Speaker 100:00 - 00:31
We actually had bot traffic pass human traffic online in the first half of twenty twenty six. In five years you might have a thousand times as much traffic on the internet as you do today. In the next six to twelve months, almost every company is going to go through some exercise like this where they're gonna cut a bunch of their their team. The underlying business model of the internet for the last twenty eight years has remained basically the same, which is it's largely advertising. The problem is bots don't click on ads. Speaker 100:00 - 00:31
实际上，在 2026 年上半年，线上 bot 流量已经超过了人类流量。五年后，互联网流量可能会达到今天的一千倍。接下来的六到十二个月里，几乎每家公司都会经历某种类似的过程：裁掉相当一部分团队。过去 28 年里，互联网的底层商业模式基本没变，本质上主要还是广告。而问题在于，bots 不会点击广告。
Speaker 100:31 - 00:34
Over the next five years, the business model of the Internet is gonna change radically. Speaker 100:31 - 00:34
在未来五年里，Internet（互联网）的商业模式将发生根本性的变化。
Speaker 200:34 - 00:50
Hi. I'm Matt Turk. Welcome to the Mad Podcast. Today, I'm excited to welcome Matthew Prince, cofounder and CEO of Cloudflare, one of the Internet's most important companies that sits in front of a very large portion of global Internet traffic. We start the conversation by unpacking a huge milestone. Speaker 200:34 - 00:50
大家好，我是 Matt Turk，欢迎来到 Mad Podcast。今天我非常高兴邀请到 Matthew Prince，他是 Cloudflare 的联合创始人兼 CEO。Cloudflare 是互联网中最重要的公司之一，位于全球很大一部分 Internet 流量的前端。我们先从拆解一个重大的里程碑开始这场对话。
Speaker 200:50 - 01:25
For the first time, the number of bots and AI agents has outnumbered humans on the Internet. We also cover how Cloudflare has become an AI infrastructure company and dig into Cloudflare workers, AI gateway, and agent security. We also don't shy away from some of the more controversial topics, including Cloudflare's recent decision to rightsize its team to adapt to the new AI era. I should add that Matthew is a wonderful storyteller, and he shares some truly wild and hilarious episodes from Cloudflare's journey from scrappy startup to the massive public company it is today. Please enjoy this fantastic conversation with Matthew. Speaker 200:50 - 01:25
这是第一次，Internet 上的 bots 和 AI agents（AI 智能体）数量超过了人类。我们还会讨论 Cloudflare 如何成为一家 AI 基础设施公司，并深入聊聊 Cloudflare workers、AI gateway 和 agent security（智能体安全）。我们也不会回避一些更有争议的话题，包括 Cloudflare 最近决定精简团队规模，以适应新的 AI 时代。我还要补充一点，Matthew 特别会讲故事，他分享了 Cloudflare 从一家灵活顽强的 startup（创业公司）成长为如今大型上市公司的过程中，一些真正疯狂又爆笑的经历。请享受这场与 Matthew 的精彩对话。
Speaker 201:27 - 01:48
I want to start with something that feels like a very crucial important moment in the history of the Internet, what you tweeted a couple of weeks ago and you said effectively that the number of bots and machines on the Internet now outnumbers humans. So can you unpack that and explain what that was? Speaker 201:27 - 01:48
我想先从一件听起来像是 Internet 历史上非常关键的重要时刻的事情开始，也就是你几周前发的那条 tweet（推文）。你当时的意思大致是，现在 Internet 上的 bots 和机器数量已经超过了人类。所以你能展开讲讲，解释一下那到底是什么情况吗？
Speaker 101:48 - 02:33
Yeah. So I mean there have always been bots that ran around online because the way Google works is they scrape the Internet and they then build their index based on that. And historically throughout most of the history of Cloudflare, it's about 20% of traffic on the Internet was bots and that held fairly consistent over a period of time until until about two years ago when we started to see a real rise in bots driven mainly by AI and and the LLMs and everything that was going on behind those. And last fall, in the fall of twenty twenty five, I was asked at what point will automated traffic pass human traffic? And we pulled the data and we sort of extrapolated out the trends and we thought that it was going to be the end of twenty twenty seven. Speaker 101:48 - 02:33
是的。我的意思是，线上一直都有 bots，因为 Google 的工作方式就是抓取整个 Internet，然后在此基础上建立它的索引。而从历史上看，在 Cloudflare 存在的大部分时间里，Internet 上大约 20% 的流量都是 bots，这个比例长期以来都相当稳定。直到大约两年前，我们开始看到 bots 真正明显上升，这主要是由 AI、LLMs（大语言模型）以及其背后发生的一切推动的。到去年秋天，也就是 2025 年秋天，有人问我，自动化流量什么时候会超过人类流量？于是我们拉取了数据，做了一些趋势外推，当时我们的判断是会在 2027 年底发生。
Speaker 102:33 - 03:13
Fast forward to South by Southwest this year, March 2026 and I was asked the same question again and we looked at the data again and we're like, oh wow, it's growing a lot faster than we expected. It's moved now to the first half of twenty twenty seven. And then, you know, I was pretty shocked when my team came to me just a few weeks ago and they said, won't believe it, but we actually had bot traffic pass human traffic online in the first half of twenty twenty six. So I think what that shows is just how fast these different systems, again almost all driven by AI, are growing and they're really driving so much more traffic across the internet. Speaker 102:33 - 03:13
快进到今年的 South by Southwest，也就是 2026 年 3 月，我又被问到了同样的问题。我们再次看了数据，反应是，哇，它的增长速度比我们预期快得多。原先的时间点已经提前到 2027 年上半年。然后，几周前我的团队来找我时，我真的相当震惊。他们说，你肯定想不到，但实际上，在 2026 年上半年，线上 bot 流量就已经超过了人类流量。所以我认为，这说明了这些不同系统——而且几乎全都是由 AI 驱动的——增长得有多快，它们确实正在推动整个 Internet 上出现多得多的流量。
Speaker 203:13 - 03:17
And what was the denominator? Was it like a HTTP request? Speaker 203:13 - 03:17
那你们用的分母是什么？比如说，是按一次 HTTP request（请求）来算的吗？
Speaker 103:17 - 03:51
Yeah. So so Cloudflare sits in front of, you know, a huge percentage of the Internet and so we and it's in a super representative sample. And and so as a result, we can always look at what are the general trends across the Internet and we have a site we call radar. If you go to radar.cloudflow.com, this tracking a number of different things and one of those is how much human traffic is there, how much bot traffic is there and in in the last couple months it it the the bot traffic exceeded the human traffic. And I and I I think that if you extrapolate this out, you know, in five years, I think you'll see a thousand times more bot traffic than human traffic. Speaker 103:17 - 03:51
对。Cloudflare 处在互联网中非常大一部分流量的前面，所以我们看到的是一个极具代表性的样本。因此，我们总能观察整个互联网的总体趋势。我们有一个叫 radar 的网站，如果你去 radar.cloudflow.com，它会追踪很多不同的指标，其中一个就是有多少 human traffic（人工流量），有多少 bot traffic（机器人流量）。而在过去几个月里，bot traffic 已经超过了 human traffic。我认为，如果把这个趋势外推，五年后你会看到 bot traffic 比 human traffic 多一千倍。
Speaker 103:51 - 04:05
And if I had to take kind of an over under bet on that, I'd take the over bet that it's just gonna it's it's growing at such an exponential rate that that you're just gonna see an enormous amount of the Internet that's driven by by by these by these various bots and agents and other systems. Speaker 103:51 - 04:05
如果非要让我对这个做一个 over-under bet（大小盘下注），我会押 over。我认为它会继续增长，而且是以如此指数级的速度增长，以至于你会看到互联网中有极其庞大的一部分，是由这些各种各样的 bots、agents（智能代理）以及其他系统驱动的。
Speaker 204:05 - 04:10
And when we say bots, do we mean agents? Do we mean crawling bots? What Speaker 204:05 - 04:10
那我们说 bots 的时候，指的是 agents 吗？指的是 crawling bots（爬虫机器人）吗？什么——
Speaker 104:10 - 04:38
do I we mean I think that agent, bot, crawler are all synonyms. It all means the same thing. It depends on whether you want to describe it in a sort of normatively positive way in which case you call it an agent. If you want to describe it kind of the ickiest way you call it a crawler or a scraper, but it's the same thing behind the scenes. It's it's a it's a machine that's accessing resources as opposed to kind of a human eyeball driving a browser accessing those those resources. Speaker 104:10 - 04:38
我的意思是，我认为 agent、bot、crawler 都是同义词，表达的是同一回事。区别只在于你想用一种带有规范性正面色彩的方式来描述它，那你就叫它 agent；如果你想用一种最让人反感的说法，你就叫它 crawler 或 scraper（抓取器）。但在底层，它们其实是同一样东西：本质上就是一台机器在访问资源，而不是一个人类的眼睛盯着浏览器、由人来驱动浏览器访问这些资源。
Speaker 104:38 - 05:28
And and again in this particular case, you still have a whole bunch of this sort of scraper crawler hacker traffic that hasn't changed that much. You still have some of the sort of bots like Google bot that are that are building search indexes that hasn't changed that much. If anything has gone a little bit down. But the thing that's really driving this is what I think we would all call agents, which is you know if if I go to ChatGPT or I go to Claude or I go to you know, any of these systems and and I'm shopping for a digital camera, you're see, you know, thousands of times more site visits for that same activity, that same job to be done of of shopping for a digital camera than if I did it myself. I I might visit five sites if I'm, you know, personally trying to figure out what digital camera to buy, whereas my agent might visit 5,000 and that's what's really driving this increased demand. Speaker 104:38 - 05:28
再说回这个具体情况，仍然有大量这类 scraper、crawler、hacker traffic（黑客流量），这一部分变化并不大。也仍然有一些像 Google bot 这样的 bots 在构建搜索索引，这部分变化也不大；如果真有变化，甚至还略微下降了一点。真正推动这一切的，是我认为我们都会称之为 agents 的东西。比如说，如果我去 ChatGPT，或者去 Claude，或者去这些系统中的任何一个，我想买一台 digital camera（数码相机），那么为了完成同样一件事——也就是选购一台 digital camera——你会看到比我亲自去做时多出成千上万倍的网站访问量。我要是自己研究买什么 digital camera，可能只会访问 5 个网站；而我的 agent 可能会访问 5,000 个，这才是真正推动需求增长的原因。
Speaker 205:28 - 05:43
Yeah, and what does that change from, I guess, an infrastructure perspective? Like what you just described, 5,000 sites versus five. In some ways, that sounds inefficient. I mean, it's efficient, for me as a user ultimately because I guess the best answer. What does that change? Speaker 205:28 - 05:43
对，那从基础设施的角度看，这会改变什么？比如你刚才说的，5,000 个网站对比 5 个网站。从某种意义上说，这听起来挺低效的。我的意思是，对我这个用户来说，最终它是高效的，因为我能得到更好的答案。那它到底改变了什么？
Speaker 205:43 - 05:45
In terms of requirements for the infrastructure? Speaker 205:43 - 05:45
你是说对基础设施需求而言吗？
Speaker 105:45 - 06:11
Yeah. Think, I mean, first of all, there there if if this trend continues, there's just gonna be an enormous amount of additional demand on on the Internet. And I go back to, you know, during COVID where over the course of two weeks we saw Internet traffic double. You know, if if if our projections are right, that's gonna seem quaint very soon. You know, in five years, you might have a thousand times as much traffic on the internet as you do today. Speaker 105:45 - 06:11
对。我是说，首先，如果这个趋势持续下去，互联网就会承受海量新增需求。我会想到 COVID 期间，当时在两周时间里，我们看到互联网流量翻了一倍。你知道，如果我们的预测是对的，那种增长很快就会显得不值一提。五年后，互联网上的流量可能会是今天的一千倍。
Speaker 106:11 - 06:52
