April 30, 2026 – In the era of AI where expenditures often soar into the hundreds of billions of dollars, a significant portion of these costs is funneled into GPUs. NVIDIA’s AI graphics cards have become an indispensable choice for major companies, with the company enjoying a staggering gross profit margin of over 70% from selling these cards.
NVIDIA has emerged as the most profitable “shovel – seller” in the AI gold rush. However, this has prompted tech giants like Google, Amazon, and Microsoft to develop their own chips in an effort to cut costs. So, how does NVIDIA convince these customers to keep buying its graphics cards? Well, after some careful calculations, it turns out that Jensen Huang’s often – repeated mantra holds true: the more NVIDIA cards you buy, the more you actually earn.
NVIDIA conducted a comparison between its latest Blackwell graphics cards and the previous – generation Hopper cards. When it comes to deployment costs, the GB300 NVL72 from the Blackwell series costs 2.65perhour,whiletheH200fromtheHopperseriescosts1.41 per hour, effectively doubling the price. But here’s the catch: the performance per dollar also doubles.
The real difference lies in token generation speed. The GB300 can generate 6,000 tokens per second, while the H200 can only manage 90 tokens per second, a gap of 65 times. In terms of power efficiency, the GB300 can produce 28 million tokens per megawatt (MW), while the H200 can only produce 54,000 tokens per MW, a 50 – fold difference.
After all these calculations, it is found that the cost per million tokens for the GB300 graphics card is 35 times lower than that of the H200 graphics card.
Of course, NVIDIA’s comparisons always come with certain conditions. For instance, the Blackwell graphics cards support the NVFP4 operator and the MTP multi – token prediction technology. The presence or absence of these technologies can create a performance gap of ten to twenty times. But then again, it’s only natural that a new – generation graphics card supports more advanced technologies, so such a comparison might not be entirely out of line.
After all, in the gaming graphics card arena, NVIDIA often uses its more advanced DLSS technology to outperform previous – generation cards by a large margin. If you were to compare the raw rasterization performance, the improvement of flagship cards might only be around 30%, or even less.
