Shenzhen Speed: A Brief Taste of AI from China
- Sharon Gai
- 2 days ago
- 7 min read
There's "China speed," the pace at which platforms get built, scaled, and exported across the Pacific. And then there's "Shenzhen speed," which is fast enough to surprise the Chinese themselves. I just spent a week there at a Prosus tech conference. Beyond participating as one of the keynotes, the bigger takeaway was listening to the Chinese AI landscape unfold with founders speaking on building robots, AR glasses, and frontier models.
The programming was very thoroughly put together, ranging from robotics, to wearables to frontier model companies. Here’s a highlight on some of the most memorable moments.
We opened with, of course, a dance routine from four humanoid robots. People know Unitree well, after making its debut on Amazon.com, but few had heard of Galbot, $3B+ valuation, the highest-valued private humanoid company in China.
Apparently that’s mostly what they are good for these days. (unless working at a factory and lifting boxes) one use case I saw that was interesting was being a practice partner for playing tennis. Clever! Throughout the conference, they strolled around, high-fiving attendees.
This particular one was like a sales assistant, he could grab me different products behind his shelf
I spoke to the Galbot founder later, and asked when he thought they would be able to achieve a humanoid robot in every household. He said it will take a while, and more data needs to be collected for an all purpose robot. He believes the average Chinese consumer will not be willing to pay for just one that does one chore. Pricing will likely be around the price of a small car, and the idea is to have the humanoid robot clean dishes, do laundry and do general cleaning, an all in one butler and so far, the humanoid robots aren’t good enough to do that yet.
During the event, I interviewed one such robot and asked if he’d like to take my job.
He refused to respond.
Chinese LLMs: A Crowded, Open, Cheaper Frontier
One of the presentations was from MiniMax, the Shanghai-based company behind the Talkie chatbot and the Hailuo video model. MiniMax is one of China's "Six Tigers," a group of foundation model startups that also includes Zhipu AI (now Z.ai), Moonshot AI, Baichuan, 01.AI, and StepFun. DeepSeek, despite the headlines, sits outside this group entirely. MiniMax has fared particularly well in the West. Both Zoom and Notion use it under the hood, and if you've ever set up models in OpenRouter, you've likely seen developers vouching for it and for Moonshot's Kimi alongside it.
But the more useful exercise is zooming out. As of April 2026, the Chinese LLM frontier looks like this:
· DeepSeek V4 Pro sits at the top of most Chinese leaderboards. It scores 87 on BenchLM, just one point behind GPT-5.4 at 88, and trails Gemini 3.1 Pro at 93 by six. A year ago that gap was much wider.
· Kimi K2.6, released by Moonshot on April 20, 2026, became the first open-weight model in the world to beat GPT-5.4 on the SWE-Bench Pro coding benchmark. It is a one-trillion parameter MoE model, with 32B active parameters, available on Hugging
· Alibaba's Qwen 3.5 has the broadest lineup. Qwen 3.5 9B, at $0.10 per million input tokens, beats much larger 120B parameter models on graduate-level benchmarks.
Three things are worth flagging about this list.
The first is that almost every model on it is open weight. Zero of the top Western models are. This is the structural divergence I'd watch most closely. American labs are building moats. Chinese labs are giving the model away for soft power influence into the Global South.
The second is price. DeepSeek V3.2 delivers roughly 90% of GPT-5.4 quality at one-fiftieth the cost. According to a RAND analysis cited earlier this year, Chinese models were operating at 16 to 25% of the cost of comparable American models. This is the part that should make Western enterprises pay attention. Cheaper does not mean inferior. It means the unit economics of building AI products on Chinese infrastructure are starting to look very different from building them on American infrastructure.
The third is specialization. There is no single "best" Chinese model the way OpenAI's GPT line tried to be a single best Western model for a few years. For coding, Kimi K2.6. For math and lowest cost, Step 3.5 Flash. For balanced general use, DeepSeek. For long context (one million tokens natively), Qwen 3.6 Plus or MiMo V2.5 Pro. For agentic reasoning, GLM-5.1. Chinese labs appear to have tacitly agreed not to fight each other on every dimension, and the result is that the frontier has multiple peaks instead of one.
The MiniMax presenter ended with a flywheel argument that I keep thinking about. The flywheel goes: model intelligence improves, willingness to pay rises, profit funds more compute, more compute trains a better model, repeat. The piece I'd add is that the Chinese flywheel is missing the "willingness to pay" arc and substituting cost compression in its place. They are not asking enterprises to pay more for smarter intelligence. They are racing to give them comparable intelligence for ninety percent less. Whether that flywheel spins as fast as the Silicon Valley one is the question I'd put on the table.
