Mistral 融资35亿美元，三星因AI内存利润增长18倍

三星因为AI内存需求激增，利润预计将暴涨18倍。

Samsung is set to post an 18x profit jump because of AI memory demand.

Mistral AI正筹备一笔35亿美元的融资，公司估值达到230亿美元。

Mistral AI is looking at a $3,500,000,000 raise and a $23,000,000,000 valuation.

他们的收入已突破4亿美元，而Midjourney现在正要求迪士尼、环球和华纳兄弟公开各自的AI使用情况，因为眼下官司一堆，大家都在指责Midjourney用了受版权保护的内容。

Their revenue has hit over $400,000,000, and Midjourney is asking Disney, Universal, and Warner Brothers to disclose their own AI use because there's a bunch of lawsuits and everyone is accusing Midjourney of, you know, using copyrighted content.

所以他们现在想把局面扭转过来。

So they're trying to flip the tables here.

Station F的FAI加速器即将重启。

Station f's FAI Accelerator is coming back.

他们的首期项目拿到了3400万美元的种子轮资金，接下来会发生什么，我们拭目以待。

Their first cohort got $34,000,000 in pre seed funding, and we're gonna see what's happening next there.

Fable Five刚刚登上了KernelBench排行榜的榜首。

Fable five has just reached the top of the KernelBench mega.

这是一个基准测试，而就在这个时候，AI代理在真实自由职业任务中的能力翻了四倍。

That is a benchmark, and this is at the same time that AI agents are quadrupling, going four x on real freelance work, what they're actually able to accomplish.

所以我们来聊聊Fable Five的能力，这个新基准到底说明了什么。

So we're gonna talk about some of the capabilities of Fable five, what, you know, this new benchmark is saying.

目前，Claude Cowork是我最爱的效率工具，只要告诉它要完成什么任务，它就能控制你的Chrome浏览器，访问电脑上的所有文件，帮你把需要登录的事务全搞定，简直太厉害了。

Currently, Claude Cowork is my absolute favorite tool for getting things done, being able to tell it to accomplish a task and it being able to control your Chrome browser, use all of the files on your computer, and basically get everything done that you need to with logged in sessions is absolutely incredible.

给它加上MCP，是我做过最棒的改进。

Adding MCPs is the best thing I've ever been able to do for it.

Claude唯一的缺点是不支持图像、音频或视频的AI功能。

The only problem with Claude is that it doesn't have image, audio, or video AI capabilities.

当然，它在智能和编程方面非常出色，但我手头的项目太多，需要各种素材——做网站要图片，做广告要视频，做播客更离不开音频，我有不少全自动播客项目，音频需求特别大。

Of course, it's super smart and great at code, but I need so many different, assets for all of my projects that I'm working on, whether I'm making websites and I need images inside of them, or I'm doing, ads and I need, video creation for all of that, or if I'm doing things with podcasting, I have a lot of AI autonomous podcast projects and I need audio.

我为自己的创业项目aibox.ai开发了一个MCP，让你可以直接在Claude里生成音频、图像和视频。

I built an MCP for my startup aibox.ai that lets you get all of the audio, image, and video right inside of Claude.

今天早上，我和Claude聊了会儿。

So this morning, I was talking to Claude.

我需要为一个播客做封面，就让它调用AI Box的图像生成器。

I needed to make a podcast cover for something, and I said, go use the AI Box image generator.

你可以用 OpenAI 的 Image 1.5，它啥都能生成。

Use, you know, OpenAI image 1.5, which could say anything.

对吧？

Right?

或者说 Gemini 3.5 Pro，随便哪个都行。

You could say, Gemini 3.5 pro, whatever.

我就跟它说，去生成一张图片，做个播客封面，还详细告诉了它我平时做播客封面用的提示词。

And I said, go and make an image, you know, go make a podcast cover, and I explained exactly the the prompt that I was used for making podcast covers.

它直接调用 MCP，连上 AI Box，生成图片，发了个链接给我，我立马就能下载。

It just used the MCP, connected to AI Box, generated the image, gave me a link, and I could go download it immediately.

