Deepseek Is Cheap. People Like Cheap
October 1, 2025
![green-dino_thumb_thumb[1]](https://www.arnoldit.com/wordpress/wp-content/uploads/2025/09/green-dino_thumb_thumb1.gif "green-dino_thumb_thumb[1]")
This essay is the work of a dumb dinobaby. No smart software required.
I read “Deepseek Has ‘Cracked’ Cheap Long Context for LLMs With Its New Model.” (I wanted to insert “allegedly” into the headline, but I refrained. Just stick it in via your imagination.) The operative word is “cheap.” Why do companies use engineers in countries like India? The employees cost less. Cheap wins out over someone who lives in the US. The same logic applies to smart software; specifically, large language models.
Cheap wins if the product is good enough. Thanks, ChatGPT. Good enough.
According to the cited article:
The Deepseek team cracked cheap long context for LLMs: a ~3.5x cheaper prefill and ~10x cheaper decode at 128k context at inference with the same quality …. API pricing has been cut by 50%. Deepseek has reduced input costs from $0.07 to $0.028 per 1M tokens for cache hits and from $0.56 to $0.28 for cache misses, while output costs have dropped from $1.68 to $0.42.
Let’s assume that the data presented are spot on. The Deepseek approach suggests:
- Less load on backend systems
- Lower operating costs allow the outfit to cut costs to licensee or user
- A focused thrust at US-based large language model outfits.
The US AI giants focus on building and spending. Deepseek (probably influenced to some degree by guidance from Chinese government officials) is pushing the cheap angle. Cheap has worked for China’s manufacturing sector, and it may be a viable tool to use against the incredibly expensive money burning U S large language model outfits.
Can the US AI outfits emulate the Chinese cheap tactic. Sure, but the US firms have to overcome several hurdles:
- Current money burning approach to LLMs and smart software
- The apparent diminishing returns with each new “innovation”. Buying a product from within ChatGPT sounds great but is it?
- The lack of home grown AI talent exists and some visa uncertainty is a bit like a stuck emergency brake.
Net net: Cheap works. For the US to deliver cheap, the business models which involved tossing bundles of cash into the data centers’ furnaces may have to be fine tuned. The growth at all costs approach popular among some US AI outfits has to deliver revenue, not taking money from one pocket and putting it in another.
Stephen E Arnold, October 1, 2025
Comments
Got something to say?
-
- Subscribe to Beyond Search
Feature archive
News archive
-
