Name: 無需 Nvidia 芯片！亞馬遜為 Anthropic 新建的人工智能數據中心規模巨大 (No Nvidia Chips Needed! Amazon’s New AI Data Center For Anthropic Is Truly Massive)
Uploaded: 2025-11-02T13:34:10.000Z
Duration: 16 min 56 s

It's kind of surreal how fast this happened.

感覺有點不真實，事情發生得這麼快。

情態副詞

It was all just dirt fields and nothing really there from the time we put the shovels into the ground when we actually have large customers building that has been in record breaking in speed and scale.

從我們動手鏟土開始，這裡就只是一片泥土地，什麼都沒有，但現在我們已經有大型客戶在這裡蓋起了創紀錄的建築，速度和規模都令人驚嘆。

Behind me is the start of what will be the largest cluster of non Nvidia chips in the world.

我身後是全球最大的非 Nvidia 晶片群的起點。

And Amazon's newest data center, Megaproject, now fully operational here in New Carlisle, Indiana.

這是亞馬遜最新的數據中心，Megaproject，現在已在印第安納州新卡萊爾全面啟用。

These seven completed buildings replaced empty cornfields fast about a year and are filled with hundreds of thousands of Amazon's own custom Trainium 2 chips.

這七棟已完工的建築取代了空蕩蕩的玉米田，大約一年時間就快速完成，裡面塞滿了數十萬顆亞馬遜自家的客製化 Trainium 2 晶片。

And soon Trainium 3 across the US entirely devoted to running AI workloads for a single customer, OpenAI rival Anthropic.

很快地，全美國將會佈滿 Trainium 3，專門為單一客戶、也就是 OpenAI 的競爭對手 Anthropic 運行 AI 工作負載。

So they're already running about 500,000 chips in Indiana today.

所以他們今天已經在印第安納州運行了大約 50 萬顆晶片。

And in fact it's going so well that they've actually doubled down on that order and we expect jointly we'll be running more than a million trainium 2 chips by the end of this year.

事實上，進展非常順利，他們已經加倍訂單，我們預計到今年年底，我們將聯合運行超過一百萬顆 Trainium 2 晶片。

We got the first on camera tour inside what Amazon's dubbing Project Rainier.

我們獲得了首次鏡頭拍攝，進入了亞馬遜稱為 Project Rainier 的地方。

And it's coming online amid hundreds of billions in AI infrastructure build out announcements compute demand is at an all time high, but

在數千億美元的 AI 基礎設施建置公告中，它正逐步啟用，運算需求創下歷史新高，但

some are concerned about a possible overbuild.

I think they're still building on a proposition that we may not end up seeing be realized.

我認為他們仍然基於一個我們最終可能無法實現的假設在建置。

It's just crazy how much building is going on.

These deals all sound great on paper, but

they only materialize when they're actually racked and loaded and usable by the customer.

只有當它們實際安裝、載入並可供客戶使用時，才能實現。

Here in Indiana, this is just the first phase.

在印第安納州這裡，這只是第一階段。

With two more campuses under construction now for a total of 30 buildings on this 1200 acres.

還有兩個校園正在建設中，總共會有 30 棟建築，佔地 1200 英畝。

It'll consume 2.2 gigawatts of electricity, enough to power more than a million homes and millions of gallons of water, which has some local leaders and residents worried.

它將消耗 2.2 吉瓦的電力，足以供應超過一百萬戶家庭，以及數百萬加侖的水，這讓一些地方領導人和居民感到擔憂。

It's just difficult to keep losing farmland.

I was for some of the growth and my friends tried to tell me, you can't let them come in because once they get their toe in there they'll want more.

我本來支持一些成長，而我的朋友們試圖告訴我，你不能讓他們進來，因為一旦他們踏進來，他們就會想要更多。

As Amazon plans out more facilities for Project Rainier in Mississippi and beyond, we sit down with top leaders to ask how this Indiana site came together so quickly, where all that power will come from and how Anthropic will train its AI here entirely without Nvidia.

