Name: 10 分鐘解讀人工智能競賽 (AI Competition explained in 10 minutes)
Uploaded: 2025-10-21T13:26:27.000Z
Duration: 10 min 25 s

In order to talk about how the US and China are competing for AI, let's frame it in these layers.

為了談論美國和中國在 AI 上的競爭，我們來分層說明。

To create AI applications, you need large language models.

要創造 AI 應用程式，你需要大型語言模型。

And to innovate large language models, you need massive server infrastructure.

而要創新大型語言模型，你需要龐大的伺服器基礎設施。

And that infrastructure depends on a supply of advanced chips, and supply requires semiconductor production.

而這些基礎設施依賴先進晶片的供應，供應則需要半導體生產。

And production is only possible with advanced design.

生產則只有在先進設計的基礎上才可能。

And practically, at every layer of this, you have varying levels of government intervention between the US and China in order to compete for AI.

實際上，在這個的每一層，美國和中國之間都有不同程度的政府幹預，以爭奪 AI。

所以，讓我們從這個問題開始。

If the competition between US and China is this intense, why don't I really see it?

如果美中競爭如此激烈，為什麼我感覺不到？

Most of us are in the application layer where we create and use applications like ChatGPT, Cloud Code, NotebookLM and Manis.

我們大多數人都在應用層，在這裡我們創造和使用像 ChatGPT、Cloud Code、NotebookLM 和 Manis 這樣的應用程式。

And these applications leverage large language models.

而這些應用程式都運用了大型語言模型。

And contrary to popular opinion, competition at the application layer is actually higher than any of the layers above.

而且跟一般人想的不一樣，應用層的競爭其實比上面任何一層都來得激烈。

And that's because of a concept called barrier to entry.

這是因為一個叫做「進入門檻」的概念。

In other words, anybody can create AI applications for practically free, but

也就是說，任何人都可以免費創造 AI 應用程式，但是

not everyone can compete with LLM providers like OpenAI, Anthropic, Meta and Deepseek.

並不是每個人都能跟 OpenAI、Anthropic、Meta 和 Deepseek 這些大型語言模型供應商競爭。

For example, Meta's Llama 3.1 model, which is an LLM that's 405 billion parameters in size, was released in July 2024.

舉例來說，Meta 的 Llama 3.1 模型，這是一個擁有 4050 億參數的大型語言模型，於 2024 年 7 月發布。

In order to train this massive model, it requires 38 septillion math operation, which in the AI industry this math operation is called FLOPS, which stands for floating point operations.

要訓練這個龐大的模型，需要 38 千兆兆次的數學運算，在 AI 產業中，這種數學運算稱為 FLOPS，也就是浮點運算。

So in order to do 38 septillion operations to train the Llama 3.1 model, you're going to need a lot of computing.

所以，為了進行 38 千兆兆次的運算來訓練 Llama 3.1 模型，你需要大量的運算能力。

Precisely, you need about 4,4486 years just to train Llama 3.1 with one H100 GPU, which is considered the highest grade GPU for training LLMs.

精確來說，光是用一顆 H100 GPU 來訓練 Llama 3.1，就需要大約 4486 年，而 H100 GPU 被認為是訓練大型語言模型的最高等級 GPU。

And for reference, the H100 GPU costs around 25 to 40 thousand dollars a piece.

作為參考，一顆 H100 GPU 的價格大約是 2.5 到 4 萬美元。

So how do we shrink down the 4486 series to something more tangible?

那麼，我們要如何把 4486 年這個數字縮小到比較實際的範圍呢？

Thanks to transformer architecture, we're able to run most pre training in parallel, basically by throwing more money at it.

幸好有 transformer 架構，我們可以透過投入更多資金來平行處理大部分的預訓練。

Meta used 16,000 H100 units just to train this model, which cost them around 400 to $640 million in equ just to train llama 3.1 model.