And that's gonna mean more, you know, servers, more network infrastructure, CPUs, more GPUs, more memory, all of the things that have to serve these bots are are gonna be very important. And it's also gonna mean we have to figure out some new business model in order to fund that. Someone has to pay for it and and you know the the business model of the Internet historically has been ads and bots don't click on ads. So it's it's going to be something different going forward and you know, I I think the most interesting question that in in the world today is over the next five years, the business model of the Internet is gonna change radically and what it changes to is totally undefined at this point. Speaker 106:11 - 06:52
这就意味着，更多的服务器、更多的网络基础设施、更多的 CPU、更多的 GPU、更多的内存——所有那些为这些 bots 提供服务的东西，都会变得非常重要。这也意味着，我们必须想出某种新的商业模式来为这一切买单。总得有人付费，而你也知道，历史上 Internet 的商业模式一直是广告，但 bots 不会点击广告。所以，接下来一定会变成别的东西。而且我认为，当今世界里最有意思的问题，就是在未来五年内，Internet 的商业模式将发生剧烈变化，但它究竟会变成什么，现在还完全没有定论。
Speaker 206:52 - 07:00
Yeah. This whole new vocabulary It used to be clicks and views and what is it going to be now? Like actions or? Speaker 206:52 - 07:00
对。这是一整套全新的词汇。以前讲的是 clicks 和 views，那现在会变成什么？比如 actions 之类的吗？
Speaker 107:00 - 07:31
We've got to figure out exactly what that is and it's going to, you know, change a number of things. Like, you know, what is a brand in an agentic world? You know, a brand historically has been, you know way as a as a shortcut to tell humans kind of this is something that has a certain set of expectations will have a certain set of experience associated with it. If I if I see a Walmart logo on the side of a building, I know what it's gonna be what like when I walk in. If I see a Marriott logo, I see a McDonald's logo, I know what that's going to be like. Speaker 107:00 - 07:31
我们必须准确弄清楚那到底是什么，而这也会改变很多事情。比如说，在一个 agentic world（agent 主导的世界）里，brand（品牌）到底是什么？历史上，brand 本质上是一种给人类看的“捷径”，用来告诉人们：这个东西会对应一组特定的预期，也会带来某种特定的体验。如果我看到一栋楼侧面有 Walmart 的 logo，我大概就知道走进去会是什么样；如果我看到 Marriott 的 logo，看到 McDonald's 的 logo，我也知道那会是什么感觉。
Speaker 107:31 - 08:11
You know, the the thing about the thing about bots, they have infinite amount of patients in order to discover you know everything that that might be might be right. So the basic idea of what a brand is going forward is going to be very different than what it has been in what had been the very much human driven kind of commerce and Internet experience. And and again, I'm not sure I'm not sure that I have I have ideas of what that might look like and and I think that Kleffler will play some role in in helping figure that out. But I think it's all going get discovered over the next five years and that's incredibly exciting. Speaker 107:31 - 08:11
你知道，bots 的特点在于，它们有无限的 patience（耐心）去发现一切可能正确的东西。所以，未来 brand 的基本概念，将会和过去非常不同；而过去的商业与 Internet 体验，本质上是非常以人为驱动的。再次说明，我也不确定那具体会是什么样子；我确实有一些想法，也觉得 Kleffler 会在帮助理清这件事上发挥某种作用。但我认为，这一切都会在未来五年里被逐步探索出来，而这让人无比兴奋。
Speaker 208:11 - 08:20
Alright, fascinating. I'd love to spend a little bit of time on Cloudflare 01/2001. If I put Cloudflare in front of my website, what happens? Speaker 208:11 - 08:20
好，太有意思了。我想花一点时间聊聊 Cloudflare。假如我把 Cloudflare 放到我的网站前面，会发生什么？
Speaker 108:21 - 09:38
Yeah, so, know, what Cloudflare started out as was how can you take a firewall, like a security product and put it in the cloud. The original idea came from, you know a realization that you know software was moving the cloud, servers were moving in the cloud, therefore all the security and networking infrastructure had to move to the cloud as well. And so that was the basic premise that that we started with and so if you signed up for Cloudflare fifteen years ago and you put us in front of your website, what you were effectively doing is is essentially putting a bouncer in front of the website that could check anyone who was trying to come to it and then stop the bad ones and then because we were we we needed to make sure we didn't slow things down, we would also then, take a lot of load off of your website, and and be able to, you know, for for good visitors deliver the content significantly faster. So the original idea of Cloudflare was sort of performance and security in a very easy to adopt package. We knew as a business that in order to succeed we eventually had to sell our services to you know major banks and governments and health all the things that we are now count as customers. Speaker 108:21 - 09:38
是这样的，Cloudflare 最初要做的是：怎样把 firewall（防火墙）这类安全产品搬到 cloud（云）里。最初的想法来自一个判断：software 正在迁移到 cloud，servers 也在迁移到 cloud，因此所有安全和网络基础设施也必须迁移到 cloud。那就是我们创业时的基本前提。所以，如果你在十五年前注册了 Cloudflare，并把我们放在你的网站前面，你实际上做的事情，本质上就是在网站门口放了一个“门卫”，它可以检查所有试图访问你网站的人，拦住那些坏的。与此同时，因为我们必须确保不会把速度拖慢，所以我们还会帮你的网站卸掉大量负载，并且能够让正常访客更快得多地拿到内容。所以，Cloudflare 最初的理念，是把 performance（性能）和 security（安全）打包成一个非常容易采用的产品。作为一家公司，我们也知道，要想真正成功，最终必须把服务卖给大型银行、政府、医疗机构——也就是今天已经成为我们客户的这些组织。
Speaker 109:38 - 10:29
But in the early days, those things didn't they were never going to trust us because we were a dopey little startup and had, you know, we had no infrastructure, we had no data to actually provide the security services. And so the thing that we did was we created a free sort of stripped down version of the service with the idea of making it available to everyone. What was interesting, we expected it was going to be a bunch of like, you know, SMBs and startups and and and individuals that would sign up. But it turns out that, you know, if you think of it sort of x y axis and you know, the x axis is is budget and the y axis is is like security issues. There's almost a perfectly diagonal line that the smaller your budget is the fewer security issues you have, the bigger your budget is the more security issues you have. Speaker 109:38 - 10:29
但在早期，这些机构根本不可能信任我们，因为我们只是个傻乎乎的小 startup（初创公司），没有基础设施，也没有足够的数据来真正提供安全服务。所以我们做的一件事是，做了一个免费、功能有所精简的版本，目标是让任何人都能用。有意思的是，我们原本以为注册的会是一堆 SMBs（中小企业）、startups 和个人用户。但后来发现，如果你把这件事想成一个 x y 坐标系：x 轴是预算，y 轴是安全问题，那么几乎存在一条完美的对角线——预算越小，安全问题通常越少；预算越大，安全问题通常越多。
Speaker 110:29 - 11:09
And so it turned out that the we even SMB startups wouldn't sign up even though it was free. But two organizations kind of defy that or two two groups defy that. One was was Hacker Kids because they're all always attacking each other and so overnight every hacker kid in the world signed up for for Cloudflare and and and then they were attacking each other trying to take us down. And then the other were you know, sort of human rights organizations, humanitarian organizations, nonprofits. Because if you're doing any of that work, turns out nine times out of 10 you're pissing somebody off and oftentimes it's you know a government or something that will then attack you. Speaker 110:29 - 11:09
结果就是，连 SMB 和 startup 其实都不太会来注册，哪怕它是免费的。不过有两类组织算是例外，或者说有两类群体打破了这个规律。一类是 Hacker Kids，因为他们总是在互相攻击，所以几乎一夜之间，全世界的 hacker kid 都注册了 Cloudflare，然后他们一边互打，一边还试图把我们打下来。另一类则是 human rights organizations（人权组织）、humanitarian organizations（人道主义组织）和 nonprofits（非营利组织）。因为如果你在做这类工作，十有八九都会惹恼某些人，而很多时候那甚至会是某个政府之类的对象，接着它就会来攻击你。
Speaker 111:09 - 11:32
So overnight we had like every humanitarian organization in the world, every hacker kid in the world using Cloudflare. They were all, know, and we all of a sudden came under just massive attacks. And so we were constantly trying to figure out how do we stay in front of it? How do we actually continue to to exist and thrive in this world? And that that forced us to build, you know, a number of different things. Speaker 111:09 - 11:32
于是几乎一夜之间，全世界几乎所有的人道主义组织、所有黑客小子都在用 Cloudflare。他们都在用，你知道，然后我们也突然开始遭受大规模攻击。所以我们一直在想，怎么才能始终领先一步？我们到底要怎样才能继续在这个世界里生存并壮大？而这也逼着我们去构建很多不同的东西。
Speaker 111:32 - 12:25
So we, you know, built a registrar product so we could register domains because we had used a registrar and we came this close to our domain getting stolen, which would have been a disaster for us. We had to build DDoS mitigation, which initially we didn't think we wanted to build at all. We had to build, you know, a DNS infrastructure because we couldn't get anyone to support us. Needed to build our own sort of VPN because everybody else was too slow or not secure enough. And eventually we had to build our own developer platform where you could where our team could turn around and actually take the ideas that they had and deploy them in a way that was that was able to be you know incredibly limited sandboxed and other things because you know at some point we got big enough that if we had an issue it was it was incredibly you know, disruptive to the world in terms of what we're doing. Speaker 111:32 - 12:25
所以我们后来做了一个 registrar（域名注册商）产品，这样我们就能自己注册 domain（域名），因为我们之前用过一家 registrar，而我们的域名差一点就被偷走了，那对我们来说会是一场灾难。我们不得不做 DDoS mitigation（DDoS 缓解），一开始其实我们完全没觉得自己想做这个。我们还不得不搭建 DNS infrastructure（DNS 基础设施），因为根本没人愿意支持我们。我们还需要做自己的某种 VPN，因为别人的要么太慢，要么不够安全。再后来，我们还不得不做自己的 developer platform（开发者平台），这样我们的团队才能把他们的想法快速变成部署出去的东西，而且这些部署还必须是被严格限制、经过 sandboxed（沙箱隔离）等处理的，因为到某个阶段我们已经大到一旦出问题，就会对世界造成非常严重的扰动——就我们所做的事情而言，影响会非常大。
Speaker 112:25 - 13:07
And so in the end the real story of Cloudflare has been, know, we started with this idea that was pretty simple which was how do you put a firewall in the cloud and then that and then we sort of created so many problems over time for ourselves that we had to solve. And then in the process of solving those problems, we've built a whole series of products. And so Cloudflare today unlike when we started, you know, I mean I have a hard time articulating exactly what it is that we do. At some level we, know, we're sort of the next generation sort of AWS or Google Cloud where we've rethought how that works. We still provide security and performance across that. Speaker 112:25 - 13:07
所以归根结底，Cloudflare 真正的故事是：我们一开始的想法其实很简单，就是怎么把 firewall（防火墙）放到 cloud（云）里；但后来随着时间推移，我们某种程度上给自己制造了越来越多必须解决的问题。而在解决这些问题的过程中，我们做出了一整套产品。所以今天的 Cloudflare，和刚开始时已经完全不同了。说实话，我现在都很难准确概括我们到底是做什么的。在某种层面上，我们有点像下一代的 AWS 或 Google Cloud，只不过我们重新思考了这一切该怎么运作。与此同时，我们仍然在其上提供 security（安全）和 performance（性能）。
Speaker 113:07 - 13:37
But we have additional products that take advantage of the network that we have in order to provide things like a cloud based VPN and other things. And and you know, we're increasingly because of the massive percentage of the Internet that we sit in front of, we're increasingly sort of playing a role in figuring out what's that business model of of the Internet of the future. And so our our mission is to help build a better Internet and I think anything that sort of is in fulfillment of that mission is is the sort of problem that we're always thinking about. Speaker 113:07 - 13:37
