AI 事实还是虚构：寓言禁令、Tokenmaxxing、SaaS 末日——与 Ara Kharazian 对谈

公司平均在每位员工身上花在人工智能上的钱到底有多少？

How much are companies actually spending per employee on AI?

人工智能是赢家通吃的游戏吗？

Is AI winner take all?

SaaS已经完蛋了吗？

Is SaaS dead?

Anthropic在和白宫闹出那场风波之后，是不是彻底完了？

And is Anthropic screwed after the Fable five dust up with the White House?

接下来，我们请Ramp的首席经济学家阿拉·哈拉扎尼来帮我们理清这些说法，分辨哪些是事实，哪些是夸大其词。

Let's get to the bottom of these questions, separating fact and fiction with Ramp lead economist Ara Kharazian right after this.

不同人说法不一，但有人说企业里高达80%到95%的人工智能项目都失败了。

Depending on who you ask, between 8095% of enterprise AI projects fail.

要想让人工智能真正为你所用，光靠增加令牌数量是没用的。

To get AI to work for you, you don't need more tokens.

你需要的是更优秀的人才。

You need better people.

它将强大的专有工具与经验丰富的资深工程师结合起来。

A pairs powerful proprietary tools with senior engineers who've seen it all.

有了这种组合，你的项目就不会停滞、不会偏航、也不会失败。

That combination means your project doesn't stall, doesn't drift, and doesn't fall.

它能顺利上线。

It ships.

无论你是需要快速推向市场的初创公司，还是面临复杂遗留问题的企业，Aboard 都能迅速提供你业务真正需要的解决方案。

Whether you're a startup that needs to get to market or an enterprise with complex legacy challenges, Aboard delivers exactly what your business needs fast.

Aboard 是你实现人工智能转型的合作伙伴。

Aboard is your partner for AI transformation.

访问 aboard.com，我们一起打造点什么吧。

Visit aboard.com, and let's build something together.

欢迎收听《大科技播客》，一档深入探讨科技世界及其延伸领域的深度对话节目。

Welcome to Big Technology Podcast, a show for cool edit and nuanced conversation of the tech world and beyond.

人工智能行业里谣言满天飞，各种说法层出不穷，而最好的应对方式，就是回归数据本身，把事实和虚构分开。

There are a lot of rumors flying in the AI industry, a lot of narratives flying in the AI industry, and what better way to attack these than to look at the actual data and separate the fact from fiction.

那我们今天就来聊聊这个话题。

Well, we're gonna do it today.

我们会讨论公司每个员工在AI上的投入有多少，以及所谓的‘token最大化’是不是真有其事；同时我们也会关注本周的新闻——白宫禁止Fable Five，实际上是对Anthropic模型实施了出口管制，这会不会对它造成长期影响？我们可以看看Anthropic上次和国防部闹矛盾时发生了什么。

We're gonna talk, of course, about how much companies are spending per employee on AI and whether tokenmaxing is a thing, but we're also gonna get into the news of the week, which is whether the White House's Fable five ban effectively putting export controls on Anthropic's model will have a long term damage looking at what happened the last time Anthropic had a dust up with the Department of Defense.

今天我们请到了阿拉·哈拉扎尼安。

We're joined by Ara Kharazian.

他是Ramp公司的首席经济学家。

He is the lead economist at Ramp.

他在Ramp经济学实验室发表了很多出色的研究。

He's publishing some great stuff at the Ramp Economics Lab.

我长期关注他的文章，非常高兴能邀请他来到节目。

I've been a reader of his work for a long time, and I'm thrilled to welcome him to the show.

阿拉，很高兴见到你。

Ara, great to see you.

很高兴能再次和你交流。

Great to talk to you again.

第一次上节目，很高兴能来。

Great to be on the show for the first time.

是的。

Yes.

相信这只是开始，以后还会有更多次。

First of many, I'm sure.

那我们先直接聊聊，政府这项Fable Five出口管制对Anthropic业务会带来什么影响？

So let's talk right off the bat about what we can anticipate the impact of the government's fable five export controls on Anthropic's business to be.

对吧？

Right?

因为之前我们已经见过类似的情况，当时国防部把Anthropic列为供应链风险。

Because we've seen a version of this before with the Department of Defense naming Anthropic a supply chain risk.

这次的规模显然更大。

This is obviously on a bigger scale.

就算他们最终撤销禁令，也可能已经造成了一些影响。

And, you know, even if they go back on the ban, there might be some impact.

我老是管它叫禁令。

I keep calling it a ban.

放松出口管制，这本质上就是禁令。

Go back on the export controls, which is effectively a ban.

这里可能还是会有一些影响。

There may be some some impact here.

那我们能期待什么呢？

So what can we expect?

你提到今年春天国防部把Anthropic列为供应链风险，这个例子确实最贴近，能帮我们判断这对Anthropic的业务或市场接受度会有什么影响。

Well, you're right to look at the Department of Defense decision from earlier this spring as the closest recent example that we might use to inform how it's going to affect Anthropic's business or business adoption going forward.

咱们回头看看春天，当时国防部把Anthropic定为供应链风险。

I mean, let's go back to the spring when this Department of Defense labeled Anthropic as supply chain risk.

在大多数行业，对大多数软件供应商来说，一旦被贴上这种标签，企业通常就不会再继续使用这家供应商了。

Usually, in most industries, for most software vendors, when you are labeled as such, businesses are not likely to continue to use that vendor going forward.

所以，如果从最根本的角度来看，我们原本会预期企业会停掉Anthropic的订阅，新客户也不愿签约，企业会明确避开Anthropic——一方面是因为政府提到的安全顾虑，另一方面也是为了能继续参与政府的合同项目。

So if you're just gonna think about this from first principles, we would have expected businesses to shut off their Anthropic subscriptions for new businesses to not wanna sign up, for businesses to explicitly not want to use Anthropic going forward, both because of a true security concern that the government is citing, and because they might wanna engage with the government on government contracts.

但今年春天实际情况并不是这样。

That's not what happened earlier this spring.

恰恰相反，今年春天我们看到企业对Anthropic的采用率反而大幅上升。

If anything, this spring was when we saw Anthropic's adoption really accelerate with businesses.

