Oracle's cloud business is powering xAI, OpenAI, and Nvidia — and the company is going cash-flow negative to fund the build-out

Summary

Oracle's cloud infrastructure unit has become the company's primary growth engine, powering AI training for OpenAI, xAI, and Nvidia while also securing major consumer customers like TikTok. The shift marks a dramatic reversal for a company whose founder, Larry Ellison, dismissed cloud computing as "complete gibberish" in 2008.

The turnaround has made Ellison the world's second-richest person, with Oracle's market cap at $654 billion. Oracle is paying a steep price to capture this opportunity: the company recorded negative annual cash flow for the first time since 1990, driven by aggressive data center expansion.

Oracle's infrastructure buildout

Oracle is building massive compute clusters to serve AI training workloads. The company is currently operating a tens-of-thousands-chip cluster for Nvidia near Singapore and running xAI's infrastructure from Utah. OpenAI represents the largest commitment, having signed deals for more than 5 gigawatts of computing power under the codename Stargate. Data centers are scheduled for completion by early 2027, with servers operational by summer 2026. Facilities are being designed for hybrid workloads: training now, inference later, assuming that at some point model development plateaus and focus shifts to deployment.

One new megasite in West Texas will consume over $1 billion annually in power alone. Oracle chose to run gas generators rather than wait for utility grid connections. The facility, being developed by Digital Bridge's Vantage Data Centers in Shackleford County near Abilene, will have a compute capacity of 1.4 gigawatts, roughly 40 percent larger than other single-site clusters.

Bare metal and strategic positioning

Oracle's path to dominance in infrastructure relied on a strategic bet made years ago. When the company was deciding between competing internal approaches to cloud, Ellison backed a bare-metal strategy championed by a team of former Amazon engineers. This approach of renting dedicated servers rather than virtualized capacity proved prescient for GPU-heavy AI workloads. The decision to focus on smaller data centers also allowed Oracle to expand into countries where AWS and Azure had not yet established presence.

Oracle also benefited from aggressive sales tactics and a simplified, lower-cost pricing model. Sales teams directly targeted AWS customers, emphasizing that Amazon's cloud bundled expensive side projects like satellites into their bills. The pitch resonated.

Key wins include Zoom, which needed capacity during the pandemic, and a $2 billion commitment from Uber in 2023. TikTok became a major customer whose infrastructure bill quickly surpassed $1 billion in run rate and soon eclipsed Oracle's entire cloud business. TikTok's video-intensive workloads forced Oracle to build AI-focused infrastructure capabilities that later became commercially valuable for training applications.

Execution risks and profitability concerns

Oracle is now hiring aggressively to deliver on these commitments. More than 600 workers have joined from Amazon in the past two years, attracted partly by Oracle's hybrid work policy as Amazon enforces mandatory office returns. About 23,000 employees now report through Clay Magorick, OCI's president and a potential successor to the 81-year-old Ellison.

The infrastructure business carries structural challenges. Training is lower-margin than inference because it requires more advanced semiconductors and precision cooling. The aggressive timelines for OpenAI have also created supply chain pressure. Costs for power transformers, cooling tents, and other critical components have risen sharply due to tariffs and vendor leverage. Oracle's cash flow turned negative, though "just barely," suggesting the company is managing burn but with limited margin for error.

Uncertainty remains about whether these massive training clusters will eventually be profitable. AI model development may reach diminishing returns faster than infrastructure vendors assume, leaving Oracle with expensive capacity and no buyers.