Everyone talks about NVIDIA's chips like they're the moat. They're not. The moat is software, and it's twenty years deep.
CUDA launched in 2006. Nobody cared for a decade. Then deep learning caught fire, and every researcher on Earth had already spent years writing CUDA kernels. That early bet compounded into something no competitor can replicate with a faster chip or a fancier interconnect. Four million developers. Forty thousand organizations. Two decades of production code that runs inference at scale every single day.
The financial output is staggering. FY2026 revenue hit $215.9 billion, up 65% year-over-year. For context, FY2024 was $60.9B. FY2025 was $130.5B. This isn't linear growth — it's a staircase, and NVIDIA keeps climbing. First-quarter FY27 guidance sits at $78 billion, a number that would've seemed unhinged two years ago. The stock trades at $197.58 with a market cap of $4.79 trillion and a forward P/E of 15.48x, which is absurdly cheap for that kind of growth. Gross margins hold at 74%. Operating margins at 65.6%. These are monopoly margins, and they're earned by software lock-in, not silicon pricing power alone.
Data Center makes up 91.5% of revenue. That's the concentration risk. A single line item carries the entire thesis. If AI capex slows, if hyperscalers finally succeed at vertical integration, if an inference revolution moves workloads off GPUs — the revenue cliff is steep. NVIDIA is betting everything on AI staying a capital-intensive buildout war. So far, they've been right.
But the challengers aren't imaginary. AMD's MI400X, expected around 2028, targets 10x the performance of the MI300X inside the Helios rack with EPYC Venice. It's a serious play if AMD can ship it on time and, more importantly, if they can get the software stack functional enough that enterprise shops feel safe migrating. That second part is where AMD has historically failed. Building a chip is an engineering problem. Building a developer ecosystem is a cultural problem, and culture moves slower than silicon.
The hyperscalers are the real long-term threat. Google's TPU v7 Ironwood delivers 4,614 TFLOP/s in 256-chip clusters and is engineered almost exclusively for inference, where Google controls the full stack from data center to model to user. Amazon's Trainium2 is already shipping and powers Anthropic's Project Rainier, the training cluster behind Claude. These aren't hobby projects. They're strategic weapons designed to reduce dependence on NVIDIA's pricing power.
Here's the thing, though — they're all attacking inference. Inference is where CUDA is weakest, where the software lock-in is shallowest, where customers are most willing to experiment with alternatives. Google's TPUs shine at running models, not training them. Amazon's Trainium is optimized for specific workloads, not general-purpose GPU compute. NVIDIA's real fortress is training, and training is where the developer habit is deepest.
That habit is the flywheel. Every new framework, every new library, every new optimization lands on CUDA first. PyTorch, TensorFlow, JAX — all tuned for NVIDIA silicon. The ecosystem grows, the talent pool grows, the tooling grows, and switching becomes exponentially harder. When a startup hires three ML engineers, all three already know CUDA. That's not lock-in imposed by a vendor. That's lock-in imposed by the labor market.
Can a motivated team port a model to alternative hardware? Sure. Will it cost them six months of engineering time, three weeks of debugging obscure memory bugs, and a rewrite of half their data pipeline? Also yes. And that's the moat — not the silicon, not the architecture, but the cost of breaking twenty years of muscle memory across four million developers.
The bull case is that switching cost compounds faster than silicon commoditization erodes it. The bear case is that inference, which is growing faster than training, is less defensible because workloads are smaller, models are more portable, and hyperscalers control their own destiny. Forward PE at 15.48x suggests the market already prices in some deceleration. The question isn't whether growth slows — it will — but whether it slows gracefully into a durable platform business or crashes into a wall of well-funded ASICs.
There's also a subtle dynamic worth watching. Every time a hyperscaler announces a custom chip, NVIDIA responds with tighter integration, better libraries, more developer perks. AMD ships MI300X, NVIDIA drops a CUDA update. Google unveils a new TPU, NVIDIA announces Blackwell Ultra. The rhythm is predictable and it's working. Revenue tripled from FY2024 to FY2026 while competition heated up. That's not coincidence — that's the flywheel doing its thing.
For now, NVIDIA owns the infrastructure layer of AI, the way Microsoft owned the desktop OS layer in 1998 or Google owned search in 2010. Generational assets don't get built every year. Just don't confuse ownership with permanence.
Disclosure: The Signal holds no position in NVDA. Positions may change. This is not financial advice.