Meta 光是訓練這個模型就用了 16,000 個 H100 單位，光是訓練 Llama 3.1 模型就花了他們大約 4 億到 6.4 億美元。

So a simple math should tell us that Meta's investment of 400 to $640 million to scale up to 16,000 H100 units to run 38 septillion flops in parallel should take them around three months, rather than 4,486 years.

所以簡單的數學告訴我們，Meta 投入 4 億到 6.4 億美元，擴充到 16,000 個 H100 單位來平行處理 38 千兆兆次的運算，大約只需要三個月，而不是 4,486 年。

So as you can see, the barrier to entry is quite high, meaning most of us probably won't be starting a company to compete against companies like Meta, OpenAI, and

所以各位可以看到，進入門檻相當高，這表示我們大多數人大概不會創辦一家公司來跟 Meta、OpenAI 和

But China has a lot of companies that can compete at this level, like Alibaba, Deepseek, Apple, Baidu, Moonshot, and

但中國有很多公司可以在這個層級競爭，像是阿里巴巴、Deepseek、Apple、百度、Moonshot 和

And as we saw with the Llama 3.1 model, training requires a lot of computing.

而且就像我們看到 Llama 3.1 模型一樣，訓練需要大量的運算能力。

In other words, for Chinese companies to compete at the LLM layers, they need access to as many GPUs as possible.

換句話說，中國公司要在大型語言模型層級競爭，就需要盡可能多地取得 GPU。

For example, if Meta only had half of what they actually had, so 8,000 instead of 16,000 H100, the training time for the llama 3.1 would have been around six months, as opposed to three.

舉例來說，如果 Meta 只有實際擁有的 GPU 的一半，也就是 8,000 顆 H100 而非 16,000 顆，那麼訓練 Llama 3.1 模型的時間大約會是六個月，而不是三個月。

And in the AI industry, a lot can happen during that three months that they would have taken them to train the model longer.

而且在 AI 這個產業，這三個月的時間裡，可能會發生很多事情，讓他們花更長的時間來訓練模型。

And this is where we start to go from the LLM layer to the infrastructure infrastructure layer.

這就是我們開始從大型語言模型層級，進入到基礎設施層級的地方。

Because in order to innovate LLMs, you need the infrastructure to support the innovation of LLM.

因為要創新大型語言模型，你就需要有基礎設施來支援大型語言模型的創新。

So if you put yourself in the shoes of Chinese companies where you have to compete with companies like Meta to innovate the next generation LLM, it's in your best interest to put your hands on as many GPUs as possible.

所以如果你設身處地為中國公司著想，必須要像 Meta 這類公司一樣競爭，來創新下一代的 LLM，那麼你最關心的就是盡可能多地取得 GPU。

And that's exactly what we're seeing from these companies.

這正是我們從這些公司身上看到的。

Meta's infrastructure is estimated to be about 350,000 H100 GPUs.

據估計，Meta 的基礎設施大約有 350,000 顆 H100 GPU。

OpenAI is planning to build Stargate, which is supposed to run up to 2 million GPUs when it's completed.

OpenAI 正在計畫建造 Stargate，預計完工後可以運行高達 200 萬顆 GPU。

And XAI released tens of billions of doll on their Colossus facility in Memphis that can run up to 350,000 GPUs.

XAI 在他們的田納西州曼菲斯 Colossus 設施上投入了數百億美元，該設施可以運行高達 350,000 顆 GPU。

Unlike the US counterparts like Meta and OpenAI, China doesn't have the same level of infrastructure with advanced chips like H100.

跟美國的 Meta 和 OpenAI 不同，中國並沒有同樣等級、配備先進晶片（像是 H100）的基礎設施。

And this is largely due to heavy restrictions that the US has put on China.

這很大程度上是因為美國對中國實施了嚴格的限制。

And in order for Chinese competitors like Alibaba, Deepseek, Moonshot, ByteDance to compete against US counterparts like OpenAI, Anthropic, Google and Xai, you need GPUs in the magnitude of hundreds of thousands, if not millions of them.

