It’s unclear to me how stringent this will be enforced since Japan is a major intermediate player in supply chain. Does this mean Ga won’t be exported to Japan?

This could affect far more than just Japan. Also, is China ready for possible Japanese retaliation? Dual-use list can be found here
Includes Ga, Ge, Sb, Scandium, yttrium, Lutetium, dysprosium, terbium, gadolinium, samarium & whole bunch of chemicals.
+ upstream material like UHMWPE, graphite, high end Al/Ti alloy, complex fibersxkzj.mofcom.gov.cn/tzgg/art/2026/…

CXMT has filed updated 2025 guidance as part of its IPO filings. It had generated revenues of 8.3B, 9.1B, 15.4B RMB from 2022-2024 & 24.2B in 2025H1.

Now, it expects 55-58B for 2025 & turning profit for the 1st time.
Likely up to $3B in rev in Q4, making it bigger than SMIC. CXMT has 3 12-inch fabs (~20k wpm each) in Hefei & Beijing. It plans to raise 29.5B RMB off IPO out of 34.5B it needs for DRAM tech upgrade, production line upgrade & R&D.

CXMT invested 15.2B in R&D from 2022-2024, 36.6% of its revenue. It seeks to keep that at ~30% going fwd.

ByteDance's Doubao/Seed team has unveiled Seed1.8.
While it is a huge improvement over 1.5, it does noticeably trail Gemini-3 & GPT-5 on most of the LLM benchmarks.

They do seem to focus on Search & "Economically Valuable Fields", where they are SOTA based on benchmarks.


Among other benchmarks, they do quite well on the Visual & Video generation as well as VLM (not a surprise if you've seen their rankings on artificial analysis for this), but not so competitive on coding.

-> Their unique data set from apps are a huge advantage in video & visual

As Gemini caught up to GPT, Google AI Ecosystem is getting a huge boost. What if as Marc says, AI is commodified? As in, OpenAI cannot maintain indefinite lead & command high premium vs Google & Anthropic?

Google can undercut OpenAI/Anthropic until those 2 go out of business.
Using IBM chief Krishna's comment, each GW of DC cost $80B, which means 100GW of DC requires $8T of capex + opex (like electricity cost, taxes & staff) + interest. Interest alone is $800B
5 yr cycle -> $1.6T depreciation/yr.
+ opex & engineers -> over $3T in cost per yr for AI.

As expected, US's latest National Security Strategy doc embraces old model & rejects new model. We already know about China's dominance in the "New 3" (EVs, Solar & Batteries), but 15th FYP show China's full embrace of Hydrogen & Biomanufacturing. How new replaces the old model.

https://twitter.com/ruima/status/1997793877689053359

China is seeing rapidly declining demand for diesel & gas. Diesel HDT ratio has shrunk to 40% recently as EVs & LNG ones exploded. In pax vehicle mkt, NEV penetration is also at 60% for Nov.

Effect on WTI px, Permian rigs & US mfg is for all to see

https://x.com/tphuang/status/1992750250738516064

Welcome to the "Doubao Phone" - Collaboration bw ByteDance & Nubia/ZTE which uses a new BD OS.

Already hailed by some as the 1st real AI-centric phone. Everything here tied to Douyin account & Doubao AI features. AI Agent simply listens to your command & carries out the task Alibaba's great competition to this is its new Quark AI phone which uses Qwen.

In this world, Android becomes less important & the foundational AI model becomes the OS driving everyone.

https://x.com/AlibabaGroup/status/1996129986902983141

Just how large is China's compute & overall AI infrastructure? A good place to start is just by looking at Doubao/ByteDance numbers, since they are the largest player in the China mkt.
Grew from 4T tokens/day in 2024/12 to 30T tokens/day last month. Same ballpark as Google's 43T If we look at In-App AI in China, the top players are Baidu, Douyin (Doubao here), Wechat, Tencent news, Alipay, Gaode, Weibo, Douyin light speed, Baidu fast & Quark.

All major search apps in China also query AI just like Google does in US. Wide spread AI-based searches.

