Robinhood CFO Debuts Second Public Venture Fund, WRITER’s New AI Model, Neolab Compute Crunch

欢迎大家收看Information的TITV节目。

Welcome, everyone, to the Information's TITV.

我是Akash Pasricha。

My name is Akash Pasricha.

今天是8月13日，星期四。

It is Thursday, August 13.

今天节目里，Robinhood的首席财务官将在纽约证券交易所和我们一起聊聊他们新推出的公开交易风险投资基金。

Today on the show, Robinhood CFO joins us from the New York Stock Exchange to discuss their new publicly traded venture fund vehicle.

接着我们会深入探讨为什么Neocloud旗下的CoreWeave和Nebius在最新财报发布后股价大涨，以及这些数字对AI基础设施意味着什么。

We'll then dive into why Neocloud's CoreWeave and Nebius are surging following their latest earning results and what the numbers signal for AI infrastructure.

我们还会看看WRITER推出的新AI模型，它旨在降低运行智能体的成本。

We'll also take a look at how a new AI model from WRITER is designed to cut the cost of running agents.

最后，我们的AI和机器人记者Rocket Drew会聊聊Neolabs如何应对算力紧张的问题。

And to close out the show, our AI and robotics reporter, Rocket Drew, will share a bit about how Neolabs are dealing with the compute crunch.

今天的节目会很精彩，我们马上开始吧。

It's going to be a great show, so let's get right on into it.

Robinhood今天在纽约证券交易所推出了他们的第二只公开交易的创投基金。

Robinhood is debuting its second publicly traded venture fund vehicle today on the New York Stock Exchange.

我想请首席财务官Shiv Verma来聊聊这只基金背后的考量，以及公司最近的业绩。

I want to bring on CFO Shiv Verma to talk about the calculus behind this fund and also the company's recent results.

Shiv，欢迎再次来到我们的节目。

Shiv, welcome back to the show.

很高兴你能来。

It's great to have you here.

Neol，谢谢你的邀请。

Neol, thanks for having me.

非常高兴能再次回来。

Very excited to be back.

那我们来聊聊你们推出的这第二只创投基金吧。

So let's talk about this second venture fund here you guys are debuting.

第一只基金里有一些大牌公司，比如Databricks、OpenAI、Merkur和RAMP。

The first one had some big names in it, Databricks, OpenAI, Merkur, and RAMP.

那这第二个基金里现在都有什么？

What is in this second fund now?

对，当然。

Yeah, absolutely.

第一个基金主要关注的是后期成长型公司。

So the first fund was focused on late stage growth companies.

你提到的很多公司都是这类。

Many of them you mentioned.

我们从客户那里听到的是，他们其实想接触早期阶段的公司。

What we heard from customers is they actually want exposure to early stage companies.

所以这个基金做的就是，和全球最大的孵化器之一YCC合作，专注于早期阶段。

So what this foot does is it focuses on early stage in partnership with YCC, one of the biggest incubators in the world.

所以我们通过和第一个基金类似的工具，让大家有机会接触到一种新的资产类别。

And so we're giving access to a new asset class using someone of the same vehicles as the first foot.

现在我想聊聊这个基金的结构，因为这些公司规模都比较小。

Now I wanna get into the structure of the fund because these are smaller companies.

有一件事让我很注意，就是这里好像有20%的业绩费，第一个基金里没有这个。

One thing that stood out to me looks like there's a performance fee on there, 20%, which it didn't have in the first fund.

跟我讲讲你们为什么决定加这个。

Walk me through the decision to add that.

好问题。

Yeah, great question.

传统的风险投资通常都有管理费和业绩费。

So traditional venture has a management fee and a performance fee.

因为这个基金更接近传统的风险投资，所以我们觉得也应该有业绩费。

Given this fund is much more similar to a traditional venture, we thought it'd be important to have that as well.

但重要的是，业绩费只在实际收益上收取。

Now it's really important to remember the performance fee only is charged on realized gains.

所以如果我们投中了下一个Airbnb、Stripe、DoorDash之类的公司，并且这个公司在这个基金里实现了流动性事件，那我们才会收这个费用。

So if we have the next Airbnb, Stripe, DoorDash, etcetera, and that comes into this fund and has a liquidity event, then we will charge the fee.

所以我们采用了风险投资的模式，但我们现在提供每日流动性。

And so we took the venture model, and now we have daily liquidity.

你不需要是合格投资者。

You don't need to be accredited.

所以我们认为这只基金比传统风投更好。

And so we think this is a better fund than traditional venture.

