When Jensen Huang took to the stage at GTC 2026 last month and unveiled what Nvidia is calling the Blackwell Ultra architecture, he did so with the quiet confidence of a man who has seen the future and spent three decades building towards it. The new chip, promising 15 petaflops of FP8 performance in a single GPU, represents the latest salvo in what has become the most consequential industrial competition of our era.
The artificial intelligence chip market, valued at $67 billion in 2025, is projected to reach $340 billion by 2030. These are numbers that attract serious attention — and serious capital. Every major technology company, sovereign wealth fund, and venture partnership on the planet is now directing resources toward the question of who will manufacture the silicon that powers the next generation of intelligent systems.
At the moment, the answer remains Nvidia. The Santa Clara company controls an estimated 80% of the market for AI training chips — a dominance so complete that it has drawn scrutiny from competition authorities in the United States, the European Union, and China simultaneously. CUDA, Nvidia's proprietary software ecosystem, is the invisible moat that has proven more durable than any hardware advantage: switching away from it requires rewriting vast quantities of code and retraining the engineers who depend on it.
Yet AMD is mounting a challenge more credible than any it has previously managed. Under Lisa Su's leadership, the company has transformed from a perpetual also-ran into a genuinely competitive force. Its MI350 accelerator has found significant traction in inference workloads — the process of running trained models at scale — where its price-to-performance characteristics offer a compelling alternative to Nvidia's premium pricing.
The more disruptive threat, however, may come from an unexpected direction: the hyperscalers themselves. Google's TPU program, now in its sixth generation, has quietly accumulated a performance lead in specific workloads that Nvidia's general-purpose architecture cannot match. Amazon's Trainium and Inferentia chips are finding increasing internal adoption across AWS. And Microsoft, through its investment in OpenAI and its own silicon programme, is developing capabilities that could eventually reduce its dependence on external suppliers.