这背后是因为去年Cloud Code成功发布，Anthropic终于开始吸引越来越多非技术用户。

It was coming off the heels of, you know, the successful launches of Cloud Code last year, finally starting to move into a more popular posture with non technical users.

就在上个月，根据Ramp Data的数据，Anthropic已经成为美国企业使用最广泛的AI模型。

And just this past month, we showed that Anthropic is now the most popular AI model used by US businesses, according to Ramp Data.

所以，之前那些预设的假设都没能成立。

So, several of those assumptions didn't come to be.

我觉得主要有两个原因。

I think two main reasons why.

第一，大多数企业根本没把国防部的这个标签当回事。

One is that most businesses didn't really seem to take the Department of Defense's label very seriously.

说白了，大家心里想的是：国防部是这么说，但这依然是目前最好的模型，而且现在还是。

You know, it was kind of like, okay, yeah, the Department of Defense is saying this, but this was one of the best models available, still is the best model available.

企业里用的人很多。

It's popular with businesses.

第二个原因是，国防部在接下来几周里接连承认，会对那些坚持使用Anthropic的企业内外部都给予豁免，这严重削弱了它的公信力。

Second reason is that Department of Defense lost a lot of credibility over the continuing weeks when it acknowledged that it would be issuing exceptions both internally and externally for businesses that wanted to use Anthropic anyway.

所以，如果回看今年春天的例子，事实可能恰恰相反——国防部的供应链风险标签反而推动了Anthropic在企业中的普及。

So, if you're going to go back to the example from earlier this spring, if anything, it's probably the case that the Department of Defense's supply chain risk labeling accelerated Anthropic's adoption with businesses.

而如今，这个标签让Anthropic在和OpenAI的竞争中占据了非常独特的地位。

And this label now, anything puts it on a very interesting competitive standpoint with OpenAI.

哪家的模型更厉害？

Who has the better model?

大概是那个被联邦政府认为强大到必须严加管控的模型。

Probably the one that the federal government has suggested is so powerful, it must be controlled.

这种品牌效应也很强。

And and there's brand strength in that.

没错。

That's right.

我们这个节目经常讨论Anthropic如何强调安全，以及它在联邦政府要求它做某些事时坚持自己的立场，这些举动反过来引发了政府的一系列反应，这一切是不是一种营销手段。

So we talk on this show all the time about whether Anthropic's positioning around safety and the fact that it's, you know, taking this steadfast approach when the federal government has asked it to do something, and that's led to some of these actions by the federal government, whether that's marketing.

不过咱们先不管它的初衷。

Let's put the intent aside, though.

不管它是不是有意为之，这种局面最终可能反而推动了它的业务发展。

It could end up being a boost for its business, you know, whether it wanted this to happen or not.

我老在想，今天是星期一，我们正在录音，每次用Claude的时候，它都会弹出提示说：‘Fable目前不可用’。

I keep thinking about the fact that, at least we're recording Monday, and I keep thinking about the fact that, you know, when you're in Claude, it says, hey, Fable is not available right now.

对吧？

Right?

确实如此。

And Yeah.

你在X上能看到好多帖子，人们都说：‘那五分钟的Fable体验，让我看到了它真正的潜力。’

You're seeing all these posts on x about people who are like, oh, those five minutes with Fable really give me a glimpse into what this could be.

你想想，等它重新上线时，这种‘禁果效应’肯定会把大家对Fable的兴趣推到极致——毕竟，连政府都觉得这模型太强大，不能随便让人用，用户的好奇心只会更旺。

And you'd imagine that when it gets turned back on, that forbidden fruit effect just kinda drives the interest in Fable, you know, the the model too powerful for the government to let you let let you use, the interest will probably just go through the roof.

不过我们也要注意到，过去六个月里，使用Anthropic模型的体验，就是经常遇到模型离线的情况，用户已经慢慢习惯了。

Well, let's also note that part of Anthropic's the product experience of using Anthropic models, at least over the last six months, has been getting really comfortable with models being down all the time.

因为Anthropic的算力限制太严重，用户经常碰到崩溃提示，或者直接看到‘模型当前不可用，请稍后再试’这类API错误信息，这已经成了家常便饭。

Because Anthropic has so many compute constraints that it is very common to be a user of Anthropic's models and to be hit with a a crash notice or something that just says, hey, the models are down right now or try again later on API error.

按理说，这种情况应该会把大量用户推到OpenAI的模型上去。

And yet, we would normally expect that that would drive a lot of users over to OpenAI models.

OpenAI的模型算力问题要少得多，稳定性也高不少。

OpenAI models are, you know, OpenAI doesn't have nearly as much of the compute issues that Anthropic does.

但过去几个月，我们并没看到用户大规模转向OpenAI的现象。

And yet, we haven't seen that kind of switching behavior yet over the past couple months.

也许有少数原本用Claude写代码的用户，转去试了OpenAI的Codex，但这种情况非常有限。

So maybe some marginal switching happening from people who would who are using quad code now switching over to OpenAI's codex.

但说真的，作为研究者和用户，我不禁想问：Anthropic这种‘禁果效应’还能撑多久？总有一天，那些真正需要干活、付了钱就指望模型稳定运行的用户，会忍无可忍的。

But if anything, you know, it makes me wonder both as a researcher and user of these models, how long is this forbidden fruit effect of Anthropic's model is gonna last before it starts to turn users off who really just need to get to work and use the models that they are paying for and and have them work as expected.

是的。

Yeah.

你是经济学家，你来说吧。

I mean, you're the economist.

我不清楚。

Don't know.

有没有什么经济理论能解释，为什么用户会继续用一个总出问题的供应商，而明明另一个功能差不多、甚至更稳定的供应商就在那儿？

Is there an economic theory that explains why people would stick with a vendor that has constant interruptions even if there's another one with, you know, just as good or on par capabilities that doesn't have Yep.

这种不稳定的状况？

The turbulence?

我们发现，这些模型的用户黏性比我们原先想的要高得多。

Well, what we found is that these models are a little bit stickier than we thought they would be.

我们总说这些模型是标准化产品，换起来应该很容易。

And we talk about them being these commodities that, oh, you can just switch between one model and the other.