那

How

考虑到这些公司规模小得多，你们是怎么计算这只基金的资产净值的？

do you go about calculating the net asset value for this fund given that the companies are so much smaller?

我是说，这些公司，你知道，我们说的是创业公司。

I mean, y c companies, you know, this is the startup world we're talking about here.

有时候很难给它们估值。

So it's hard to put a value on them sometimes.

那你们具体是怎么操作的？

So how do you how do you mechanically do that?

对。

Yeah.

问得好。

Great question.

我们投资的每个基金通常都是通过SAFE来进行的。

So every fund that we invest in is typically done through a SAFE.

SAFE是针对早期创业公司的。

So SAFE are done for early stage companies.

这意味着估值要等到公司未来融资时才能确定。

This means the valuation is not set until the company raises at a future round.

我们还会设定一个上限，这样如果他们估值很高，我们的投资者就能获得更好的回报。

We also put a cap on it so that if they rage at a large valuation, our investors get better.

我们有一个独立的估值委员会，会审查每一家公司，然后还有第三方估值服务来帮忙估值。

We have an independent valuation committee that looks at every single company, and then we have a third party valuation service that helps value them.

所以我们尽力确保客户得到的是实际的净资产值。

And so we do our best to make sure that customers are getting actual NABs.

我们每个季度都会报告，而且是在SEC注册的。

We then report every quarter, it's SEC registered.

所以即使这些是早期公司，也有一个非常严谨的流程来尝试给它们估值。

So even though these are early stage companies, there's a very deliberate process to try to value them.

那你们最终在这些小公司里会占多少股份呢？

And how much ownership do you end up taking in these small companies then?

是的，这是个好问题。

Yeah, it's a great question.

通常来说，我们可能持有0.5%到1%的股份。

So typically, we might have half a percent to 1% ownership.

对于早期公司来说，这其实已经相当可观了。

And so for an early stage company, that's actually quite large.

想想那些从Y Combinator出来的公司，比如DoorDash、Stripe或者Airbnb，它们上市时市值能达到几十亿甚至几百亿。

If you think about companies that have come out of Y Combinator, such as DoorDash or Stripe or Airbnb, they can go public in the billions or tens of billions of volume value.

所以很早就拿到股份，真的非常有力。

So getting a ownership stake very early on is is really powerful.

这些公司都是AI公司吗？还是给我们举几个例子？

Is it all AI companies or give us some of the names here?

所以公司投资组合是多样化的。

So the company's a mix of different portfolios.

其中有一些是AI公司。

So it is some AI companies.

我们也会发布一些信息或观点。

We put out some, information or perspectives.

大约50%的公司专注于AI或AI基础设施。

It's about 50% are focused on AI or in AI infrastructure.

还有一些公司做的是真正的硬科技，也就是制造实体产品。

There's also companies doing real hard tech, so building physical things.

有金融科技公司，也有医疗健康公司。

There's fintech companies, there's healthcare companies.

所以覆盖面很广。

So it kinda spreads the wide gamut.

对我们来说，我们只是在每一批项目中寻找最好的公司。

And for us, we're just trying to look for the the best companies in each individual batch.

太好了。

Great.

那么，Shiv，这是一个令人兴奋的基金，我想祝贺你它的首次亮相。

Well, Shiv, it's an exciting fund, and I wanna congratulate you on the debut of it.

等你们宣布第三只基金的时候，我们会再请你回来。

We'll have you back on when you guys are announcing the third fund whenever that's coming.

这里是TITV，刚才的是Robinhood的首席财务官Shiv Verma。

That is Shiv Verma, the CFO of Robinhood here on TITV.

CoreWeave和Nebius的业务都在蓬勃发展。

CoreWeave and Nebius' businesses are both booming.

CoreWeave本季度的收入比去年翻了一倍多。

CoreWeave revenue more than doubled in the quarter from last year.

Nebius的收入增长了五倍以上。

Nebius revenue is up more than fivefold.

关于这些业绩，我想请来Jefferies的高级分析师Brent Thill。

For more on those results, I want to bring on Brent Thill, senior analyst at Jefferies.

Brent，欢迎回到节目。

Brent, welcome back to the show.

很高兴你能来。

It's great to have you here.

谢谢邀请。

Thanks for having me.

好的。

Okay.

那我们先从CoreWeave开始吧。

So CoreWeave, let's start there.

他们的业绩有什么让你印象深刻的？

What stood out to you about their results?

CoreWeave的需求依然非常强劲。

CoreWeave continues to see massive demand.