Will AR Glasses Replace Cel Phones?
Next we heard from Rokid, an AR glasses startup that sells something to Meta Ray Bans, the glasses that light up when you’re recording. I tried it on myself. And Rokid has some basic functions: navigation while on a bike, driving, walking; playing music; reading from a teleprompter. This means next time I do a keynote, I can wear these glasses and just read, not need to remember anything. No no, I probably won’t do that. I’d still like to use my brain to think while I can. Its other unique function is translation. If you were in a meeting or speaking with someone who you couldn’t understand, whatever the other person is saying, is then displayed in front of you in your most comfortable language. I thought that would be pretty useful. Having tried the Meta Ray Ban before, I thought the two were pretty comparable. The price gap, though, is real. Rokid's display glasses retail for $599 internationally, against $799 for the Meta Ray-Ban Display. Rokid’s goal is to replace the cell phone. It’s not that I don’t think it will. It’s more that a TON of more functionalities will need to happen before it does.
We’re in the early stages. The current version of these glasses seem to like the first version of the iPhone, a prototype. As time goes on, there will be more functionalities.
Last, I thought Genspark’s founder had an interesting presentation. His big point was that the future is already here, but not evenly distributed. He fully believes the autonomous organizations is already here, except most people are just still stuck in the chatbot era. He marks $1000/month as the magic spend to be deemed a power user of AI. Less than that, and you’re just an amateur user. I’m not sure I agree there.
The State of AI in China: A Wider View
Stepping back from the Prosus floor for a moment, here's what the data actually says about where China is.
The headline finding from Stanford's 2026 AI Index is that the US-China AI performance gap has effectively closed. As of March 2026, on the Arena leaderboard, the top US model leads the top Chinese model by just 2.7%. The four leading global models (Anthropic, xAI, Google, OpenAI) are separated by fewer than 25 Arena points, and DeepSeek and Alibaba trail only modestly. Six months ago, this would have been a much wider gap. Six months from now, it might not exist.
But the more interesting story is the volume. Between 2022 and 2025, China's total model releases grew from 151 to 849. The US released 50 notable models in 2025, China released 30, and Alibaba alone now sits in the global top three for model output, behind only OpenAI and Google. China has more than 700 generative AI services officially registered, with hundreds of millions of users interacting with them inside super-apps. This is the part Westerners tend to miss. China is not trying to win with one DeepSeek or one Qwen. It is flooding the field.
Then there is deployment, which I think is the most underreported piece of the entire China AI story. Roughly 34% of job functions at Chinese companies are already fully integrated with AI tools, compared with 30% globally. The "AI+" initiative, released by China's State Council in August 2025, sets a target of 70% AI penetration in key sectors by 2027 and 90% by 2030. The 15th Five-Year Plan, passed in March 2026, projects that AI-related industries will exceed 10 trillion yuan, around $1.4 trillion, by 2030. While the US debate is fixated on the race to AGI, China is running a different race: deployment at industrial scale.
Compute and capital still favor the US, and by a lot. American private AI investment hit $285.9B in 2025, more than 23 times China's $12.4B. The US has an estimated 5,427 data centers, more than ten times any other country. But the comparison understates Chinese spending, because state guidance funds (an estimated $184B deployed into AI firms between 2000 and 2023) do not show up in private investment figures. And on industrial robotics, China installed 295,000 industrial robots in 2024 alone, compared to 34,200 in the US. That is the lead I felt walking through the Prosus floor.
The constraint China is genuinely working under is chips. US export controls have cut off access to top-end NVIDIA datacenter GPUs, which is why Chinese model builders have leaned so hard into Mixture-of-Experts architectures, efficiency gains, and domestic accelerators like Huawei Ascend. The phrase I kept hearing from founders in Shenzhen was a version of "do more with less," which, as the title of my book suggests, is a phrase I have some affection for.
My final thoughts
The most groundbreaking, needle-moving new models still come from the US, Mythos, for example. But what the Chinese are great at doing is taking something that is working and streamlining it and making it more accessible. It makes me think of how the US used to be flooded Chinese made goods because China was able to make it cheaper. I wonder if the same will happen in the intelligence era. At first, we are all using US intelligence, but as time goes on, as the Chinese become better and better at reducing cost, eventually some consumers and some enterprises shift to using Chinese intelligence. Or possibly a division of some work on US models, some on Chinese models to save costs.
Today’s target is tomorrow’s baseline. I remember that from my Alibaba days. And if things developed quickly back then, I can’t imagine the speed that China is on now when it comes to AI.
--
Hello! I’m Sharon Gai, author of How to Do More with Less: Future-Proofing in an AI-Driven World a keynote speaker on AI and its effects on workers and the future of work.
Comments