从模型本身的角度看，也许确实可以这么想。

And maybe at the model level, you really can think of things like that.

但说到企业里实际使用AI的员工，更好的模型并不一定就能让他们换平台。

But in terms of how many employees of firms are actually using AI, you know, better models don't necessarily create the switching for for employees and users.

关键是这些模型周边的产品体验。

It's about the product experience around those models.

你知道，Claude Code 能这么成功，不是因为它用的是 Claude 的模型，而是因为它深度融入了你的工作流程，以一种非常高效、主动的方式运作——它是第一个让工程师能独立完成多步骤任务的工具，而不用一直像哄小孩一样去引导聊天机器人。

You know, Claude code was so successful not because it was powered by the models of Claude, but because it was was integrated into your workflows in a very effective and agentic way, that it was the first model and first experience that allowed a engineer to execute on multi step tasks without, you know, babying a chatbot the entire time.

所以，模型的细微改进确实重要也有帮助，但这些并不是 Cloud Code 能快速增长的全部原因。

And so such that incremental improvements of the model are important and helpful, but they weren't the whole story for the actual growth of Cloud Code.

因此，你可以想象，正是这种对工具和所处软件环境的依赖，让 Cloud Code 和 Codex、OpenAI、Anthropic 之间形成了某种粘性。

And so you can imagine that to be driving some of the stickiness between Cloud Code and codex, OpenAI and Anthropic, is that people get really used to the tools, the software that they're in.

用户更喜欢其中某一个提供的使用体验。

They like the experience that one provides over the other.

我觉得还存在一种品牌效应：Anthropic 那套 AI 安全的立场，不管你喜不喜欢，确实有一批人很欣赏他们至少在表面上认真对待 AI 影响的态度，不管他们是不是真能当好这个守护者。

I also do think there is going to be a sort of branding effect, where Anthropic's AI safety posturing, you know, love it or hate it, there are people who really do like that they're the company that at least postures itself as being thoughtful about the effects of AI, whether or not they are the right guardians of that.

对。

Yeah.

联邦政府说不定无意中帮了 Anthropic 一把，通过强化它的安全主张，让它显得更可信，从而制造出一种‘禁果效应’。

It's possible that the federal government is sort of, you know, disrupting Anthropic, giving it a helping hand by sort of making its safety messaging seem more legitimate and, again, giving it that forbidden fruit effect.

好的。

Alright.

再问你一个经济学角度的问题，然后我们就进入你的数据。

One more economics style, you know, thought question for you, then we can get into your data.

最近有报道说，OpenAI打算大幅降价。

You know, there have been recent reports that OpenAI is looking to drop prices drastically.

我们在周五的节目里聊过，我的想法是，他们其实是在考虑客户的长期价值。

And we talked about it on the Friday show that my thought was basically, they are looking at lifetime value of potential customers.

如果他们降价，就能让使用者习惯像Codex这样的工具，然后利用你提到的这种黏性留住他们。

And if they were to drop their prices, they could get people used to using, let's say, a codex and then keep them with the stickiness that you're talking about.

所以对他们来说，先降价吸引用户，再指望他们留下来，这其实是个挺聪明的策略。

And so therefore, it would be a good move on their end to say, we're gonna drop prices to win people over, and then hopefully, they'll stick with us.

你怎么看？

Your thoughts.

这部分想法很自然。

Part of it, think, very natural.

这个市场竞争太激烈了，往往几个月内，一个新玩家就能超越原本很受欢迎的厂商。

This is an extremely competitive market where it's very common to see in a matter of months, a newcomer take the lead over a relatively popular provider.

软件行业以前就见过这种情况，但在人工智能领域尤其明显。

You know, we've seen this in software before, but it's especially true in acute in AI.

比如 Cursor 和 GitHub Copilot 之间的竞争就是个例子。

I mean, we saw this with Cursor versus GitHub Copilot.

几年前，当 AI 代码补全刚出来的时候，GitHub Copilot 是企业用户的首选工具。

You know, when coding agents came out, at least when AI Code Autocomplete came out two, three, four years ago, GitHub Copilot was the enterprise tool.

那时候几乎人人都在用它。

That's what everybody used.

说起来也不奇怪，它背后还有微软支持。

Unsurprisingly, it was also backed by Microsoft.

结果 Cursor 一出来，大概一年内就抢走了大部分市场。

Cursor comes out, and within about a year, it has the majority of the market.

接着，Quad Code 又出现了，现在它的市场份额已经反超了。

Then, by the way, quad code comes out, and then now quad code has majority of the market.

后来，OpenAI 和 Anthropic 之间也发生了类似的情况。

And then so we saw a similar story with OpenAI Anthropic.

根据 Ramp 的数据，我们通过旗舰研究项目 Ramp AI Index，持续追踪美国企业使用人工智能的情况，特别是那些通过订阅或代币直接付费的企业比例。

You know, in Ramp's data, we have been tracking using our flagship research, Ramp AI Index, the share of firms in The United States that are using AI, at least paying for it through subscriptions or tokens directly.

然后我们进一步细分，看看企业们具体在为哪些模型付费。

And then we break it out by which model they're paying for.

在人工智能商业化的大部分时间里，也就是从 2023 年开始，OpenAI 明显占据主导地位，美国大约有 20% 到 40% 的企业都在为 OpenAI 付费。

And for most of AI's commercial existence, 2023 onwards, OpenAI was clearly the dominant player, somewhere hovering between 20 to 40% of businesses in The US were actively paying for OpenAI.

这个比例一直没什么大的变动。

And it wasn't really budging much.

增长非常缓慢，尤其是 2023 年和 2024 年以后。

Was just a very gradual increase, particularly 2023, 2024 onwards.

当时几乎没人关注 Anthropic，虽然它在技术圈里挺受欢迎，但整体上并不被视为市场上一个广为人知的竞争对手。

No one was really thinking about Anthropic, which was popular with technical users, but otherwise wasn't this sort of broadly understood competitor in the market.

到了 2025 年下半年，我们看到使用 Anthropic 模型的企业比例逐月稳步上升。

Second half of twenty twenty five, we see month over month percentage point increases in the share of firms that are using Anthropic's models.

