Jensen Huang opened CES 2026 with a striking claim: the computing industry is undergoing a $10 trillion modernization, funded by hundreds of billions in VC capital and enterprise R&D budgets shifting from classical methods to AI.
The technical specifications alone are staggering. Vera Rubin's 240 terabits/second bandwidth. Cosmos's trillion-mile simulations. DeepSeek R1 proving open models have reached the frontier.
But the most significant announcement wasn't about performance—it was architectural.
Every breakthrough NVIDIA demonstrated accelerates a fundamental shift: AI infrastructure requires data centralization at a scale that creates entirely new infrastructure requirements.
Key Takeaway (Before the Details)
AI infrastructure enables data convergence that was previously impossible.
That convergence creates a new critical infrastructure layer: secure vector embeddings and context memory at scale.
Traditional encryption architectures can't match AI throughput. Encryption-in-use becomes the enabling layer for regulated enterprise deployment.
The Architectural Shift: Why NVIDIA's Breakthroughs Demand New Security Infrastructure
NVIDIA's keynote revealed three capabilities that transform how enterprises must think about data architecture—and create demand for complementary security infrastructure.
1) Agentic AI Systems Enable Unified Context Memory
Huang demonstrated agentic systems that reason across multiple data sources, use tools, and maintain persistent memory across sessions.
The personal assistant example—emails, calendars, home cameras, task lists—illustrates the architectural pattern clearly.
Agents deliver value by unifying previously siloed enterprise data.
The business implication: organizations adopting NVIDIA's agentic AI capabilities will centralize fragmented data into vector databases, memory layers, and knowledge graphs.
ServiceNow, Palantir, and Snowflake aren't just integrating NVIDIA's stack—they're becoming aggregation layers for enterprise data at unprecedented scale.
This centralization enables new capabilities. It also creates new infrastructure requirements.
2) Test-Time Scaling Creates Persistent Storage Requirements
The shift from one-shot inference to reasoning models (like DeepSeek R1) introduces massive token growth per query—on the order of 5x annually.
Huang revealed that Vera Rubin's KV cache architecture adds 16TB of context memory per GPU to handle this explosion.
The architectural implication: this context memory contains sensitive business data that becomes persistent, indexed, and optimized for retrieval.
Every conversation. Every query. Every intermediate reasoning step.
A financial services firm using agentic AI for fraud detection isn't just analyzing transactions—it's building a permanent, searchable vector record of customer behavior.
For regulated enterprises, this persistent storage creates new compliance and security requirements that traditional architectures weren't designed to handle.
3) Open Models Democratize AI—and Multiply Infrastructure Demand
Huang celebrated DeepSeek R1 as proof that open models have reached the frontier, enabling "every company, every industry, every country" to participate.
For infrastructure providers, this represents a massive expansion opportunity.
Instead of a handful of hyperscalers deploying foundation models, thousands of organizations are now centralizing sensitive data—but without hyperscaler security budgets or dedicated security teams.
A mid-market fintech running Nemotron 3 for customer support needs enterprise-grade security at mid-market economics.
A hospital deploying BioNeMo for clinical decision support must meet HIPAA requirements without Google's threat intelligence budget.
NVIDIA's democratization of AI performance creates demand for democratized security infrastructure that scales to these new deployment patterns.
The Next Frontier Benchmark: Encryption at Vera Rubin Scale
The keynote devoted significant time to performance benchmarks:
• Vera Rubin: 5x peak inference performance, 3.5x training throughput
• Spectrum X: 25% higher throughput for AI Ethernet workloads
• MV-Link 6: 240 TB/s cross-sectional bandwidth (roughly 2x global internet traffic)
These breakthroughs create an architectural possibility that wasn't previously viable: encryption-in-use at production AI scale.
Traditional encryption strategies face fundamental limitations at this performance threshold.
Encryption-at-rest: NVIDIA's confidential computing features protect models during deployment—a critical foundation layer. The natural complement is protecting vector embeddings during the retrieval operations that Vera Rubin's KV cache architecture now makes performance-viable at scale.
Application-layer encryption: Vector similarity search requires mathematical comparison operations. Even if private data is stored in encrypted form, traditional approaches decrypt before similarity calculations. At Vera Rubin scale, milliseconds of exposure per query multiplied by millions of queries creates persistent attack surface.
The reconstruction risk: Vector embeddings aren't one-way hashes—they're mathematically reconstructable. Research has shown modern embeddings can recover original text at high accuracy. When centralized vector databases contain millions of embeddings representing customer communications, medical records, or financial transactions, they become high-value targets.
The performance breakthroughs NVIDIA demonstrated make a new approach architecturally feasible: encrypted search at GPU speeds, maintaining sub-100ms latency while eliminating plaintext exposure during retrieval.
Quantifying the Business Impact: What Centralization Enables (and Requires)
To understand the infrastructure opportunity, translate NVIDIA's technical announcements into enterprise deployment scenarios.
Scenario 1: Healthcare AI Using Cosmos for Medical Imaging
Use case: Hospital deploys Cosmos for radiology report generation.
Data centralization: 10M imaging studies converted to embeddings for similarity search.
