NVIDIA's Contested Reign
NVIDIA has ruled the AI chip market since the beginning of the deep learning revolution. In 2026, Jensen Huang's company still controls over 80% of the datacenter GPU market. But this dominance, long uncontested, faces coordinated attacks from all fronts.
The current situation recalls Intel in the 2000s: an apparent monopoly position masking structural cracks. Customers—hyperscalers, AI companies, governments—desperately seek alternatives. Not from technical dissatisfaction, but from strategic necessity: depending on a single supplier has become an unacceptable risk.
Established Challengers
AMD has made a remarkable comeback with the MI300 series. Its chips offer performance comparable to NVIDIA's H100 for certain workloads, at a lower price. Microsoft and Meta have publicly adopted AMD for part of their infrastructure, legitimizing the alternative. ROCm software, long AMD's Achilles heel, has improved considerably.
Intel struggles more. Gaudi 3, its dedicated AI chip, exists but hasn't found its market. The company is betting on a foundry strategy, manufacturing others' chips rather than its own. A relative admission of failure on the design front.
Google with its TPU v5 remains a major force, but these chips are only accessible via Google Cloud. It's both a strength—perfect integration with the ecosystem—and a limitation—no direct purchase possible for companies wanting their own infrastructure.
The Startup Wave
The real danger for NVIDIA may come from specialized startups:
Cerebras offers wafer-scale chips, giant chips covering an entire wafer. For training very large models, the approach offers unique advantages in memory bandwidth.
Groq bets on ultra-fast inference. Its Language Processing Units (LPU) display record latencies, ideal for real-time applications. The 2024 public demonstration created a sensation.
SambaNova targets enterprises with a full-stack approach, combining hardware and software in an integrated offering.
Graphcore (acquired by SoftBank) continues developing its IPUs, with particular focus on innovative architectures like sparse transformers.
The Geopolitical Factor
The chip war is also a geopolitical war. American export restrictions to China have created a parallel market, where local players like Huawei (with its Ascend chips) try to fill the void.
Taiwan, where TSMC manufactures nearly all advanced chips, remains the major tension point. A disruption of Taiwanese production would paralyze the global industry. This dependency pushes the United States and Europe to invest massively in local production capabilities, but results won't be visible until the end of the decade.
Technical Trends
Several technical innovations are redefining the market:
Advanced packaging: integrating multiple chips in a single package (chiplets) allows bypassing Moore's Law limits. NVIDIA uses it in Blackwell, AMD in MI300.
High bandwidth memory (HBM): the bottleneck is no longer compute but memory. HBM3e and soon HBM4 become critical differentiators.
Optimized inference: after the training race, the industry focuses on efficient inference. Specialized architectures for this use case gain importance.
Quantization: running models in reduced precision (FP8, INT4) without significant quality loss allows using less powerful chips.
Stakes for Enterprises
For AI-using enterprises, this boiling market is an opportunity. Prices drop under competitive pressure. Options multiply. Negotiating power against NVIDIA increases.
But complexity increases too. Choosing the right architecture for each workload requires growing expertise. Software lock-in risk (CUDA at NVIDIA) remains a major decision factor. Multi-platform optimization becomes a sought-after skill.
Conclusion
NVIDIA remains the undisputed king, but the crown wobbles. By 2028, the market will probably be more balanced, with 2-3 viable alternatives for each use case. Competition will benefit everyone: lower prices, accelerated innovation, less dependence on a single player. For AI chip buyers, it's the best moment in a decade.