需求已经超过了供应。

The demand is outstripping supply.

像微软、亚马逊、谷歌，大多数超大规模云服务商都说客户在拼命要求增加供应。

Like Microsoft, Amazon, Google, most of the hyperscalers have been saying that customers are screaming for more demand.

所以，华尔街说AI已经完了这种说法，其实挺可笑的。

So this concept that we're we're done with AI is kinda silly in terms of what Wall Street's been saying.

需求环境依然非常强劲。

The the the demand environment remains incredible.

CoreWeave说他们基本上已经卖光了。

CoreWeave says they're basically sold out.

我觉得最突出的是，他们正在缩短合同期限。

And I think what what really stood out is they're shortening their contract durations.

这些合同上他们看到了更好的单位经济性，因为他们能提高价格。

They're seeing better unit economics on these contracts because they're able to raise price.

而且他们提价的时候，客户并没有反对。

And as they raise price, customers aren't pushing back.

客户反而说，我们还要更多。

They're saying, we'll take more.

所以利润率在上升。

So margins are going higher.

而且我觉得，Dawn，华尔街之前有个担忧，就是他们可能赚不到钱，可能只是收入增长，但利润上不去。

And I think, Dawn, there was a fear that Wall Street had that they were not going to be able to make money, that this would just be a top line revenue acceleration with no bottom line profit.

但CoreWeave告诉你的恰恰相反。

And what CoreWeave is telling you is that's that's exactly the opposite.

你看到的是收入在加速增长。

You're seeing accelerating revenue.

利润率也很不错。

You're seeing good margins.

其他所有超大规模云服务商也都是这样。

And we saw that across all the other hyperscalers.

微软、亚马逊和谷歌都实现了更快的增长和更高的运营利润率。

Microsoft, Amazon, and Google saw accelerating growth and higher operating margins.

所以我觉得在这个AI时代，我们看到的，就是那种担心，嘿。

So I think what we're seeing in this age of AI that on the fear of, hey.

这真的很贵。

It's really expensive.

你永远赚不到钱。

You'll never make money.

我觉得四大巨头刚刚告诉你这不是真的，增长和利润是可以兼得的。

I think that the the big four just told you that it's not true, that you can have growth and profits together.

我觉得CoreWeave另一个有趣的地方是，他们的需求正在扩大。

I think the other thing that's interesting for CoreWeave is that, they're seeing broadening demand.

不只是来自那些超大规模云服务商。

It's not just coming from the hyperscalers.

显然，他们为Meta建设。

Obviously, they build for Meta.

他们为微软建设。

They build for Microsoft.

但他们实际上开始看到像卡特彼勒这样的其他客户，以及一些较小的公司也来建设。

But they're actually starting to see, other customers like Caterpillar and, other smaller companies come in and build.

他们建立了一个按需推理的业务，基本上你可以直接通过他们按需启动推理服务。

They built a on demand inference business where you can just basically turn up inference through them on demand.

这个业务是新的。

That business is new.

它正在蓬勃发展。

It's booming.

他们在这方面的需求非常旺盛。

They're seeing great demand on that side.

所以没错。

So Right.

总的来说，你知道，数字真的很好，预期也很好。

Overall, you know, really good numbers, good expectations.

华尔街之前的态度确实非常消极。

Wall Street's view was was really negative going in.

你怎么看……我知道你并不覆盖Nebius，所以我不会特别问你关于那些业绩的问题。

How do you I I I know I don't think you cover Nebius, so I won't ask you specifically about those results.

你怎么比较这两家公司？

How do you compare and contrast these two companies?

我是说，Nebius的营收比上季度增长了450%，我记得是这样。

I mean, Nebius, their results revenue was up 450%, I think it was, compared to last quarter.

看起来他们在盈利指标等方面还处于更早期的阶段。

They are a bit earlier, it looks like, in their build in terms of their profitability metrics, stuff like that.

我是说，你怎么对比这两家公司？

I mean, how do you compare and contrast these two companies?

对。

Yeah.

我觉得，现在CoreWeave会说，他们是在为全球最顶尖的公司建设，不是说其他竞争对手不好，但他们的观点是，他们是在为最大的那些客户服务。

I mean, I I think right now what CoreWeave will say is that they're bringing they're building for the best companies on the planet, planet and and nothing against, the other competitors, but their view is that we're building for the the biggest of the big.

