Nvidia Built An AI Moat. Can Rivals Find The Drawbridge?

Nvidia (NASDAQ:NVDA) is the defining company of the AI boom. Its chips power almost every major frontier model and data center build out. The markets believe this dominance will last. That’s why the stock trades at roughly 38x FY’25 earnings and about 25x FY’26. These multiples imply more than strong performance. They imply durable, recurring revenue from AI infrastructure. Revenues for this year are projected to come in at about $215 billion, with the number expected to cross $300 billion next year. Margins are even more eye-popping at about 50% net level, 60% at the operating level and 70% at gross margin level.

Image by Jacek Abramowicz from Pixabay

What’s Driving Results, And What Are The Risks?

AI budgets are exploding. Companies view AI as a generational platform shift. Capex is stretched. Cash burn is being tolerated by investors. Every hyperscaler is racing to build “AI factories” of 10,000 to 100,000 GPUs. Demand for top-end chips has exceeded supply for more than two years. Nvidia sits at the center of all of it. Its chips are fastest. Its interconnect is fastest.

The question is not about today. Today is solved. The question is whether this structure holds once:

• The competition catches up – rivals like  AMD’s (NASDAQ:AMD) are getting more competitive with their GPUs, Cloud computing players are doubling down on custom chips.

• Investors start pressing Nvidia’s largest customers to show real, measurable AI profitability — something still largely missing.

• AI moves from compute heavy model training to much more cost-sensitive inference

Ask yourself – Is holding NVDA stock risky? Of course it is. High Quality Portfolio mitigates that risk.

Nvidia’s Moat: Systems, Not Just Silicon

Many think Nvidia’s moat is the chip. It’s not. The moat is the system.

This matters because modern AI isn’t about a single GPU. It’s about tens of thousands of GPUs acting as one. That requires extreme parallelism. That requires ultra-low-latency connections. And that requires stable, optimized software that scales. Competitors can build fast chips. But building end-to-end systems that scale to frontier-model level clusters is another game Nvidia sells an “AI factory” as a package It sells the silicon, the high-powered GPUs, NVLink/NVSwitch interconnect Networking (InfiniBand), the CUDA Software stack and carries out Cluster-level orchestration.

So, Are There Switching Costs?

For CPUs, switching is fairly trivial. x86, ARM, new instruction sets—software adapts. Compilers catch up. The workloads can migrate.

AI is different. Why?

Training code is written for specific CUDA kernels. Distributed training frameworks depend on Nvidia’s instruction patterns. Memory management is custom. Parallelism behavior is tuned. Model engineers build around Nvidia’s performance quirks.

Switching entails

• Rewriting large parts of the training stack

• Re-optimizing models

• Possibly rebuilding distributed training infrastructure and reconfiguring networking

• Revalidating the reliability of the overall system, at scale

This is expensive. Measured in months of engineering time and tens or hundreds of millions of dollars. For hyperscalers, it’s even higher because clusters run continuously; downtime is revenue lost.

The CUDA System

Does Inference Change Things?

Training has favored Nvidia’s superfast and versatile chips. Inference – which is essentially running the trained models to new data in real time, at scale, will account for the vast majority of AI compute in the long-run. Inference is still GPU-heavy for now because models keep growing, and flexibility matters, but things could very well change as inference becomes the dominant AI workload. Why?

• Costs per chip and margins matter more, given the scale of queries.

• Power efficiency matters more

• Latency matters more

• Cost per token becomes central

The economics could flip in favor of custom chips. Nvidia’s share could drop as hyperscalers build their own ASICs.  Custom silicon from the likes of Alphabet, Meta, and Amazon (NASDAQ:AMZN), could begin to fill their respective server farms sooner than we realize.  See Google’s TPUs Can Trump Nvidia. 

Can Nvidia Sustain This?

Short term: Yes.
Medium term: Likely.

Long term: Nvidia’s lead will very likely narrow.

Inference economics favor specialized silicon, and Nvidia’s biggest customers including Google and Amazon are building just that.  Open-source alternatives to CUDA, from the likes of AMD, will mature. More importantly, hyperscalers will not tolerate 70% to 80% gross margins on their largest AI cost item (GPUs) forever. As models stabilize and workloads become more predictable, the industry will prioritize cost efficiency over peak performance. And once multiple viable ecosystems are gradually developed, the gravitational pull toward a multi-supplier strategy becomes unavoidable.

If Nvidia’s margins compress or there are stronger than expected market share gains by rivals, its not just earnings that could take a hit. Investors will also reassess Nvidia’s earnings multiple, and this could lead to a valuation reset.

The Trefis High Quality (HQ) Portfolio, with a collection of 30 stocks, has a track record of comfortably outperforming its benchmark that includes all 3 – the S&P 500, S&P mid-cap, and Russell 2000 indices. Why is that? As a group, HQ Portfolio stocks provided better returns with less risk versus the benchmark index; less of a roller-coaster ride, as evident in HQ Portfolio performance metrics.

Invest with Trefis Market-Beating Portfolios

See all Trefis Price Estimates