而像阿里巴巴、Deepseek、Moonshot、ByteDance 這些中國的競爭者，要想跟 OpenAI、Anthropic、Google 和 Xai 等美國公司競爭，就需要數十萬，甚至數百萬顆 GPU。

In October 2022, the US banned Nvidia's export of H100 and A100.

2022 年 10 月，美國禁止了 Nvidia 出口 H100 和 A100。

In October 2023, US banned the H100 chips, and in January 2025 they further restricted H20.

2023 年 10 月，美國禁止了 H100 晶片，並在 2025 年 1 月進一步限制了 H20。

Until recently, the US government agreed to share Nvidia's 15% of their revenue to Chinese sales.

直到最近，美國政府才同意分享 Nvidia 在中國銷售收入的 15%。

So, as you can see in the infrastructure layer, China is faced with this barrier to entry, which means their infrastructure doesn't as easily support innovation in the LLM layer as the US does.

所以，就像你們在基礎設施層看到的，中國面臨進入門檻的問題，這意味著他們的基礎設施不像美國那樣容易支持大型語言模型層的創新。

Which brings us to the next layer, which is the supply layer.

這就帶我們到了下一個層級，也就是供應層。

China has many companies that are willing to compete with the US in the LLM layer.

中國有很多公司願意在大型語言模型層與美國競爭。

Their supply of chips is being throttled by the us, leading to a weaker infrastructure.

他們的晶片供應受到美國的限制，導致基礎設施較弱。

So the natural question is this, if China doesn't have access to these chips, how are they able to compete with the US?

所以很自然的問題是，如果中國無法取得這些晶片，他們要如何與美國競爭？

Deepsea, for example, used the inferior GPU, the H800 card that were bought previously to the ban, to train the Deepsea V3 model.

例如 Deepseek 就利用了禁令生效前購買的較差的 GPU，也就是 H800 顯示卡，來訓練 Deepseek V3 模型。

And even with the lower grade GPU with lesser capacity, DeepSeek released the V3 model on December 26, 2024.

即使是使用容量較小的低階 GPU，Deepseek 也在 2024 年 12 月 26 日發布了 V3 模型。

And this has completely shocked the world because China essentially proved that they can innovate with less.

這完全震驚了全世界，因為中國實際上證明瞭他們能以較少的資源進行創新。

So even with the extreme restrictions that the US has put on China, China was still able to produce Deep Seq V3 and Deep Seq R1.

所以，即使美國對中國實施了嚴格的限制，中國仍然能夠生產 Deepseek V3 和 Deepseek R1。

That went head to head against OpenAI01, which was a huge breakthrough in the US.

這與 OpenAI 的模型正面交鋒，這在美國是一項巨大的突破。

And to put things into perspective, the Llama 3.1 model that we covered earlier took nearly three months with 16,000 H100 GPUs that required 38, septillion operations.

為了讓大家有個概念，我們之前介紹過的 Llama 3.1 模型，花了將近三個月的時間，動用了 16,000 個 H100 GPU，進行了 38 千兆兆次運算。

And compared to this, Deep SEQ not only used H800 GPUs, which trails behind the H100, but

相比之下，Deepseek 不僅使用了落後於 H100 的 H800 GPU，

And the training only required 3.8 septillion operations, which is 10 times less than the llama 3.1.

訓練也只進行了 3.8 千兆兆次運算，這比 Llama 3.1 少了 10 倍。

The implication of this is quite large on the infrastructure layer, because if you can train an LLM more efficiently, the cost burden in the infrastructure layer is significantly less demanding.

這對基礎設施層的影響非常大，因為如果能更有效率地訓練大型語言模型，基礎設施層的成本負擔就會大大減輕。

So from earlier we Talked about how OpenAI is investing in Stargate, that has capacity up to 2 million GPUs and XAI's huge investment in Memphis.