但我们还有更多产品，利用我们现有的 network（网络）去提供例如 cloud based VPN（云端 VPN）之类的能力。并且，随着我们所处位置已经挡在互联网前方的流量占比越来越大，我们也越来越在参与思考：未来互联网的 business model（商业模式）到底应该是什么。所以我们的 mission（使命）是帮助建设一个更好的互联网，而我认为，任何有助于实现这个使命的事情，都是我们一直在思考的问题。
Speaker 213:37 - 13:53
Yeah, the concept of network is fascinating and seems very central to the story. And it's a network of users, but it's also physical network. Right? Like, what what you guys did initially was to create this edge network. Like, how does one even go about building that as a young startup? Speaker 213:37 - 13:53
对，network（网络）这个概念很迷人，而且看起来也是这个故事里非常核心的一点。它既是用户网络，也是物理网络。对吧？你们最初做的，其实就是建立这个 edge network（边缘网络）。那一家年轻的 startup（初创公司）到底要怎么着手去搭这样的东西？
Speaker 113:53 - 15:00
I mean, in the very beginning, you know, we when we when we launched, we had servers in five cities around the world. The Chicago, Ashburn, Virginia, San Jose, California, Amsterdam, and Tokyo. Except Tokyo sort of, if we had been totally honest, deserved an asterisk because we couldn't totally figure out the routing and if we got it wrong, whole bunch of traffic from The US ended up hitting hitting Tokyo and so we would literally turn the Tokyo servers on and off throughout the day depending on, you know, where who is who is awake around the world. Today fast forward and you know we're in over 50 cities worldwide, we're in over a thousand data centers worldwide and and we will go from, you know, the minimum deployment is sort of like a rack of servers in some places we'll have hundreds of racks of servers just depending on what's what's needed there. But I think that that was the that was, you know, a huge focus behind the scenes was how do you figure out how to get into all of these places around around the world. Speaker 113:53 - 15:00
我是说，在最开始，我们上线的时候，在全球五个城市有服务器：Chicago、Ashburn, Virginia、San Jose, California、Amsterdam 和 Tokyo。不过如果完全诚实一点，Tokyo 其实应该打个星号，因为我们当时还没完全搞明白 routing（路由）；一旦搞错了，就会有一大堆来自 The US 的流量跑去打到 Tokyo，所以我们那时真的会根据一天当中全球哪些地方的人醒着，来反复开关 Tokyo 的服务器。到了今天，时间快进之后，你知道，我们已经遍布全球 50 多个城市、1000 多个 data center（数据中心）；部署规模小的地方可能最低就是一整柜服务器，而有些地方我们会放几百柜服务器，具体取决于当地需要什么。但我觉得，幕后一个巨大的重点始终是：你怎么才能进入世界上所有这些地方。
Speaker 115:01 - 15:45
And it was it's a series of kind of stories of how did we create enough value to you know the ISPs around the world that the Internet providers for consumers that they would invite us to be into their their buildings and you know it's a number of different things. Remember Telecom Pakistan was this incredibly difficult organization to work with and there's a lot of people in Pakistan so there was there was a lot of traffic that was coming from it. It was all hitting our Frankfurt Data Center and and I kept saying we should we should put servers in Pakistan and they said and our team kept saying you don't understand. They're incredibly difficult. They've kicked Facebook out, they kicked YouTube out, they kicked Akamai out, they're never gonna let us in. Speaker 115:01 - 15:45
这背后其实是一连串故事：我们是怎么创造出足够的价值，让全世界的 ISP（互联网服务提供商），也就是给消费者提供网络接入的那些公司，愿意邀请我们进入他们的大楼。这里面有很多不同的情况。我记得 Telecom Pakistan 就是一个特别难打交道的组织，而 Pakistan 人口很多，所以那边来的流量也非常大。这些流量当时全都打到我们的 Frankfurt Data Center。我一直在说，我们应该把服务器放进 Pakistan，而团队总是说，你不明白，他们特别难搞。他们把 Facebook 赶出去过，把 YouTube 赶出去过，也把 Akamai 赶出去过，他们绝不可能让我们进去。
Speaker 115:45 - 16:18
And then out of the blue one day, we get an email from the guy who's the head of networks at Telecom Pakistan and he says, hey, we is there any way that we can get Cloudflare gear into our into our into our network and into our system? And I went to the team, was like, I thought you said these guys were difficult and so we we got the guy who ran networks on the phone and I was like, oh, you know, we'd we'd heard you were kind of difficult. Why why do you want us there? He said, I'm a huge cricket fan. And the Pakistani National Cricket team is a Cloudflare customer. Speaker 115:45 - 16:18
结果有一天，完全出乎意料地，我们收到一封邮件，发件人是 Telecom Pakistan 的网络负责人。他说，嘿，有没有什么办法能把 Cloudflare gear（设备）接进我们的网络、接进我们的系统？我就去找团队，说，我还以为你们说这些人很难搞。于是我们就和那位负责网络的人通了电话，我就说，哦，你知道，我们之前听说你们挺难打交道的。那你们为什么想让我们进去？他说，我是个超级 cricket（板球）迷。而 Pakistani National Cricket team 是 Cloudflare 的客户。
Speaker 116:18 - 16:45
They were a free customer. And I'm sick of the latency of backhauling the traffic to Frankfurt. So set of servers and we'll we'll install them in in our network. And so the the secret to Cloudflare success has always been how do we provide, you know, a huge number of services for free to a bunch of people and then it it ends up being that there's somebody in that group which is which is really important. We do other things as well. Speaker 116:18 - 16:45
他们原本是免费客户。而且我受够了把流量回传到 Frankfurt 所带来的延迟。所以他们说，那就部署一组服务器吧，我们会把它们安装到我们的网络里。Cloudflare 成功的秘诀一直都是：我们怎样为大量用户免费提供海量服务，而最终这群人里总会有一些人是非常重要的。我们也还做其他事情。
Speaker 116:45 - 17:20
So, know, if you if you type in, you know, www.amazon.com, it turns out that, you know, Amazon gets to choose where w w w points. VeriSign, which runs .com, gets to choose where amazon.com points. The question is who gets to decide where .com points? And the answer is, there are these 13 root servers that are out there, that that underlie the Internet. And when I when I say a server, people picture a single box, but in these cases, they're often, you know, thousands of machines that are that are sitting behind this. Speaker 116:45 - 17:20
所以，你如果输入 www.amazon.com，实际上，Amazon 可以决定 www 指向哪里。运营 .com 的 VeriSign 可以决定 amazon.com 指向哪里。问题是：谁来决定 .com 指向哪里？答案是，互联网底层有 13 台 root server（根服务器）。我说到 server（服务器）时，人们会以为是一台单独的机器，但在这些场景里，背后往往其实是成千上万台机器。
Speaker 117:21 - 17:54
And they're run by organizations that you'd be sort of surprised by like USC has one, the University of Maryland has one, NASA has one and the reason there are 13 is because that's the amount that you can stick in a in a standard packet. So there will never be a fourteenth, root server. And and the organizations that run them will probably never change as well because this has become this incredibly political topic. Like China wants one, Russia wants one. And and and so as a result, the University of Maryland will probably continue to run it even though, know, why the University of Maryland? Speaker 117:21 - 17:54
它们由一些会让你有点意外的组织运营，比如 USC 有一台，University of Maryland 有一台，NASA 有一台。而之所以是 13 台，是因为标准 packet（数据包）里能装下的就是这么多。所以永远不会有第 14 台 root server。而且运营这些 root server 的组织大概率也永远不会变，因为这已经成了一个极具政治性的话题。比如 China 想要一个，Russia 想要一个。所以结果就是，University of Maryland 很可能会一直运营它，尽管你会想，为什么偏偏是 University of Maryland？
Speaker 117:54 - 18:23
We provide the infrastructure for two of the 13 root servers. We don't run the root servers themselves, but they they are on our network. And that's incredibly important because anytime you send an email, anytime you click on a link, anytime you use an app on your phone, the very first thing that happens is a is a request to one of these root servers. And so, we provide those services, we provide them for free, and and we do it for two of the 13. We've been asked to do it for more, but we think two is sufficient. Speaker 117:54 - 18:23
我们为 13 台 root server 中的两台提供基础设施。我们并不直接运营 root server 本身，但它们部署在我们的网络上。这一点极其重要，因为每次你发一封邮件、每次你点开一个链接、每次你用手机上的 app，最先发生的事情之一就是向其中一台 root server 发出请求。所以，我们提供这些服务，而且是免费提供；13 台里面，我们承担了其中两台。也有人请我们做更多，但我们觉得两台已经足够了。
Speaker 118:24 - 19:01
And and that's an again one of the things that when we go to an ISP anywhere in the world, we say, hey, would you like one of the root servers to be running inside your network? And they're like, absolutely, that'll improve everything. And then in the process of getting that in, that then helps us, you know, all of the other services that we provide, run on that same same infrastructure. And so because we've done the hard work to make it so that every service at Cloudflare can run on any one of our servers. When we can get into these places, it means that we then have a really privileged access to kind of the fundamental core of how the internet works. Speaker 118:24 - 19:01
这同样也是一件事：当我们去世界任何地方找 ISP（互联网服务提供商）时，我们会说，嘿，你们愿不愿意让一台 root server 在你们的网络内部运行？他们都会说，当然愿意，这会让一切都变得更好。而在把这件事部署进去的过程中，也会反过来帮助我们——我们提供的所有其他服务，都运行在同一套基础设施上。所以，正因为我们做了那些艰难的工作，确保 Cloudflare 的每一项服务都能在我们任意一台服务器上运行，当我们进入这些地方时，就意味着我们获得了一种非常特殊的能力，可以触及互联网运行方式中最基础、最核心的部分。
Speaker 219:01 - 19:12
It's amazing, right? Like we all take the internet for granted. It's so unbelievably interesting to hear this, right? Like how it all works. So that's a physical network. Speaker 219:01 - 19:12
这太惊人了，对吧？我们都把互联网视为理所当然。但听到这些真的让人觉得不可思议，尤其是了解它到底是怎么运作的。所以，这是一张物理网络。
Speaker 219:12 - 19:23
Do you think of Cloudflare as a network effects kind of business currently around, you know, whether that's a data network effect or otherwise? Or is that more of a VC way of looking at the world? Speaker 219:12 - 19:23
你现在会把 Cloudflare 看作一种具有 network effects（网络效应）的业务吗？无论那指的是 data network effect（数据网络效应）还是什么别的？还是说，这更多只是 VC 的一种看世界的方式？
Speaker 119:23 - 20:16
No, mean I think networks inherently have network effects and you know as we see as we have gotten bigger our rate of growth of who uses us has continued to accelerate. So I think there are multiple different network effects that that are stacking on top of each other. But but one of them is, you know, we provide a VPN service and and a and a and one of the most popular DNS services in the world one point one point one point one one. Everyone who uses one of those things then they're inherently going to get access to everything else on Cloudflare a little bit faster. And so as more people use kind of one side of our network, it's just it just makes the other side of our network much more valuable and as more people use the other side of our network it makes the other side of network more valuable. Speaker 119:23 - 20:16
不，我的意思是，我认为 network 天生就具有 network effects。而且我们也看到，随着我们变得更大，使用我们服务的人数增长速度还在持续加快。所以我认为，这里有多种不同的 network effects，正一层层叠加起来。不过其中一个例子是：我们提供 VPN 服务，也提供全球最受欢迎的 DNS 服务之一——one point one point one point one。每个使用这些服务的人，都会天然地更快访问到 Cloudflare 上的其他一切。所以，当更多人使用我们网络的一侧时，网络的另一侧就会变得更有价值；而当更多人使用网络的另一侧时，这一侧也会变得更有价值。
