If you think the AI trade is over, NVIDIA CEO Jensen Huang just made a compelling case that we are barely out of the starting gate. Speaking live from CES 2026 in Las Vegas, Huang outlined a future where the entire global computing infrastructure, valued at over $10 trillion, is being rebuilt from the ground up. For value investors, this isn't just about the next fast chip; it represents a fundamental structural shift in how the world processes information, creating a massive, multi-year replacement cycle for legacy hardware.

Jensen Huang (Founder & CEO, NVIDIA): Huang continues to position NVIDIA not merely as a component manufacturer, but as a "full-stack" computing platform. His core thesis is that traditional software coding is being replaced by AI training and reasoning, necessitating a complete overhaul of data centers, networking, and industrial design software.

Huang argues that we are witnessing two simultaneous shifts: the move from CPUs to GPUs (accelerated computing) and the move from coding software to training AI. This creates a massive replacement cycle for existing infrastructure.

For investors, the crucial number here is $10 trillion. That is the estimated value of the last decade's computing infrastructure that needs to be modernized. Huang notes that hundreds of billions of dollars in venture capital and R&D budgets are shifting entirely toward this new stack. This suggests that NVIDIA's demand isn't a temporary spike, but a long-term structural upgrade of the world's hardware.

The headline announcement was the Vera Rubin platform (named after the astronomer who discovered dark matter). With Moore's Law slowing down, NVIDIA can no longer rely solely on shrinking transistors to get faster. Instead, they are using "extreme co-design," building six distinct chips that work as a single supercomputer.

The specs highlight a massive engineering lead. The new Rubin GPU features 5x the floating-point performance of the previous Blackwell generation, despite having only 1.6x the transistors. Furthermore, the system is highly efficient: the new NVL72 rack runs on 45°C warm water cooling, eliminating the need for energy-intensive chillers. This creates a "time to market" moat, allowing companies to train massive 10-trillion parameter models 4x faster than before.

One of the biggest hidden costs for data centers is the "buffer" capacity they must reserve for power spikes. AI workloads typically spike power usage by ~25% instantly, forcing companies to buy more power infrastructure than they usually need.

The Vera Rubin architecture introduces "power smoothing," which effectively caps these spikes. For a CFO at a major cloud provider, this is a massive selling point. It means they don't have to waste capital on 25% "buffer" electricity infrastructure. This lowers the Total Cost of Ownership (TCO) for NVIDIA customers, making it incredibly difficult for cheaper competitor chips to compete on actual operating costs.

We are moving past simple chatbots to "Agentic systems" that can reason, plan, and use tools. Huang discussed "test time scaling", the idea that AIs now "think" before they answer.

For investors, this creates a compounding demand for compute power. As models get smarter, they require significantly more processing power per query, not just for training. Huang noted that the number of tokens generated increases by roughly 5x per year. This ensures sustained demand for NVIDIA's GPUs even after the initial training phase is over.

A major pain point for developers is "context", an AI's ability to remember long conversations or massive datasets. As context grows, it clogs up the network.

NVIDIA introduced a solution using the BlueField-4 processor to handle the "KV cache" (the AI's working memory). This offloads massive amounts of data from the expensive GPU to a dedicated storage platform within the rack. This enables "lifelong learning" for AI agents, allowing them to remember every interaction forever, a critical feature for enterprise adoption that competitors have yet to solve at this scale.

Huang made a bold claim: "NVIDIA today is the largest networking company the world has ever seen." He is pushing their Spectrum-X Ethernet platform, which turbocharges standard data center cabling for AI.

The investment thesis here is cross-selling. Huang argues that if their networking switch improves a $50 billion data center's efficiency by even 10%, that saves $5 billion, essentially making the networking gear "free." This allows NVIDIA to capture massive revenue from networking hardware, moving beyond just the GPU and into the cables and switches that connect them.

Investors often worry that tech giants like Google or OpenAI might become too verticalized and stop buying NVIDIA chips. However, Huang heavily emphasized NVIDIA’s support for the "Open Model" ecosystem, specifically mentioning DeepSeek R1.

By democratizing the software layer, ensuring that powerful open-source models are available to everyone, NVIDIA ensures decentralized demand. If the software is free and accessible to every startup and nation, the only scarce resource remains the NVIDIA hardware required to run it. This prevents any single software company from bottlenecking the market.

A major barrier to AI adoption in healthcare and finance is data privacy. Banks cannot risk sending proprietary data to a cloud processor if there is any risk of exposure.

Huang explicitly noted that the entire Vera Rubin stack is "confidential computing safe." Every bus, link, and connection is encrypted, even during computation. This unlocks the regulated enterprise market, arguably the largest pool of untapped AI spending, by allowing banks and hospitals to use cloud AI without fear of data leaks.

Finally, NVIDIA is aggressively moving AI out of the computer and into the physical world. Huang introduced Cosmos (a physics-based world model) and Alpamayo (a thinking model for self-driving cars).

This is a strategy for deep vendor lock-in. By partnering with industrial titans like Siemens, Cadence, and Synopsys, NVIDIA is embedding its CUDA-X libraries into the software used to design chips, factories, and robots. The Mercedes-Benz CLA, launching soon with NVIDIA's full stack, is just the first domino. This positions NVIDIA to power the automation of the physical economy, a market potentially larger than the digital one.

NVIDIA is demonstrating that it is no longer just a hardware component supplier; it is the architect of the next global industrial platform. By integrating six custom chips into a single system, solving complex economic problems like power smoothing, and securing the physical industrial sector, they are building a moat that goes far beyond raw processing speed.

So What?

For the value investor, the "Vera Rubin" launch confirms that NVIDIA is successfully decoupling its performance gains from the slowing pace of Moore's Law. The $10 trillion infrastructure reset provides a long runway, while new features like confidential computing and physical AI open up vast, previously untapped markets in finance and heavy industry.

For more of my insights on this topic, be sure to follow me.

I am not a licensed financial advisor, and the information shared here reflects my personal investment decisions and opinions only. This content is for informational and educational purposes and should not be construed as financial, investment, or trading advice. Past performance is not indicative of future results. Investing involves risks, including the potential loss of capital.