ChinAI #321: Shallow, Narrow, and Slow — the reality of China's DeepSeek adoption

Greetings from a world where…

you don’t want to “buy the wooden box and return the pearls” [买椟还珠]

…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).

Feature Translation: The latest notes on DeepSeek’s all-in-one machine

Context: Have you heard the latest? DeepSeek has spread fast and wide across China! As the Financial Times reported in February, “DeepSeek's advances have sparked a nationwide push in China to deploy its large language models everywhere from hospitals to local governments.” Lea Thome, for China Brief, cites a Chinese think tank report that boasts: “72 provincial and city governments had fully deployed DeepSeek as of early March.” And don’t forget about the Chinese military! Sources say they’ve also become early deployers of DeepSeek. Put it all together, and now we get LinkedIn posts that extol “how China has rapidly scaled a homegrown generative AI nationwide.”

This week’s feature translation (link to original Leiphone article) provides a reality check. Has anyone stopped to ask a very simple question: How do Chinese state-owned enterprises, local governments, and military actors actually implement large language models like DeepSeek? Hint: it’s not through software as a service. Instead, they buy a turnkey appliance called an “all-in-one machine” [一体机] equipped with DeepSeek (e.g., H3C’s LinSeer Cube¹ or Inspur subsidiary IEIT’s product pictured below). As Erica Zhao reports for Leiphone, the market for these all-in-one machines has significantly cooled.

This IEIT all-in-one machine supports DeepSeek’s 671B “full-parameter” model

Key Takeaways: Since the frenzy of reports about fast and wide DeepSeek deployment, these all-in-one machines have not diffused past early adopters and attracted few repeat customers.

• From the article: “Leiphone asked many industry insiders, and the answers were all negative. Only one leading complete-system vendor had repeat sales — repeat purchases for two machines from a research institute. ‘Everyone is wallowing about in the pit where the first batch of users landed, and new users are more cautious,’ lamented Jianing, a complete-system vendor sales representative.”

• Yang Fei, an all-in-one machine salesperson, tells Leiphone that the demand for large-scale deployment has faded. As Zhao writes, “At the beginning of March, everyone was still complaining about the low conversion rate of business opportunities, but since April, even the business opportunities of ‘only consulting, no orders’ have dropped from hundreds of opportunities a day to more than ten. In various WeChat groups, it became rare to see users asking and discussing all-in-one machines...”

What happened? Let’s go through three factors in turn: 1) the gap between press release and actual adaptation; 2) barriers to fine-tuning with local data; 3) disregard for ROI amidst initial deployments

1. Deployment takes time! It can take months to conduct reliability testing, perform on-site debugging, and educate clients. As Zhao writes, “In the overwhelming and bustling news of ‘xx all-in-one machine and xx application have completed adaptation’ some time ago, there is also a lot of watered-down content. This is a common trial-and-error strategy in the industry - after reaching a deal, advertise first…and then truly adapt, customize development and deploy according to user needs.” Please, please, please: Don’t take fluffy announcements at face value.

2. Deployment is also very difficult, especially when it comes to integrating application-specific data. Consider the case of hospitals. All-in-one machine vendors recounted that they would often be called to meetings with hospital leaders and receive no questions after their demonstrations. From the article: “(all-in-one machines deployed in) large hospitals may not be able to support enough concurrent users; and outmoded/backward informatization in small hospitals may become a barrier to the entry of all-in-one machines.”

3. The early adopters — state-owned enterprises and local governments — were just just trying to check a box, so they could brag that they had adopted DeepSeek. They didn’t take into account whether a full-parameter or distilled version of DeepSeek was better for ROI. Also, they didn’t pay much attention to the software layer. Put simply, this is not the “fast-acting, seemingly almost biological process” of diffusion that best fits general-purpose technologies, as I have argued in my research.

   1. Let’s get even more granular here. If you look at the vendors that sold the most all-in-one machines (e.g., H3C and Inspur subsidiary IEIT Systems), it’s the established leaders with guanxi connections to the SOEs, financial institutions, and other large buyers. Zhao writes, “The crucial moat in the all-in-one machine business that the market struggled to figure out in the first half of the year turned out to be customer relationships.”

   2. Furthermore, why is DeepSeek being spread through a combined software-hardware package, instead of sold as a software service? One reason is that the Chinese market doesn’t have strong consciousness in paying for software (partly due to prevalence of software piracy in the 2000s).

If all-in-one machines do serve as a sustainable route of LLM diffusion in China, it will be a slow and steady path.

• Zhao emphasizes, “There are also real implementation cases of all-in-one machines, but this type of implementation requires more specific industry models and local deployment combined with user data.”

• Huang Zhong, a company leader who has deeply engaged with all-in-one machines over the past two years, points out that for many application sectors, you need a multi-modal configuration of large models and many different small models good at different tasks: “One Deep Seek is obviously not enough.”

FULL TRANSLATION: The latest notes on DeepSeek’s all-in-one machine: the full-parameter version is not selling well, and the market is grabbing scattered low- and mid-end models

ChinAI Links (Four to Forward)

Must-listen: The Audiobook version of Technology and the Rise of Great Powers is out!

A few of you have told me (in a kind way, I hope) that you read my book before bed. Now, to wind down, you can listen to it too.

Should-read: You don’t have to be America or China to win in AI, says Rishi Sunak

Former Prime Minister Rishi Sunak penned clearsighted op-ed in The Economist that used GPT diffusion theory as a guidepost for the UK’s AI strategy.

Should-read: Why are there no massive Chinese SaaS companies?

In getting a better sense of why the Chinese market is more averse to paying for software as a service, I learned a lot from Lillian Li’s Substack post.

Should-read: The Summer of Free AI-gency

By Katie Baker for The Ringer, this was a fun crossover between two of my main interests: AI talent and NBA free agency.

Thank you for reading and engaging.

These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.

Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).

Also! Listen to narrations of the ChinAI Newsletter in podcast format here.

1. For example, H3C markets its LinSeer Cube DR4000 as a “LLM AI appliance that delivers turnkey full-stack deployment, domain-specific deep customization, and end-to-end full-lifecycle services.” https://www.h3c.com/en/d_202506/2541937_294554_0.htm