Breaking

Cheap Chinese Open-Weight Models Squeeze AI Hyperscalers' Pricing Power as 2026 Capex Tops $1 Trillion

September 12, 2026 at 04:07 EDT

Foundation Models
Infra & Chips
Open Source

A wave of inexpensive, high-performing open-weight AI models out of China is forcing investors and enterprises to question whether the world's biggest cloud providers can keep charging premium prices for computing power, even as those same companies pour more than $1 trillion into AI infrastructure this year.

September 2026 · AI Infrastructure

The Open-Weight Squeeze on AI Hyperscalers

Cheap, high-performing open models from China are undercutting API pricing even as the biggest cloud providers pour more than $1 trillion into AI infrastructure this year.

$1T+

2026 hyperscaler AI capex

30%

OpenRouter share for OpenAI, Google and Anthropic models, down from about 70%

$0.14 vs $5.00

Price per million input tokens: DeepSeek V4 Flash vs GPT-5.5

2.8T

Parameters in Moonshot's Kimi K3

Share of OpenRouter activity running on closed US models (OpenAI, Google, Anthropic)

70%

a year ago

30%

now

Price per million input tokens

$0.14

DeepSeek V4 Flash

$5.00

GPT-5.5

GPT-5.5 costs roughly 36 times more per million input tokens than DeepSeek V4 Flash.

Enterprises are switching

AT&T's open-model use rose from 20% to 40%, targeting 60% and up to 80% cost savings
DoorDash, Siemens and Airbnb have shifted workloads toward open models
Latham & Watkins buys its own Nvidia hardware to run open models in-house

Limits to the threat

Cloud-hosted open models still consume data-center capacity
Scaling and electricity assumptions behind displacement estimates remain untested
Pressure falls hardest on AI labs' margins, not hyperscaler infrastructure itself

"Open weight models are closing the capability gap with frontier models at a fraction of the cost."

— Jawad Jahan, senior analyst, Juniper

Six major hyperscalers are projected to spend $1.3 trillion on capex in 2027 — with only one forecast to post positive free cash flow that year.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.