这才是 Claude 最大的突破。

This is the biggest unlock for Claude.

如果你想试试我为 Claude 做的这个 MCP，去 aibox.ai/claude 就行。

If you wanna try out the MCP that I've created for Claude, go to aibox.ai/claude.

我会把链接放在简介里，每月八美元就能用上所有这些功能。

I'll leave that in the description, and you get access to all of that for, like, eight dollars a month.

你可以在Claude里直接处理所有的图片、音频和视频。

You can have all of the image, audio, video inside of Claude.

随时通过AI Box的MCP调用它。

Call it at any time with AI Box's MCP.

这简直太厉害了，特别实用，希望能让Claude为你打开新世界。

It's absolutely incredible, super useful, and I hope that unlocks Claude for you.

三星预计今年的利润将是去年的18倍，而最主要的原因就是AI，特别是对一种叫高带宽内存（HBM）的需求暴增。

Samsung is expecting to have profits that are 18 times higher than they were a year ago, and, of course, the biggest reason is AI, but specifically, the demand has exploded for something called high bandwidth memory or HBM.

你可以把它理解成超高速内存，层层堆叠，直接集成在AI芯片上。

So, essentially, you could think of this as ultra fast memory, and it's stacked into layers and attached directly into AI chips.

NVIDIA和AMD都在用这种技术。

NVIDIA and AMD are both using this.

这相当于把内存直接焊在处理器上，让AI模型能飞快地处理海量数据。

So this is like memory that's bolted straight onto their processors, and it essentially lets AI models process tons of data really fast.

正是它带来了这种速度优势。

It gives them that speed.

几年前，整个内存行业彻底崩盘了。

So a couple years ago, the entire memory industry had completely crashed.

2023到2024年，价格一直在跌，因为产能严重过剩。

I think 2023 to 2024, prices were falling because there was way too much supply.

他们当时建得太多了。

So they basically overbuilt.

但突然之间，OpenAI、微软、谷歌、亚马逊、Meta这些公司全都疯狂采购AI硬件，一次就买几十万颗芯片。

But then all of a sudden, we had this perfect moment where OpenAI, Microsoft, Google, Amazon, Meta, all of them started buying up AI hardware in just insane quantities, hundreds of thousands of chips at a time.

整个市场就这样彻底反转了。

And so the whole market has switched now.

之前是产能过剩，现在变成了严重短缺，市场行情一路飙升。

All of a sudden, instead of this big oversupply that we had before, we now have a massive undersupply, and so the market is skyrocketing.

所有生产这种内存的公司都赚得盆满钵满。

And any of these companies that are building this memory are making a ton of money.

三星最初落后于最大的竞争对手——SK海力士。

So Samsung was initially behind, I think, the biggest competitor, which is SK, SK Hynix.

这是来自韩国的公司，我觉得我以前在其他播客里提到过他们，这次他们也从中获益巨大。

This is from South Korea, I think I've talked about them on other podcasts before, but they are also benefiting from this a ton.

现在三星也紧追其后。

And so now Samsung is kind of coming right behind them.

他们正在大力提升产能，其中有两点特别关键。

Now they are ramping up production, and there's two things in particular.

一个是HBM4，这正是他们眼下财务暴增的主要推动力之一。

There's the h HBM four, and that's basically becoming one of the biggest drivers of their kind of financial boom that they're seeing right now.

这款产品现在特别抢手。

That one's really popular.

这之所以重要，是因为HBM的利润极高。

And the reason why this matters is that the HBM is super profitable.

它的售价远超传统计算内存，而全球能大规模量产的公司只有三家。

It sells for way more than any traditional compute memory, and only three companies in the world can manufacture at scale.

三星、SK海力士，还有美光——我觉得美光总部在爱达荷州，我也在之前的播客里聊过他们，因为他们现在简直一路狂飙。

So Samsung, SK Hynix, and Micron, which I think Micron is out of, Idaho and, you know, I've also talked about them on the podcast before because they are absolutely blowing up.