所以，从他们的角度来看，打个比方，他们造的是约翰迪尔拖拉机，而其他人造的是割草机和除草机，面向大众市场。

And so, the analogy what would would be from their perspective is they built the John Deere tractor, and and others are building weed whippers and lawnmowers, for for the for the mass market.

这没什么不对的。

So nothing wrong with that.

只是方法不同而已。

Just a different, approach.

记住，CoreWeave服务并支持着世界上一些最复杂的企业，比如Meta、微软、谷歌。

Remember, CoreWeave serves and powers some of the most sophisticated enterprises in the world, Meta, Microsoft, Google.

我可以列一下名单。

I'll go through the list.

他们网站上还列出了其他一些他们支持的公司。

There's a handful of others that are on their web page of companies that they power.

所以他们一直与少数几家非常大的客户绑定在一起。

So they have been tied to a smaller number of very, very large customers.

所以，我觉得，你知道，有不同的策略，显然我们看到的是有多个供应商的空间，对吧？

And so, I think, you know, there's different strategies, and obviously, what we're seeing is there's room for multiple vendors, right?

你也看到甲骨文进入了这个市场。

You've seen Oracle get into this market

对。

Right.

还帮助了OpenAI。

And help the OpenAI.

你也看到很多其他公司采取了不同的策略。

You've seen many others take different approaches.

我觉得我们看到的是，这些策略大多都在奏效，因为需求环境太好了，而且供应根本跟不上。

I think what we're seeing is the all most of these approaches are working because the demand environment's so good, and there is no there is no supply.

那我们就来谈谈那个大问题吧。

So So let's get I mean, let's get to the big question then.

现在需求确实很大。

I mean, the demand is great right now.

供应却很少。

Supply is low.

这对Neocloud来说都是好消息，尤其是在定价方面。

It's all great news for the Neocloud, certainly in terms of pricing.

你觉得这些公司的这种增长能持续多久？

How sustainable do you think this growth will be for their companies?

目前来看，我们并没有看到任何放缓的迹象。

Right now, we're seeing no slowdown.

所有供应商的资本支出都在全面增加。

CapEx goes higher for all the vendors across the board.

我们看到需求回升的模式一直很稳定。

We're seeing consistent patterns of of the pickup in demand.

看看企业级AI的部署，其实规模还很小。

When you look at enterprise AI deployments, it's tiny.

比如，大型企业里AI代理的部署数量还是个位数。

Like, we're talking like single digit agent deployment at large enterprises.

对吧？

Right?

科技公司们，你知道，都在笑。

The tech companies, you know, all laugh.

他们觉得，你这是什么意思？

They're like, what do you mean?

我们到处都在用AI。

We AI's everywhere.

比如说，你走进一家大银行、保险公司或者大型航空公司。

Well, like, you go into a big bank, an insurance company, a big airline.

我的意思是，他们在AI部署上只是刚刚起步。

I mean, like, they've scratched the surface in AI deployment.

而且我觉得我们反复说过，跟踪科技行业二十五年，从来没有哪个董事会讨论过Web 3、比特币，或者向云计算的迁移，他们都没提过。

And I think what we've said repeatedly is that having followed tech for twenty five years, there's no boards that ever talked about Web three, Bitcoin, you know, the move to cloud, that none of them talked about it.

现在他们全都在谈论AI。

They're all talking about AI.

从金融服务到医疗保健，你到处看看，都是这样。

From financial services to healthcare, you know, you go across the board.

所以我觉得我们正处于部署的早期阶段。

So I think we're in early stage of deployment.

记住，这个过程是缓慢的。

Remember, this is glacial.

这不是一夜之间就能发生的。

It doesn't happen overnight.

那些受监管的大型企业不能直接跳进AI的池子里。

Big, big enterprises that are regulated can't just jump in and cannibal into the AI pool.

他们进展得非常缓慢。

They do it very slowly.

而且，我们今天看到的需求，基本上只是试探性地迈出一小步，这已经表明这种趋势会持续好几年。

And because the demand we're seeing today, basically on putting one toe into the water, this gives you a signal that it's gonna be sustainable for multiple years.

那么，我们现在是不是处于泡沫之中？

Now, are we in a bubble?

未来三到五年，我们会不会过度建设？

Are we gonna overbuild three, four, five years out?

我觉得这才是争论的焦点。

I think that's the debate.

现在谁也看不清楚，因为目前所有东西都卖光了。

No one can see that yet because we right now, everything sells out.

价格还在往上涨。

Prices are going higher.

客户看到了投资回报。

Customers are seeing there's an ROI.