上个月，Anthropic在企业采用率上正式超越了OpenAI，一下子成了焦点。

Coming to the forefront last month when open when Anthropic overtook OpenAI and actual business adoption.

现在，有大约41%的美国企业使用Anthropic，而使用OpenAI的企业占39.5%。

So now Anthropic sits at about 41% of US firms are using Anthropic 39.5% of firms are using OpenAI.

Anthropic的增长势头还在继续。

Anthropic is still growing.

OpenAI则基本持平，没什么大变化。

OpenAI is relatively flat.

就连那些最早采用AI的行业，Anthropic的增长也在加速，而OpenAI依然没什么起色。

And then even in the sectors that are early adopters of AI, we're seeing that growth continue to grow and accelerate, while OpenAI holds relatively flat.

对，没错。

Right.

这是一个非常动态的市场，按理说所有玩家都会拼命竞争。

So this extremely dynamic market where you could expect all of the involved players to want to compete with each other.

但我认为，现在大家只盯着OpenAI和Anthropic，其实还有很多被低估的玩家。

But I actually think that, you know, right now the focus on OpenAI and Anthropic, but there are a lot of players that are underrated.

我觉得谷歌被严重低估了。

Google, I think is extremely underrated.

它可能会成为这场竞争中最大的赢家，但现在没人提到它。

And think might end up being one of the big winners here that no one's talking about.

明白了。

Okay.

不过简短说一下，价格战或者价格打压的问题。

But then briefly, the price the price war or the the price undercuts.

你觉得光靠降价就能打破用户的黏性吗？

Do you think that's enough to dislodge the stickiness?

我知道你没有相关数据，但肯定有个公式能说明价格如何影响

I know you don't have the data on it, but there has to be some formula out there about, you know, the price plays into

使用率。

usage.

我觉得降价确实够用。你看，迟早会迎来一个转折点，企业会持续要求更好地控制AI支出。

Well, so I do think it's gonna be enough to I think, look, we're gonna come to a head at some point where we know businesses keep demanding some better control over AI spend.

你知道吗，以前有一阵子大家都在搞‘令牌最大化’，觉得得尽可能多花钱买令牌。

You know, we went through a sort of tokenmaxing era where everyone was talking about, okay, we need to spend as much as possible on tokens.

但现在进入了一个新阶段，企业开始说：我们得管住令牌的开销，至少得弄清楚该花多少。

And then now we're in this sort of new era where businesses are saying, hey, we need to actually rein in token spend or at least understand where it should be.

开销可以继续涨，但至少我们得掌握主动权。

It can continue to rise, but we at least need some control over what it is.

无论是OpenAI还是Anthropic，都没推出能让企业真正管理令牌支出的产品，也没设计任何机制来鼓励企业控制成本。

Neither OpenAI nor Anthropic have built products that allow firms to actually manage their token spend, nor have they built products that incentivize firms to keep those costs under control.

说白了，OpenAI和Anthropic反而希望企业花得越多越好。

If anything, OpenAI and Anthropic are incentivized to have firms spend as much as possible.

到目前为止，它们主要靠降价来竞争。

So, so far, you know, they can compete on price reductions.

但企业真正想要的，其实是某种程度的掌控力。

But really what firms are asking for is some degree of control.

比如说，能不能开发个功能，让我们能智能分配任务，把每个任务自动分配给最高效、最划算的模型。

You know, maybe that means, hey, build us something that allows us to smart route tasks over to the most performant, but also most efficient model for that task.

其他竞争对手已经提供了这样的功能。

Other competitors are offering that.

但通常不是OpenAI和Anthropic。

It's not usually open eye and Anthropic though.

所以你的数据显示，就像你之前说的，Anthropic在企业支出上已经超过了OpenAI。

So your data shows that like again, like you said, Anthropic has overtaken OpenAI with business spend.

简短说一下，Ramp经济学实验室受到的批评，不管有没有道理，是说你们研究的主要是使用Ramp卡、依赖Ramp服务的公司，这类公司往往偏向初创企业和科技型公司，不能代表整个经济生态。

Very briefly, Ara, the criticism of the Ramp Economics Lab, whether well placed or not, has been that, yeah, you're looking at, you know, companies with Ramp cards, using Ramp, tends to lean towards, you know, sort of startups and tech forward companies, and it's not representative of the full economy.

你怎么看？

Your thoughts.

我通常这么看：我们的数据集在各个行业中的分布其实相当均衡。

The way I normally think about this is that so we have actually a pretty good distribution in our dataset across sectors.

不过，不管哪个行业，使用我们平台的企业本身都更偏向科技前沿，而且它们本来就会用Ramp这类工具来管理支出。

However, no matter the sector, the businesses on our platform are inherently more tech forward, and that they're using something like Ramp to manage their spend in general.

我觉得这反而是个优势，原因有几点。

I do see that as a strength for a couple reasons.

第一，人工智能是一种非常新颖且尚在初期阶段的技术，其他数据集根本无法有效追踪。

One, AI is this very new and nascent technology, and one that is not effectively tracked by other datasets.

政府的数据集也一样跟踪不到位，它们在调查企业AI支出时方法本身就有问题，而且所覆盖的企业本身就有选择偏差。

It's not effectively tracked by government datasets either, which have their own set of criticisms, as far as how they are surveying firms about AI spend, and also the firms they are serving themselves being a little bit self selected.

实际上，AI支出恰恰更集中在这些技术前沿的企业身上，所以如果你想真正了解企业如何使用和投入AI，反而更应该关注这些走在前沿、带头推动这项技术的公司。

AI spend, if anything, is skewed over toward these tech forward businesses, such that if you do want to understand how businesses are spending and operating on AI, it actually behooves you to look at these very forward thinking businesses that have been leading this charge.

它们更有可能是这项技术的早期使用者。

They are more likely to be early adopters of this technology.

至于它们是否代表了美国普通企业的平均水平——我们很清楚，它们并不代表。

And whether or not they are representative of the average firm in The United States, we know they are not.

未来几年，大多数普通企业很可能会越来越接近这些前沿公司，而不是现在这种普遍状态。

It's more likely than not that the average firm will look more like these firms in a couple years than they look like the average firm today.