Speaker 120:16 - 21:10
And so I think those things are very much compounding and then going forward, know, as we think about you know what is is gonna what is the future business model the Internet going to look like, I think that's going to be another place where it's going to make it incredibly compelling for for as many people as possible to be be on our network. But it's you know it's interesting. It takes about five minutes to sign up for Clefler's basic services. It takes about you know ten seconds to leave and so well I think that there are things that are always driving what it is that's the value behind Cloudflare. We have to constantly live up to providing that value otherwise you know our there's there's customers aren't locked in in any way and if we if we're not providing service if we're not creating value, you know the whole thing could go away you know very quickly. Speaker 120:16 - 21:10
所以我认为这些因素都在不断叠加。再往前看，当我们思考未来的商业模式、Internet 的未来形态会是什么样时，我觉得那也会成为一个极具吸引力的点，让尽可能多的人愿意加入我们的网络。不过，这也很有意思。注册 Cloudflare 的基础服务大约只要五分钟，而离开大概只要十秒钟。所以，虽然我认为一直都有一些因素在驱动 Cloudflare 的价值，但我们必须持续兑现这种价值；否则，客户并没有以任何方式被锁定，如果我们没有提供服务、没有创造价值，整件事你知道，可能会非常快地消失。
Speaker 221:10 - 21:22
But at least from a security standpoint why would you I guess you could always go to a competitor but like you benefit from the collective wisdom of knowing all the DDoS attacks in the world at any point in time. Speaker 221:10 - 21:22
但至少从安全角度看，为什么要离开呢？我想你当然总是可以转去竞争对手那里，但你能受益于这种集体智慧：在任何时刻都知道全球所有的 DDoS attack（分布式拒绝服务攻击）。
Speaker 121:22 - 22:29
And that's and again I think that's another one of the network effects that we have is that you know we act almost like an immune system and so an attack against any one of our customers benefits all of the customers that that are out there. The story that Michelle, my co founder wouldn't want me to tell because it's a little it's a little it's a little blue, I guess, was that early on I remember we we used have a bell that would go off anytime anyone signed up for for Cloudflare and and the bell went off and we all would run back to our, you know, laptops to see what was who had signed up and and this was back in in 2010 or maybe early twenty eleven. And and I remember just being like my I mean my eyes got wide and I and I I blushed because it was a Turkish prostitution website that had signed up. And and he was like, oh, okay. I guess we need a like a complete handbook or somebody just sort of say, know, there might be some kind of colorful things that sign up for Cloudflare. Speaker 121:22 - 22:29
而且，这同样也是我认为我们拥有的另一种 network effect（网络效应）：我们几乎就像一个免疫系统，因此针对我们任何一个客户的攻击，都会让所有其他客户受益。关于这点，有个故事，Michelle——我的 co-founder（联合创始人）——大概不希望我讲，因为稍微有点“荤”，算是吧。早期我记得，我们曾经有一个铃，只要有人注册 Cloudflare，它就会响。铃一响，我们就都会跑回自己的 laptop（笔记本电脑）前，看看是谁注册了。这大概是在 2010 年，或者 2011 年初。我记得当时我眼睛都睁大了，脸也红了，因为注册的是一个 Turkish prostitution website。然后我们就想，哦，好吧，我猜我们需要一本完整手册，或者至少得有人提醒一句：注册 Cloudflare 的，可能会有一些“很有颜色”的东西。
Speaker 122:29 - 22:57
But but then what happened next is the bell went off again and it was another Turkish prostitution site and then another one and then another one. And by the end of the first week, like over a thousand Turkish prostitution sites had signed up for Cloudflare. It was really strange because Probably lucky fit. I mean, we didn't have we didn't have a sales team, we hadn't market, we'd we hadn't translated the website into Turkish. There was I mean, was it was just like and it wasn't like there were a bunch of like, I don't know, Bulgarian, you know, escort sites that were signing up for whatever reason was Turkish. Speaker 122:29 - 22:57
但接下来发生的是，铃又响了，又是一个 Turkish prostitution site，然后又一个，又一个。到了第一周结束时，已经有一千多个 Turkish prostitution sites 注册了 Cloudflare。这真的非常奇怪，因为大概只能说是运气碰上了。我们的意思是，我们当时没有 sales team（销售团队），也没做 marketing（市场推广），也没有把网站翻译成土耳其语。我的意思是，这一切就像是——而且也不是说有一堆，比如我不知道，Bulgarian escort sites 在注册；不知为什么，偏偏就是 Turkish。
Speaker 122:57 - 23:30
So we, you know, curiosity got the better of us. And so we contacted one of the webmasters behind us. We set up a Skype call with them and it was interesting. Was first of we got on the phone and he was just effusively thankful for Cloudflare and he said, you know, you might not know this about Turkey but as you go for the Western Turkey, it's very European, they don't love what we do but they kind of tolerate it. But as you go further east it becomes a very much more conservative, much more Muslim country and they think what we're doing is just the epitome of evil. Speaker 122:57 - 23:30
所以你知道，好奇心还是占了上风。于是我们联系了其中一位 webmaster（网站管理员），还和对方开了一个 Skype call。挺有意思的。首先，我们一接通电话，他就对 Cloudflare 表达了极其热情的感谢。他说，你们可能不了解 Turkey，但在西部 Turkey，它非常欧洲化，人们虽然不喜欢我们做的事，但某种程度上还能容忍；可越往东走，这个国家就越保守、越 Muslim，他们觉得我们做的事简直是邪恶的极致。
Speaker 123:31 - 24:26
And so he said listen, when someone reads TechCrunch, they heard about you, we signed up, we are always getting these attacks we assume from kind of sort of more sort of Muslim fundamentalists, kind of Eastern Turkish people. You protect us from this tax, by the there's not a lot of money in Turkish escorts, so we're using the free version of your service and thank you very much. And so we have machine learning systems that will categorize attacks and so we labeled this style of attacks the TE which stood for Turkish escort attacks. And so anytime we'd see you know that this the system would identify you know the TE attacks are coming. Fast forward a year later and I was in our office late on a I think Thursday or Friday afternoon the phone rings and it's a Dutch gentleman. Speaker 123:31 - 24:26
所以他说，你听着，当有人在读 TechCrunch 时，他们听说了你们，我们就注册了。我们一直会遭受这些攻击，我们猜大概是来自更偏 Muslim fundamentalists、偏 Turkey 东部的那些人。你们保护了我们不受这些 attacks 的影响。顺便说一句，Turkish escorts 这行没什么钱可赚，所以我们用的是你们服务的免费版，非常感谢。于是我们有一些 machine learning（机器学习）系统会对攻击进行分类，所以我们把这种类型的攻击标记为 TE，也就是 Turkish escort attacks。这样一来，每次系统识别出 TE attacks 要来了时，我们就知道了。时间快进到一年后，我那时某个星期四或星期五下午很晚还在办公室，电话响了，是一位 Dutch 先生打来的。
Speaker 124:26 - 24:48
He's calling from Baku, Azerbaijan and he says the contest is tomorrow. We're completely offline. Everyone has told us you're the only ones who can who can help. And I was like what contest? And I grew up in Park City, Utah like I, you know, hadn't traveled out out of the country until until college and you know wasn't wasn't very familiar with, with things that happen around the rest of the world. Speaker 124:26 - 24:48
他从 Baku, Azerbaijan 打来电话，说比赛明天就要开始了，我们现在完全离线。所有人都告诉我们，只有你们能帮忙。我当时就想，什么比赛？我是在 Park City, Utah 长大的，你知道，我直到上大学才出国，所以对世界其他地方发生的事情并不是很熟悉。
Speaker 124:48 - 25:00
So when he said Eurovision, I said, what's that? Right? And he's like, no, it's a really big deal. And I was like, a song contest? Like, it's like American Idol with nationalism, whatever. Speaker 124:48 - 25:00
所以当他说 Eurovision 的时候，我说，那是什么？对吧？然后他说，不，这是件大事。我就说，一个歌唱比赛？就像带着民族主义色彩的 American Idol，之类的吧。
Speaker 125:00 - 25:21
Just, you know, sign up for the services. Like, no, we need help. And I'm like, no, just whatever sign up, you'll be fine. What had happened was the finals in Eurovision, one of the six contestants was transgender and Baku, know, Azerbaijan is again a very relatively conservative Muslim country. And some people thought that this was incredibly offensive that this was happening there. Speaker 125:00 - 25:21
就是，你知道，去注册这些服务就行了。可实际情况是，不，我们需要帮助。而我当时还在想，不，随便注册一下就好了，你们不会有事。事情的起因是 Eurovision 的决赛里，六位参赛者中有一位是 transgender（跨性别者），而 Baku，也就是 Azerbaijan 的首都，又是一个相当保守的穆斯林国家。于是有些人觉得这件事发生在那里，是极其冒犯性的。
Speaker 125:21 - 25:50
So they launched pretty significant attacks against them and knocked them offline. I came in the next morning and we had a bunch of French engineers and their eyes were like saucers and they're like, do you know who signed up last night? And I'm like, oh yeah, that that that song contest. And they're like, no, no, you don't understand what a big deal it is. And for those of you who are also ignorant Americans in the audience, Eurovision gets more like viewers than the Super Bowl. Speaker 125:21 - 25:50
所以他们对其发起了相当大规模的攻击，把他们打到 offline（下线）。第二天早上我到办公室时，看到一群 French engineers，眼睛瞪得像盘子一样大，他们说，你知道昨晚是谁注册了吗？我说，哦对，就是那个歌唱比赛。他们说，不不，你根本不知道这事有多大。顺便说一句，对于在场同样无知的美国听众来说，Eurovision 的观众人数比 Super Bowl 还多。
Speaker 125:50 - 26:22
Like it's it's this it is it's American Idol with nationalism. It's crazy. And and so and so we protected them and we looked at like the attacks and and the system is identifying them and it goes t e t e t e t e. Exact same pattern, exact same attacker. Fast forward a year after that, get a frantic call from the CTO of JPMorgan and he's like, we're under attack, you know, they they they're they're taking offline a bunch of our kind of consumer facing services. Speaker 125:50 - 26:22
它就像是带着民族主义色彩的 American Idol。非常疯狂。于是我们保护了他们，我们去看那些攻击，系统在识别它们时显示：t e t e t e t e。完全相同的模式，完全相同的攻击者。时间快进到一年后，我接到 JPMorgan 的 CTO 打来的一个惊慌失措的电话，他说，我们正在遭受攻击，他们正在让我们一批面向消费者的服务 offline（下线）。
Speaker 126:22 - 27:06
We've heard you guys are really good at at this. You need to bring your best solutions engineers and best salespeople and be in New York first thing Monday morning. We didn't have salespeople and our one network guy didn't own a suit so we like over the weekend we bought him a suit and we all fly out Sunday night and we're in the JPMorgan offices first thing like 6AM on Monday morning and like I remember they have this whole room, it's a war room where they're they're dealing with this and they slide this stack of logs across the table and this guy on our team, his name is Sheree, he looks at and goes, Oh, know these guys. It's the Turkish escort attackers. What it turns out is it actually wasn't anyone in Turkey. Speaker 126:22 - 27:06
我们听说你们在这方面非常厉害。你们得带上最好的 solutions engineers（解决方案工程师）和最好的 salespeople（销售人员），周一一早就出现在 New York。可我们那时根本没有 salespeople，而且我们唯一的 network guy（网络工程师）连西装都没有，所以周末我们给他买了一套西装，周日晚上大家一起飞过去，周一早上大概 6 点第一时间就到了 JPMorgan 的办公室。我记得他们有一个专门处理这件事的房间，是个 war room（作战室）。他们把一大摞日志从桌子那头推给我们，我们团队里有个人叫 Sheree，他看了一眼就说，哦，我认识这帮人，是 Turkish escort attackers（土耳其 escort 攻击者）。结果后来发现，其实根本不是 Turkey 的人。
Speaker 127:06 - 28:04
