Oracle bets private data beats public models

The database giant promises AI training without data exposure. Ellison's $500 billion wager on enterprise secrets.

Larry Ellison wasn't in the room. Security concerns kept Oracle's 81-year-old founder piped in from an offsite location during the company's AI World conference. But his message landed with precision: The next phase of artificial intelligence won't run on public internet data—it'll run on the private information locked inside corporate databases.

The pitch is simple. Large language models reach their ceiling when trained only on publicly available content. Real value comes from connecting those models to proprietary enterprise data—customer records, financial details, operational secrets—without exposing that information to model providers like OpenAI or Anthropic. Oracle claims it can thread that needle.

Key Takeaways

• Oracle's RAG technology keeps enterprise data in-house while letting AI models query it—no exposure to OpenAI or competitors

• Abilene data center delivers 1.2 gigawatts across 400 hectares with 500,000 Nvidia GPUs—one of world's largest AI facilities

• Tight coupling with Nvidia and $300B OpenAI contract creates dependency risks if either partner stumbles or AI demand weakens

• Vertical AI targeting core processes shows promise while 95% of horizontal AI projects fail to reach production deployment

This repositioning marks a genuine shift for a company that spent decades selling database licenses. Now Oracle positions itself as essential infrastructure for an AI era where competitive advantage flows from training models on information competitors can't access.

What's actually new

Oracle announced two products aimed at making this vision operational. The AI Data Platform combines data storage with integrated AI services, letting companies build what Oracle calls a "data lakehouse"—structured warehouses merged with flexible data lakes—and run AI models directly against that infrastructure.

The technical mechanism is RAG, or Retrieval Augmented Generation. Oracle's system converts enterprise data into vectors—mathematical representations AI models can process—but keeps those vectors inside Oracle's database. Models query specific information without seeing the full dataset. An example from Oracle's own sales operations: AI models analyzed anonymized customer data to predict which clients would buy which products in the next six months, generating automated email campaigns with personalized references.

"It's fascinating that it can solve such a problem so quickly," Ellison said, describing AI determining where Oracle's sales teams should focus their next two quarters.

The second announcement: Multicloud Universal Credits. One contract lets enterprises use Oracle databases and AI services across Amazon Web Services, Google Cloud, Microsoft Azure, or Oracle's own infrastructure. Dave McCarthy, research VP at IDC, called it "the rocket fuel that could accelerate broad adoption of Oracle's multicloud services."

The strategy's clear: Oracle isn't competing with cloud giants—it's positioning as their complement. An "and" cloud, not an "or" cloud, as the company frames it internally.

The infrastructure arms race

Behind the business proposition sits massive physical infrastructure. Abilene, Texas is where Oracle, OpenAI, and SoftBank are constructing one of the world's largest data centers. Scale: 1.2 gigawatts—enough to power a million four-bedroom apartments. The facility sprawls across 400 hectares. Up to 500,000 Nvidia GPUs will populate eight massive buildings. About 3,500 workers show up daily to build it.

What differentiates these facilities, according to Neil Sholay, Oracle's VP for AI in EMEA speaking with implicator.ai in Las Vegas, is what the company calls "Second Generation Cloud" architecture. While competitors designed infrastructure a decade ago and expanded incrementally, Oracle built for high-performance computing from the start.

The core concept: bare metal. Applications run directly on hardware without virtualization layers between them and processors. "That means application and GPUs or CPUs are so close you can barely see the gap between them," Sholay explains. "You get incredibly high performance."

Then there's RDMA—Remote Direct Memory Access. A server in Frankfurt can reach into a Paris server's memory directly. No copying data back and forth. For training runs that last days or weeks, that reduction in latency matters. So does reliability when a single failure can waste millions in compute.

Oracle skipped the custom chip race. While Google built TPUs and Amazon developed Trainium processors, Oracle bet on optimizing infrastructure around Nvidia's latest generations—currently H100, soon Blackwell. Sholay claims Oracle runs "generally twice as fast for half the cost." The customer list includes OpenAI, Meta, Elon Musk's xAI, and Nvidia itself.

The dependency calculation

This strategy carries substantial risk. Oracle is tightly coupled with few key partners—primarily Nvidia and OpenAI. The reported contract between Oracle and OpenAI: $300 billion. If OpenAI hits economic or regulatory turbulence, Oracle sits on massive data center capacity without its anchor tenant.

