2026-01-07 03:22:47

I heard that a certain domestic company might order 2 million H200 chips, with a total value of around 300 billion RMB. If this happens, it would be over 50 billion USD, directly surpassing NVIDIA's current inventory of 700,000 units. If this deal goes through, TSMC would need to work overtime to produce more than 1 million chips.

But the key point—this is just a letter of intent, and it might also be just hype for publicity. The real crucial factor is waiting for approval from the regulatory authorities. This amount of money is too staggering; anyone in that position would have to be very cautious.

Honestly, this situation has two sides. On one hand, if this order is truly diverted to domestic GPU companies, it could turn a bad situation into a good one, bringing astronomical revenue to local chip firms. On the other hand, training, developing, and deploying large models is extremely time-sensitive; no one can wait around. Buying ready-to-use computing power is very reasonable.

Looking at the trend for 2026, new applications like Agent and vibe coding are set to emerge. The token consumption for these is on a completely different scale—much more intense than simple chat Q&A. If these complex demands go live, the computing power requirements will skyrocket exponentially in minutes.

By the way, H20 is indeed somewhat underwhelming now—its training performance is only about one-sixth of H200, and it’s not much cheaper. While inference costs are decent, training is really not up to par. In contrast, H200’s training capability remains quite strong; many leading US companies are still relying on it for their main computations.

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.

18 Likes

Reward
18
9
Repost
Share

Comment

0/400

LiquidationSurvivor

· 01-09 23:22

2 million pieces? Wake up, can the regulators pass this? Nonsense. Speaking of H200 training capability, it's really impressive, but the probability that this deal falls through... do you bet or not? It's another intent order and hype, the story is quite elaborate. When Agent is activated, the computing power demand skyrockets, which is the real bottleneck. H20 is just a semi-finished product, its training performance is terrible, yet some still praise it. 300 billion is not a small number; even I would think carefully before acting. Instead of waiting for domestic breakthroughs, it's better to use ready-made solutions. Speed is king. TSMC overtime? They've been at full capacity for a while; you're overthinking it. This matter looks like just a paper story; a single word from regulators and it's all over.

View OriginalReply0

FloorSweeper

· 01-09 00:41

ngl this smells like classic paper hands cope... 2M chips at $500B? that's not an order, that's a negotiation flex. regulators gonna kill this before it even gets legs anyway lol

Reply0

DeFiCaffeinator

· 01-09 00:25

300 billion? Regulatory approval is still nowhere in sight, it's just hype. Let's wait until it actually materializes.

View OriginalReply0

SmartContractPhobia

· 01-07 03:51

2 million pieces? That's a joke, how can we pass the regulatory hurdle? No matter how nicely you put it, it's just an intent order. Let's wait until the approval comes down. H200 is currently the hot commodity; no one can replace it. If it really happens, TSMC will have to work overtime like crazy. The demand for computing power is skyrocketing; no chips are enough. The opportunity for domestic GPUs has arrived, but with limited time, using ready-made ones is very realistic. Purely hype-driven, I don't really believe it. H20 is indeed a burden; training performance is really disappointing. Spending 300 billion requires caution; don't talk nonsense. Once the Agent application is launched, the demand for computing power will explode.

View OriginalReply0

DAOTruant

· 01-07 03:47

Don't take the intention sheet too seriously; once regulatory approval is granted, everything will be over, and capital loves to hype things up this way. If this deal actually materializes, domestic chips might have some hope; otherwise, it still depends on Nvidia's stance. H200 is indeed powerful, but the problem is time can't wait. Whoever has the stock now is the boss. Starting to compete for agents only in 2026? It should have started long ago. The computing power arms race has already begun, hasn't it? Spending 50 billion on it, TSMC might just take off directly, but the premise is that this thing can really happen.

View OriginalReply0

QuorumVoter

· 01-07 03:45

2 million pieces? The moment this number comes out, it's already outrageous. The real test is whether the regulatory approval will go through. An intent order can be so hotly traded, a typical pattern of rookie investors getting exploited. Let's wait until it actually lands. H200 training capability is indeed strong, but the road to domestic replacement is still long. The demand for computing power truly takes off in 2026. Token consumption is indeed more complex than it seems. Over 50 billion dollars? That's terrifying. TSMC will have to work overtime until next year.

View OriginalReply0

ser_ngmi

· 01-07 03:30

2 million pieces? If it really crashes down, the US would be scared to death. I'm just worried that in the end, it's all talk and no action. This deal of over 50 billion USD, only when the regulatory hurdle is cleared can we consider it a win. Let's not boast prematurely. H200 training capability is indeed excellent, but the key is when domestic GPUs can catch up. Wait, if this order really materializes, won't TSMC's production capacity be directly maxed out? Every sentence is correct, but the key still depends on whether we can get the approval. Otherwise, it's all just a fantasy.

View OriginalReply0

MevTears

· 01-07 03:25

2 million units? Come on, the regulatory approval process can drag on forever. H200 is indeed powerful, but we need to wait for domestic solutions to mature. An intent order is just an intent order; don't hype it up too much. The shortage of computing power will inevitably explode; whoever has the chips in 2026 will win. H20 really underperforms; it's better to grit your teeth and go with H200. If this order really goes through, TSMC will have to work overtime like crazy haha. Wait, is this trying to bypass restrictions? How could regulators approve that? With that Agent setup, the demand for computing power will really be explosive. H200's training capability is indeed strong; the top American companies still rely on it. Spending 300 billion? Does anyone dare to be so extravagant?

View OriginalReply0

SerumSquirrel

· 01-07 03:23

2 million pieces? Just saying that number is already ridiculous, and regulators will definitely clamp down... Wait, if you're really going to invest that much money into H200, why not just invest in domestic products directly? After all, it's all about burning money. H20 is indeed garbage, training is weak and extremely expensive, I've been saying that for a long time. With tokens consuming so much in 2026, the existing inventory simply won't be enough. If this deal really goes through, TSMC will have to work overtime until the end of time... The approval hasn't even come through yet, and they're already hyping it up—typical news game. The computing power gap is right here; no one can wait, they need to buy now. Top American companies still rely on H200 to support them; there's no other choice.

View OriginalReply0