Exclusive-Nvidia sounds out TSMC on new H200 chip order as China demand jumps, sources say

The report tells us some things. What the report says is that we need to understand it in a way. The report is trying to explain something to us. So what does the report say in words?

Reuters reports that Nvidia is talking to TSMC (Taiwan Semiconductor Manufacturing Co.) about placing an additional order for its H200 AI chips, because Chinese demand has jumped sharply.

The headline numbers in the story are big:

Chinese tech firms have placed orders for “more than 2 million” H200 chips for 2026, according to sources Reuters spoke to.

Nvidia reportedly has about 700,000 H200-class units in inventory (including some GH200 “Grace Hopper” superchips).

A source told Reuters that TSMC’s work on expanded output would likely start in Q2 2026.

So the basic situation is:

China really wants a lot of H200 supply. The thing is, Nvidia does not have that much available now. So Nvidia is trying to get space to make these things so they can give China what they want. At the time Nvidia also has to make sure other places get the Nvidia GPUs they want. China wants a lot of H200 supply and Nvidia has to find a way to meet that demand, for H200 supply.

2) Why does China want the H200 badly all of a sudden. The H200 is something that China needs. They want it very much. China is trying to get the H200. It seems like they really need it. The question is why China wants the H200 badly now. What is it, about the H200 that China wants much. China wants the H200. They are trying to get it.

(A) This is an improvement for China from what they were getting before. China is now getting something better. The things China was getting were not that good. Now China has something that is really an upgrade. China is very happy, with this development.

For much of the last two years, Chinese customers have had access to restricted / “China-compliant” Nvidia chips rather than Nvidia’s top-tier AI accelerators. Reuters notes that Chinese buyers see H200 as a significant upgrade compared with chips currently available to them, and it specifically contrasts H200 with the H20, a downgraded chip designed for China.

Reuters also reports that sources claimed H200 delivers roughly “six times” the performance of the H20 (as perceived by these customers), which helps explain why Chinese firms are eager to buy H200 even at a higher price.

(B) The “permission window” might be open again—temporarily

A big driver is policy uncertainty. Reuters says the Trump administration recently allowed exports of H200 to China with a 25% fee, reversing the prior ban.

When companies think a window of opportunity may open or close they usually hurry to get the things they need. This kind of thinking can make demand for things go up fast because no company wants to be the one that waits and then misses out on what they need. Companies do not want to be left when it comes to getting the supply they need so they rush to get it and this is what happens when companies think about the window of opportunity, for supply.

Chinas internet giants are making artificial intelligence systems. They are doing this to get a lot of intelligence capacity. The artificial intelligence capacity of Chinas internet giants is going to be really big. Chinas internet giants are working on this every day to make their artificial intelligence capacity even bigger. Chinas internet giants want to have the artificial intelligence capacity, in the world.

Reuters says the bulk of the 2+ million chip orders comes from major Chinese internet companies.

It also points to a separate report relayed by Reuters: ByteDance may spend about 100 billion yuan (around $14B) on Nvidia chips in 2026, if sales are allowed.

The number that ByteDance has might be right. It might be wrong but one thing is clear: Chinas top platforms, like ByteDance are spending a lot of money on computers to train and use big artificial intelligence models or big AI models for short and these big AI models are what ByteDance and other Chinas top platforms are focusing on.

3) What is the H200. Why is the H200 valuable to people? The H200 is something that a lot of people’re interested, in. So what makes the H200 so special? Why do people think the H200 is worth paying attention to?

The Nvidia H200 is part of the Nvidia Hopper generation, which’s the same family as the Nvidia H100. The Nvidia H200 is made to make Artificial Intelligence training and Artificial Intelligence inference go faster. It also helps with high performance computing. The Nvidia H200 does this to make these tasks easier and faster for people who use the Nvidia H200, for Artificial Intelligence and computing.

Nvidia highlights one core advantage: the H200 is positioned as the first GPU with 141 GB of HBM3e memory and 4.8 TB/s memory bandwidth, which matters a lot for large language models (LLMs) that are bottlenecked by memory capacity and bandwidth.

The memory upgrade is a reason why the H200 is appealing to people who want to run bigger models or serve more users per GPU. The H200 is attractive because of this memory upgrade. People who want to run models or serve more users, per GPU like the H200 for this reason.

H200 vs GH200 (mentioned in the Reuters report)

Reuters says that Nvidia has two kinds of things in their inventory:

standalone H200 chips, and

some GH200 Grace Hopper superchips, which combine Nvidia’s Grace CPU with Hopper GPU architecture.

Nvidia’s own GH200 pages describe server configurations like GH200 NVL2 connecting two superchips with NVLink, delivering very large memory pools and bandwidth.