HW isn't the only one in China working hard at optical module network or SuperNode or OCS.

Xizhi Tech showcased its LightSphere X distributed OCS all optical SuperNode @ WAIC 2025.

It partnered w/ Biren Tech & its BR166L GPU + ZTE to connect together 2000 GPUs in 1 SuperNode.
InnoLight is the biggest player in the optical module industry. It's rumored to be Google's sole OCS supplier & will also supply 1.6T modules to Google & Nvidia + domestic customers. Eoptolink's 1.6T optical module has also been validated by Nvidia & its 800G supplies Google.

HW's Atlas 950/960 SuperPoD are truly innovative products. It offsets China's disadvantage in advanced node chip fabrication by increasing memory, optimizing for FP4/8 & building large super node using HW's networking expertise.

More smaller AI die connected like it's all in 1. HW utilizes an all optical network built w/ OCS (optical circuit switch). OCS is not new. Google built super nodes w/ 4096 TPUs.

HW tries to scale up this further w/ OXC (optical cross connect) to address limitation of increasing complex web of connections as # of nodes grow.

Laser is a cornerstone tech for other advanced techs like fiber optics, Lidar, laser cutting/cleaning machines & ofc laser interferometer.

China was weak in high-end laser until recent, but has now built full supply chain. Big players are Han's laser, HGTech & Wuhan Raycus.

Requires upstream breakthroughs in optical/electronic material, optical/laser chips, digital control, solid state laser, mechanical system, power source & gas laser machines.

Different types of lasers require different upstream supply chain. See below for domestic ratio of 1-3kw to 10kw+ from 2019 to 2023:

In 2019, China produced 20% of laser it needed above 6kw. By 2023, it produced 86% of its 10kw+ demand, 77% of 6-10kw & 99% in sub-6kw laser mkt.

all this happened while demand exploded across the board.

We are in 2025Q4 & HW has already showcased Atlas-960 SuperPoD w/ 2.2 TB/s interconnect & 34 PB/s total interconnect + 9.6 TB/s memory speed + the full system-level performance.

ppl on X are using metrics like PP/$ to measure AI-chips. Spend less time thinking about per chip cost or performance.

Think about things on a system level. Why does HW emphasize the all optical connections bw cabinets? Why does HW create its own HBM for inference? What about the push for SiPo chips w/ self generating light for optical modules? What about HW's new UB Mesh protocol?

How is HW able to do 15488 card SuperNodes, but Nvidia can only do 72 or 144?

https://x.com/tphuang/status/1983543440651497923

AI chip cost is a tiny portion of overall AI DC buildout cost. America is never going to be competitive vs China in cost per 100 EFLOPS of compute.

Especially not when companies like Meta are building their AI servers in the home of CJNG cartel & wasting expensive AI chips.

Next in my series of Wassenaar breakthroughs is on oscilloscope + related Signal Analyzers/Generators & Network Analyzers.

SiCarrier oscilloscope unveiling is the trigger behind this series of threads.
Why are these machines so important & what kind of breakthroughs did we see?

1st of all, oscilloscopes are really important in semi industry. You need 60GHz band one to test 7nm chip communication & a 90GHz one to test 3 to 5nm.

You need it for Serdes testing, especially as AI chip interconnect require ever fast speed. 224G requires 50x # of test vs 54G

Final part of my series on China's semi material progress in past yr are regarding to Diamond & AlN, the other 2 4th gen semi material.

Diamond is basically the ideal material for heat sink in high power applications. GaO/N epitaxy-on-diamond IC has significant power advantages In the past yr, we have seen both Xian Jiaotong University & ZS Tech achieve 2-inch wafer mass production tech. This allows the start of commercial application for diamond substrate in higher cost chips.

Next in my series of Wassenaar breakthroughs is ceramics & Ceramic matrix composite (CMC) materials.

China succeeded in advancing its own SiCf/SiC + Cf/SiC CMCs in the past couple of yrs + SiN fiber application & continuous fiber production.