隨著 Amazon 在 Mississippi 及其他地區為 Project Rainier 規劃更多設施，我們將與高層領導人坐下來，詢問印第安納州這個據點是如何如此快速地建立起來的，所有的電力將從何而來，以及 Anthropic 將如何完全在沒有 Nvidia 的情況下在這裡訓練其 AI。

Just over a year ago, all of this was farmland.

就在一年多前，這一切都還是農田。

Now it's an operating AI data center, one of the world's largest.

現在這是一個營運中的 AI 資料中心，是世界上最大的之一。

AWS CEO Matt Garman spoke with CNBC's MacKenzie Cigalos about the project.

AWS 執行長 Matt Garman 與 CNBC 的 MacKenzie Cigalos 談論了這個計畫。

This is not some future project that we've talked about that maybe comes alive.

這不是一個我們談論過、或許會實現的未來計畫。

This is running and training their models today.

這就是他們今天正在運行和訓練他們的模型。

Cornfields, the data centers almost overnight.

玉米田，資料中心幾乎在一夜之間出現。

In late 2022, around the time OpenAI launched ChatGPT, Amazon was shopping for a site for some big AI plans.

在 2022 年底，差不多是 OpenAI 推出 ChatGPT 的時候，Amazon 就在尋找一個地點來實現他們龐大的 AI 計畫。

Talking to utility companies like American Electric Power.

跟 American Electric Power 這類電力公司洽談。

Ashley Savio helped steer the deal through local subsidiary Indiana Michigan Power.

Ashley Savio 協助透過當地子公司 Indiana Michigan Power 促成了這筆交易。

We introduced the concept of an Indiana data center to Amazon and they basically said, what do you have?

我們向 Amazon 提出了印第安納州資料中心的構想，他們基本上就問，你們有什麼？

So we pointed Amazon to that location and the volume of extra high voltage transmission lines that exist, as well as a very substantial substation.

所以我們向 Amazon 指出了那個地點，以及現有的大量超高壓輸電線路，還有一個非常重要的變電站。

By the spring of 2023, Amazon was making site visits.

到了 2023 年春天，Amazon 就開始進行實地考察。

A year later, the $11 billion deal was official, the largest capital investment in Indiana history.

一年後，這項 110 億美元的交易正式拍板定案，成為印第安納州史上最大的資本投資案。

Amazon broke ground in September 2024 and seven buildings began full operation in October 2025.

Amazon 在 2024 年 9 月動工，七棟建築在 2025 年 10 月全面啟用。

這件事發生得這麼快，感覺有點超現實。

當地的大筆回饋金讓事情變得順利許多。

The county granted Amazon more than $4 billion in property and technology tax exemptions over 35 years.

該郡在 35 年內授予亞馬遜超過 40 億美元的財產和技術稅務減免。

State legislation passed in 2019 will save Amazon some $4 billion more over 50 years.

2019 年通過的州立法將在 50 年內為亞馬遜節省約 40 億美元。

In exchange, the state is going to have a GDP improvement of over a billion dollars.

作為交換，該州將使國內生產總值（GDP）增加超過 10 億美元。

And we brought in close to 9,000 jobs already.

我們已經帶來了近 9,000 個工作機會。

While those are temporary construction related jobs, Amazon said it will create 1000 long with at least 600 of them paid above the county's average wage.

雖然這些是臨時的建築相關工作，但亞馬遜表示將創造 1,000 個長期職位，其中至少有 600 個的薪資高於該郡的平均工資。

It was all just dirt fields and nothing really there.

那裡原本只是一片泥土田地，什麼都沒有。

Site leader Josh Salabedra's been building data centers with Amazon for 14 years.

營運主管 Josh Salabedra 在亞馬遜建造數據中心已經 14 年了。

He moved to Indiana from the West coast in 2024 to oversee the build, hiring four general contractors to make it happen.

他於 2024 年從西海岸搬到印第安納州，負責監督建設，並聘請了四家總承包商來完成這項工作。

Never seen us do a project this quickly and that's been the most exciting part of it.

從來沒見過我們這麼快地完成一個專案，這也是最令人興奮的部分。

We also have a second campus that we were working on as well.

我們還有另一個園區也在進行中。

I think that that's the customer demand right now, right?

我想這是現在客戶的需求，對吧？

And as we saw hey coming and machine learning coming, we changed it to a different building type.