所以，如果你想放眼未来、提前洞察趋势，这种数据集反而更值得参考。

So I think if you want to be forward looking, you want to look into the future a little bit, you probably want to use this kind of data set.

说句实话，我觉得在很多方面，我们其实低估了AI的普及速度。

For what it's worth, I actually think that in many ways we underestimate adoption.

好的。

Okay.

行，我们后面再详细聊方法论。

All right, we'll get into more methodology later.

不过既然这些公司都是走在前沿的，咱们不妨深入看看你的数据，因为关于AI资源浪费的说法不少，但你的数据提供了一个不太一样的视角，我觉得值得好好讨论一下。

But given that these are the forward looking companies, let's get into some more of your data because I think that there have been some narratives about token waste that your data has a little bit of a different perspective on that I think we should we should just discuss.

你之前提到过，成为一家AI先行公司需要付出的成本。

So you've talked about what is the cost of being an AI pilled company.

每个员工每月是7449美元。

It's $7,449 per employee per month.

你说过，前1%的公司每月在每个员工身上花这么多，而前10%的公司每月每人花611美元。

So you said the top 1% of firms spend that much per unemployed per month, and the top 10% spend $611 per employee per month.

而中位数的公司，每月每人只花11美元，差不多就相当于一个企业版ChatGPT账号或者云服务的订阅费。

And the median firm spends just $11 like the cost of an enterprise seat on enterprise chat GPT or a cloud subscription.

一个月花7400美元在一项技术上，对单个人来说，这在科技领域确实相当罕见。

Now $7,400 a month is pretty high, on a technology, like a sing for a single person to spend that much on a technology seat or license is somewhat unheard of.

不过，当我们看到那些新闻标题，说某家公司一不小心就花掉了五亿美元，你的数据其实展现了一个不一样的画面——公司们现在并没有肆无忌惮地花钱。

However, when you think about the headlines that we've been seeing a company left caught on and spent a half billion dollars in a month, Your data actually presents somewhat of a different picture that companies aren't sort of spending unrestrained right now.

他们更像是在试探水温，而不是直接全身心扑进去，把算力令牌当废纸一样挥霍。

They seem to be, know, sort of dipping their toe in the water as opposed to going all the way in and, you know, spending tokens like they're going out of business?

说实话，AI支出是我们数据里增长最快的开支类别。

Well, look, AI spend is the fastest growing spend category we've ever observed in our data set.

可能是有史以来企业支出中增长最快的之一，当然这得看你怎么定义‘企业’，要是往前追溯到远古时代，那另当别论。

Probably one of the fastest growing spend categories for businesses ever, depending on how far back you go into what a business is defined as in prehistoric times.

我真想不到还有什么开支能比这个涨得更快。

I couldn't imagine any spend ramping faster than this.

没错。

Yeah.

那还能是什么别的东西呢？

What which what could it be?

正是如此。

Exactly.

从2025年1月到2026年5月，也就是上个月，每家企业的AI代币支出增长了15倍。

And so since January 2025 through May 2026, so last month, per business spend on AI tokens is up 15 x.

而且这还只是针对那些本来就在用AI的企业。

And that's amongst firms that were already spending on AI.

但与此同时，对大多数公司来说，AI支出本身其实并不算大。

At the same time, AI spend itself isn't really that meaningfully large for most businesses.

所以虽然增长很快，但对那些AI支出最高的25%企业来说，这项开支只占非薪资支出的2%左右。

So it's grown a lot, but for the top quartile of firms, that are spending on AI, top 25%, it's only about 2% of business spend excluding payroll.

如果算上薪资，那占比大概只有1%。

It'd be at 1% if you were to include payroll.

所以说，增长确实非常快。

So it's grown a lot.

正因如此，我们才不断听到企业高管们担心：我该怎么管理这种增长？

That's why we get all these concerns from, you know, company executives about, you know, how do I manage this growth.

但就实际金额而言，它还是相对很小的。

But as far as its actual level, it's relatively small.

你会发现，当人们说企业正在削减AI支出时，可能只是在某些部门这么做，比如更谨慎地选择使用的模型，或者确保团队没有失控的预算。

So you'll notice, you know, when people talk about firms pulling back on AI spend, You know, they might be pulling back on AI spend in some parts of the firm, they might be more mindful about which models they're using or making sure that teams don't have uncontrolled budgets.

但如果你真去看看最近几个月企业的AI支出，就上个月，环比仍然增长了14%。

But if you actually look at firms spend on AI in the last couple months, just last month, it still increased 14% month over month.

所以我们的平台上有明确的数据表明，企业正在做出更注重成本控制的决策。

So there are there's clear evidence on our platform to that firms are making more cost discipline decisions.

我们注意到，越来越多的AI支出正从OpenAI和Anthropic转向这些开源平台。

You know, we've seen an increase in AI spend being routed away from OpenAI and Anthropic and over to these sort of open source platforms.

上个月，DeepSeek是增长最快的供应商之一。

Last month, DeepSeek was one of the fastest growing vendors on ramp.

但通过这些开源渠道的AI支出，占比依然很小。

And yet, it's still a very small share of AI spend that is actually going through those rails.

总体来说，使用开源平台的企业本身就不多。

It's a relatively small share of businesses that are using open source platforms in general.

而绝大多数的支出仍在持续增长。

And the vast majority of spend happening is still rising.

这些成本控制措施确实重要，但只是在边缘做调整，根本赶不上AI支出持续上升的速度。

So those cost discipline measures are important, but they're really just occurring on the margin, and they're not happening fast enough to pull back the rising slope of AI spend.

是的。

Yeah.

你觉得会不会有那么一天，有些公司花在AI上的钱，比花在每个员工身上的还多？

Do you think that there's gonna be a moment where some firms start to spend more on AI than they do spend, you know, on they spend more on AI per employee than they spend on employee.

比如，我不确定这有多准，但听说有人算过，运行Fable或者Mythos API每小时要花600美元。

For instance, I don't know how accurate this was, but I think like we're actually trending in this direction where someone figured out how long how much it would cost to run the fable or the mythos API.

要是算上一整年，那就是120万美元。

And they found out it was something like $600 an hour, where if you like multiply that over a year, it's 1,200,000.

