Nvidia Blackwell: Towards a Drop in AI Token Prices

Concerns about the costs of AI tokens are set to ease thanks to a major technological advancement expected this year. A CEO of an AI infrastructure company, who preferred to remain anonymous, revealed that new, more powerful and efficient AI models will emerge. These innovations are expected to make AI tokens more abundant and significantly cheaper. Tokens, which are the basic units used by models to process information, are essential for measuring and pricing AI usage.

However, the reduction in concerns regarding tokenmaxxing could have varied effects. Users might either cut back on their consumption or, conversely, embark on a new frenzy of token consumption. In any case, a price drop seems inevitable. Already, some AI model providers have begun to lower their rates, and others are considering following this trend.

Sam Altman, CEO of OpenAI, recently expressed his concerns about the high costs of AI. He emphasized that the startup is exploring several ways to help users get more value while spending less. This trend is already visible in the data. A token spending index, tracked by Silicon Data, dropped from 2.06 at the end of May to 1.75 on June 10. According to Carmen Li, CEO of Silicon Data, this could indicate a widespread decline in token prices across many AI models.

Blackwell: A Technological Revolution

The drop in token prices is primarily driven by a new wave of technologies in AI data centers. Nvidia's Blackwell GPUs, currently being deployed en masse, are essentially supercomputers that will enable the training and execution of new AI models more efficiently by the end of the year.

These systems required considerable installation time due to their need for water cooling and other complex configurations. Nevertheless, the return on investment could be immense. SemiAnalysis, an AI research firm, compared Nvidia's Blackwell system, the GB 300 NVL72, to the older Hopper HGX 200 system. The new system generates 6,000 tokens per second, which is 65 times more than the old one.

Despite their efficiency, the new Blackwell offerings consume even more electricity than previous systems. In terms of energy consumption, Blackwell produces 2.8 million tokens per megawatt, compared to 54,000 from Hopper, which is 50 times more. Electricity prices are rising, partly due to these energy-hungry AI data centers. The cost of generating one million tokens has dropped from $4.20 with Hopper to just 12 cents with Blackwell, representing a 35-fold reduction.

As 2026 progresses, new AI models will increasingly be trained and executed on these Blackwell systems. This evolution is expected to lead to a massive increase in the number of tokens generated at low cost. AI model providers will then be able to lower their prices, making AI more accessible to a broader audience.