Compliance requirements:
• Average healthcare breach cost: $10.93M (IBM Cost of Data Breach 2024)
• HIPAA fine range: $100–$50,000 per record
• Potential exposure: $100M–$500M+ for 10M records
• Reputational damage: 61% of consumers would switch providers after a breach (KPMG 2024)
Infrastructure requirement: Encryption-in-use becomes table stakes for production deployment.
Scenario 2: Fintech Using Nemotron 3 for Fraud Detection
Use case: Bank deploys agentic AI analyzing transaction embeddings for fraud patterns.
Data centralization: 500M transactions/year stored as vectors in persistent caches.
Compliance requirements:
• Average financial services breach: $6.08M (IBM Cost of Data Breach 2024)
• Regulatory fines (PCI-DSS): $5,000–$500,000 per incident plus mandated monitoring
• Revenue impact: breach events drive lasting declines in transaction volume (Javelin Strategy)
Infrastructure requirement: Customer-controlled encryption keys (BYOK/HYOK) for regulatory compliance.
Scenario 3: Enterprise Using Physical AI for Autonomous Systems
Use case: Logistics company deploys physical AI for warehouse robotics and real-time decisions.
Data centralization: IoT telemetry, supply chain data, vendor contracts, video feeds—all transformed into embeddings.
Risk multiplication factors:
• Autonomous systems require 24/7 uptime (not periodic batch processing)
• Persistent context windows extend operational windows
• Multi-modal inputs increase data sensitivity and compliance scope
Infrastructure requirement: Security architecture that doesn't sacrifice the performance NVIDIA's hardware enables.
The Complementary Infrastructure Layer: Encryption-in-Use at AI Scale
NVIDIA demonstrated that AI infrastructure can scale beyond global internet traffic, train enormous models faster than entire industries can reorganize, and deploy reasoning systems in production.
This creates architectural demand for security infrastructure that matches that scale and performance.
The solution requires rethinking encryption architecture to work with, not against, GPU-accelerated workloads.
Forward-Secure Cryptographic Indexing
Encrypt embeddings before they enter the database, and perform similarity search directly on encrypted representations.
• Eliminates plaintext exposure during search
• Maintains sub-100ms latency via GPU-accelerated cryptographic operations
• Scales to billion-vector datasets without sacrificing the throughput NVIDIA's hardware enables
CyborgDB integrates with NVIDIA infrastructure to provide this layer, maintaining <15% latency overhead while ensuring data remains encrypted during query execution.
Customer-Controlled Key Management (BYOK/HYOK)
Enterprises maintain cryptographic control via HSMs or cloud KMS.
• Database administrators can't decrypt data even with full system access
• Insider threat mitigation requires both system access and independent key theft
• Compliance frameworks can verify independent key custody
Metadata-Aware Optimization
Encrypt what matters most—sensitive embeddings—while keeping low-sensitivity metadata accessible for system performance and operations.
This architectural approach enables NVIDIA's performance breakthroughs to deploy in regulated environments that traditional encryption approaches blocked.
The Enterprise Reality: Adoption Timelines Are Non-Negotiable
Security teams may want to slow AI adoption until security infrastructure matures.
That fails on business grounds.
Competitive pressure: adoption timelines are measured in quarters, not years.
Regulatory acceleration: emerging AI frameworks don't reduce adoption—they mandate security controls.
Talent expectations: top ML teams won't join organizations stuck in pilot purgatory.
The organizations that deploy encryption-in-use infrastructure alongside NVIDIA's performance stack will capture the full value of AI modernization without the security-adoption tradeoff.
The Infrastructure Stack CES 2026 Makes Possible
NVIDIA demonstrated the performance layer that makes AI infrastructure viable at scale:
• Vera Rubin and MV-Link 6 for throughput
• Spectrum X for AI networking
• Confidential computing for model protection
Open models (DeepSeek R1, Nemotron 3) democratized the model layer, enabling deployment across thousands of organizations.
The security infrastructure layer—encryption-in-use for vector databases—becomes architecturally viable at this performance scale, enabling regulated enterprises to deploy AI at the speeds NVIDIA demonstrated.
Together, these layers enable the $10 trillion modernization Huang described: AI infrastructure that delivers both breakthrough performance and enterprise-grade security.
The Opportunity the $10 Trillion Modernization Creates
The organizations that recognize this architectural shift early will build sustainable competitive advantages.
NVIDIA solved the performance problem. Open models solved the access problem.
The decade's biggest infrastructure opportunity is solving the security-at-scale problem that makes AI deployment viable in regulated industries: healthcare, finance, government, and enterprise.
The techniques already exist. The performance is proven. The demand is accelerating.
The barrier isn't physics—it's deployment prioritization.
The enterprises that deploy encryption-in-use infrastructure before their first vector database breach will still be operating when AI infrastructure reaches the full scale NVIDIA demonstrated at CES 2026.
About Cyborg
Cyborg provides encryption-in-use for vector databases, enabling enterprises to deploy AI at scale without exposing sensitive embeddings during search operations. CyborgDB integrates with NVIDIA infrastructure to maintain <15% latency overhead while ensuring data remains encrypted during query execution.
Learn more at cyborg.co.