4) The money: pricing, modules, and the grey market

Reuters has some specific details about prices from the people they talked to. These details, from the sources are very specific which is not usually what Reuters does. Reuters typically does not include specific pricing details but this time they did according to the sources.

Nvidia has priced H200 around ~$27,000 per chip (for the variants planned for China).

An eight-chip module is expected to cost about 1.5 million yuan, versus about 1.2 million yuan previously for the H20 module (before it became unavailable).

Sources told Reuters this H200 module pricing is about a 15% discount vs grey-market alternatives, which were cited at over 1.75 million yuan.

This is really important because it tells us why people still want to buy things even when they cost a lot of money: demand, for these things can go up a lot.

If official supply is available, it may still be cheaper and safer than grey-market channels.

Big companies like to buy things in a way so they can get support when they need it and they like to have warranties too. This way they can also get things delivered on time. That is very important for big companies, like these. Big companies want to know when they will get the things they buy and they want to make sure they get the things.

5) Why TSMC is central to the story

Nvidia does not make its super powerful computer chips. The company depends on places like TSMC to build Nvidia graphics processing units or Nvidia GPUs for short. Nvidia needs these companies, called foundries to actually make the Nvidia chips.

Reuters notes that H200 uses TSMC’s 4-nanometer process.

So if Nvidia wants more of the H200 supply it has to make sure it gets this H200 supply. Nvidia needs to do this to have more H200 supply available.

wafer capacity at TSMC, and

enough downstream capacity across packaging, memory supply (HBM), and system integration.

The real problem is not just making the chip itself. This is where things get stuck. The chip is a part of it but it is not the only thing that slows us down when we are making the chip.

Modern Artificial Intelligence Graphics Processing Units are not, like Central Processing Units. They are really huge. They need very advanced packaging. Modern Artificial Intelligence Graphics Processing Units also depend on bandwidth memory stacks.

Even if a company that makes chips can make more of them the supply chain can have a lot of trouble with:

advanced packaging capacity,

HBM memory availability,

substrate supply,

testing and system assembly.

That is why sounding out TSMC is a deal: it means Nvidia is thinking about the whole process and trying to make sure they get the most important parts first. Nvidia wants to make sure they have what they need from TSMC. This is important, for Nvidia because they need TSMC to make their products.

6) The timeline: what happens and when does it happen with the timeline? The user wants to know what is going on with the timeline and when things are going to happen with the timeline.

Reuters gives a sequence:

Step 1: Initial shipments from existing stock (early 2026)

Reuters previously reported (Dec 22, 2025) that Nvidia told Chinese clients it aims to begin shipping H200 to China before the Lunar New Year holiday in mid-February—but only if Beijing approves.

That earlier Reuters report said initial shipments could total 5,000 to 10,000 chip modules, equivalent to roughly 40,000 to 80,000 H200 chips.

Step 2 is when we get some room to make more things and that will start a bit later which is, in the second quarter of the year 2026 when the extra manufacturing capacity begins.

The new Reuters “Exclusive” then says Nvidia asked TSMC to start production of additional chips, with work expected to start in Q2 2026.

So even if approvals happen the extra volume from production does not arrive right away. The extra volume, from production is something we will see in 2026 it is not something that will happen next week.

7) The biggest problem is that we might not get approval, from Beijing. This is a big risk because we need Beijing to say yes. The thing is, Beijing approval is not something we can count on. Beijing approval is really important. We have to remember that Beijing approval is not guaranteed.

One of the most important lines in Reuters is that Beijing has yet to greenlight shipments/purchases of H200 chips.

China would not want to hesitate if the computer chips are really powerful. The thing is, powerful computer chips can do a lot of things. So why would China want to wait and not use these computer chips? China needs computer chips to do many things like make new technology and help their economy. If the computer chips are really powerful China should use them to make their country stronger.

Reuters explains the problem that Chinese officials are thinking about:

Foreign computer chips from countries can really help make Artificial Intelligence development go faster now.

These foreign chips are very good, at doing lots of things which is what Artificial Intelligence needs to get better.

So these foreign computer chips can really accelerate Artificial Intelligence development now.

…but easy access might slow domestic chip industry momentum, because buyers will prefer Nvidia if it’s available.

Reuters also mentions a proposal: requiring bundling H200 purchases with a ratio of domestically produced chips, to keep demand flowing to local suppliers.

The idea of bundling things is a very old trick that companies use. It is like saying “You can buy the things from other countries. But you have to help the companies, in your own country too.” This bundling idea is a thing that people do in business. The bundling idea is used to help the ecosystem.