现在真正能生产这种芯片的公司就三家，而且大规模制造难度极高。

So there's really just three main companies, that are making this right now, and it's really hard to manufacture this at scale.

所以这三家公司都在疯狂扩张。

So all of them are growing in a crazy amount.

美光这家公司在以前可能没多少人听说过，但因为现在内存供应极度紧张，它的身价一路飙升，已经跻身全球最大的公司之列。

Micron is not a company I think that was a is a household name, but it's becoming one of the biggest companies in the world because of just this huge supply pinch and shortage, and so the prices are just going up a ton.

三星也在努力减少对内存业务的依赖。

Samsung is also trying to become less dependent on memory.

他们打算进军定制AI芯片领域。

They're gonna be expanding into custom AI chips.

据说他们目前正在和Anthropic洽谈，计划开发专用的AI处理器。

They're also reportedly in discussions right now with Anthropic about building specialized AI processors.

他们还明确表示，希望进一步扩大半导体制造业务。

And they said specifically that they want to grow their semiconductor manufacturing business as well.

所以，现在对三星来说，形势简直太棒了。

So, I mean, right now, obviously, it's super exciting for them.

他们从内存市场的爆发中获益良多，但也清楚这不可能永远持续下去。

They're seeing a ton of benefit from all of the memory exploding, but they know that this isn't gonna last forever probably.

所以多元化是个相当明智的策略。

So the diversification is a pretty smart move.

Mistral AI 现在打算以 230 亿美元的估值融资 35 亿美元，几乎是其 2025 年 9 月 C 轮融资价格的两倍。

Mistral AI right now is looking to raise $3,500,000,000 at a 23,000,000,000 valuation, and that is nearly double its September 2025 series c price.

如果你还不知道的话，他们总部在法国，今年二月他们的经常性收入已经达到了 4 亿美元。

So they're based in France, if you didn't know already, and their recurring revenue just hit $400,000,000 in February.

这比一年前的 2000 万美元增长了整整 20 倍，简直不可思议。

That's about 20 x the $20,000,000 that they had a year ago, which is pretty crazy.

我觉得特别有意思的是，我们最近经常听到一个话题，尤其是特朗普政府一度试图限制 Fable Five 之后，虽然后来又撤回了。

And I think what's interesting about them in particular is we we hear talk a lot, I think especially after Trump kinda banned Fable five or or the Trump administration, whatever, had to get pulled back for a little bit before it went out.

尤其是在欧洲，很多人在讨论 AI 主权和 AI 基础设施，这比其他地方要频繁得多。

We had a lot of different players, especially in Europe, talking about AI sovereignty and AI infrastructure more than I think I've seen other places.

我之前在播客里也邀请过像 Cohere 这样的大厂创始人，他们也经常谈这个话题。

And I've had, you know, some of the big founders from, like, Cohere on the podcast who talks a lot about that as well.

但我认为，对于美国以外的国家来说，人工智能主权是个非常重要的议题，它们希望把投资放在能帮助自己实现这一目标的企业上。

But I think AI sovereignty is a huge deal for countries that are not in The United States, and they wanna make sure that they're putting their investments into companies that will help them achieve that.

Mistral 就是个很好的例子，因为它们提供开源模型，人们可以在本国境内、依托本地的数据中心运行这些模型，真正获得这些国家和企业所追求的‘人工智能主权’。

And Mistral is a good one because they have open source models so people can run those, you know, locally inside their own country with, you know, data centers inside their own country and and really get that quote unquote AI sovereignty that a lot of these companies are our countries are looking for.

目前 Mistral 表示，今年目标是实现十亿美元的年度经常性收入，并计划投入四十亿美元在法国和瑞典建设数据中心，同时推出 Mistral Compute——一个由英伟达提供技术支持的欧洲计算平台，今年就会落地，紧跟 XAI、OpenAI、谷歌等公司布局计算平台的步伐。

Right now Mistral says that they're targeting $1,000,000,000 in annual recurring revenue this year, and they're committing $4,000,000,000 to build data centers in France, Sweden, and they're also launching launching Mistral Compute, which is an European NVIDIA powered platform, and they're gonna be doing that, this year, following after what it seems like XAI and OpenAI and Google and all of them are kind of doing with their compute platforms.

