![Weekly Dose of China Tech [9.7.2026]](https://substackcdn.com/image/fetch/$s_!I9Cq!,w_1200,h_675,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F5f011b0b-5f75-4840-ba61-bddd8f98958b_1600x840.png)
Hi friends,
Hope things have been good since we last talked.
A lot of this week’s tech story happened below the software layer.
The headlines are still filled with models and agents, but the more important competition is increasingly taking place underneath them: electricity, memory, chips, data centers, and the financing needed to build all of it.
Ulanqab is a good example. China is building AI infrastructure at a scale where the problem is no longer simply finding enough accelerators. It is finding enough power to keep them running, and designing data centers that can work around the grid itself.
Memory is becoming another bottleneck. CXMT is putting its new LPDDR6 into Xiaomi’s flagship foldable, but the bigger opportunity may not be smartphones at all. If AI keeps pushing bandwidth requirements higher, the same technology could eventually find its way into PCs, cars, robots, and servers.
And underneath all of this is a familiar question: who is willing to spend first, before the economics are fully proven?
That question runs through everything from Enflame’s IPO and ByteDance’s new debt to Z.ai’s rapid revenue growth and China’s increasingly aggressive robotics buildout.
The AI race is becoming less about the model on the screen.
It is increasingly about everything that makes the model run.
Anyway, let’s take a look.
This Week Features... #
Ulanqab: China’s Bet That AI’s Real Bottleneck Isn’t Chips, It’s the Grid
In Ulanqab, Inner Mongolia, one of China’s fastest-growing computing hubs, dozens of new data-center projects are being built alongside massive wind and solar resources. The region’s planned and operating capacity jumped from 3.3GW to 12.5GW in less than a year, while facilities such as Star River are being designed around their own generation, storage, and power-management systems rather than simply waiting for the public grid to catch up.
The US is running into the same problem from the other direction: enormous AI demand colliding with a grid that was never built for it. China’s answer has increasingly been to build energy and compute infrastructure together.
That does not make Ulanqab an automatic winner. Capacity is expanding faster than utilization, and the economics of “green” power are more complicated than the headlines suggest.
This week’s feature looks at why the next AI bottleneck may be electricity rather than silicon — and why the companies that benefit most may be the ones selling power, batteries, cooling, and networking rather than GPUs.
CXMT’s LPDDR6 Bet Is Bigger Than a Smartphone
CXMT is making headlines with the arrival of its LPDDR6 memory in Xiaomi’s new flagship foldable, but the phone may just be the first test bench for a much larger ambition.
The key opportunity is bandwidth. Faster memory is becoming increasingly important for on-device AI, where performance can be limited by how quickly a processor can move data rather than by the processor itself.
CXMT is therefore trying to get its LPDDR6 into production early, using Xiaomi’s own XRING platform to work through the qualification process. If it succeeds, the same technology could eventually move from phones into PCs, cars, robots, and AI servers as those markets look for more bandwidth-efficient memory.
The interesting part is that CXMT does not need to beat Samsung or SK hynix on every benchmark immediately. It needs to establish customers, prove yields, and build enough production capacity before AI-driven memory demand moves into a much larger market.
This week’s feature looks at why a memory chip designed for a phone could become part of China’s broader strategy to build an AI hardware stack at home, why memory may become one of the next major bottlenecks in AI.
The News… #
(I) Z.ai Is Growing Fast and Still Bleeding Cash
Z.ai’s first-half revenue jumped 400% year-on-year to RMB 954 million ($142 million), driven almost entirely by a shift from on-premise deployments to cloud APIs and subscriptions: cloud-based revenue alone grew 2,736% to RMB 825 million ($123 million), or 86.5% of the total. Gross profit rose 164% to RMB 252 million ($37.4 million), but gross margin fell from 50% to 26.4% as cloud inference racked up far higher compute costs. The company posted a $308 million net loss for the six months, with R&D spending up 33.6% to RMB 2.13 billion ($317 million).
The growth underneath is real: ARR hit $1.6 billion by the end of August, twice domestic rival MiniMax’s $800 million, on more than 7.4 million registered users. But the margin compression is the more important number. Moving from custom deployments to cloud subscriptions is supposed to be the path to scale. So far for Z.ai, it’s also the path to a bigger bill.
(II) JD Wants to Build the “4S Dealership” Network for Robots
JD.com is building RoboBase facilities designed to handle almost everything a robot needs after it leaves the factory: sales, delivery, maintenance, assembly, R&D, and real-world data collection. Projects are already underway in Guangzhou, Shanghai, and Wuxi, with 80 bases planned across China over the next five years.
The bet fits JD’s existing strengths better than most companies chasing humanoid robots. It already runs a nationwide logistics and warehouse network with more than 900,000 employees, and the logic here is that once robots reach mass production, the bottleneck won’t be building them. It’ll be delivering, maintaining, and continuously improving millions of machines already out in the field, work that looks a lot more like running a dealership network than running a factory.
