[News] Why GPU Rental Prices Keep Falling — and What It Says About the AI Boom

Is the AI bubble about to burst? The question has sparked increasing concern. According to Financial Times, AI hardware poses a striking paradox — it’s extremely costly to buy but cheap to rent.

The report notes that when NVIDIA’s B200 GPU debuted in late 2024, it likely sold for around $500,000. Yet by early 2025, the same chip could be rented for about $3.20 an hour, with prices sliding further to $2.80 per hour by last month.

NVIDIA refreshes its chip architecture every two years, so one might assume that the best-funded data center operators would lure customers with steep discounts on older hardware.

Yet, as Financial Times notes, that hasn’t been the case. Despite falling GPU rental prices, hyperscalers such as Amazon’s AWS, Microsoft’s Azure, Google, and Oracle have kept their rates largely unchanged—widening the gap between these giants and a growing field of smaller competitors. The report suggests that these new entrants are the ones pushing average rental prices lower.

Why Smaller GPU Providers Are Slashing Rental Prices

The report suggests smaller providers are cutting prices because most GPU-as-a-service users are AI start-ups and research institutions already tied to major hyperscalers. Meanwhile, typical enterprises seeking chatbots or summarization tools rely on ready-made models from OpenAI or Anthropic, paying by the token instead of renting GPUs. As a result, GPU-as-a-service firms are barely breaking even, with few customers left outside the hyperscaler ecosystem.

Overestimated GPU Market and Looming Shakeout

The report highlights that NVIDIA’s A100 GPU, priced at approximately $199,000 at its 2020 debut, would require around $4 per hour in utilization over a five-year lifespan to break even. Average rental rates were about $2.40 per hour at the time, declining to roughly $1.65 today. Yet, the report notes, the figure is distorted by hyperscalers maintaining rates above $4 per hour, even as smaller competitors cut prices to $0.40.

Given this dynamic, the report points out that the size of the GPU market may be overestimated. Moreover, an inevitable wave of data-center consolidation is expected to force many AI start-ups out of the market, as they struggle to afford the real cost of computing power.

TSMC Sees Token Surge Fueling AI Chip Demand

On the other hand, TSMC Chairman and CEO C.C. Wei voiced confidence in AI’s momentum. As noted by TechNews, during the earnings call he said the number of tokens required for AI computation is “almost doubling every three months.” When asked why token growth has outpaced the company’s AI-related revenue CAGR, he explained that as customers move to more advanced nodes, chip performance rises significantly, allowing each generation to process far more tokens. He added that this dynamic supports steady growth in overall wafer demand.

Read more

• AI Data Center Capex Breakdown and Future Outlook
• [News] TSMC Unfazed by NVIDIA GPU Curbs in China, Eyes Mid-40%+ AI Revenue CAGR Through 2029

(Photo credit: FREEPIK)