the same pattern as with the chips: Oracle is selling its layer within other companies’ infrastructure.

Oracle’s share price now reflects a specific, dated commitment. The margin above 40% and EPS above $21 in FY30 are measurable milestones against which the market will assess each quarter individually.

Any delay in ramping up capacity or deterioration in the economics of renting third-party silicon will therefore show up more quickly than before, when the story was simply about contracts.

Meta finances AI data centers through the market: $95 billion offering

Meta $META announced on October 26 an issuance of equity instruments worth up to $95 billion, earmarked exclusively for AI infrastructure and what the company calls personal superintelligence. It is the largest capital-raising operation in the company’s history and marks a shift in how computing capacity is financed.