It turns out it was originally a student in Iran who who again saw what was happening in Turkey and thought that was it was repugnant. So you know launched attacks against them, then saw the Eurovision contest launched attacks against them. That got him recognized and he actually got brought into the the Iranian military apparatus, sort of forward cyber or that offensive cyber capabilities at which point they attacked a bunch of US banks which is how they came after this and the same guy to this day is running all of Iranian cyber attacks. And so when you see like all of the things where they were creating like the sort of Lego style movies, it's this one guy who's who started out as attacking you know Turkish escorts. Again that's I think shows kind of the power of Cloudflare serving everyone helps us learn from even these sort of weird corners of the internet. Speaker 127:06 - 28:04
后来发现，最初是 Iran 的一个学生，他同样看到了 Turkey 发生的事，觉得那非常令人不齿。于是他对他们发起了攻击，之后又看到 Eurovision 比赛，也对他们发起了攻击。这让他出了名，后来他实际上被吸收进了 Iranian military apparatus（伊朗军事体系），进入某种前沿 cyber（网络）或 offensive cyber capabilities（进攻性网络能力）部门。再之后，他们攻击了一批 US banks（美国银行），也就是这样才找上了这件事。而同一个人直到今天仍在负责伊朗的所有 cyber attacks（网络攻击）。所以当你看到后来那些他们制作的、像 Lego 风格的小电影时，背后其实都是这一个人——而他最开始不过是在攻击 Turkish escorts（土耳其 escort 服务）。再说一次，我觉得这恰恰说明了 Cloudflare 为所有人提供服务的力量：它让我们即使从互联网这些奇怪的角落里，也能学到东西。
Speaker 128:04 - 28:13
And as a result today, know, almost every major financial institution in the world, almost every government, you know, relies on us in order to stay safe online. Speaker 128:04 - 28:13
结果就是，今天，几乎世界上所有主要的金融机构、几乎所有政府，都依赖我们来保障其在线安全。
Speaker 228:14 - 28:15
And and hilarious. Speaker 228:14 - 28:15
而且这还特别好笑。
Speaker 128:15 - 28:40
And Michelle will be mortified when she's she's like you need to stop. Was I Allen and Company, the big investment bank, invited me to one of their conferences they do in in Arizona and I told this story and I haven't been invited back since. So so there you go. Speaker 128:15 - 28:40
Michelle 听到这里一定会尴尬死，她会说你得停下了。我当时是不是在 Allen and Company？那家大型投资银行曾邀请我去参加他们在 Arizona 办的一场会议，我讲了这个故事，从那以后他们就再也没邀请过我。所以，事情就是这样。
Speaker 228:40 - 29:00
Okay. Well, thanks for sharing. You know, the fact that you found your initial, customer, sort of segments in like Hacker Kids and, human right organizations, you know, when you when you think about like VC slides and and and raising money, that's not exactly your first obvious, know, Silicon Valley kind of playbook. Speaker 228:40 - 29:00
好的。嗯，谢谢你的分享。你知道，你们最初找到的客户细分群体，像 Hacker Kids 和 human rights organizations（人权组织）这样的对象；当你想到 VC（风险投资）路演材料和融资时，这显然并不是那种一眼就能想到的、典型的 Silicon Valley 式 playbook（打法）。
Speaker 129:00 - 30:05
No, it interesting I remember VC early on said okay you guys have gotten really popular but how are you gonna expand outside of you know the Bay Area? And I remember sort of saying and person was just completely incredulous. I was like, we're really big in like, you know, Tunisia but we don't have anyone in the Bay Area. And so it turned was slow, wherever sort of, you know, it was more lawless, those were the places that we initially got, you know, most of our traffic and it actually took a really long time for us to win sort of the tech natives, the YC crowd, you know, the the folks that were were that would be sort of what you would think would be our natural audience. Whereas, you know, I mean we were, again, if you went to Congo, you know, 50% of the top 100 local websites in Congo, you know, were using Cloudflare because it was a because the further you were actually away from sort of the tech epicenters, the the more valuable what we provided was. Speaker 129:00 - 30:05
对，不，这很有意思。我记得早期有 VC 说，好吧，你们现在已经很受欢迎了，但你们打算怎么扩展到 Bay Area 之外呢？我记得我当时大概这么回答，而对方完全不敢相信。我说，我们在 Tunisia 之类的地方其实做得非常大，但在 Bay Area 反而一个客户都没有。结果就是，发展速度比较慢的地方，或者说更“无法无天”一点的地方，恰恰是我们最初获得大部分流量的地方。而且，我们花了非常长的时间，才赢得那些 tech natives（原生技术圈用户）、YC crowd，也就是那类你原本会觉得应该是我们天然受众的人。相反，你知道，如果你去 Congo，当地排名前 100 的网站里，大概有 50% 都在使用 Cloudflare，因为你离那些技术中心越远，我们提供的东西就越有价值。
Speaker 230:05 - 30:11
Yeah. Goes to show there's not one way of doing things and no one playbook. Speaker 230:05 - 30:11
是啊，这说明做事并不只有一种方式，也不存在唯一的 playbook（标准打法）。
Speaker 130:11 - 30:53
No. And and again, I I think, you know, it's and this this I say this to people and it sometimes kind of frustrates them, but it it like it wasn't like we wrote grand strategy memos on why we built a developer platform. We built it because we needed it, but but that's then turned out to be incredibly valuable kind of to defining what what Cloudflare is is going forward. And so at some level the story of Cloudflare has always been, you know, create these really thorny hard technical problems, solve them for ourselves and then in the process that that that builds on what it is that that, really drives the business going forward. Speaker 130:11 - 30:53
对。而且还是那句话，我觉得——我也经常这样跟别人说，虽然有时这会让他们有点沮丧——事实并不是我们写了什么宏大战略备忘录，来论证为什么要做 developer platform（开发者平台）。我们是因为自己需要，所以才去做；但这后来却变得极其有价值，甚至定义了 Cloudflare 未来会成为什么。从某种层面上说，Cloudflare 的故事一直都是：遇到这些非常棘手、非常困难的技术问题，先为自己把它们解决掉，而在这个过程中，又不断夯实那些真正推动业务持续前进的东西。
Speaker 230:53 - 31:15
Yeah. And speaking of, VCs, there was, some like fun, back and forth on X, maybe a couple weeks ago now. I guess without necessarily re litigating the specifics precisely because you had a little bit of a different kind of business, what was your path and what were some of the, I guess, lessons learned in terms of like fundraising? Speaker 230:53 - 31:15
好的。说到 VC，几周前你们在 X 上好像还有过一段挺有意思的来回讨论。我想，不一定要把当时的细节重新争论一遍；但正因为你们的业务类型有点不一样，你们当时走的是怎样一条路？以及在 fundraising（融资）这件事上，你学到了哪些经验教训？
Speaker 131:16 - 31:48
You know I think first of all I mean we ended up getting to work with just some incredible investors over time you know Carl Ledbetter, Scott Sandell, Brad Burnham at USV. You know, they they they really matched kind of our style. And what what our style tended to be was we thought of venture capitalists primarily as, know, sources of money and, you know, mostly sort of we wanted when when we needed money, we wanted to be able to to access capital. Speaker 131:16 - 31:48
你知道，我觉得首先一点是，随着时间推移，我们最终有机会和一些非常了不起的投资人合作，像 Carl Ledbetter、Scott Sandell、Brad Burnham at USV。你知道，他们的风格和我们的风格其实非常契合。而我们的风格通常是：我们主要把 venture capitalists（风险投资人）看作资金来源；基本上，当我们需要钱的时候，我们希望自己能够获得资本。
Speaker 231:48 - 31:50
Like no thought leadership. Speaker 231:48 - 31:50
比如说，不需要什么 thought leadership（思想领导力）？
Speaker 131:50 - 32:24
But we yeah. We we really just weren't like, we didn't need help hiring. We didn't we we didn't really want a buddy. We sort of wanted like, we want people who you know would push us from time to time would sort of say, hey, have you thought about what's you know over the horizon? But weren't you know in our business being like, well I think you should price it at you know $22.99 not you know $21.99 mean like we did there was like we were sort of like we know our business better than anyone else and so like if you believe in us, do that and get out of the way. Speaker 131:50 - 32:24
对，我们确实就是这样。我们并不需要别人帮我们招人，也并不真的想要一个“伙伴”式的人。我们想要的是这样的人：他们偶尔会推动我们一下，会提醒一句，嘿，你有没有想过地平线之外会发生什么？但不会深入干预我们的业务，比如说，“我觉得你们应该把价格定在 $22.99，而不是 $21.99。”我的意思是，我们当时的想法就是：我们比任何人都更了解自己的业务，所以如果你相信我们，那就支持我们，然后别挡路。
Speaker 132:24 - 33:10
We saw lots of other varieties of that, know, people who you know really wanted to be kind of operators and again I think there's not one way of doing this for some founders that's incredibly valuable to them and they want that. And so you know what we had, talked to almost everyone that was out there and there were lots of people that just thought you know we were complete idiots. Guy at like bunch of people at Emergence Capital like Santi and Gordon who I who I love. There was one guy who was just like, you're an idiot and we're never going to invest in you. And again, I get it, we kind of were. Speaker 132:24 - 33:10
我们还见过很多其他类似的情况。你知道，有些人真的很想成为那种 operator（运营操盘手），而且我还是觉得，这件事并没有唯一正确的做法；对一些 founders（创始人）来说，那种安排对他们非常有价值，他们也确实想要。所以我们当时几乎和市面上所有相关的人都聊过，也有很多人觉得我们完全是白痴。像 Emergence Capital 的不少人，比如 Santi 和 Gordon，我很喜欢他们。其中有一个人就直接说，你是个白痴，我们永远不会投你。说实话，我也能理解，我们某种程度上确实挺像的。
Speaker 133:10 - 33:55
But that's okay because we found people who did believe in us and have, you know, done incredibly well coming along the journey with us and it made sense. But we were an odd you know company, know, think that if you look at the three of us that founded it, most companies get founded by by people who are friends with each other. Michelle Lee and I, we were not friends. It wasn't we were enemies, I think we had a ton of respect for each other but we were just radically different people. To this day like, you know, Michelle like she has a giant pink piano in her living room and her idea of a good time is at night to sit down with her kids and she plays piano, her husband plays guitar and the kids sing along. Speaker 133:10 - 33:55
但这也没关系，因为我们找到了那些相信我们的人，而且他们一路和我们同行，最终也做得非常好，这样就说得通了。不过，我们本来就是一家很特别的公司。我觉得如果你看我们三个创始人，大多数公司都是由彼此是朋友的人创办的。Michelle Lee 和我当时并不是朋友。也不是说我们是敌人，我觉得我们彼此非常尊重，但我们就是截然不同的人。直到今天，Michelle 的客厅里还有一架巨大的粉色钢琴，她理想中的快乐时光，是晚上和孩子们坐在一起，她弹钢琴，她丈夫弹吉他，孩子们跟着唱。
Speaker 133:55 - 34:58
That sounds like the lowest circle of hell to me. And but but again like that's but we we're we've obviously you know we've become friends over time because we you know built this together but we were radically different as people and I think that some investors when they looked at us they were like, How does this Motley Crue fit together? But what was powerful was because we were so different and the way Michelle describes it is she's like when you're starting a company, want to you want it to be like a Venn diagram where you know you've got big circles and they have just enough overlap that you can like still communicate with the other people but that your lanes are incredibly clearly defined. And so you know in the you know sixteen now, almost seventeen years that we've been at this, like Michelle and I have never really had a fight. Because and because fundamentally, like what I do and what she does are just in, you know, we we each have our own lanes and we can we can we can succeed in those those lanes. Speaker 133:55 - 34:58
