The End of AI's Price Inversion: Alibaba and Baidu Raise Compute Prices 34%

Last week I saw a piece of news that didn’t initially catch my attention — Alibaba Cloud and Baidu Intelligence Cloud both announced price hikes for AI compute services.

This week, chatting with friends who build AI applications, I found everyone was doing the math: Is the assumption that “AI models keep getting cheaper” still valid?

What’s the “Price Inversion”?

Let me break this down.

Between 2024 and 2025, the AI industry had a peculiar phenomenon: model prices were falling, but compute costs were rising. Using large model APIs became incredibly cheap. But the underlying GPU compute? NVIDIA H100 rental prices surged nearly 100% in 2024.

This is the “price inversion” — upstream expensive, downstream cheap, middlemen subsidizing everything.

Why Did the Inversion End?

Several factors converged:

First, too many models, too much competition. Domestic LLM vendors grew from dozens in 2024 to over a hundred in 2025. API price wars became ruthless. Nobody was profitable, everyone was fighting for market share.

Second, GPU costs didn’t drop. NVIDIA cards remain expensive, domestic alternatives are still climbing, and data center electricity and maintenance costs keep rising.

Third, investor appetites cooled. Starting late 2025, VC enthusiasm for AI dampened. The subsidy-based market-grabbing playbooks stopped working.

So Alibaba raising prices 34% on April 18th, Baidu 30% — this isn’t coincidental. It’s the entire industry returning to rationality.

DeepSeek V4 Leaks Hit the Same Day

Interestingly, DeepSeek V4 architecture details leaked simultaneously — Mega MoE architecture, 1.6 trillion parameters, active expert count jumping from V3’s 256 to far more.

Two things are happening at once: underlying model capability continues evolving, while compute costs rise. What equilibrium price will these opposing forces create?

My Take

AI’s “cheap era” may be pausing. The old internet playbook of “burn cash for scale, raise prices once dominant” doesn’t work in AI.

But this isn’t necessarily bad. Prices returning to rationality means the industry shifting from speculation back to business fundamentals. AI applications that genuinely reduce costs and increase efficiency will survive; those surviving on subsidies will be eliminated.

For us ordinary developers, compute costs may need to become a serious factor in technical decisions going forward.