它们已经收购了不少基础设施领域的初创公司。

They have acquired a bunch of infrastructure startups.

其中一家叫 QIABE。

One in particular is called QIABE.

它们还与 MGX、英伟达和法国公共投资银行合作，在巴黎附近打造一个人工智能园区。

They have a joint venture with MGX, NVIDIA, and BPI France to build an AI campus near Paris.

我觉得它们真的想证明，巴黎和法国整体上正努力成为人工智能的中心。

I think they're really trying to show and I think Paris and France in general is really trying to show that they are kind of the center for AI.

它们正在全力以赴，争取站到这个领域的最前沿。

They're they're really trying to get on top of all of that.

现在Mistral的收入来源主要集中在企业和政府客户，而不是面向普通用户的聊天应用。

Mistral's revenue mix right now, I think, really skews heavily to enterprise and government, so it's not consumer chat.

对吧？

Right?

大多数人用的还是ChatGPT或者Claude来做聊天。

Most people are using ChatGPT or Claude for that.

他们是在照搬Palantir的模式。

They're copying Palantir's playbook.

他们有一支前线工程师团队，帮助各家公司把自己的基础设施跑起来。

They have a bunch of forward deployed engineers that are helping different companies run models on their own infrastructure.

他们还推出了一款叫Forge的产品。

They have something called Forge.

所以他们确实是在效仿这种前线工程师的策略。

And so they're they're kind of copying this forward deployed engineer strategy.

Midjourney正在要求迪士尼、环球和华纳兄弟披露它们内部的AI训练做法，这是眼下一场重大版权诉讼的一部分。

Midjourney is asking Disney, Universal, and Warner Brothers to disclose their internal AI training practices as part of a big copyright lawsuit that's happening right now.

这些制片公司其实都在起诉Midjourney，我觉得他们是想反咬一口——如果Midjourney能证明这些公司自己也在用未经授权的内容训练模型，毕竟它们都有各自的AI项目，那他们就能说：你看，这不过是行业常态。

All of those different studios are actually suing Midjourney in particular, and I think they're trying to, like, flip the tables because, basically, if Midjourney can prove that all of those studios are training on unlicensed content too because all of them are have different AI programs, then I think they say, look, this is just the industry norm.

你们自己也都在这么做。

You guys are all doing it as well.

这可不是我们一家初创公司才这样，他们完全有可能让整个案子被驳回。

This isn't just us as a startup, and they're they could probably try to get this whole thing thrown out.

这三家制片公司加起来的市值大约有四千亿美金。

The three studios combined have a market cap of about $400,000,000,000.

去年他们一起起诉了Midjourney，因为Midjourney当时成功生成了巴特·辛普森和达斯·维达的形象。

They've all sued Midjourney last year because I think it was Midjourney was able to reproduce Bart Simpson and Darth Vader.

他们就说：你都能生成这些角色，那明显就是拿我们的独家数据在训练，不管是迪士尼的、环球的，还是巴特·辛普森和达斯·维达这些角色。

They're like, look, if you can produce that, then you're obviously training on our own, you know, Disney special data or whatever universal special data, our our characters, Bart Simpson and Darth Vader.

总之，这场官司已经拖了很久。

And so anyways, this lawsuit's been going on for a while.

我觉得最终这些公司都会拿到赔偿。

I think they're all expected to get a payout.

但我觉得，到这个时候，Midjourney 反过来想把责任推回给他们。

But I think at this point, Midjourney is trying to just point the finger back at them.

法官之前把工作室的披露范围限制在面向消费者的 AI 生成内容上。

Judge previously limited the studio disclosure to AI usage that produced consumer facing content.

Midjourney 希望取消这个限制，以便获取所有内部的 AI 工作流程和训练数据。

Midjourney wants that boundary erased to access all internal AI workflows and training data.