What are their applications & where is China making the breakthroughs? CMC using Cf/SiC is widely used in aerospace industry in remote sensing satellite cameras.

Such material can achieve near-zero thermal expansion coefficient + improve dimensional stability of space camera components.
Low density can reduce overall weight of camera -> lighter launch cost & increased sat payload capacity

CXMT shared @ IEEE 2025 ASICON that it is launching LPDDR5X products w/ thickness of just 0.58mm.

Its product lineup including 12/16/24/32GB option w/ speed of 8533/9600/10677 Mbps. This matches what the big 3 offers to the mkt. Likely indicates CXMT has succeeded in 14nm node. Samsung started mass production of LPDDR5X last yr on its 12nm node w/ 0.65mm thickness w/ 12/16GB options, so CXMT has surpassed that offering w/ thinner package & larger memory options.

This is rapid progress for CXMT who only started producing 6/12GB LPDDR5 early last yr.

Next in my series of Wassenaar breakthroughs is Carbon Fiber & other Fibers + Composites.

Toray is the industry leader & has most recently developed T1200 grade CF, but export control basically only allows T300 to be exported.

In the past yr, China has almost caught up to Toray
Yesterday, I posted about CSST group reaching 95% yield & stable production of its T1100 grade CF product ZA70.

High end CF is critical for modern aerospace industry. Next-gen COMAC airliner is dependent on it. Same w/ the rocket launchers & aero engine.

https://x.com/tphuang/status/1981767919336833175

First on the list of broken choke-holds is High speed Analog-to-Digital Converters (ADCs)

Essential in RF, oscilloscopes, monitoring & control, software testing & digital imaging.

A $30B mkt growing toward $50B & controlled by TI & ADI. This is the crown jewel of Analog chips See below for Wassenaar Arrangement guidelines for ADC & DACs.

Chart provides better visual overview. Basically, nothing over 1.3GSPS for 8-bit resolution, 600MSPS for 10-bit, 400MSPS for 12-bit, 250MSPS for 14-bit & 65MSPS for 16bit+.

2019 HW entity listing was start of US/China tech war. It was further ramped up in Oct 2022 & then in 2023 & 2024 by Biden. Many wonder why China did not react harder until recently?

I will write a series of threads to explain China progress over past few yrs & why it waited so long. Today, China called for decisive progress in key tech in advanced semi, machine tool & instrumentation for 15-5 period.

Its tech priorities also match that of Wassenaar Arrangement. In fact, China has broken many chokehold in past yr. Now, it no longer fears Western export control.

NY times reported on China's significant cost advantage in NPP construction.

How is NPP construction decreasing in cost in China? Well, China really focused on scaling, localizing & clustering supply chain for both cost & natsec reasons. I will discuss the work it has done in this direction. A look at China's nuclear plants show a wide variety of 2nd/3rd gen NPP from France, Canada, Russia & US. It was originally most CPR-1000 variants, but new projects are mostly Hualong-1 or AP1000 derivatives.

Scaling up limited designs significantly lower cost of construction.

Assuming this is the real EU reasoning, while understandable, this is more alarming for Chinese investment than caving into US pressure.

I will explain history of Wingtech investment, current fab situations & why this is so disturbing & must be addressed for future China/EU tech cooperation.

https://x.com/tphuang/status/1979228463128043583

Wingtech acquired Nexperia for 14.7B RMB back in 2017 & through investment in its R&D, OSAT facility in Dongguan & attracting Chinese customers (many in EV sector), it was able to turn profit by 2022 & pay off debts by 2024.

Keep in mind its R&D increased during this time.

Reasons are far more than just Helium. It has been a banner yr for China's tech, financial, trade & military development to overcome possible US sanctions.
+ Tariffs hasn't worked:
In Aug, China's export went up YoY despite 33% drop to US. A list of their breakthrough 👇

https://twitter.com/RnaudBertrand/status/1977266253837251066

On helium, they broke through on 6N9 grade ultrapure He from NG, which is needed for semi, aerospace & cutting edge science.

China expects to be more than 50% self sufficient by 2028 & can import the rest from Qatar + other countries.