In the past three years, the explosive expansion of AI has propelled Nvidia to a market valuation of $5 trillion, with this year’s net profit even surpassing the combined sales of Intel and AMD. This is not a coincidence, but rather a moat built from technology, products, and ecosystem. The dominance of the Blackwell and Hopper generations of accelerator cards in the market is primarily due to their ability to turn GPUs into infinitely scalable “supercomputer units.” Blackwell’s training performance is 2.5 times better than Hopper, and the chip itself has grown too large to be manufactured as a single unit, requiring two massive crystals to interconnect and form a “logically single” structure. This scale and clustering capability is the most irreplaceable factor in training large models.
However, what truly sets the gap is not performance, but the ecosystem. Nvidia does not just sell chips; it sells NVLink, clusters, software, CUDA, and delivery systems. Hopper and Blackwell enable thousands of GPUs to work as a single machine, backed by over a decade of system engineering accumulation from the gaming GPU era, creating a “system-level lock-in” that AMD and Intel cannot currently catch up to. You can create a GPU with similar performance, but it is challenging to develop an equally mature ecosystem in the same timeframe.
Of course, competitors are not without their moves. AMD’s MI300 series has already secured large orders from OpenAI and Oracle, and the new generation MI450 is set to ramp up production; Intel has taken a more direct approach, temporarily abandoning competition to collaborate with Nvidia, combining efforts in PCs and data centers. However, the reality is that Nvidia still holds a staggering 90% market share in the data center GPU market and has committed to releasing a new flagship every year, meaning competitors are facing a high-speed train that is always accelerating.
Adding to this reality is the fact that demand is far from peaking. Microsoft, Amazon, Google, and Meta have collectively announced multi-billion dollar investments in computing power, and OpenAI is also “madly hoarding cards.” Nvidia even anticipates that data center revenue will reach $500 billion over the next five quarters, forcing analysts to revise their models upward. The only real threat to Nvidia may not come from AMD or Intel, but from geopolitical factors—U.S. bans and restrictions on China have led Nvidia to write down $5.5 billion in inventory this year, while China’s countermeasures have abruptly halted H20’s sales.
Therefore, the underlying logic of this era is quite clear: It is not about who can create a faster chip, but who can provide a complete set of “computing power system engineering” that allows enterprises to train, run, and scale their models. In this regard, Nvidia is still far ahead.