对我来说，那听起来像地狱的最底层。不过话虽如此，我们后来显然还是随着一起创业、共同建设这家公司，慢慢变成了朋友，但作为人，我们本质上非常不同。我觉得有些投资人在看我们的时候会想，这个 Motley Crue（杂牌军）到底怎么凑到一起的？但真正有力量的地方在于，正因为我们如此不同——Michelle 的说法是，创业时你希望团队像一个 Venn diagram（维恩图）：每个人都是很大的圆，彼此有刚刚好的重叠，这样你们还能互相沟通，但各自的赛道又定义得极其清晰。所以在这十六、快十七年的过程中，Michelle 和我几乎从来没真正吵过架。因为从根本上说，我做的事和她做的事就是分属不同领域；我们各自有自己的赛道，也都能在自己的赛道里做成功。
Speaker 134:58 - 35:52
And even though, you know, even though I am good at some things, she is good at a completely different set of things and then that just really drives a lot of what I think has caused us to be successful over time and we now get founders that are coming to us for advice all the time and one of the most common questions we get is like how did you guys split up responsibilities? And you know, I I it depends on how snarky a mood I'm in. But if you're asking that question, probably means you have the wrong co founders because fundamentally like what you want is is to cover as much surface area as possible and for each person to have a very distinct lane. And so in our case, you know, Lee was the technical genius that was that just was heads down writing code. It didn't want to be on the board, didn't want to meet with investors, was sort of, you know, but he he but it but it absolute technical genius. Speaker 134:58 - 35:52
而且尽管我擅长一些事情，她擅长的却是完全不同的一套能力，而这确实极大推动了我们这些年来的成功。现在也经常有 founders（创始人）来向我们请教，其中一个最常见的问题就是：你们当初是怎么划分职责的？至于我的回答，就看我当时有多想挖苦人了。但如果你在问这个问题，可能往往就意味着你找错了 co-founders（联合创始人），因为从根本上说，你想要的是尽可能覆盖更大的能力面，而且让每个人都拥有非常明确的赛道。就我们而言，Lee 是技术天才，他就是埋头写代码，不想进 board（董事会），不想见投资人，总之就是那种人——但他绝对是技术天才。
Speaker 135:53 - 36:30
Michelle was sort of the operations person who you know, like would take in a thing that worked once and figure out how to scale it to you know a million times. And so you know ended up running a lot of like our go to market organization, those things that you need to have real process and scale. And then I'm a good storyteller and so and so like in this, you know, the case of the VC thing, Vinod Khosla, we he's thinking of investing in us. I am doing most of the talking because I'm the good storyteller. Michelle's, you know, thinking about I mean, she was a she literally was a Six Sigma black belt. Speaker 135:53 - 36:30
Michelle 更像是运营负责人，她能把一个只成功过一次的东西，想办法扩展到一百万次。所以后来她负责了我们很多 go-to-market（市场推进）组织相关的事情，也就是那些真正需要流程和规模化的部分。而我比较会讲故事，所以在 VC（风险投资）那类场景里——比如 Vinod Khosla 在考虑要不要投我们的时候——基本上是我在负责大部分表达，因为我是那个更会讲故事的人。Michelle 当时想的则是另一套东西——她可是货真价实的 Six Sigma black belt。
Speaker 136:30 - 36:53
Like, she's she's a process person and and she's incredibly good at it. I'm terrible at that. But like she's doing that and Lee's like, why am I at this dinner? Like seriously, could be I have so many more productive things going on if I was just back writing code. And so again, I don't I don't blame people for saying like, you know, if you're pattern matching off, you know, everyone's got to be a good storyteller. Speaker 136:30 - 36:53
她就是那种流程型的人，而且对此极其擅长。我在这方面就很糟。但她负责那个，Lee 则会想：我为什么要来参加这顿晚饭？说真的，如果我回去写代码，我手头有一大堆更高产的事情可以做。所以还是那句话，我并不怪别人会说，如果你是按某种 pattern matching（模式匹配）来判断，那似乎每个人都得是个会讲故事的人。
Speaker 136:53 - 37:06
That just wasn't what we were but I actually think there's real strength in that because you want different people with different perspectives and different skills. And I think that's you know part of why we were such a strong founding team. Speaker 136:53 - 37:06
但我们本来就不是那样。不过我其实认为这里面有真正的优势，因为你就是需要拥有不同视角、不同技能的不同的人。我觉得这也是为什么我们会成为一个如此强的创始团队的重要原因之一。
Speaker 237:06 - 37:36
Okay awesome. Let's talk about the AI stuff at Cloudflare from a product standpoint because you guys have built so much and I think the world is starting to realize but I was struck, I'm like I knew some of the things you build but as I was prepping for this like it's just like such a list. So walk us through the general strategy and maybe timeline of like when you started building for AI and what you built. Speaker 237:06 - 37:36
好，非常好。我们来聊聊 Cloudflare 的 AI 相关业务，从产品角度来讲，因为你们已经做了非常多东西。我觉得外界现在也开始意识到这一点了；但让我印象很深的是，我之前就知道你们做过其中一些东西，可在我为这次采访做准备时，才发现那真的是长长一串清单。所以请你带我们梳理一下整体战略，也可以顺带讲讲时间线：你们是什么时候开始为 AI 构建产品的，以及具体都做了哪些东西。
Speaker 137:37 - 38:15
Yeah, so I think that the real kind of beginning of it starts back in well I mean at some level Cloudflare has always been an AI company. I remember pitching Benchmark actually and and I said, know, Cloudflare is an AI company and like Matt Kohler rolled his eyes so like that I was like I'm never gonna say that ever again. And that was back in like 2010 or something. And but at some level, that's always been true. Again, Turkish escorts, like that's a machine learning system that's recognizing patterns and then and and that whole thesis was if you get enough traffic going through our system and you can run machine learning systems against it Speaker 137:37 - 38:15
对，所以我觉得这件事真正的开端其实可以往前追溯。某种程度上，Cloudflare 一直都是一家 AI 公司。我记得我当年向 Benchmark 做融资推介时，曾说过，Cloudflare 是一家 AI 公司，然后 Matt Kohler 还翻了个白眼，所以我当时就想，我以后再也不这么说了。那大概是在 2010 年前后。不过从某种层面上看，这一直都是真的。再比如 Turkish escorts，这本身就是一个 machine learning（机器学习）系统，在识别模式，而整个核心论点就是：如果有足够多的流量经过我们的系统，并且你可以对这些流量运行 machine learning 系统，
Speaker 238:15 - 38:15
Yeah. Speaker 138:15 - 39:32
You can do interesting things. So so like it's always been kind of in our DNA and if you if you actually kind of look at where the real start of Cloudflare was, it it came from Paul Graham before Y Combinator used to throw a used to throw a conference at MIT called the MIT Anti Spam Conference and and it was it was really a machine learning conference because spam was this perfect machine learning place and you know like Jon Graham coming who's now on our board but was our CTO for a really long time like he was a machine learning guy and so like it was always kind of in our in our core but but we would but we would not have really described ourselves. I think that the modern era of kind of Cloudflare and AI started really in 2017 and what had happened was we were deploying software in the same way that everyone was deploying software, you know, and but but our network kept growing and then the the size and importance of the customers that we had kept growing. And it it was terrifying because if you pushed when you push code out it, you know, there was always a risk you'd make a mistake and take things down and we had had done that and I remember getting a call from the CEO of FedEx at the time and we would we were in the midst of an outage and he was like when are you gonna be back online? Speaker 138:15 - 39:32
你就能做出一些很有意思的事。所以这其实一直都写在我们的 DNA 里。如果你真的去看 Cloudflare 最初真正的起点，它其实来自 Paul Graham——在 Y Combinator 之前，他常常在 MIT 办一个叫 MIT Anti Spam Conference 的会议，而那实际上本质上是一个 machine learning 会议，因为 spam（垃圾信息）是 machine learning 非常理想的应用场景。还有像 Jon Graham——他现在是我们董事会成员，但曾长期担任我们的 CTO——他本身就是做 machine learning 的。所以这一直是我们的核心组成部分，只不过我们当时并不会真的这样定义自己。我觉得 Cloudflare 与 AI 的现代阶段，真正开始于 2017 年。那时发生的情况是，我们部署软件的方式和其他所有人一样，但我们的网络规模不断扩大，我们客户的体量和重要性也不断上升。这让人非常害怕，因为每次把代码推上线时，始终都有出错并导致服务中断的风险，而我们也确实发生过这种事。我记得当时接到 FedEx CEO 的电话，那时我们正处于一次 outage（故障）之中，他问我：你们什么时候能恢复在线？
Speaker 139:33 - 40:21
And and I and I was like, any second now, we're so sorry, but you're not even a customer. And he's like, no no, but we rely on a bunch of your customers and some of our planes can't land if you're off if you're offline. And they could technically land but they had a checklist of procedures and one of these things was to check against you know some systems that were on us and that was just terrifying to us. And so we thought we've got to build a new way of developing and deploying software which can be extremely lightweight, it can scale you know automatically to whatever whatever size you need. It can be you know rolled out to a single customer or a group of customers or to all of our customers in you know really fine grain controlled way. Speaker 139:33 - 40:21
我当时说，随时，真的非常抱歉——可你们甚至都不是我们的客户。然后他说，不不，但我们依赖你们很多客户，而如果你们离线，我们有些飞机就没法降落。严格来说飞机当然还是能降落，但他们有一套操作检查清单，其中一项就是要核对一些部署在我们这里的系统。那对我们来说实在太可怕了。所以我们就想，我们必须建立一种全新的软件开发和部署方式：它要非常轻量，能够自动扩展到你需要的任何规模；它要能以非常细粒度的方式发布给单个客户、一组客户，或者我们所有客户；
Speaker 140:21 - 41:09
You can pull it back very quickly and it needs to be cost effective because we're you know we're still running and trying to you know build this as efficiently as possible and that gave rise to the Cloudflare Developer Platform which we call Cloudflare Workers. I wish I could say that we saw you know what was coming in AI and we did in some cases back in 2020 either late twenty one or early twenty two, I can't remember. We did a partnership with this, you know, graphics processor company to put GPUs at the edge of our network in order we said to be able to do a bunch of interesting things including inference. Issued a press release, did did a whole thing. It was crickets. Speaker 140:21 - 41:09
它还必须能够非常快速地回滚，而且成本要足够可控，因为我们当时仍在运营公司，也在努力尽可能高效地建设这些能力。这最终催生了 Cloudflare Developer Platform，我们称之为 Cloudflare Workers。我倒希望自己可以说，我们当时已经看到了 AI 将要到来的趋势；在某些情况下，确实是这样。大概在 2020 年、2021 年底或 2022 年初——具体我记不清了——我们和那家 graphics processor 公司合作，把 GPU 部署到我们网络的 edge（边缘），目的是做很多有意思的事情，其中就包括 inference（推理）。我们还发了新闻稿，做了整套宣传。结果毫无反响。
Speaker 141:09 - 42:12
No one cared And I'm like, that's surprising. And that graphics chip company was Nvidia and hilariously like two or three years later in 2024, we we and you can line this up. We actually issued the exact same press release, just changed the date and our stock doubled. Like and so and so I think we've always kind of been in this space thinking about it and I think because we are so close to users around the world and we are so connected to everything else, it makes us just a very natural place to take some of these AI tasks. So you know inference we're really really good at and and you know almost all of the models you can now use Cloudflare to directly access and you get better performance, you get better responsiveness, we can deliver at a much lower lower cost than than most most anyone else. Speaker 141:09 - 42:12
根本没人关心。我当时想，这倒挺让人意外的。而那家 graphics chip 公司就是 Nvidia。很有意思的是，两三年后的 2024 年，我们——你可以去对照时间——实际上发了几乎一模一样的新闻稿，只是改了日期，然后我们的股价就翻了一倍。所以我觉得，我们其实一直都在这个领域里思考这些事情。而且因为我们离全球用户都非常近，又与其他一切都高度连接，这让我们天然成为承载一些 AI 任务的理想位置。所以比如 inference，我们就做得非常非常好；现在几乎所有模型，你都可以通过 Cloudflare 直接访问，而且性能更好、响应更快，我们还能以远低于大多数其他人的成本来交付这些能力。
Speaker 242:12 - 42:15
And those are open source models that Speaker 242:12 - 42:15