所以，当你看这个趋势时，你觉得我们真会走到一步，让公司花在每个员工身上的AI费用，超过他们付给员工的工资？

So so so what do you think when you think about the trajectory, do you think we're gonna get to a place where people end up spending more on AI per employee than they spend on employees themselves?

我觉得肯定会有公司走到这一步。

I imagine some will.

对。

Yes.

对一些公司来说，我觉得这么做非常合理。

Some for some firms, I imagine it makes a lot of sense.

对吧？

Right?

但对普通公司而言，真的很难判断该花多少钱才合适。

But for the typical firm, there's really it's really hard to benchmark where you should be.

目前顶尖的1%公司，每月在AI上的投入高达每人7500美元。

And with so the top 1% of firms spend $7,500 per person per month on AI.

但这只是顶尖的1%。

But that's the top 1%.

你可以想象，这群公司技术含量极高，它们用AI不仅是为了员工日常使用，还可能用于公司底层基础设施。

So you can imagine that's a pretty tech heavy group that actually may include a lot of firms that are ultimately using AI not just for the employees usage, but also for the underlying infrastructure of the firm.

也许它们已经开发了很多内部工具。

Maybe they've built a bunch of internal tools.

对吧？

Right?

所以整体上就平衡了。

So it all gets balanced out.

而软件工程师的花费通常是两倍，每月一万五千到一万六左右。

And software engineers are double that, typically 15,000 or close to 16,000 a month.

不过，这还得看公司具体情况。

Well, again, would depend on the firm.

因为如果你看看我们平台上典型的公司——再说一遍，这是Ramp的数据。

Because you look at the typical firm on our platform, and again, this is Ramp.

这是一个相对技术前沿的平台。

So relatively tech forward platform.

对吧？

Right?

而中位数公司的支出，每人每月只有大约十一美元。

And the median firm is only spending about $11 per employee per month.

没错。

Right.

所以你看，这其实是个聊天工具的订阅服务。

So, you know, that's bear that's a chat subscription.

这属于OpenAI和Anthropic那种基础档的订阅，可能稍微贵一点点。

That's like one of the low level OpenAI and Anthropic subscriptions, maybe a little bit more on the margin.

这也再次提醒我们，现在还处于非常早期的阶段——绝大多数企业都还没跟上，这也彻底改变了我们的研究方式。

So, you know, it's another reminder of sort of how early we are, In that, the vast majority of firms and this has transformed our research approach.

过去一年半，我的研究一直集中在估算人工智能的经济影响上。

Because for a long time, the last year and a half, my research is focused on trying to estimate the economic impact of AI.

但如果在生产率数据里根本找不到痕迹，人们还是不确定AI最终能带来多大收益，那我们只能从最基础的问题入手：到底有多少企业在用AI？

But if there's no way to find that in productivity statistics, if people are still not sure what the ultimate gains of AI is gonna be, then really the only way to start is, hey, how many firms are using AI?

所以，我们最初版的Ramp AI指数，就是想回答这个问题。

And so our original version of Ramp AI Index is just that.

简单说，就是看看美国有多少企业真的在购买AI服务，而且是月复一月地持续付费，至少能由此判断：这东西到底有没有价值？

It's hey, what is a share of firms in The US that are even buying it and buying it month over month, to try to get some way of at least approaching the question, hey, is this valuable?

现在，超过一半的企业已经在使用AI了，54%的企业至少在某种形式上使用或为AI服务付费。

Now, more than 50% of firms are using AI, 54% of firms are using AI in some way or at least paying for it.

所以问题也开始转变了，不再只是谁在用AI、有没有价值，而是他们到底怎么用？

So, question gets to transform a little bit, not just who's using AI, is it valuable, but how are they using it?

他们在这上面花了多少钱？

How much are they spending on it?

怎样才算真正有效地使用这些模型，并在组织里顺利落地？

What does it mean to be an effective user of these models and to deploy it effectively through your organization?

因为AI真正有意思的地方在于，它的使用分布很不均衡，某些行业比其他行业更倾向于采用它。

Because that's what's really interesting about AI too is that it is unevenly distributed, and that certain sectors are more likely to adopt it than others.

这些产品本身的设计和推广也不均衡，最先进、最高效的AI应用，往往都是为特定岗位设计的，比如编程助手。

The products themselves are also unevenly designed and distributed, and that some of the best, most advanced usages of AI are designed for certain job categories, coding agents.

这显然是目前最直接、最能提升生产力的AI商用方式。

It's like the most obvious commercially advanced way that you could use AI to be productivity enhancing.

但对大多数公司来说，这种功能根本不存在。

And yet that doesn't exist for most other firms.

你可能会想，其他很多岗位其实也能提升效率，但市面上的产品还没发展到能让普通用户一眼看懂怎么用的程度。

You know, maybe there's productivity enhancements you can find in a lot of other jobs, but the products themselves are not well developed to make that clear to like the average user.

所以，人工智能的分布并不均衡。

So, AI is unevenly distributed.

因此，企业要想弄清楚什么是AI的正确使用方式，并以此为标杆，最终会很困难，尤其是因为AI带来的效益，比如生产力提升，不太可能在头几个月就显现出来。

And so, it's ultimately gonna be difficult as a firm to identify and benchmark against what good usage of AI means, especially because the effects of AI, the productivity gains of AI, are not likely to show up in your first couple months.

在组织内部，甚至在个人工作流程中推行AI，显然都需要一个学习过程。

There's clearly a learning curve to implementing AI throughout your organization and even through your own personal workflow as an employee.

除了学习曲线之外，还有一个最低采用门槛。

And then beyond the learning curve, there's also this sort of minimum threshold of adoption.

我不觉得有人会指望，只要每个人都有个聊天机器人，就能带来巨大的经济效益。

Where I don't think anyone really expects you to get massive economic gains from everyone having a chatbot.

但要是每个人都能为自己的工作定制专属的代码工具，那就更有吸引力了。

But everyone having their own cloud code for their job is much more compelling.

可如果你只是用了一个月两个月的聊天机器人，就觉得AI没用，然后就放弃了，那你是看不到这一点的。

But you wouldn't know that if you're just using a chatbot for like a month or two and then you write off AI because it's like, you know, what's the point of this?

