AI companies aren't raising prices for users while Nvidia and Micron book windfall profits

Makers of AI chips and memory are posting record profits on the generative AI boom. But the companies that build the models themselves are in no hurry to pass rising infrastructure costs on to end users, and that is changing the pricing logic of the entire industry.

Windfall profits for hardware suppliers

For several quarters in a row, Nvidia has posted margins that are unusual even by the standards of the semiconductor industry's best periods. Memory makers have now joined it. Micron Technology is reporting comparable results amid frenzied demand for HBM (High Bandwidth Memory), specialized memory for AI accelerators.

SK Hynix, Samsung and Micron together cannot expand HBM production faster than demand from data centers is growing. Analysts forecast that the shortage will last until at least the end of 2026. That means component prices will stay high and suppliers' margins will keep rising.

• HBM3E is a key component for next-generation GPUs (Nvidia H200, B200)
• HBM prices have more than doubled over the past two years
• Production capacity is expanding, but it still lags demand
• The shortage is forecast to last through all of 2025–2026
• New players (Broadcom, Marvell) are entering the segment but have not yet changed the situation

Investors, not users, are closing the gap

Despite rising costs, major AI model developers are generally holding or cutting the price of API access. Competition at this level is so intense that openly passing costs on to customers is risky: a company would simply lose market share to a cheaper rival.

Infrastructure costs are currently covered by three factors. The first is optimization of the models themselves: more efficient architectures lower the cost of inference. The second is economies of scale: the more requests run through a single cluster, the lower the unit cost. The third is venture subsidies: investors are putting in billions in expectation of future monetization, covering current losses.

OpenAI, by analysts' estimates, still operates at a loss at the operating-cost level. Anthropic receives investment from Amazon and Google that effectively subsidizes developers' access to its models. This creates artificially low prices in the market that do not reflect the real cost of compute.

Who ends up paying

The AI industry's value chain works paradoxically today: hardware makers make money, model developers run at a loss or near zero, and end users get services at subsidized prices. The question is how long this can continue.

If the market consolidates and only 2–3 large players survive, the price war will subside and the cost of services will start to rise. If a technological breakthrough arrives, costs will fall naturally. If large corporate customers start moving to their own models and infrastructure, that would shift the entire balance of power. For now, venture capital is paying for the "banquet," but its patience is not infinite.

What this means

The current pricing situation benefits everyone building AI-based products right now: API costs are historically low relative to the real cost of compute. For businesses this is a window of opportunity: while the market subsidizes the infrastructure, it makes sense to build up AI expertise and AI-based products as much as possible. A repricing is inevitable; the only question is when.