AI Compute Price Surge: The End of Cheap Training Costs

On April 18, I saw someone complaining in an AI developer group: “A100 rental prices went up again—last year I could negotiate down to $2/hour, now I can’t even get $3.”

This wasn’t an isolated case. I checked GPU prices across several major cloud platforms and found compute price increases are now widespread. AWS p4d instances (A100) are up about 30% from early this year, and one major domestic cloud provider’s V100 instances have jumped nearly 50%.

Several factors are driving these increases. First, supply-demand imbalance. The LLM arms race shows no signs of cooling—everyone’s scrambling for GPUs. Second, geopolitical factors—export controls have raised acquisition costs for high-end GPUs. There’s also an often-overlooked factor: electricity costs. AI data centers consume staggering amounts of power, and global energy prices are broadly rising.

What does this mean for the industry?

The most direct impact: smaller teams are being squeezed out. Training a GPT-3-class model already cost millions of dollars—now it might double. Can’t afford cards or rental fees? You’re stuck relying on big tech APIs, which means losing technical independence.

But viewed another way, this might also be a “bubble-bursting” process. Over the past two years, AI has seen too many “bandwagon” projects—no real technical substance, just PowerPoints and concept pitches to raise funding. With rising compute costs, that game becomes unsustainable. The market will naturally filter for genuinely competitive players.

The rise of companies like DeepSeek is emblematic. Their models match OpenAI’s performance, but training costs are reportedly a fraction of the price. How? Architecture optimization, engineering prowess, data efficiency—these “soft skills” become core competitive advantages in an era of scarce compute.

For ordinary developers, my advice: first, prioritize inference optimization over training from scratch; second, pay attention to emerging domestic chips and edge computing solutions; third, learn to do more with less—this may be an essential skill for the coming years.

The era of cheap compute may truly be over.