但你看看你研究里那些曲线图，对吧？

But you look at the curves that you have in your research, right?

这只是一个数字，就像三根竖起来的冰球杆一样。

And it's just a number, it just like three hockey sticks, right?

如果你看看那些使用AI最顶尖的1%公司、前10%公司，以及中位数公司的每月人均支出，你会发现前两个几乎是平缓上升的线，但突然之间，三者全都急剧飙升。

If you look at the spend per employee per month of the top 1% of companies using AI, the top 10%, and the median company using AI, it's like legitimately like a like a lightly sloping line and then all of a sudden it shoots up in all three of those categories.

所以你其实是在试图找到答案——这项技术到底哪里有价值？

So you're trying to you've you said you're trying to get to the answer of, you know, how is is this technology valuable?

我很想听听你的看法，这些数据到底意味着什么？虽然它既有定量的成分，也有定性的部分，对吧？

And so I'd I'd love to hear your perspective on what these numbers mean Even though it's more, I guess, quantitative or qualitative and quantitative, right?

是的。

Yeah.

我们看到这些支出增长，是不是说明公司已经在AI上获得了回报？

Do the fact that we're seeing these spend increases mean that companies are seeing an ROI on AI?

还是说，这仍然只是出于害怕错过的心理？

Or is it still potentially in the sort of FOMO stage?

我认为，企业根本没有理由仅仅因为义务感或害怕错过就花这么多钱。

Well, I really am of the school of thought that businesses have no reason to be spending this much money just out of a sense of obligation and FOMO.

我明白，如果我们说的是人买股票，这种情况我能理解。

You know, I get it if we're talking about people buying stocks.

但企业对软件做出相当大的投资，这可不是随便说说的——他们不仅投入了大笔资金，每个月还在持续增加支出。

Right, but like companies making fairly large investments in software, you know, that you can just talk about it externally without making not only the large investments in software, but month over month increases in how much they are spending.

所以我一般认为，如果企业真这么做了，那肯定是从中发现了实际价值。

So I'm generally of the school of thought that if firms are doing this, they must be finding some value out of it.

不过，我们也可以从一些量化数据里寻找这方面的证据。

But there are some places that we look quantitatively for that evidence.

这些数据也让我们意识到，AI市场和大多数软件市场很不一样。

And also informs our thinking that this is different from most software markets.

比如，AI投入最高的那些企业，从不只依赖一家供应商。

So one is that most advanced, the most advanced spenders on AI don't lock in with one vendor.

这跟我们通常对软件的认知完全不同——比如你用客户关系管理软件，一般就选一个长期用。

Right, so this is fundamentally different from how we typically think of software, where it's like, hey, if you're using a CRM, you're gonna use one CRM.

也许你一开始会试用几家供应商，但最终还是会锁定一家。

Maybe you'll like experiment with a couple providers, but ultimately you're gonna sign with one.

AI的情况可不是这样。

That's not the case with AI.

AI支出排名前1%的公司平均会用八个供应商，而中位数可能只用两个。

The top 1% of spenders on AI use eight vendors on average, whereas the median maybe uses two.

这里说的供应商，特指大型语言模型提供商，可能还包括一些AI基础设施公司。

And that's vendors fairly narrowly defined as LLM providers, and maybe some AI infrastructure companies.

而且这也不只是简单的试用。

Now, it's also not just experimentation.

确实有一部分是持续在试错，因为AI模型不断推出，更新又快又频繁，如果你的公司正在有效使用和部署AI，那最好能付费接入所有主要模型公司，这样就能随时切换到最适合当前任务的模型。

You know, there's some amount of it that's always like the sort of continuing experimentation, where it's like, okay, the AI models come out, but they also change so much so frequently, that if you are an organization that is using and implementing AI effectively, you probably wanna have paid access to all the major model companies, so that you can switch to the most effective model for whatever task makes sense.

或者当新模型上线时，能快速评估它是否适合某个工作流程。

Or when a new model comes out, you can see if it makes sense for this workflow or that workflow.

对前1%的公司来说，这才是真正懂AI的用法。

That's what being a good AI user often means to these firms in the top 1%.

但对很多企业，以及负责采购AI的业务人员来说，这种想法还很陌生。

But that is an unfamiliar idea for many businesses and frankly, business people at firms who are buying AI in charge of procurement.

我们公布OpenAI和Anthropic的采用率时，经常有人问这个问题。

I often get the question when we report OpenAI versus Anthropic adoption rates.

我该选OpenAI还是Anthropic？

Should I buy OpenAI or Anthropic?

但如果你天天用这些模型，听到这种问题其实会很惊讶，因为你会直接说：你应该两个都试试，最好两个都能用。

And then when you which, you know, if you're someone who uses these day to day, these models day to day, it's like a very surprising question, because you would just say, well, you should try both, you should probably have access to both.

同时使用两个平台其实没那么贵。

It's not that expensive to have access to both.

所以这个问题本身听起来根本不像个合理的问题。

And so the question itself doesn't really register as a sensible question.

但如果你用传统软件采购的思维来看待AI，就很容易冒出这种疑问。

But if you are applying the typical common practices of software procurement to AI, you will find yourself asking that.

我觉得这只是一个完全不同的市场，大家还没适应过来。

I think it's just a fundamentally different market that the people are not used to.

而这正是推动大量支出背后的原因。

And that's what ends up driving a lot of the spend.

但我认为，如果企业得不到实际好处，他们是不会跟上这种高级AI应用的步伐的。

But I don't think firms would move along that advanced AI adoption curve if they weren't getting some benefit.

你总不能一直不停地换新供应商吧。

You know, you're not going to keep signing up for new vendors.

而且，你大概也不会一直续费给那些供应商。

You're also probably not going to keep renewing vendors.

从我们的数据来看，那些持续使用AI服务的企业，续约率每年都在上升，说明他们更倾向于长期使用已有的供应商，而不是频繁更换。

And if anything in our data set, we see that renewal rates increase year over year with firms that are So more they're more likely to stick to the vendors that they've been using as opposed to switch so frequently.

关于这一点，我有个问题。

One question about that.

这其实挺让人惊讶的。

It's kind of remarkable.