我们看到积压订单创下历史新高，远远超过了投入的资本支出。

We're seeing backlog and record highs far exceeding the CapEx dollars that are going in.

所以当收入超过资本支出时，你怎么能说这是泡沫呢？

So when you have revenue that's exceeding the CapEx dollars, how how do you say it's a bubble?

而且，当美敦力的CEO说医生下午眼睛疲劳，从不做检查时，你怎么能说这是泡沫？

And how do you say it's a bubble when you have the CEO of Medtronic saying doctor's eyes get tired in the afternoon, never get a checkup.

AI永远不会累。

AI never gets tired.

它会拯救生命。

It's gonna save lives.

航空公司在天空中拯救生命。

Aviation companies saving lives in the sky.

金融服务公司让他们的金融代理更高效。

Financial services firms making their financial agents more productive.

对。

Right.

目前还没有任何信号。

Let's Like no signals yet.

所以如果出现了产能过剩，而这些新云公司会成为核心，那你主要看哪些数据来寻找担忧的迹象呢？

So if there is overcapacity that is built and these Neo Cloud companies would be at the center of it, I mean, what numbers do you look at in terms of looking for signs of concern?

我的意思是，是不是运营成本开始比……好吧，我在这里停顿一下，因为现在盈利能力不是首要任务。

I mean, is this operating costs starting to rise faster than well, I mean, I'm pausing here because right now, profitability is not the priority here.

但具体你会看哪些指标来判断，嘿，产能过剩开始成为人们的问题了？

But what metrics specifically do you look at to say, hey, overcapacity is starting to be an issue for people?

想想看，当你看到这个，你说价格在下降，全年我们听到第一季度、第二季度到第三季度的价格都在下降。

Think when you look at this and you say pricing goes lower, all year, we've heard pricing from q one, q two into q three.

现在价格在上涨。

Now pricing's going higher.

对吧？

Right?

那价格是在暴跌吗？

So is pricing falling out of the sky?

不是。

No.

是在上涨。

It's rising.

需求情况怎么样？

How's the demand picture?

也在上涨。

It's rising.

从营收增长加速来看，你看到了什么？不仅仅是Neolabs，还有亚马逊、谷歌、微软、甲骨文。

What are you seeing in terms of revenue growth acceleration, not not just for the Neolabs, but for Amazon, Google, Microsoft, Oracle.

它们都在加速增长。

They're all accelerating.

你看看实际收入进账和支出之间的差距。

You look at the actual split between revenue intake and dollar spent.

这种情况我们反复见过。

We've seen this repeatedly.

这个差距大得离谱，简直可以开辆大卡车横着穿过去。

There's a huge you can you can drive a semi, you know, sideways through this.

根本不在一个量级上。

It's not close.

所以我们现在认为，这些信号表明，如果收入超过了建设成本，作为财务分析师，我们看到的是回报不错。

So we think right now, those are the signals, which are if your revenue is exceeding that of the cost to build, as a financial analyst, what we're seeing is there's a good return.

而且每家公司都展示了运营利润率的改善。

And every single company has showed you improving operating margin.

当这种情况开始停滞，当需求增长停滞，当你进入企业市场，开始推广企业AI，然后大家都说这不行，这简直是笑话，是骗人的把戏，这些智能体该干活的时候却在打高尔夫，到时候又会出现失控的智能体。

When that starts to stall, when the demand picture stalls, when you get into the enterprise and you start rolling on enterprise AI and everyone's like, this is this isn't working or this is a joke and this is hocus pocus and these agents are off, you know, playing golf when they should be at work, again, we're gonna have rogue agents.

那时候我们就要开始担心了。

Then we start to concern to concern.

我们目前没有看到任何这样的迹象。

We don't have any signals of that.

对。

Right.

最后一个问题，Brent。

Last question for you, Brent.

有没有迹象表明CoreWeave近期会减少对NVIDIA的依赖？

Any sign that CoreWeave is gonna diversify away from NVIDIA in the near future?

你有什么预期吗？

Any expectation from you?

我觉得他们没有理由把自己的小船从豪华游艇上解下来。

I don't think that they have any incentive to disconnect their their rowboat from their massive luxury yacht.

现在完全没有理由分开。

Zero zero reason to disconnect at this point.

不过话虽如此，你知道，还有其他船也在靠近呢。

But but, I mean, that being said, I mean, there's, you know, there's other boats on the horizon too.

对吧？

Right?

他们正在……我是说，不是吗？

They're get I mean, no?

我觉得

I think

我觉得CoreWeave说过他们是中立的。

what I think what CoreWeave has said is they're agnostic.