隨著我們看到 AI 和 Machine Learning 的發展，我們改成了不同的建築類型。

Big tech companies have been trying all sorts of structures to help them build as fast as possible, like tents.

大型科技公司一直在嘗試各種結構，像是帳篷，來幫助他們盡快建造。

Meta's put up as it builds a data center in Ohio.

Meta 在俄亥俄州建造數據中心時也是這樣。

In Indiana, AWS has changed to another building type, adding efficiency and newer tech like liquid cooling for the next buildings under construction.

在印第安納州，AWS 改用了另一種建築類型，為接下來正在建造的建築增加了效率和像液態冷卻這樣的新技術。

Now it's not just fast, but it is secure and reliable AWS infrastructure.

現在它不僅快速，而且是安全可靠的 AWS 基礎設施。

And so it's not a tent out there in a field.

It is a industrial Enterprise scale data centers.

而是工業級的企業規模數據中心。

Building out the total 30 buildings here, Amazon says will take at least two more years.

亞馬遜表示，在這裡建造總共 30 棟建築，至少還需要兩年時間。

When complete, it's expected to consume about 2.2 gigawatts of electricity and millions of gallons of water, which is why local resident Dan Caruso had 100 of these signs made.

完工後，預計將消耗約 2.2 吉瓦的電力和數百萬加侖的水，這就是為什麼當地居民 Dan Caruso 製作了 100 個這樣的標誌。

They're uncertain whether the power grid is going to be able to supply Amazon, GM and the town.

他們不確定電網是否能供應亞馬遜、通用汽車和這個鎮。

General Motors and Samsung are also building in New Carlisle, a $3.5 billion battery plant begun in 2024.

通用汽車和三星也正在新卡萊爾建造一座價值 35 億美元的電池廠，該廠於 2024 年開始動工。

At Amazon, some 4,000 construction related workers come to the site daily to make the speedy build out possible.

在亞馬遜，每天約有 4,000 名與建築相關的工人來到工地，以實現快速的建設。

That's a lot for the roads around this 1900 person town.

這對這個只有 1900 人的鎮上的道路來說，負擔很重。

It's gotten very dangerous on Highway 20 and Highway 2 and the county has responded to that with more patrols.

20 號公路和 2 號公路變得非常危險，縣政府也是以增加了巡邏。

Our police department for our town has been taxed because they do respond.

我們鎮上的警察部門壓力很大，因為他們確實需要出動。

Also, that's the biggest concern is safety.

Amazon is paying $7 million for highway improvements and county officials told CNBC it's negotiating up to $15 million more for road improvements near the site.

亞馬遜正在支付 700 萬美元用於公路改善，縣政府官員告訴 CNBC，它還在就近 1500 萬美元的道路改善進行協商。

Amazon is also paying $114 million for upgrades to other utilities like a water treatment plant and sewer lines.

亞馬遜還將支付 1.14 億美元用於升級其他公用事業，例如水處理廠和汙水管線。

Although an AI data center is different, AWS has been building data centers for its millions of cloud computing customers for nearly two decades and infrastructure services VP Prasad Kalyanaraman has been at Amazon that whole time.

雖然 AI 資料中心有所不同，但 AWS 近二十年來一直在為其數百萬雲端運算客戶建置資料中心，而基礎設施服務副總裁 Prasad Kalyanaraman 也在亞馬遜工作了這麼久。

他在西雅圖接受了 Cigalos 的訪問。

When you want to build some capabilities for genai which are unique to genai, like the network or for instance liquid cooling and so on, those are additions on top of foundational capabilities that we built over a long period of time.

當你想為 genai 建置一些獨特的功能時，例如網路或像是液體冷卻等等，這些都是我們長期以來建置的基礎功能之上的額外功能。

And that's why we're able to build at the scale that we are able to speed and scale are key.

這也是為什麼我們能夠以這樣的速度和規模來建置，速度和規模是關鍵。

As Amazon joins the crowded race to build Data Centers for AI Computer Meta's 2 GW Hyperion Data center in Louisiana and OpenAI's 1 GW Stargate site in Texas are two of the biggest that are being built Now.