(III) China’s State AI Fund Buys a Sliver of Kuaishou’s Kling
China’s National AI Industry Investment Fund agreed on August 31 to invest roughly $206.5 million into Kuaishou’s Beijing-based Kling AI unit, taking about 1.14% of Kling’s enlarged registered capital and securing repurchase rights as part of the deal.
It’s a small stake by percentage, but notable for who’s writing the check. A national-level fund putting state money directly into one of China’s leading video-generation models, rather than into chips or data centers, is a sign of how far up the AI stack government capital is now willing to go.
(IV) You Can Now Buy AI Credits on Taobao Like a Phone Plan
Z.ai opened a flagship store on Tmall on September 2, letting consumers buy its Coding Plan subscriptions, Lite, Pro, and Max, on monthly, quarterly, or annual terms, alongside AI hardware including computers with built-in agents and devices running Alibaba’s Qwen models.
It’s a small move, but it points at something bigger. Instead of topping up credits through a developer platform, Chinese consumers can now buy AI subscriptions the same way they’d buy a phone plan or a streaming service, discounts included during shopping festivals. Getting AI onto Tmall next to everything else people already buy is a real step toward turning it into a mass-market product rather than a developer tool.
(V) Enflame Prices Its IPO Well Below the Market’s Recent Enthusiasm
Enflame priced its STAR Market IPO at about $20.97 per share, targeting $902 million in gross proceeds. The company builds cloud AI chips for data centers, with Tencent alone accounting for 83.79% of 2025 sales, and remains unprofitable, pricing at a 61.8x diluted 2025 price-to-sales ratio, below the peer average.
That peer average is the interesting part. Moore Threads gained 425.46% on its debut, MetaX gained 692.95%, and Biren gained 75.82%. After watching Unitree’s stock get pulled 45% below its post-IPO high on nothing but STAR Market trading mechanics a few weeks back, Enflame’s more conservative pricing looks less like caution and more like a company trying not to become the next name investors point to when the pricing anchor swings the other way.
(VI) A $399 Open-Source Robot Just Became a Showcase for Chinese Parts
Hugging Face’s $399 Microduck, an open-source desktop biped meant to let any developer run reinforcement learning on physical hardware, has drawn more than 10,000 pre-orders against a 20,000-unit target. Its controller runs on Rockchip’s RK3566 processor, and Unitree is offering its S288 brushless digital servo as a substitute for South Korea’s ROBOTIS XL330.
If Microduck hits its sales target, it becomes more than a hobbyist robot. It’s a low-cost, high-volume proof point that Chinese robotics components can slot into a Western open-source project as the default choice, not just the cheap alternative.
(VII) Huawei and Xiaomi Are Launching Foldables Hours Apart
Huawei will unveil the Mate XT 2 at 2:30 p.m. on September 7, followed by Xiaomi’s 18 Fold at 7 p.m. the same day. Both phones are expected to run in-house processors, and Xiaomi has confirmed its XRING O3 will pair with CXMT’s newly mass-produced LPDDR6, making the 18 Fold the first commercial smartphone platform built on it.
The timing looks less like coincidence and more like two companies racing to own the foldable conversation before Apple ships one of its own. For CXMT, the stakes are higher than a launch-day headline: the 18 Fold is the real-world test bed for whether its LPDDR6 gambit, and the AI-server ambitions riding on it, actually holds up in a shipping product.
(VIII) Tencent’s WorkBuddy Is Pulling Away, For Now
Tencent used its latest ecosystem event to announce that WorkBuddy will connect to microphones, AI glasses, and AI recording cards, turning spoken instructions directly into completed tasks, alongside a new financial-services edition for insurers, asset managers, and banks. Industry users can also package their own workflows into shareable “Skills.”
QuestMobile put WorkBuddy’s July monthly active users at 6.58 million, against 230,000 for Alibaba’s Qodework, a lead of more than 25 times. That’s a real head start in China’s AI-office race, though a lead built on being first to market is usually the one most vulnerable to a fast follower with deeper pockets.
(IX) ByteDance Just Made Borrowing $30 Billion Look Easy
ByteDance has reportedly expanded a planned dollar loan from $20 billion to $29.6 billion after unusually strong demand from banks, a deal that would rank as Asia’s second-largest dollar loan this year, though it remains unsigned. The proceeds are earmarked mainly for general corporate purposes, separate from the up to $70 billion in capital spending ByteDance is reportedly considering this year for AI infrastructure.
The pricing is the tell: the loan’s opening margin is 68 basis points over SOFR, down from 85 basis points on ByteDance’s previous offshore loan. Borrowing more, and getting a better rate for it, isn’t something banks do for a company they’re worried about.
(X) Moonshot AI Quietly Files for a Hong Kong IPO
Moonshot AI, the company behind Kimi, has confidentially filed an A1 form with the Hong Kong Stock Exchange, officially kicking off its IPO process, while simultaneously raising what’s likely its final pre-IPO round at a $50 billion pre-money valuation. That’s roughly eight times where the company stood less than a year ago: Kimi was valued at about $4.3 billion at the end of 2025, above $20 billion by this May, and $35 billion post-money after a July round that followed the launch of its K3 model. DeepSeek, still privately held, is widely expected to list sometime in the first half of next year.