比如，他们在大家讨论Kimmy之前就发布了关于Kimmy的新闻稿。

Like, they put a press release out about Kimmy, you know, before anyone's talking about Kimmy.

他们接受所有模型。

Like, they they embrace all models.

他们接受所有半导体基础设施、所有硬件、所有液冷，无论你想在哪里建。

They'll embrace all all semi infrastructure, all hardware, all liquid cooling, wherever you want to build.

他们对自己的数据中心也持开放态度。

They're open to their own data center.

他们会和其他拥有自己数据中心的人合作。

They're going to they're going part with other people that have their own data centers.

他们是开放的。

They're open.

我觉得他们采取的是中立的态度。

And I think they're taking an agnostic view.

这就是纳德拉在微软做Azure时的做法，你知道的。

And this is what Nadella did at Microsoft, you know, when they started Azure.

保持中立的态度。

Have an agnostic view.

和所有人合作。

Work with everyone.

不要只说我们只用一种钉子来盖房子。

Don't don't just say we're only working with one nail to build a house.

你需要多种钉子。

You wanna have multiple nails.

所以，不，我没看到什么真正的变化。

So, no, I I don't I don't see a real change.

我觉得他们的看法是对的，就是拥抱所有人。

I think their view is the right view, which is embrace everyone.

但是，你想不想跟黄仁勋的游艇绑在一起？

But like, do you want to be tied to Jensen's yacht?

我愿意。

I would.

就是说，没有理由断开联系。

Like, no reason to disconnect.

对。

Right.

对。

Right.

好。

Great.

好的，Brent，感谢你来做客。

Well, Brent, I want to thank you for coming on.

这位是Jefferies的高级分析师Brent Thill，欢迎来到TITV。

That is Brent Thill, senior analyst at Jefferies, here on TITV.

Rider，一家专注于企业AI营销和销售的公司，今天发布了新的旗舰模型，名为Palmyra X6。

Rider, an enterprise focused AI marketing and sales company, released a new flagship model today entitled Palmyra X6.

它还对其软件进行了升级，并表示这将使平台对客户来说更便宜。

It also made upgrades to its software, it says, will make the platform cheaper for its customers.

为了详细解读这些，我想请来WRITER的首席执行官May Habib。

To break that all down, I want to bring on May Habib, CEO of WRITER.

May，欢迎再次来到节目。

May, welcome back to the show.

很高兴你能来。

It's great to have you here.

嘿，Akash。

Hey, Akash.

谢谢你的邀请。

Thanks for having me.

那么，Palmyra X6，我一会儿让你来解释这个名字，但先跟我们说说发布这个新模型的理由，以及它有什么不同之处。

So, Palmyra X6, I'll let you explain the name in a minute here, but tell us about the rationale here to release this new model and what's different about it.

对。

Yeah.

现在任何在企业界工作的人都知道，成本效益和数据主权是大家最关心的问题。

Anybody who's spending time with the enterprise these days knows that cost efficiency along with sovereignty is just the number one issue for folks now.

我认为，企业在采用或未采用某些技术方面的情况，确实向市场表明，企业愿意在AI上花费的金额有了新的上限。

I think the adoption of, and maybe lack thereof, of fable in the enterprise is really showing the market that we've got a new upper bound for how much businesses are willing to spend on AI.

而且大家真的不希望随着采用率的提高，支出也爆炸式增长。

And folks just do not want to have spend explode, right, as adoption does.

他们真的厌倦了追逐那些前沿基准和前沿价格。

And they're really tired of chasing these frontier benchmarks, frontier prices.

他们想要的是能够可靠地大规模完成他们关心的工作的模型。

They want models that can do the work that they care about reliably at scale.

这正是我们开发X6的初衷，也是我们构建它的方式以及围绕它打造的配套工具的原因。

And that's really what motivated X6, how we built it, and the harness that we built around it.

而且我觉得，我们真的看到了专用模型在性能和成本上给企业带来的优势，这正是他们所需要的。

And I think, you know, we're really seeing purpose built, giving that performance advantage and that cost advantage that enterprises are looking for.

它是在5.2 GLM上训练的，对吗？

And it's trained on 5.2, GLM, is that right?

是的。

Yep.

你知道，我们是第一家真正从零开始训练模型的应用型AI公司，对吧？

You know, we are the first of the applied AI companies that were really training our models from scratch, right?

我们花了好几年时间做这件事，现在也还在继续。

We spent years doing that, we're still doing that.