我亲自见过这些工作室谈论用 AI 做预可视化、分镜设计，还有像特效迭代这类工作，但他们从没公开过，自己到底用什么数据来训练这些工具。

I've personally seen all of those different studios talk about using AI for pre visualization, storyboarding, and things like VFX iteration, but they haven't really disclosed what, you know, they're training all their data in order to to create those tools with.

所以我觉得 Midjourney 是想抓住他们，发现他们用的训练数据其实和大家一样，就是整个互联网，然后说：你们不也在这么做吗？

So I think mid journey is hoping that, they'll be able to catch them in basically the the same training data that everyone's using, which is kind of the whole Internet, and say, look, you guys are doing it as well.

明白了。

Okay.

一家叫 Station F 的 FAI 加速器，今年九月要启动他们的第二期项目。

A company called Station F's FAI Accelerator is launching their second cohort this September.

他们第一期的二十家 AI 初创公司，全都成功拿到了总计三十四亿美元的种子轮融资。

They have 20 different AI startups from the first batch that all were actually able to raise $34,000,000 in pre seed funding.

这是巴黎的一个中心，他们把自己定位为欧洲版的硅谷。

This is a a Paris hub, they're positioning themselves as kind of Europe's answer to Silicon Valley.

说实话，在欧洲所有国家里，我觉得伦敦的英国和法国都做得不错。

And to be honest, I've actually of all the countries in Europe, I think London's England's doing quite well, and France is doing quite well.

他们有Mistral，这家大放异彩的AI公司，现在还想再培养更多。

They have Mistral, of their big breakout AI company, and they're trying to get more.

所以他们在努力搭建加速器和孵化器，真心想在国内培育出一批企业。

So they're they're trying to build these accelerators and these incubators and and really grow some companies inside of their country.

他们的目标是让每家初创公司在六个月内实现一百万欧元的收入。

They're targeting a million euros in revenue per startup within six months.

这其实是对一批投资者的回应，那些人总说欧洲初创公司商业化太慢，赶不上美国公司。

This is basically, they're trying to respond to a bunch of investors that are criticizing European startups saying that they're way slower than American companies to actually commercialize.

欧洲其实已经冒出不少了不起的初创公司了。

There's been a lot of incredible European startups.

比如Lovable，你就知道有多厉害了。

I mean, you got Lovable.

你看看Synthesia，做视频生成的，简直太厉害了。

You got lots of amazing, you know, Synthesia for video generation.

还有11 Labs，做音频的，也很出色。

You got 11 labs for audio.

欧洲冒出不少顶尖的AI公司，我觉得法国正想把它们全都集中到本国来。

I mean, you have some of the most, you know, some of the most prolific AI companies coming out of Europe, and I think France is trying to really concentrate them all into their country.

这个项目第一批学员里，80%都是有过创业经验的人。

80% of the first cohort that came out of this were founders that had been entrepreneurs before.

差不多三分之一的人拥有博士学位。

About a third of them have PhDs.

这种背景在欧洲和美国的创业者里其实挺常见的。

That's something that would expect a lot more out of founders coming out of Europe and America.

不过一般来说，做AI研究的人本来就多集中在这样的群体里。

That's probably less common, although when it comes to AI researchers, I mean, that's usually who's in that kind of cohort of people.

第二批学员，他们会新增六个合作伙伴。

The second cohort, they're gonna add six new partners.

他们还会加入11 Labs、Open Router、GitHub 和 HubSpot 这些公司。

They're gonna have 11 Labs, Open Router, GitHub, HubSpot that are all joining that.

他们还联合了 Anthropic、AWS、Google、Meta、Microsoft 和 Mistral AI，全都参与了这个合作计划。

They also have Anthropic, AWS, Google Meta, Microsoft, and Mistral AI that are all that are all part of their partnership program.

我很期待看到这个项目能孵化出哪些AI初创公司。

I'll be excited to see what AI startups come out of this program.

明白了。

Okay.

Fable 刚刚达成一个重要里程碑。

Fable just hit a big new milestone.

