Larry Ellison reveals why Oracle built a power plant to train AI - and what it means for developers.

Larry Ellison reveals why Oracle built a power plant to train AI - and what it means for developers.

BackerLeader posted 8 min read

Oracle's AI Vision: From 1.2 Billion Watts to Saving Lives

Oracle AI World delivered a message developers need to hear: AI isn't about replacing engineers. It's about building tools that solve humanity's hardest problems while making developers dramatically more productive.

The Infrastructure Reality

Larry Ellison opened with a stunning statistic: Oracle is building AI data centers consuming 1.2 billion watts of power. That's enough electricity to power a million four-bedroom homes.

The scale is staggering. Oracle's Abilene, Texas facility will house 450,000 NVIDIA GB200 GPUs when fully provisioned. The project started as empty land in June 2024 and began delivering GPUs in less than a year.

But here's what matters for developers: Oracle isn't just building infrastructure for infrastructure's sake. They're training more multimodal AI models than any other company - models that developers will use to solve real problems.

The Database + AI Integration

Ellison made a critical point that other cloud providers miss: publicly available training data isn't enough. AI models need access to private enterprise data to reach peak value.

Much of the world's high-value data already sits in Oracle databases. Oracle's strategy: make that private data accessible to AI models while keeping it secure and private.

The Oracle AI Database uses Retrieval-Augmented Generation (RAG) to vectorize any data - whether in Oracle databases, OCI object store, or even Amazon S3 - making it accessible for AI reasoning without exposing it to other customers.

For developers, this means:

  • No need to copy sensitive data to train models
  • AI can reason across both public and private data
  • Security boundaries remain intact
  • Models stay current with real-time data

Code Generation That Actually Works

Oracle's approach to AI code generation differs from competitors in one crucial way: every generated application is stateless, secure, and scalable by default.

Using Oracle APEX with AI generation, applications automatically include:

  • Zero security holes (the generator doesn't forget things)
  • Built-in reliability (stateless design means automatic failover)
  • Unlimited scalability (designed for millions of users from day one)
  • No single point of failure

Ellison noted that 80% of Oracle employees report AI has already improved their work quality. Internally, Oracle uses AI for:

  • Touchless expense automation
  • Invoice matching
  • AI-driven recruiting
  • Deal progression in sales
  • Real-time supply chain alerts

The productivity gains are enabling Oracle to rebuild Cerner's entire codebase, a quarter-century of healthcare software, using AI generation. All clinic operating code is already complete, with acute hospital systems planned for next year.

Real-World AI Agents

The keynotes showcased AI agents solving complex real-world problems, not toy demos.

Healthcare: Provider to Payer Integration

Oracle built an AI agent that solves a problem every healthcare provider faces: prescribe the best possible care that's fully reimbursable.

The agent:

  1. Uses RAG to access latest medical literature and patient test results
  2. Assists doctors in determining optimal care
  3. Checks insurance policies to ensure full reimbursement
  4. Handles exceptions (like UK NHS covering Ozempic only above certain BMI thresholds)
  5. Provides banks with reimbursement data to extend credit to cash-strapped hospitals

This single agent addresses clinical care, insurance complexity, and hospital financing simultaneously - automating the entire healthcare ecosystem, not just the hospital.

Biofine Technologies: Fighting Antibiotic Resistance

Paulo from Biofine Technologies described using Oracle's vector database to save lives in Brazil.

Traditional bacterial testing takes five days. Biofine's solution:

  • Extracts bacterial DNA from patient samples
  • Creates vectors from millions of DNA elements
  • Uses Oracle's vector search across 700,000 bacterial DNA records
  • Identifies bacteria and antibiotic resistance in four hours

Results: mortality rates from bacterial infections dropped from 70% to 50%. Biofine expects to save 2,000 lives in Brazil in 2025 alone.

The vector approach handles bacterial mutations, finding close matches rather than requiring exact DNA sequences. This means identifying resistance patterns even for novel bacterial strains.

The Metagenomic Testing Vision

Ellison revealed work on a metagenomic testing device that could revolutionize pathogen detection.