隨著亞馬遜加入建置 AI 資料中心的激烈競爭，Meta 在路易斯安那州的 2 GW Hyperion 資料中心和 OpenAI 在德州的 1 GW Stargate 站點是目前正在建置的兩個最大的資料中心。

Many providers of that talk about gigawatts of capacity, but

許多供應商談論的是數吉瓦的容量，但

they don't tell you when they're going to get to gigawatt scale.

他們不會告訴你他們什麼時候能達到吉瓦級的規模。

What we have with Rainier is a real thing, which you've seen physically as well.

我們擁有的 Rainier 是真實存在的，你實際上也看過了。

But what else sets Amazon's apart is it's filled entirely with its own in house custom AI chips.

但亞馬遜的另一項獨特之處是，它完全採用了自家客製化的 AI 晶片。

Behind me right now they are working on delivering these towers, two of them together is going to make up one ultra server and inside of that is 64 AI chips.

在我身後，他們正在運送這些機櫃，兩個機櫃組成一臺超伺服器，裡面有 64 顆 AI 晶片。

These are Amazon's own custom Trainium 2 chips and what's especially interesting is there are no Nvidia GPUs in there.

這些是亞馬遜自家客製化的 Trainium 2 晶片，特別有趣的是裡面沒有 Nvidia 的 GPU。

Amazon's been designing its own chips since 2013, starting with a specialized chip called Nitro.

亞馬遜自 2013 年起就開始設計自己的晶片，最初是名為 Nitro 的專用晶片。

It bought Israeli chip startup annapurna labs in 2015, where we got a tour, and

2015 年，他們收購了以色列的晶片新創公司 annapurna labs，我們也去參觀了，

by 2018 had its own ARM based server chip Graviton.

到了 2018 年，就有了自家基於 ARM 架構的伺服器晶片 Graviton。

Its first AI chip, Inferentia, came in 2019.

他們的第一款 AI 晶片 Inferentia 則是在 2019 年問世。

Amazon launched Trainium One in 2021 and Trainium Two came out last December.

亞馬遜在 2021 年推出了 Trainium One，而 Trainium Two 則是在去年 12 月問世。

AWS chips have one main keeping down costs.

AWS 晶片有一個主要目標是降低成本。

Amazon CEO Andy Jassy said in July that Trainium 2 offers 30 to 40% better price performance than rival GPU instances.

亞馬遜執行長 Andy Jassy 在七月表示，Trainium 2 的價格效能比競爭對手的 GPU 實例高出 30% 到 40%。

Availability also like there's massive backups for Nvidia chips and if you can get chips faster and not have to wait online for Jensen to give you what you need, then you could get working faster and be in much better shape.

供應方面，Nvidia 晶片有大量的備貨，如果你能更快拿到晶片，不用等 Jensen 才能拿到你需要的東西，你就能更快開始工作，狀況也會好很多。

Trainium chips are less powerful and Simpler than Nvidia GPUs, but

Trainium 晶片比 Nvidia GPU 效能較弱且較簡單，但

Amazon says by fitting twice as many in each building, it can boost compute power and keep electricity and cooling needs down.

亞馬遜表示，透過在每個建築物裡安裝兩倍數量的晶片，可以提升運算能力，並降低電力和冷卻需求。

When you are able to control the stack all the way from the lowest layer of the infrastructure, you're able to build the right set of capabilities that these model providers want.

當你能夠控制從基礎設施最底層開始的整個堆疊時，你就能建構出這些模型提供者想要的正確功能組合。

You don't need to add all the other capabilities that you don't need.

你不需要加入所有你不需要的其他功能。

Some say that training 1 and inferentia 2 instances haven't been as competitive for Genai training due to weak hardware specs or poor software integration.

有人說，由於硬體規格較弱或軟體整合不良，訓練 1 和 Inferentia 2 實例在 Genai 訓練方面沒有那麼有競爭力。

What do you say to those specific critiques?

Sure, I fully admit Trainium 1 it was our first generation chip and that absolutely was a learning chip, both from an architecture perspective as well as a Software perspective.

當然，我完全承認 Trainium 1 是我們第一代的晶片，從架構和軟體層面來說，它絕對是一個學習型的晶片。