而这些是 open source（开源）模型，
Speaker 142:15 - 42:38
But even the even the big like, know, OpenAI's, you know, big customer of ours, know, all their mobile app is built, on us. We have direct access to them, and we can get you better performance than you can if you oftentimes just go directly to them. Anthropic, again, another big customer. All of cloud.ai sits on top of us. Speaker 142:15 - 42:38
但即便是那些大的，比如 OpenAI，也是我们的重要客户，他们整个 mobile app 都构建在我们之上。我们和他们有直接连接，而且很多时候，相比你直接去访问他们，我们反而能给你更好的性能。Anthropic 也是一样，也是我们的大客户。整个 cloud.ai 都跑在我们之上。
Speaker 242:38 - 42:44
So But if I'm a developer and I'm looking to run Kini, Mistral, whatever, I can do this on Speaker 242:38 - 42:44
所以，如果我是一个 developer（开发者），想运行 Kini、Mistral 之类的模型，不管是什么，我都可以在这上面做。
Speaker 142:44 - 42:56
cloud Totally. .Run on us. And in that case, will be processed on the GPUs that are sitting in our network distributed around the world in you know hundreds of cities. Speaker 142:44 - 42:56
完全是这样。……在我们这里运行的话，这些任务会在我们分布于全球网络中的 GPU 上处理——你知道的，遍布数百座城市。
Speaker 242:56 - 43:01
That's in France, the edge network doesn't lend itself to training necessarily? Not Speaker 242:56 - 43:01
那是在 France；edge network（边缘网络）并不天然适合用来做 training（训练），对吗？并不是——
Speaker 143:03 - 43:46
today. You know I think that if you think about what you need for training, you need lots and lots and lots of machines with very powerful GPUs. We have both of those things but what we don't have is a way of like we don't have an InfiniBand, you know, network fabric that's connecting all of those machines. And so where an AWS or even in SpaceX, they're building giant data centers with lots of machines very close together hyper networked. We have just as many machines but they're spread out all around the world and as a result most training tasks were not the right place to do that. Speaker 143:03 - 43:46
现在还不是。你知道，我觉得如果你想想 training 需要什么，你需要非常非常多的机器，以及性能很强的 GPU。我们这两样都有，但我们没有的是一种——比如说——我们没有把所有这些机器连接起来的 InfiniBand 网络 fabric（网络互连架构）。所以像 AWS，甚至 SpaceX，他们在建设巨型 data center（数据中心），里面有大量彼此距离很近、并且通过高密度网络连接起来的机器。我们的机器数量一样多，但它们分散在全球各地，因此大多数 training 任务并不适合放在我们这里做。
Speaker 143:46 - 44:30
We're always testing and experimenting and I know I I don't know that that will necessarily be be the case forever. But but but what we are really good at is anything that you need kind of a beefy machine where like your phone or your laptop isn't enough and you want it to be either fast or or lower cost or, or or you have some sort of regulatory jurisdictional reason why why you want it to run, you know, in a particular region and not necessarily get set back to, you know, Ashburn, Virginia. And, and I think that's driven, you know, a huge amount of the adoption of these AI workloads, across across our network and that's by far the fastest growing part of our business. Speaker 143:46 - 44:30
我们一直都在测试和实验，我也不敢说这种情况永远都会如此。但我们真正擅长的是：凡是你需要一台比较 beefy（高性能）的机器、而你的手机或 laptop（笔记本电脑）又不够用的场景；并且你希望它要么更快、要么成本更低，或者你出于某种监管或司法辖区原因，希望它运行在某个特定 region（区域）里，而不一定被发回到 Ashburn, Virginia。并且我认为，这推动了大量 AI workload（AI 工作负载）在我们网络上的采用，而这也是迄今为止我们业务增长最快的部分。
Speaker 244:30 - 44:40
So that's workers. Another product that seems incredibly timely given the current conversation is the Gateway. Do you want to talk about that, what that does? Speaker 244:30 - 44:40
这就是 workers。另一个产品，考虑到当前的讨论，看起来时机也非常合适，就是 Gateway。你想讲讲那个吗？它是做什么的？
Speaker 144:40 - 45:41
Yeah, so I mean again this it's all driven by by us and so we needed to be able to use AI models but we're a security company where you know we have you know various regulatory requirements based on the customers that we have and so we needed some way to be able to understand like how our team was using these these models and so we built something that we you know called the AI gateway and and it allows you to do a couple of things. One is, it allows you to audit as if you're the CIO or CTO audit all of the sort of the prompts that have gone off to the various AI systems and then all the responses that you get back. And we actually have an AI system that's that allows you to very quickly surface things based on you know, what you might think of as problematic. If you've got a customer support agent and it starts acting like a Nazi, like you want to be able to see that, you know, incredibly quickly. And so we give you that auditing. Speaker 144:40 - 45:41
对，所以我的意思是，这同样也是由我们自己的需求驱动的。我们需要能够使用 AI model（AI 模型），但我们是一家 security（安全）公司，你知道，基于我们服务的客户，我们要满足各种监管要求。所以我们需要某种方式，来了解我们的团队是如何使用这些模型的，于是我们构建了一个我们称为 AI gateway 的东西，它能让你做几件事。第一，它允许你——如果你是 CIO 或 CTO——去审计所有发往各种 AI system（AI 系统）的 prompt（提示词），以及你收到的所有 response（响应）。实际上，我们还有一个 AI system，能让你非常快速地发现一些你可能会认为有问题的内容。如果你有一个 customer support agent（客服 agent），而它开始表现得像个 Nazi，那你肯定希望能非常快地看到这一点。所以我们提供这种审计能力。
Speaker 145:41 - 46:41
We also allow you to do basically you know prompt injection the way that if if you're in the way that Claude or the way that Anthropic sort of put guardrails around around Fable in order to take what was mythos and and turn it into something that they they thought they could release broadly is they put a whole bunch of prompts and say do not do you know the following things. You as a business like if you're if you're Cloudflare, you know there are certain things that we don't want AI systems to do and the AI gateway allows you to to inject that into it. And then the and then the third thing which again we built for ourselves was we need to control costs. The the cost of all these tokens just, know, it can can be just overwhelming and it's really easy for that to lose for you to lose control of that. Gateway allows you to route things intelligently where you don't necessarily have, I don't know, somebody who wants, you know, an AI agent to summarize their email. Speaker 145:41 - 46:41
我们还允许你做的，基本上可以理解为 prompt injection（提示词注入）。就像 Claude 或 Anthropic 给 Fable 加 guardrails（护栏）的方法一样：为了把原本是 mythos 的东西变成他们认为可以大规模发布的内容，他们加了大量 prompt，告诉它“不要做以下这些事”。而对于企业来说——如果你是 Cloudflare——我们也有一些明确不希望 AI system 去做的事，AI gateway 就允许你把这些要求注入进去。然后第三件事，同样也是我们为自己构建的，就是我们需要控制成本。所有这些 token（令牌）的成本，真的可能高得让人吃不消，而且非常容易失控。Gateway 可以让你智能地进行路由，不一定非得——比如说——有人想让一个 AI agent 总结自己的 email。
Speaker 146:41 - 47:04
Like you don't need the latest frontier model in order to do that. You can run it on something which is a lot simpler. And so for for our team, oftentimes they have no idea which model they're hitting. Gateway is just deciding that for them and again we built that for ourselves and we and people are always asking us like what are you guys doing around AI and we showed it off and and people like I need that and so now again that's turned into into a product. Speaker 146:41 - 47:04
做这种事你并不需要最新的 frontier model（前沿模型），用一个简单得多的模型就可以运行。所以对我们团队来说，他们很多时候根本不知道自己打到的是哪个 model；Gateway 会替他们做这个决定。再说一次，这也是我们为自己构建的。人们总会问我们“你们在 AI 方面都做了什么”，于是我们把它展示出来，大家就会说“我需要这个”，所以它现在又变成了一个产品。
Speaker 247:05 - 47:13
So continuing our tour, drawable objects, HGNSA SDK, workflows, browser runs, sandboxes. Speaker 247:05 - 47:13
继续我们的梳理：drawable objects、HGNSA SDK、workflows、browser runs、sandboxes。
Speaker 147:13 - 47:54
It's all just stuff we needed. And then it turns out other developers needed it as well. And so if you think about like how did AWS develop, it was it was Amazon's way of building essentially systems where, know, their own team could could build things very very quickly. And and that that was and that that that I I think the story of Cloudflare has very much been the same thing where we needed all of these things to build what we were doing and then that's turning out to be incredibly valuable to our to our customers and and to anyone who's trying to do anything in AI today. There we are the best place to build any kind of agentic workload. Speaker 147:13 - 47:54
这些本来都只是我们自己需要的东西。后来发现，其他开发者也同样需要。你可以想想 AWS 是怎么发展起来的：它本质上是 Amazon 为了让自己的团队能够非常快速地构建东西，而搭建出的一套系统。我认为 Cloudflare 的故事在很大程度上也是一样的：我们需要这些东西来构建我们自己在做的事情，而结果证明，它们对我们的客户，以及今天任何试图做 AI 的人来说，都极其有价值。对于构建任何类型的 agentic workload（智能体工作负载），我们都是最好的平台。
Speaker 147:55 - 48:41
An example is if you you take every knowledge worker on earth and you imagine that they're they're each gonna have one agent that is working on their behalf. It turns out that the CPU utilization of that is more, it's 40 times the current annual CPU production that's out there. If you're doing it in a traditional sort of AWS containers based way, it's just it's just too expensive to do that. And I think the idea of every knowledge worker having just one agent is insane. You're gonna have hundreds of different agents that are working on your behalf in which case we don't we just don't have the capacity to be able to provide that. Speaker 147:55 - 48:41
举个例子，如果你把地球上所有 knowledge worker（知识工作者）都算进去，并设想他们每个人都会有一个 agent（智能体）替他们工作，那么这种需求对应的 CPU 利用量，会超过当前全球年度 CPU 产能的 40 倍。如果你用传统那种基于 AWS containers（容器）的方式来做，这件事成本高得根本无法成立。而且我认为“每个知识工作者只有一个 agent”这个想法本身都很离谱。实际上，你会有几百个不同的 agent 在替你工作；那样的话，我们根本没有足够的容量来提供支持。
Speaker 148:41 - 49:45
And so what we've done is said there's got to be a better way of doing that. And so again what Workers did is it said you know if once by time we had you know physical servers and then we had VMs then we had containers, Workers said what's the next step beyond that that's even more efficient and we built all of Workers around a concept which is which everyone's familiar with because it's the tabs in your browser which are isolates and that that is a you know safe environment that is sandboxed from the other tabs that are running. But it doesn't need to have you know a whole new copy of the operating system. It doesn't need to have a whole bunch of tools and as a result you can stand it up much more quickly, you can tear it down much more quickly and for agentic workloads where your agent is effectively going to be right, whatever you're asking it to do, it's effectively going to be writing code for you all the time. It that is turning out to be kind of the perfect environment to be able to to be able to build these things. Speaker 148:41 - 49:45
所以我们做的是说，必须有一种更好的实现方式。Workers 所做的，就是在“先有物理服务器，再有 VM（虚拟机），再有容器”之后，继续追问下一步是什么，什么会更高效。于是我们围绕一个大家其实都很熟悉的概念构建了整个 Workers——也就是你浏览器标签页背后的 isolates（隔离执行环境）。那是一种安全的环境，会和其他正在运行的标签页彼此 sandbox（沙箱）隔离。但它不需要一整套新的操作系统副本，也不需要一大堆工具，因此你可以更快地把它启动起来，也可以更快地把它销毁。对于 agentic workloads 来说——因为你的 agent 实际上会一直在替你写代码，不管你让它做什么——这种环境结果证明，几乎就是构建这类系统的完美运行环境。