Current PCR testing only works for known pathogens. If your test comes back negative, doctors simply don't know what you have.

The new device:

  • Sequences everything in a blood sample
  • Identifies any pathogen (bacteria, virus, fungus) even if novel
  • Detects circulating tumor DNA for early cancer diagnosis
  • Determines antibiotic resistance patterns
  • Costs low enough for widespread hospital deployment

For developers, this represents the power of AI-assisted genomic analysis. The device would have detected COVID-19 far earlier, providing pandemic early warning.

Ellison's point: if these devices existed in hospitals worldwide, we'd never be caught off guard by novel pathogens again.

Energy Sector Transformation

Calvin Butler from Exelon described managing one of America's largest electrical grids while embedding AI across operations.

The challenge: more transformation in the next 10 years than the previous 100 years. Every technology addition to the grid introduces security risks.

AI helps Exelon:

  • Predict outages before they happen
  • Communicate restoration times accurately
  • Manage preventive maintenance schedules
  • Optimize grid investments based on demand patterns

Butler emphasized that AI deployment isn't about reducing headcount - it's about capturing data and improving customer communication. The grid becomes smarter, not smaller.

Transportation: Avis Budget Group

Marius from Avis Budget Group explained why they jumped to Oracle Database 23ai despite being on a "highly deprecated version."

Their reasoning: if you're going to upgrade anyway, go to the version where Oracle is investing. The 23ai capabilities and roadmap aligned perfectly with Avis's business needs.

Using natural language queries, Avis employees become "problem solvers, not information gatherers." Instead of clicking through dashboards to slice and dice data, they ask questions directly and get answers immediately.

The goal: reduce the lag between observing a problem and taking action. AI gives back time - the most important commodity nobody controls.

Marius's metric for success: how quickly can you get to market with the right decisions? AI dramatically compresses that timeline.

Hospitality: Marriott's Human-First Approach

Ty from Marriott emphasized that AI isn't replacing human hospitality - it's enabling it.

The problem: checking someone into a Marriott property requires navigating dozens of systems. Front desk associates spend time on "swivel chair" data entry rather than engaging with guests.

Marriott's digital transformation creates a "single pane of glass" - unifying dozens of systems so associates can focus on people, not processes. Less time on keystrokes, more time sharing local recommendations and creating authentic connections.

The AI strategy started by asking associates: "What's the most painful part of your job?" Then deploying AI to fix those pain points. The result: contagious adoption as associates experience real improvements.

The Vector Database Breakthrough

Multiple speakers highlighted Oracle's vector database as the enabling technology.

For Biofine, vectors allow matching similar (not identical) bacterial DNA, handling mutations automatically.

For Marriott, vectors enable personalized guest recommendations based on preferences and history.

For Avis, vectors power natural language queries across complex operational data.

The pattern: vectorization makes unstructured data (DNA sequences, customer preferences, operational metrics) accessible for AI reasoning. Oracle's database handles the vectorization automatically for any data source.

Multimodal AI Architecture

Ellison explained that modern AI models mirror human brain structure, multiple neural networks handling different tasks:

  • Convolutional networks for vision (edges, color, motion)
  • Vision transformers for recognition
  • Large language models for reasoning
  • Specialized networks for mathematics

Like the brain's specialized regions (visual cortex, prefrontal lobe, language centers), AI models use specialized networks working together.

Oracle's infrastructure supports these multimodal models at scale - the 1.2 billion watts isn't excessive, it's necessary for reasoning at human (or superhuman) levels.

The Code Generation Philosophy

Oracle's code generation differs from competitors in approach. Rather than generating code directly from English (which Ellison calls "curiously imprecise"), Oracle created a declarative AI generation language in APEX.

The advantage: generated applications have consistent characteristics:

  • Stateless design (automatic failover)
  • Built-in security (no forgotten edge cases)
  • Designed for scale (millions of users)
  • Zero single points of failure

This matters because generated code becomes production code. If the generator produces brittle applications, you've automated technical debt creation.

