We already know what happens to the GPU hourly price when OpenAI or Anthropic launch a new model: it doubles
This week, an analyst named Tomasz Tunguz published in X two revealing graphs. They show the evolution of what it costs AI startups to access cloud computing, and there is bad news. The cost of renting the NVIDIA B200 GPUs with Blackwell architecture has gone from $2.31 per hour in early March to $4.95 per hour this week. It is an increase of 114% in just six weeks and it has a clear cause: the arrival of new models from Anthropic and OpenAI. What the graphs show clearly. Those charts focus on the price index of Ornna cloud computing trading marketplace. The first of them covers the price of renting the B200 chips from the end of 2025 until today, and there are vertical lines showing each release of the latest models from OpenAI and Anthropic. The correlation is almost perfect: GPT-5 Codex, Claude 4.5, GPT-5.3 Codex, Claude Opus 4.7 and GPT-5.5 coincide with a jump in price indices. Every time these companies announce a new version of their frontier models, demand skyrockets, and so does the cost. If you want the best, pay (much more). The second graph shows the price difference between renting the previous generation of chips, H200 with Hopper architecture, and the new B200. The historical average of that “spread” is $1.06, but now it stands at $2.09, practically double. That means buyers—startups and AI companies—are paying a record premium for the extra memory and superior computing power of Blackwell architecture chips. Accessing the latest of the latest was already expensive. Now it is even more so. This also makes the H200 in a second class option for the most demanding models of 2026. Action and reaction. There is overwhelming logic here. When OpenAI or Anthropic release a new model, there is an explosion in inference. Developers and companies want to test them as soon as possible and integrate these models into their products (or compete with them). To do this, they need computing quickly, and a simultaneous demand is caused that unbalances the available inventory in the market for renting AI chips by the hour. The problem is that the supply of B200 does not grow at the same rate. Some companies have wanted to anticipate, and we have the perfect example in Google. He has bought all the B200s he can, and that has made these GPUs around now the 500,000 dollars on the secondary market according to analyst Jack Minor. The irony of efficiency. The curious thing is that the more efficient these chips are – and the B200s are – the more companies want to rent them at the same time to take advantage of those efficiency advantages that should lead to cost savings. What actually happens is that the scarcity of these advanced chips cancels out any theoretical savings. Long term contracts. Startups and companies that think in the short term are especially harmed in this area, because they face price jumps that are increasingly difficult to assume. Companies that signed computer rental contracts at the price then can now operate at less than half the cost of their competitors. Thinking in the medium or long term seems reasonable, although once again those who win are the hyperscalers and those companies that have managed to get hold of many B200s. And who wins even more is of course NVIDIA, which cannot cope. Few alternatives. In other markets such as energy or metals there is usually room for maneuver, Tunguz points out, but the same is not happening at the moment in the AI segment. In the oil market, for example, if the price rises 114% in six weeks, companies can buy futures, options or fixed-price supply contracts to protect their margins. In cloud computing rental, those options are much more limited. And the result is a much more volatile segment. This will go further. We are probably facing a peak in demand that will be followed by a correction: the new batch of B200 chips that arrive in the second half of 2026 are expected to cause a drop in current prices. However, that $4.95 is now the new floor, not a peak, because demand for AI computing will continue to grow faster than TSMC’s production capacity. In the absence of the supply of AI chips growing significantly – and there are certainly movements that are trying to achieve this, such as those of Google with its TPUsAmazon with its Trainium or Huawei with its Ascend—, the problem will still be there. In Xataka | Europe is taking its technological independence so seriously that it is aiming for the most ambitious goal: NVIDIA