所以我們之前談到 OpenAI 投資 Stargate，它擁有高達 200 萬個 GPU 的容量，以及 XAI 在 Memphis 的巨額投資。

If Deepsea continues, continues to innovate state of the art LLMs with less flops, it could totally undermine the US investments on the infrastructure layer that we saw on OpenAI and XAI.

如果 Deepseek 繼續以更少的運算量創新先進的大型語言模型，這可能會完全削弱我們在 OpenAI 和 XAI 上看到的美國在基礎設施層的投資。

Since China is forced to innovate with less, and they did innovate with less.

因為中國被迫以更少的資源進行創新，而且他們確實做到了。

And the implication of that could be that maybe less is the norm, which would completely weaken US's perceived strength because they control the supply layer.

這可能意味著「少即是常態」，這將完全削弱美國的優勢，因為他們掌控著供應鏈。

Regardless of the case, the layer above supply is manufacturing.

無論如何，供應鏈之上是製造業。

Meaning either way, these advanced chips need to be manufactured to continue the innovation of LLMs.

這意味著無論如何，這些先進晶片都需要製造出來，才能繼續大型語言模型的創新。

And competition in the manufacturing layer between the US and China is huge as well.

美國和中國在製造業的競爭也非常激烈。

Back in 2025, China put forward a policy called Made in China 2025, which was a movement to deleverage their dependencies on foreign countries, especially in the semiconductor layer.

早在 2025 年，中國就提出了一項名為「中國製造 2025」的政策，旨在減少對外國的依賴，尤其是在半導體領域。

Their goal was to have 70% of their core materials from Chinese domestic content.

他們的目標是 70% 的核心材料都來自中國境內。

Essentially, China wants to produce their own chips instead of buying from companies like Nvidia.

基本上，中國想要自己生產晶片，而不是向 Nvidia 等公司購買。

Not only are fabs, which are factories that make chips extremely expensive, it's also a logistical nightmare in requires extreme technical precision.

生產晶片的工廠（fabs）不僅極其昂貴，而且在後勤上也是一場惡夢，需要極高的技術精準度。

In fact, there's really only one place in the world that produce most of these chips, and

事實上，世界上只有一個地方生產了大部分這些晶片，

that's in Taiwan, by a company called tsmc.

那就是臺灣，由一家叫做 TSMC 的公司生產。

That stands for Taiwan Semiconductor Manufacturing Company.

TSMC 代表臺灣積體電路製造公司。

So China's effort to deleverage from companies like TSMC and Nvidia would mean that they would at least have to produce chips internally at the same capacity as TSMC at the minimum.

所以中國試圖擺脫對 TSMC 和 Nvidia 等公司的依賴，至少意味著他們必須以與 TSMC 相同的產能來內部生產晶片。

But most countries have given up trying due to yet another barrier to entry.

但大多數國家都因為又一個進入門檻而放棄了嘗試。

But in this case, the barrier to entry isn't just about money and capital, but

但在這個案例中，進入門檻不僅僅是金錢和資本，而是

it's also two different things things talent and equipment.

While China has more STEM graduates than the us, in fact nearly double the amount which helps in the chip design, what they lack is equipment.

中國的 STEM 畢業生比美國多，幾乎是兩倍，這對晶片設計有幫助，但他們缺乏的是設備。

More specifically a $250 million equipment called EUV lithography.

更具體來說，是一種叫做 EUV 微影的 2.5 億美元設備。

While most electronics that we use, like microwaves, door sensors and dishwashers can be made in fabs with lower grade equipment like the DUV, high performance GPUs like the H100 requires EUV to generate.

我們日常使用的微波爐、門窗感應器和洗碗機等電子產品，可以在較低階的設備如 DUV 的晶圓廠中製造，但像 H100 這類高效能 GPU 則需要 EUV 來生產。

The problem is, EUV is also near impossible in replicating due to how accurate the equipment needs to.