但这也意味着我们明白，预训练并不是应用型AI创造价值的唯一环节。

But it also means we know that pre training isn't the only place where value is created in applied AI.

我们在针对特定任务的环境上做的后训练，加上在配套工具上的投入，真正为关系管理和营销等使用场景量身打造了解决方案。

And the post training that we've done on task specific environments alongside the investment in the harness really built for relationship management use cases, marketing use cases.

这真的实现了大家想要的三重效果，对吧？

It's really just delivered this trifecta that folks want, right?

更便宜、更快、更好。

Cheaper, faster, better.

成本降低52%，任务几秒钟就能完成。

52% lower cost, tasks completing in seconds.

我们的智能体可以连续运行八小时甚至更久。

We've got agents that can run for eight hours or longer.

那这些结果是怎么通过模型实现的呢？

And how get those do results with the model?

给我讲讲这里面的具体机制吧。

Walk me through the mechanics here.

好的。

Yeah.

所以，我们要真正思考的是，一个完整的框架里应该包含什么，对吧？

So, the ability for us to really think about what goes into a harness, right?

模型是怎么获取上下文的？

How does the model pull context?

它怎么使用公司数据？

How does it use company data?

它怎么利用上下文窗口？

How does it use its context window?

它怎么在工具之间进行交接？

How does it manage handoffs, right, among tools?

它怎么应用安全护栏？

How does it apply guardrails?

你知道，一个完整的框架其实涵盖所有这些，甚至更多。

You know, a harness is really all of those things and more.

它怎么做质量检查？

How it does QA, right?

它怎么生成子代理，又怎么验证这些子代理的工作？

How it spawns sub agents, how it checks those sub agents work.

而且，当你像我们一样，真的在多年理解客户如何在我们的专业领域使用AgenTeq AI的基础上，对吧？

And when you, like us, really sit on top of years of understanding how customers want to use AgenTeq AI in the domain that we focus on, right?

销售和市场营销。

Sales and marketing.

这真的能给你带来巨大的优势，让你能够真正构建那个模型，构建那个框架。

It just gives you a tremendous leg up in being able to really build that model, build that harness.

这个框架本身实际上是模型无关的。

The harness itself is actually model agnostic.

我们构建它时就是让它模型无关的，我们的平台也是模型无关的。

We had built it to be model agnostic, and our platform is model agnostic.

但当然，和我们的模型配合效果最好。

But of course, works best with our model.

而且，模型和框架一起，在WRITER平台上效果最佳。

And the model and the harness together work best with the WRITER platform.

那么，GLM 5.2，这是来自中国的一个模型，关于人们是否对使用海外模型有所顾虑，是否想留在美国，一直有很多讨论。

Now, so, GLM 5.2, that's one of the models that comes out of China, and there's been a whole discussion around whether or not people have hesitation using models from overseas, whether they want to stay here in The US.

你自己有没有过这种犹豫？

Did you have any of that hesitation yourself?

我觉得市场在这方面已经发生了很大的变化，企业对这件事的理解也变了，对吧？

I think the market has really shifted on this, and enterprises' understanding of this has really shifted, right?

这是一个MIT许可的开源权重模型。

This is a MIT licensed open weights model.

这是目前最强的开源权重模型。

It's the strongest available open weight model.

权重、环境、后训练、框架，所有这些都是由一家美国公司在美国本土、美国基础设施上构建的。

And the weights, the environment, the post training, the harness, all of that built by a US company on US soil, on US infrastructure.

它运行在美国的基础设施上。

It's running on US infrastructure.

和原始开发者完全没有联系。

There's literally no connection to the original developers.

我觉得市场真的已经转向了。

Think the market has really turned.

这趟车已经开走了。

Ship has sailed on this.

我是说，现在没有哪家企业不在用这种方式使用开放权重模型。

I mean, there isn't a single enterprise that's not using open weight models in this way.

对。

Right.

那我们现在把话题转到开放权重模型的竞争上。这周早些时候，NVIDIA发布了他们最新的开放权重模型，或者说，是其中一个版本。

So now just shifting to the race in open weight models broadly, I mean, we had NVIDIA come out earlier this week with their latest open weight model, or a version of it, I should say.

Meta也重新加入了这场竞争，带着他们的开放权重模型。

Meta is also reentering the fray, I should say, with their open weight models.

你觉得哪家美国公司最有希望挑战中国的开放权重模型领先者？

Who do you think Which US company do you think has the best shot of challenging the open weight leaders in China?