8) Why the situation with China matters to the world not just China itself. The thing is, what happens with China is important for everyone because it affects the globe and that is why the situation, with China matters globally not just in China.

This story is not, about Nvidia and China. It also affects the supply of Artificial Intelligence chips. The global Artificial Intelligence chip supply is what is really impacted here.

Reuters points out that if Nvidia makes H200 chips to meet Chinas needs it might not have enough for other places. This is because Nvidia has to find a balance. Nvidia has to make sure it has H200 chips for everyone, which can be a problem. Nvidia is trying to figure out how to make H200 chips, for China without running out for other people who need them. Nvidia needs to balance the H200 output so it does not tighten supply in areas.

China demand,

U.S. and allied market demand,

and production priorities for its newer lines (Blackwell, and the upcoming Rubin line that Reuters mentions).

A key strategic detail: H200 isn’t Nvidia’s newest generation

Nvidia has already moved on to architectures in the broader market but the H200 remains highly valuable because it is still really good and people need it. The H200 is very important for some people. Nvidia made the H200. It is still very useful. People are still using the H200. The H200 is a product, from Nvidia.

This has been proven people say that something is proven when it is known to be true. The fact that something is proven is very important, to a lot of people. People want to know that something is proven.

This thing is used a lot, in software stacks people like to use it because it works well with different programs the software stacks really support it.

and it still delivers enormous performance—especially with its memory improvements.

That creates a dynamic: it is pretty cool to think about the dynamic that is created by this situation the dynamic is really something to consider and I think the dynamic is what makes this so fascinating.

Nvidia wants to push the world toward its platforms so people start using the newest platforms, from Nvidia. The company Nvidia is really trying to get everyone to use the platforms that Nvidia is creating.

The thing is, when we talk about computer chips like H200 we have to think about the politics and rules of countries. These things can make chips like H200 really important in certain areas. This is because H200 chips are still very good even if they are not the ones. So H200 chips can be very useful, in regions.

9) So when we think about this we have to ask ourselves who actually benefits from this situation and who loses out in the end who are the people that benefit and who are the ones that lose.

If the people, in charge say yes then the winners will be announced. The winners are the people who will get something if everything gets approved. The winners are waiting to see what happens with the approvals.

(1) Nvidia

More sales, more scale, and potentially stronger leverage across the supply chain. Reuters also quotes Nvidia arguing that blocking exports can benefit foreign competitors.

(2) TSMC and the supply chain

If Nvidia increases its orders this means Nvidia will get money from making things and from services that are related to Nvidia.

(3) China’s big AI buyers

If they can obtain H200 through official channels, they may get better pricing and reliability than grey-market purchases.

Losers / risks

(1) Global customers waiting for GPUs

Any shift in allocation can tighten supply (or delay deliveries) elsewhere. Reuters explicitly raises the possibility of tighter global supply.

(2) China’s domestic AI chip makers (short-term)

If H200 becomes widely available, it can reduce demand for local alternatives—unless bundling rules force a mix.

(3) Nvidia’s policy risk

Even if U.S. policy allows sales now (with a fee), policy can shift again. And even if the U.S. allows it, China can still delay or block approvals.

10) What to watch next (very practical checklist)

If you are following this story here are the signals that will confirm whether this story becomes a story, with volume or this story stays stuck.

Beijing’s decision on approving H200 imports/purchases (and whether bundling conditions are imposed).

First real deliveries: whether Nvidia actually ships meaningful volumes before mid-February 2026 (Lunar New Year period).

TSMC production ramp: confirmation of new H200-related capacity starting in Q2 2026.

Demand disclosures: whether major Chinese buyers hint at confirmed capex/AI infrastructure expansion (ByteDance’s reported spend is one indicator, but it’s conditional).

What happens to the GPU availability when China places a lot of orders, for Graphics Processing Units. Do the waiting times for getting a Graphics Processing Unit get longer in places if China buys a lot of Graphics Processing Units and this reduces the supply of Graphics Processing Units.

Reuters is talking about a big deal it is a high-stakes scramble. Reuters says that people are, in a hurry to get things done it is a high-stakes scramble for Reuters. This high-stakes scramble that Reuters is describing is very important.

China’s biggest tech companies want H200 in massive volume (2M+ units ordered for 2026).

Nvidia’s near-term supply is limited, so it’s approaching TSMC to expand production—with the ramp likely starting in Q2 2026.

Everything depends on approvals and politics: U.S. policy has shifted (allowing sales with a fee), but China still hasn’t approved imports, and may impose conditions like bundling domestic chips.

If this goes through, it could tighten AI chip supply globally at a time when demand is already extremely intense.