問題是，EUV 的精確度要求極高，幾乎不可能複製。

We're talking precision down to a molecular level.

我們說的是分子等級的精確度。

And there's really only one place in the world that manufactures euv and it's a Dutch company called asml.

世界上真的只有一個地方生產 EUV，那就是一家荷蘭公司 ASML。

And just like in the supply layer US was banning Nvidia from selling advanced chips to China.

就像在供應鏈層面，美國禁止 Nvidia 將先進晶片賣給中國一樣。

ASML is also banned from selling EUV equipment to China as well.

ASML 也被禁止將 EUV 設備賣給中國。

In other words, in order for China to compete with the US in chip production, the barrier to entry here is not only the cost of fabs, which is around $20 billion for factory, they also lack the equipment, such as the EUV lithography to make the chips internally.

換句話說，中國若想在晶片生產上與美國競爭，這裡的進入門檻不僅是晶圓廠的成本，一座工廠大約要 200 億美元，他們也缺乏 EUV 微影等設備來自行製造晶片。

So as you see, in every layer we see two sides of the same coin.

所以你看，在每一個層面上，我們都看到了一體兩面。

One being the barrier to entry and the other being competition.

一方面是進入門檻，另一方面是競爭。

Meaning in order for China to compete, they need to overcome every barrier to entry that the US is imposing on them.

也就是說，中國要想競爭，就必須克服美國對他們施加的每一個進入門檻。

And the US has been making this barrier gradually difficult for China, while China, in order to still compete but with less, has been proving and almost flaunting at this point by releasing models like deep seq, V3 Kimike, 2 coin, 3 quarter as open models, as if to show China isn't too worried about the bans and restrictions that the US has been putting on every layer of the spectrum here.

美國一直在逐步增加中國的難度，而中國為了在競爭中不落後，並展現其能力，已經發布了像 Deepseek、V3 Kimike、2 coin、3 quarter 這樣的開源模型，彷彿在表明中國並不擔心美國在各個層面實施的禁令與限制。

So the ultimate question is this, what does this all boil down to?

所以，最終的問題是，這一切到底歸結於什麼？

At the end of the day, everything is contingent on the very last layer, the application layer, which is where most of us live and breathe.

歸根結底，一切都取決於最後一層，也就是應用層，這是我們大多數人生活和呼吸的地方。

AI is only going to be as good as what the application come out of it.

AI 的好壞，最終取決於它能產出什麼樣的應用程式。

In other words, both the US and China are dumping billions of money in supporting companies like intel, smic, Nvidia, essentially at the top layer down, because it's all about betting on what AI can do in the application layer, like healthcare, military and finance.

換句話說，美國和中國都在投入數十億美元支持 Intel、SMIC、Nvidia 等公司，從最頂層往下，因為這一切都是在押注 AI 在應用層能做什麼，例如醫療保健、軍事和金融。

So although we don't really experience the crazy competition that happens in the upper layers, it's really about us in justifying the competition at the top by innovating cool stuff like ChatGPT, Manus, NotebookLM, Replete, Relovable, V0, Bolt, Cursor, Klein, or any other domain specific solutions in military, healthcare, transportation, science, or anywhere else where AI could fit in.

所以，雖然我們沒有真正體驗到上層發生的激烈競爭，但這實際上是關於我們透過創新像 ChatGPT、Manis、NotebookLM、Replete、Relovable、V0、Bolt、Cursor、Klein，或任何其他領域特定的解決方案，來證明頂層競爭的合理性，無論是在軍事、醫療保健、交通、科學，還是 AI 可以融入的任何地方。

So ultimately it's up to us in leveraging all the breakthroughs that we're seeing in each layers above to unlock more automation, more disruption, and

所以最終，這取決於我們如何利用我們在上面各層看到的突破，來實現更多的自動化、更多的顛覆，以及