因为这是大家都很好奇会怎么发展的竞争，对吧？

Because this is the race that everyone's curious to see how it plays out, right?

是的。

Yeah.

我觉得，没有哪家应用型AI公司会放弃参与这场竞争的野心。

I mean, look, I think not a single applied AI company, right, is not absconding from its its ambition to be in this race as well.

我认为未来几个月，我们很可能会看到企业级市场的竞争重新拉开帷幕。

I think what we are likely to see over the next few months is a real reopening of the enterprise race.

这场竞争远未结束。

It is by no means over.

我觉得，当你真正和CIO一对一交流时，对吧？

I think the enterprise, when you really get a CIO one on one, right?

他们对Anthropic的做法并不满意。

They're not happy with what Anthropic is trying to do.

他们对OpenAI给企业带来的东西也不满意。

They're not happy with what OpenAI is bringing to the enterprise.

我认为最终OpenAI是一家消费级AI公司，Anthropic是一家编码模型公司，而企业级市场实际上非常开放，赢家会非常多元化。

I think ultimately OpenAI is a consumer AI company, Anthropic is a coding model company, and the enterprise is actually really, really wide open and incredibly diverse in terms of the winners.

NVIDIA会赢，我们也会赢，应用型AI、原生AI公司都会赢。

NVIDIA is going be a winner, we're going to be a winner, The applied AI, native AI companies are going to win.

这是因为企业想要掌控权，对吧？

And it's because the enterprise wants control, right?

实验室正在构建

The labs are building

梅，你押注谁？

who's your money on, May?

我是说，有那么多令人兴奋的，你会押注谁去真正挑战中国的开源权重领导者？

I mean, there's so many exciting Who would you put your money on to really challenge the open weight leaders in China?

我觉得是美国。

I think it's The U.

的

S.

初创生态系统。

Startup ecosystem.

NVIDIA 在他们所做的事情上无疑是领导者。

NVIDIA is certainly a leader in what they're trying to do.

所以，可能是反思吧。

So, reflection maybe.

也许就是反思。

Maybe reflection.

对。

Yeah.

我觉得这场竞赛还远未结束。

I think it's still a really wide open race here.

而且我认为，对于那些真正深入了解情况的人以及正在做采购决策的企业来说，权重的来源其实没那么重要，不管它们来自中国还是美国。

And I think the provenance of the weights matter a lot less, whether they come from China, whether they come from The US, to folks who are really in the weeds here and enterprises who are making these purchasing decisions.

我想问你一个问题——你刚才提到了OpenAI和Anthropic，昨天你在LinkedIn上发了一篇关于水印的帖子，很有意思。我想知道你能不能给我解释一下，你发那篇帖子想传达什么信息？关于这些玩家如何处理水印，以及这整体上反映了他们在内容和所有权方面的什么策略？

Let me ask you a question about So you talked about OpenAI and Anthropic, and you had an interesting post yesterday on LinkedIn about watermarking, and I wondered if you could explain to me what was the message you were trying to send with that post with respect to how some of these players are addressing watermarking, and then what that really says overall about their strategy around content and ownership and stuff like that.

好的。

Yeah.

这里的背景是，欧盟的《人工智能法案》已经生效了。

The context here is in the EU, the EU AI Act has gone into effect.

水印是他们要求企业做的事情之一。

Watermarking is one of the things that they ask companies to do.

而且这确实是客户的责任，要披露他们发布的内容是否由AI生成。

And it's really on the part of the customer to disclose when they've put out content that has been generated with AI.

但这又回到了那个关于一刀切AI的评论。

But this gets back the comment on one size fits all AI.

绝大多数使用AI开发内容的企业，他们做的事情远不止按一下按钮，然后发布Claude生成的内容。

Vast majority of enterprises using AI to develop content, they're doing a ton more than pressing a button and publishing what Claude spits out.

所以，我们支持对内容进行水印标记，保持透明度。

And so, we are pro transparency on watermarking content.

显然，我们会遵守法律，但把每一个AI辅助的输出都标记为某个模型生成，这给了模型太多的功劳。

Obviously, we're gonna be complying with the law, but labeling every AI assisted output as generated by X model gives the model way too much credit.

我们正在寻求一种解决方案，让企业能够真正识别AI在其中的角色。

And we're pursuing a solution that allows companies to really identify the role, right, that AI applied.

然后，这些来源信息将完全以他们自己的品牌呈现，并与AI所扮演的角色相匹配。

And then that provenance is going to be really under their own brand and commensurate with the role that AI played.