他们的新系统在 KernelBench Mega 测试中登顶了。

Their new system topped the KernelBench Mega Benchmark.

它能生成高度优化的 CUDA 代码，速度比之前的基准快了将近19倍。

It is able to write highly optimized CUDA code that is about 19 times faster than what their baseline was before.

甚至超过了 Claude Opus。

It even beat Claude Opus.

这实际上意味着，人工智能在优化GPU运行软件方面变得越来越强了。

And, basically, that means that AI is getting a lot better optimizing the software that powers GPUs.

与此同时，AI代理在几乎所有现实世界中的工作上都变得强大得多，比如那些通过类似自由职业任务来衡量的领域——3D建模、建筑设计、动画制作等等。

At the same time, AI agents are becoming way more capable at basically every job that can be done in the real world on a benchmark that is basically like, basically, what it's doing is measuring paid freelance work, things like three d modeling, architecture, animation.

它还做了大量分析工作。

It does also a lot of, like, analysis.

短短八个月内，成功率从2.5%飙升到16.1%，真的非常惊人。

The success rate jumping from just two point five percent to sixteen point one percent in only eight months, is really impressive.

GPT-5.5的得分只有6.3%，而Fable直接以16.1%碾压了它。

GPT 5.5 only scored 6.3%, so Fable is absolutely crushing that with 16.1.

研究人员还测试了AI在耗时超过一小时的复杂计算任务上的表现。

Researchers also tested AI on much longer compute tasks that take over an hour to complete.

虽然现在这些模型在这方面还很吃力，但进步非常明显。

And while today, right now, the models are really struggling with that, they are getting a lot better.

尽管优化GPU代码可能不是每个人使用Fable 5的场景，但我真的非常兴奋，也特别佩服。

And while optimizing code for GPUs might not be everybody's use case for Fable five, I'm really excited and really impressed.

我亲自用它完成了很多Opus 4.8根本搞不定的项目。

I've personally used it to do a lot of projects that Opus 4.8 just couldn't do.

我正在开发一个应用，里面的控件让我连续两天急得直抓头发——因为用Opus 4.8根本没法让它们自定义。

I have an app that I'm building, and, the widgets on the app, I literally was just pulling my hair out for two days because Opus 4.8, I couldn't figure out how to make them customizable.

我把这些问题丢给Fable 5，结果两小时就全解决了，一切运行正常。

And I sent, I pointed Fable five at them, and in two hours, it was able to get the whole thing sorted out and working correctly.

很多人说Fable 5用起来跟Opus差不多，但我觉得它聪明多了。

So I think, like, overall, Fable five, a lot of people are like, oh, it feels just like Opus.

可能只强了百分之十五到二十，但这点差距带来的影响太大了，尤其是能让你把事情做得快得多。

I think it's much smarter, and maybe it's like 15 or 20% better, but that 15 or 20% makes a huge difference, and it's really good especially when it's, you know, just being able to get things done much faster.

很多复杂的任务，它都能轻松搞定。

A lot of complex tasks, it's great at doing them.

我快凑够200条播客评价了，真的就差一点。

Guys, I am so close to hitting 200 reviews on the podcast.

如果你还没写评价，能不能帮我留一条？这对我真的特别特别重要。

If you haven't left a review yet, it would mean the absolute world to me if you could drop a review.

我都看了。

I read them.

这些评论给了我继续下去的动力。

They give me the motivation to keep going.

我真的很感谢每一条评论。

I appreciate them all.

如果你还没写，我会非常感激。

If you haven't done it yet, I would be eternally grateful.

如果你想体验在Claude、OpenAI、Cursor或者其他支持MCP的工具里直接使用音频、图片和视频功能，记得去aibox.ai/mcp看看。

Make sure to go check out aibox.ai/mcp if you wanna get access to audio, image, video, all inside of Claude or OpenAI or Cursor or any of these other tools that take MCP.

我对这个功能特别兴奋。

Guys, super excited about that.

赶紧去试试吧。

Go check it out.

链接在简介里。

There's a link in the description.

我们下一期再见。

I'll catch you all in the next episode.