You have probably heard that Washington’s export controls completely locked China out of the advanced silicon market. You might also think Nvidia’s temporary workaround chips are the only things keeping Chinese artificial intelligence afloat.
That is not the full story anymore.
A silent, massive surge is happening inside China's domestic semiconductor sector. The newly released first-half 2026 financial forecasts of China's top graphics processing unit (GPU) developers reveal a stark reality: local chipmakers are no longer just concepts on a pitch deck. They are moving mass volumes, making money, and aggressively capturing the market vacuum left by Western restrictions.
Inside the Explosive H1 2026 Financial Forecasts
On July 16, 2026, two of China's primary domestic GPU developers, Moore Threads and Hygon Information Technology, dropped their preliminary earnings estimates for the first half of the year. The numbers are staggering.
- Moore Threads projected first-half 2026 revenue of $1.65\text{ billion}$ to $1.75\text{ billion}$ yuan ($244.0\text{ million}$ to $258.8\text{ million}$ USD). That is a year-on-year increase of $135%$ to $149%$.
- Hygon Information Technology forecast its H1 revenue between $8.5\text{ billion}$ and $9.3\text{ billion}$ yuan ($1.3\text{ billion}$ to $1.4\text{ billion}$ USD), expecting net profits to soar up to $52%$ to reach $1.83\text{ billion}$ yuan.
Moore Threads only listed on the Shanghai Stock Exchange STAR Market in December 2025. In less than a year, its market capitalization has hovered around $305\text{ billion}$ yuan ($45.1\text{ billion}$ USD). This is not speculative vaporware. The commercialization of their Kua'e AI computing cluster and the mass production of their high-end GPUs are actively generating hundreds of millions of dollars in recurring revenue.
These companies are capitalising on an unprecedented local supply vacuum. In May 2026, China expanded its tech replacement drive specifically to target AI chips. State agencies, municipal datacenters, and domestic cloud giants are now actively mandated to swap out their Western silicon.
Why Big Tech Is Walking on Two Legs
Chinese internet companies cannot afford to rely on imports that could be completely cut off tomorrow. It is too risky. This has led to a dual-sourcing strategy that is keeping the domestic semiconductor supply chain fully booked.
Look at ByteDance. The company is managing a massive computing load, especially since its Doubao AI assistant surpassed $345\text{ million}$ monthly active users. The daily average token call volume for Doubao is over $120\text{ trillion}$. To power this, ByteDance is splitting its infrastructure. They use high-end training processors like Huawei's Ascend series and Cambricon’s high-end training cards for training models, while relying on cheaper domestic alternatives for inference.
Nvidia is trying to keep its foot in the door. The US recently allowed the shipment of some H200 AI chips to China under strict, case-by-case licenses. However, these volumes are incredibly small and heavily restricted. To make matters worse, Nvidia is slashing its "white list" of buyers in Asia to crack down on gray-market smuggling.
Because of this constant friction, Chinese tech companies are giving local suppliers something they never had before: massive, guaranteed anchor orders.
The True Ceiling Is Wafer Allocation
While these sales projections are impressive, the industry faces a massive bottleneck. It is not a design problem. It is a fabrication problem.
When Moore Threads or Cambricon declare they are sold out, they do not just mean they have lots of buyers. They mean they literally cannot get enough raw silicon wafers manufactured.
[Domestic AI Chip Design Houses]
│
▼ (Orders Placed)
[SMIC / Local Chinese Foundries] ◄─── (The True Bottleneck: Limited Wafer Capacity)
│
▼ (Allocation)
- Huawei Ascend (Priority 1)
- Cambricon (Priority 2)
- Hygon / Moore Threads (Priority 3)
Because of US export bans on advanced Dutch lithography equipment, China's domestic foundries—principally SMIC—have a hard ceiling on how much advanced logic silicon they can produce each month.
Huawei gets the first bite of the apple. Its Ascend series consumes the lion’s share of SMIC's advanced packaging and lithography capacity. Other domestic players like Hygon, Moore Threads, and Cambricon have to fight over whatever capacity remains.
To counter this, China is aggressively pouring capital into its domestic supply chain. ChangXin Technology (CXMT) recently went through a massive $85\text{ billion}$ yuan IPO to fund a huge expansion of its 12-inch DRAM and High-Bandwidth Memory (HBM) production lines. Beijing's state-backed Big Fund also stepped in to become the third-largest shareholder of SMIC, cementing state control over wafer distribution.
The Battle of the Software Moat
Hardware is only half the battle. Nvidia's biggest competitive advantage is its proprietary CUDA software platform, which developers have spent nearly two decades coding on.
Chinese chip companies are fighting back by designing hardware that works out of the box with open-source software. Hygon’s Deep Compute 3 chip, for instance, has completed adaptation with $365$ mainstream large models, including popular open-source models like DeepSeek and Qwen.
The industry is moving away from proprietary, single-vendor ecosystems. When Chinese AI developers write code, they are increasingly building on top of PyTorch or custom middleware that makes it easy to switch chips instantly. If a developer can seamlessly run a model on a Huawei Ascend chip on Monday and a Moore Threads GPU on Tuesday, Nvidia’s software moat begins to dry up.
What To Do Next
If you are an investor, tech analyst, or hardware strategist, you need to adapt to this new bifurcated reality.
- Diversify your hardware footprint. Do not assume Nvidia is the only game in town. If you are deploying edge AI or inference services targeting the Asian market, start testing workloads on domestic alternatives. The performance gap on the inference side has narrowed significantly.
- Track wafer capacity, not chip designs. The real winners of the Chinese AI boom are not just the companies with the best architectures. Watch SMIC's capital expenditures and the progress of CXMT’s HBM lines. Fab capacity is the ultimate metric of success here.
- Target open software ecosystems. Avoid tying your code to proprietary hardware frameworks. Build your models using hardware-agnostic runtimes so you can migrate workloads seamlessly as supply chains continue to fragment.