对吧？

Right?

如果你看看OpenAI和Anthropic的收入曲线，就能发现它们确实符合这个趋势。

If you look at the graph of revenue that you're seeing from OpenAI and Anthropic, it is it follows that.

我的意思是，你肯定会这么想。

I mean, you would imagine.

对吧？

Right?

它也应该呈现出那种急剧上升的曲线形态。

It was gonna follow that hockey stick as well, that that type of curve shape.

有意思的是，尽管公司把预算分散在两到八个供应商之间，但这两个却能持续增长，你不觉得吗？

Isn't it interesting that even as companies spread their spend across, know, two to eight vendors that those two have been able to grow the way that they have?

是的。

Yeah.

这是因为 adoption 是按公司层面统计的，但公司内部其实差异很大，谁在用 AI、怎么用，都不一样。

Well, it's because adoption, you know, we measure at the firm level, but then within the firm, there's a lot more heterogeneity around who's using AI and how.

再往个人层面看，差异就更大了——我在这播客里不该用‘异质性’这种经济学术语，说‘差异’就行。

And then within the person, there's even more I should stop using economics terms like heterogeneity when I'm on this podcast, variation.

对。

Yes.

你知道，不同团队还处在采用曲线的不同阶段。

You know, you have different teams that are still on different parts of the adoption curve.

而且每个团队里，个人在做不同任务时，也处在采用曲线的不同阶段。

And then you have individual people within those teams that for different tasks, they're on different parts of the adoption curve.

因为大家还在摸索怎么上手。

Because people are still figuring out how to onboard.

这正是我想说的学习曲线问题。

This is what I mean about the learning curve.

公司有学习曲线，团队有学习曲线，每个人在做具体任务时，也在经历自己的学习曲线，琢磨着这事儿到底能不能用AI高效完成。

The firm is on a learning curve, the teams are on a learning curve, and then the individual itself is on a learning curve for their specific tasks, trying to figure out, hey, can this task actually be done effectively?

我有没有真的试过这个任务呢？

Have I even tried this task or not?

我不是说所有事情都能交给AI来做。

Not to say that everything can be done by AI.

我觉得不是所有事情都能。

I don't think everything can be.

但我认为，你越去尝试，就越能弄清楚它擅长什么、不擅长什么。

But I do think the more that you experiment with it, you will find what it is good at, what it's not good at.

如果它在某件事上还行，你就会慢慢学会怎么调整自己的工作流程——比如，去掉这一部分，或者删掉那一部分，这些其实已经没必要了。

And then if it's somewhat good at something, you kinda get better at understanding how do I modify my workflow so that actually, maybe I take out this part of it, maybe I take out that part of it, this part's not really necessary anymore.

哦，原来AI现在能完成80%的工作了。

Oh, now AI can actually do 80% of the job.

对吧？

Right?

但不通过实际尝试，你是发现不了这一点的。

But you don't figure that out without experimentation.

所以我觉得，这就是为什么随着时间推移，投入在AI上的预算会不断上升。

So I think that's why you see rising spend over time.

我们显然还没到达人们找到AI支出合理水平的阶段。

We're clearly not at the point at which people have found like their benchmark level of AI spend.

如果你看看我们的人均AI支出图表，就会发现，根本不存在那种‘一锤子买卖’的情况。

And if anything, if you wanna look at our charts of AI spend per person, you know, you don't see a token maxing era.

根本不存在一个点，说它涨上去之后又降下来了。

There's no point at which, oh, it went up and then it went down.

它一直在往上走。

It's like still going up.

再说， median 公司每个月才花 11 美元。

Again, median firm is only spending $11 a month.

所以还有很大的增长空间。

So there's a lot of room to grow.

但即便是前 1%，支出也还在持续上升。

But even for the 1%, it's still going up.

是的。

Yep.

好，我们先休息一下。

Alright, let's take a quick break.

回来之后，我想再多聊聊你对 DeepSeek 的观察，因为很多人以为 DeepSeek 的爆发发生在 2025 年 1 月或 2 月，之后就过去了。

When we come back, I want to talk a little bit more about what you're seeing with DeepSeek because I think a lot of people expected that that DeepSeek moment, you know, sort of happened in February, January, February 2025, and then dissipated.

但它的增长又回来了。

But it is, growing once again.

我们一会儿来聊聊这个。

So we'll talk about that.

先谈一谈模型编排，再聊一聊 SaaS。

We'll talk about model orchestration, and then we'll talk about SaaS.

对吧？

Right?

看看这个所谓的 SaaS 大崩盘，是不是真在数据里体现出来了。

Whether this SaaS pocalypse is actually being born out in the data.

我们稍后就来深入讨论这个。

So we'll do that right after this.

我们又回来了，这里是《科技大观》节目，嘉宾是 Ramp 的首席经济学家阿拉·哈拉扎尼安。

And we're back here on big technology podcast with Ramp lead economist, Ara Kharazian.

阿拉，很高兴见到你。

Ara, great to see you.

再次感谢你来到这里。

Thank you again for being here.

我们来聊聊你看到的 DeepSeek 的增长情况吧。

Let's just talk a little bit about the DeepSeek, growth that you're seeing.

你提到，从数据来看，大家越来越依赖这些廉价的开源模型。

You're saying basically what you found in the data is that, there is an increased reliance on these cheap open source models.

在你统计的热门模型榜单里，DeepSeek 排在第一位。

DeepSeek, when it comes to your list of trending models is number one.

你觉得这说明了人工智能的采用方式发生了什么变化？

What do you think that says about the way that AI is adopting?

它真的在被广泛采用吗？

Is it being adopted?

有人认为，未来会是这样：一个超级智能的底层模型做统筹决策，比如 Opus 4.8、GPT-5.5 或 6，而那些简单任务则交给开源的小模型去处理。你觉得是这样吗？

Is it that as you know, some people said the future is going to be that you have this like maybe super smart foundational model orchestrator that makes your big decisions for you, like an Opus 4.8, or a GPT 5.5 or six, and then you just sort of deliver the the more straightforward work to these smaller agents with the open source models.

我想知道，这种趋势在数据里是不是已经显现出来了。

And I'm curious if you think that that is already being borne out in the data.