The Nvidia dependency runs equally deep. Oracle ordered 50,000 AMD GPUs for 2026 as a hedge, but Nvidia remains foundational. Any supply disruption or competitive shift damages Oracle's expansion timeline.

The fundamental question: Is demand for AI compute sustainable or are we watching a bubble inflate? Ellison describes demand for AI inference as "insatiable." The counterargument carries weight. Oracle, Microsoft, and OpenAI are pouring hundreds of billions into infrastructure collectively. If global AI momentum weakens, overcapacity becomes likely—with falling prices and eroding margins following close behind.

Building gigawatt-scale data centers requires more than capital. It needs available energy, cooling systems, and regulatory approvals. "There isn't 4.5 gigawatts of uncontrolled capacity on the grid," experts note about Oracle's Stargate plans. The company will need to build its own power generation—gas turbines, possibly small nuclear reactors. Every delay gives competitors room to catch up.

The vertical bet

The real test isn't technological—it's whether enterprises actually deploy AI in core business processes. Sholay knows the statistics: 95 percent of generative AI projects fail, per studies from Gartner, McKinsey, and MIT. Companies experiment with chatbots but few initiatives reach production.

Sholay distinguishes between horizontal AI—tools distributed broadly across organizations—and vertical AI targeting core processes. A UK satellite TV provider runs 2,000 service vehicles, mixing combustion and electric. Electric vehicles cost 13 percent more to operate. Oracle and the client are building an AI agent analyzing weather, traffic patterns, electricity prices per kilowatt-hour, and charging station availability in real time to optimize routes.

"That's a critical vertical core process," Sholay says. "That's typically not done by most companies today."

The pattern holds across examples: digital research agents for financial institutions, HR assistants for supermarket chains saving 90,000 hours annually. Not general-purpose tools—targeted systems attacking specific operational bottlenecks.

Why this matters:

• Oracle's repositioning tests whether AI infrastructure advantage comes from proprietary chips or optimized architecture around commodity processors. The company wagered on the latter while competitors invested in custom silicon.

• The private data thesis suggests competitive moats shift from model capabilities to proprietary training data and secure access mechanisms. If correct, enterprises with unique datasets gain advantage regardless of which foundation model they use—making infrastructure providers like Oracle potential kingmakers.

❓ Frequently Asked Questions

Q: What is RAG technology and why does it matter for enterprise AI?

A: RAG (Retrieval Augmented Generation) converts enterprise data into vectors—mathematical representations AI models can process—while keeping that data inside the company's database. Models query specific information without seeing the full dataset. This lets companies use ChatGPT or Claude on proprietary data without sending customer records, financials, or trade secrets to OpenAI or Anthropic.

Q: How big is the Stargate project beyond the Abilene facility?

A: Stargate will ultimately span multiple locations and reach 10 gigawatts of AI computing power. Oracle, OpenAI, and SoftBank plan to invest up to $500 billion total. Beyond Abilene, five more U.S. sites are planned in Texas, Ohio, and New Mexico, plus international projects in the United Arab Emirates and Norway.

Q: Why didn't Oracle build custom AI chips like Google and Amazon?

A: Oracle bet on optimizing infrastructure around Nvidia's latest GPUs rather than developing proprietary processors. The company uses bare metal architecture (no virtualization layer) and RDMA technology to squeeze maximum performance from H100 and upcoming Blackwell chips. Oracle did hedge by ordering 50,000 AMD GPUs for 2026, but Nvidia remains central to the strategy.

Q: What's the difference between Oracle's "bare metal" and traditional cloud computing?

A: Traditional clouds run applications on virtual machines, adding a software layer between apps and hardware. Oracle's bare metal approach runs applications directly on physical servers—no virtualization in between. For AI training that runs days or weeks, this eliminates latency from copying data through virtualization layers. Oracle claims this delivers twice the speed at half the cost.

Q: How does Oracle's Sovereign Cloud address European data protection requirements?

A: Oracle's EU Sovereign Cloud runs on infrastructure physically separated from U.S. systems and operated entirely by Europe-based personnel. This ensures compliance with strict European data residency rules. The company already operates a similar sovereign cloud in the UK serving government customers, and has expanded to Germany and Spain for organizations requiring guaranteed data sovereignty.