NVIDIA’s H100 GPU, released in 2023, became the de facto standard for training large language models. By 2025, major AI companies controlled anywhere from 100,000 to 1 million H100 units. The computational bottleneck in AI wasn’t algorithms—it was H100 availability. The chip shortage was so severe that startups were negotiating multi-year prepayments to secure allocation. NVIDIA’s revenue skyrocketed. By early 2026, the company’s market cap exceeded $2 trillion, making it more valuable than oil giants and pharmaceutical companies. And everyone was trying to break its hold.
Why the Monopoly Formed
NVIDIA didn’t set out to dominate AI; it was first-mover advantage combined with high switching costs. CUDA—NVIDIA’s programming model—is the default for ML frameworks. PyTorch defaults to CUDA. TensorFlow defaults to CUDA. Moving off CUDA isn’t switching chips; it’s rewriting enterprise software.
The company’s engineering was excellent. The H100 was genuinely the best chip for LLM training. But dominance turned dominance into monopoly pricing. H100s that cost $12,000-15,000 retail were selling secondhand for $35,000-40,000. Startups were sometimes paying $100,000+ per chip to secure allocation.
The Escape Routes
- Google TPUs and custom silicon. Google, Meta, and Tesla have all built custom AI chips. These are designed specifically for their workloads and offer better performance per dollar at massive scale. But they’re not general-purpose or available commercially.
- AMD MI300X. AMD’s chip is competitive and costs less. Major adoption hasn’t come yet, but 2026 is seeing real traction. Companies building new capacity are increasingly considering AMD.
- Startup alternatives. Cerebras, Graphcore, and newer entrants are designing chips optimized for inference or specific model architectures. These won’t dethrone NVIDIA for general-purpose training. But they erode use-case-by-use-case.
- Physics fundamentals. The speed of light and thermodynamic efficiency limits eventually dictate that semiconductor gains follow Moore’s Law. NVIDIA won’t maintain 80% gross margins indefinitely.
What This Means for 2026
NVIDIA’s grip is loosening. Not dramatically—the company will remain a powerhouse. But the shortage is easing. Prices are declining. Alternatives are becoming viable. The company’s valuation reflects maximum hype. In a few years, the narrative will shift from “who can break NVIDIA’s monopoly” to “why did NVIDIA’s margin compress.” Both will be true. Monopolies that are visible and grievous attract capital and talent toward disruption. NVIDIA built the most important component in AI. It might also have built the target on its back.