Medical Devices and Robotics

Ellison revealed Oracle's work on medical robotics, arguing robots will be better surgeons than humans - not because they're smarter, but because they have better tools:

  • Microscopic vision (seeing individual cells without microscopes)
  • Perfect hand-eye coordination
  • Ability to cut between healthy and cancerous cell layers

For Mohs surgery (removing facial cancer lesions), robots can see exactly where cancer ends and healthy tissue begins, enabling more precise removal with better cosmetic results.

The implication for developers: AI-assisted medical devices represent a massive opportunity. The challenge isn't just training models - it's building reliable, safe systems where errors have life-or-death consequences.

The Greenhouse Vision

Ellison described Oracle's robotic greenhouses - buildings that are actually inflatable structures held up by positive air pressure.

The advantages:

  • 90% less water than traditional farming
  • Grows food year-round
  • Locates near population centers (fresher food, lower CO2 transport)
  • No human contamination of growing areas
  • Robots move plants as they grow

For developers, this represents AI-orchestrated physical systems. The robots decide when to move plants, when to harvest, how to optimize growing conditions.

The same structure could serve as a Martian habitat - fold it up, transport it on a SpaceX rocket, inflate it on Mars. Ellison wasn't entirely joking about that possibility.

Climate Engineering with AI

Perhaps the most ambitious vision: using AI to engineer crops that manage atmospheric CO2.

Wheat modified by AI produces 20% more grain per acre while converting more CO2 into calcium carbonate (the mineral in coral reefs). The CO2 gets removed from the atmosphere permanently.

If deployed globally, modified wheat could reduce atmospheric CO2 from 440 parts per million to 400 - essentially managing the climate through enhanced photosynthesis.

The engineering challenge: designing plants that:

  • Produce more food
  • Consume more CO2
  • Convert CO2 to stable minerals
  • Maintain nutritional value

This isn't theoretical - Ellison's Wild Bio (part of Oxford's Ellison Institute) has already created wheat that produces 20% more grain per acre.

Autonomous Drones

Oracle built an air traffic control system for medical specimen delivery drones.

The system uses RFID tags to maintain chain of custody while protecting privacy - nobody knows whose blood sample is in transit, but the system ensures it reaches the right lab and results return to the right doctor.

Beyond medical delivery, the drones can:

  • Detect forest fires via infrared cameras
  • Identify arsonists who set fires
  • Follow suspects in police chases (safer than high-speed pursuits)
  • Search for lost hikers

For developers, this represents the challenge of building reliable autonomous systems operating in physical environments with safety implications.

What This Means for Developers

Oracle's AI World message: the infrastructure is ready. The models are trained. The database can vectorize your private data securely.

Now what are you going to build?

The examples from the keynotes show the pattern:

  1. Identify a complex problem (antibiotic resistance, healthcare reimbursement, grid management)
  2. Connect AI models to relevant private data via vector search
  3. Build agents that orchestrate multi-step processes
  4. Deploy at scale using Oracle's infrastructure

The competitive advantage won't come from having access to AI models - everyone will have that. It will come from:

  • Speed of implementation
  • Quality of private data integration
  • Sophistication of agent orchestration
  • Reliability of production deployment

Oracle's bet: developers who can connect AI models to real business data and build reliable agents will solve problems that were previously intractable.

The 1.2 billion watts of power isn't about showing off. It's about giving developers the computational foundation to build applications that change industries, save lives, and solve humanity's hardest problems.

That's the real AI revolution - not chatbots, but tools that let developers tackle challenges we couldn't solve before.

1 Comment

0 votes

More Posts

How I Built and Compared a Custom CNN vs MobileNetV2 for Boat Image Classification

Ezinne Anne Emilia - May 23

Most AI projects fail because companies try to boil the ocean instead of solving one simple problem.

Tom Smith - Sep 9

Oracle reveals why Python notebooks won't run your enterprise AI - and what will.

Tom Smith - Oct 14

The Possibility of Training a Multimodal AI for Cryptocurrency Auto-Trading Decisions

Muhammed Shafin P - Jul 28

Optimizing Neural Network Architectures for Improved AI Efficiency and Scalability

Aditya Pratap Bhuyan - Mar 19
chevron_left