Speaker 149:45 - 50:01
And again, I wish it was because we were so clever and and saw the future, but I think a lot of it is it was it was what we needed to build for ourselves and it's turning out to be just a very powerful platform for for anyone who's doing doing anything kind of on the modern Internet. Speaker 149:45 - 50:01
还是那句话，我真希望这是因为我们特别聪明、提前看到了未来，但我觉得更大程度上，这只是我们为了自己必须构建出来的东西，而现在它正证明自己是一个非常强大的平台，适用于任何在现代 Internet 上做事情的人。
Speaker 250:01 - 50:08
That looming CPU shortage crisis, by the way, it seems to be an under somewhat under discussed topic. Speaker 250:01 - 50:08
顺便说一句，那个正在逼近的 CPU 短缺危机，似乎是一个讨论得还不算充分的话题。
Speaker 150:08 - 50:55
I mean these agents like you need a lot of GPUs, we all know how it, but like CPU utilization is is going through the roof. And and that's going to be, know, in a bunch of different a bunch of different directions. And so we're constantly trying to think about, you know, how can we improve the underlying kind of efficiency of the network. So you know like a while back three or three or four months now, we rewrote WordPress in in Rust in order to in order to deliver it and called it Em Dash because we used to play on the fact that we used AI to do it and people are like, you know, why'd you do that? And and we did it because like there's a huge community that supports WordPress. Speaker 150:08 - 50:55
我是说，这些 agent 确实需要大量 GPU，这一点大家都知道；但 CPU 利用率其实也在飙升，而且会在很多不同方向上持续上升。所以我们一直在思考，怎样才能提升底层网络的整体效率。比如说，三四个月前吧，我们用 Rust 重写了 WordPress 的实现来交付它，并把它叫作 Em Dash——因为我们之前还拿“这是用 AI 做的”这件事开过玩笑。很多人会问，你们为什么要这么做？我们这样做，是因为有一个非常庞大的社区在支持 WordPress。
Speaker 150:55 - 51:31
WordPress should be successful going forward but there's no way that WordPress is going to be able to scale for what's coming going forward. It's just, know, it's a it's a PHP piece of software. Again, amazing for what it was but it costs like, you know, 3 or $4 at current traffic levels in order to run. If traffic explodes a thousand times, like people aren't going to be putting WordPress sites up because it's gonna cost 3 or $4,000 in order to be able to to run these things. So we've got to reinvent a bunch of these fundamental technologies that are out there. Speaker 150:55 - 51:31
WordPress 未来当然应该继续成功，但按未来将要到来的规模，它不可能继续扩展下去。它终究只是一个 PHP 软件。再次强调，对它所处的时代而言，它已经非常了不起了；但在当前流量水平下，运行它大概要花 3 到 4 美元。如果流量暴涨一千倍，人们就不会再去部署 WordPress 网站了，因为那样运行这些东西的成本会变成 3000 到 4000 美元。所以，我们必须重造一大批现有的基础技术。
Speaker 151:31 - 52:13
Not not not not because not not because we're like, oh, we want it, you know, we need to own it. We open source, we give it, you know, we give it away, but because we believe that our mission is to help build a better Internet and what the the tidal wave that is coming if we don't get in front of it, we don't upgrade some of these fundamental systems, know, it's it's going to it's going to make you know a lot of the Internet just break because we have to have an infrastructure that can actually scale and that's going to be you know, how do we how do we take the the user interfaces and the things that people know and are familiar with and have an ecosystem around them but then port them into ways that they're just significantly more cost effective and efficient to run. Speaker 151:31 - 52:13
不是不是不是不是因为，不是因为我们会觉得，哦，我们想要它，你知道，我们需要拥有它。我们会把它 open source（开源），我们会把它给出去，你知道，我们会免费提供它；而是因为我们相信，我们的使命是帮助建设一个更好的 Internet（互联网）。如果即将到来的这股巨浪我们不提前应对，如果我们不升级这些基础系统，你知道，Internet 的很多部分都会直接坏掉，因为我们必须有一套真正能够 scale（扩展）的基础设施。问题在于，我们要怎么把用户已经知道、熟悉、并且围绕其形成了生态系统的那些 user interface（用户界面）和事物，迁移到运行成本显著更低、效率显著更高的方式上。
Speaker 252:13 - 52:19
In terms of breaking, what does an agentic internet mean in terms of security? Speaker 252:13 - 52:19
说到“坏掉”，agentic internet（代理式互联网）在安全性方面意味着什么？
Speaker 152:19 - 53:19
Yeah, mean it's security is going to be, I think for the next two years, we're going to see just a series of really scary things happen online. Know, probably the worst kind of bug that was out there for a long time was this thing called Log4j where you could send a very simple command and compromise almost any any server running anywhere. You know, I think for for the next, you know, a hundred and four day a hundred and four weeks, so two years, you're gonna see a Log4j like vulnerability every single week. Because like and again, we were we were part of Glasswing and and got to use, you know, Mythos to look at software. We get early releases of the various OpenAI models. Speaker 152:19 - 53:19
对，我觉得在接下来的两年里，安全会变成一件非常可怕的事，我们会看到一连串真的很吓人的线上事件发生。你知道，过去很长一段时间里，可能最糟糕的一类 bug（漏洞）就是那个叫 Log4j 的东西：你只要发出一个非常简单的命令，就几乎可以攻陷任何地方运行的任何 server（服务器）。我觉得在接下来的，你知道，104 天——不，104 周，也就是两年里——你会看到每周都出现一个类似 Log4j 的 vulnerability（漏洞）。因为，就像我说的，我们也参与过 Glasswing，并且用过 Mythos 来分析 software（软件）。我们还能提前拿到各种 OpenAI model（模型）的早期版本。
Speaker 153:19 - 54:05
We've built our own security models. These models are incredible at finding vulnerabilities and they're gonna find them like crazy and I think security is going to be it's gonna seem really scary for the next little bit. There's gonna be a bunch of VCs that invest a bunch of money because you're gonna see just, you know, an explosion in cybersecurity companies that are stopping this. But what's interesting is I think it's gonna be a little bit of a head fake because I think that two years from now what's going to happen is software is going to get just massively better as a result. So internally at Cloudflare, you know we like last year we had an outage that was incredibly embarrassing and awful and we said we have to get better. Speaker 153:19 - 54:05
我们也构建了自己的安全模型。这些模型在发现漏洞方面非常惊人，而且会疯狂地找到漏洞。所以我觉得，接下来一小段时间里，安全会显得非常吓人。会有一堆 VC（风险投资）投入大量资金，因为你会看到大量 cybersecurity（网络安全）公司爆发出来，专门阻止这些问题。但有意思的是，我觉得这可能有点像个 head fake（假动作），因为我认为两年之后会发生的是：software 会因此变得大幅更好。所以在 Cloudflare 内部，你知道，去年我们发生过一次 outage（宕机），那次事故非常尴尬，也非常糟糕，于是我们说，我们必须做得更好。
Speaker 154:05 - 54:54
And so we actually built an agent that not only reviews every code release that we we send out, but every configuration change. So if you go into the cloud flow dashboard, push a button that generates a file that then gets pushed out to the rest of our network. And we now have an agent that's checking all of those files every code release and it's trained on ten years of incidents. We have big public incidents that you know make the news, but there's there's sort of a background noise of small incidents over time and like one of those sort of just audible gasp moments that we had was a guy who runs our infrastructure engineering team who was presenting at our all hands meeting and he said okay here's sort of our background noise of incidents and then there's a cliff and then it goes like this. He's like what do you think happened there? Speaker 154:05 - 54:54
所以我们实际上构建了一个 agent（智能体），它不仅会审查我们发出的每一次 code release（代码发布），还会审查每一次 configuration change（配置变更）。也就是说，如果你进入 Cloudflare dashboard（控制台），按下一个按钮，就会生成一个文件，然后这个文件会被推送到我们网络的其他部分。现在我们有一个 agent 在检查所有这些文件、每一次代码发布，而且它是基于十年的 incident（事故）数据训练出来的。我们有一些重大的公开事故，会登上新闻；但除此之外，随着时间推移，其实一直还有一种小事故构成的“背景噪音”。我们当时有一个几乎让人倒吸一口气的时刻：负责我们基础设施工程团队的一个人，在全员会上做展示，他说，好，这里是我们事故的背景噪音，然后这里突然断崖式下降，接着就变成这样了。他说，你们觉得那里发生了什么？
Speaker 154:54 - 56:04
And that was the moment that we released this agent that's checking in. I think one of the things that people don't appreciate and don't talk about enough with AI is, yeah, AI is you know good at replacing a bunch of tasks that the humans do and it doesn't tire and it doesn't need breaks and it doesn't you know have have have to sleep. But the really interesting thing is it's also uncorrelated to all the other biases that are within an organization. A team of people working together will tend to develop a set of biases and that doesn't mean that AI doesn't have biases, but they're uncorrelated with the rest of the team. And so it's actually incredibly good at measuring things and what I think you're gonna see is that if it's gonna seem like software is incredibly scary for the next two years and then it'll kind of just fade away and we're gonna all be releasing software which is so much better than it ever has been before and so it's like our uptime and our reliability and our performance has gone up you know an order of magnitude over the course of the last year and it's all because we're using agents in order to more securely you know build build our systems and I think everyone is is going to be doing that going forward. Speaker 154:54 - 56:04
那就是我们发布这个检查流程 agent 的时刻。我觉得人们没有充分意识到、也没有被足够多讨论的一点是：AI（人工智能）当然擅长替代人类做的一大堆任务，而且它不会累，不需要休息，也不需要睡觉。但真正有意思的是，它和一个组织内部其他偏见之间是“不相关”的。一个一起工作的团队，往往会逐渐形成一套共同偏见；这并不是说 AI 没有偏见，但它的偏见和团队其他人的偏见并不相关。所以它实际上特别擅长做测量和校验。我认为你将会看到的是：software 在接下来的两年里会显得极其吓人，然后这种感觉会慢慢淡去，而我们最终发布的软件会比过去任何时候都更好。所以你知道，在过去一年里，我们的 uptime（在线率）、reliability（可靠性）和 performance（性能）都提升了一个 order of magnitude（数量级），而这一切都是因为我们在用 agents（智能体）以更安全的方式来构建我们的系统。我觉得接下来每个人都会这么做。
Speaker 256:04 - 56:35
We are on the on that topic let's go into how you guys use AI internally. You have some interesting stats that you published, just, a few weeks ago in April, where you said the company has 93% of R and D employees, using AI coding tools, 3,683 internal users, whatever $241,000,000,000 tokens. How did you get there? How did you create that culture for a company that was started in 2009? And '2, how do you measure success? Speaker 256:04 - 56:35
既然说到这个话题，我们就来聊聊你们内部是怎么使用 AI 的。几周前，也就是 4 月，你们公布了一些很有意思的数据：公司里 93% 的 R&D（研发）员工都在使用 AI coding tools（AI 编码工具），有 3,683 名内部用户，还有 241,000,000,000 个 token（词元）。你们是怎么走到这一步的？对于一家创立于 2009 年的公司，你们是如何建立起这种文化的？第二，你们又是如何衡量成功的？
Speaker 156:35 - 57:05
We were selling picks and shovels and sort of the AI Gold Rush but we were actually pretty cautious users ourselves. We'd run experiments but but we were sort of not kind of all in on it and we had a bunch of people internally that were that were skeptical. A guy named Kenton Varda who built the Workers' Platform, this incredible engineer. You know, we're not supposed to talk about 10x engineers but Kenton is clearly a 10x engineer. He's just he's he's he's he's just a genius. Speaker 156:35 - 57:05
我们当时是在 AI Gold Rush（AI 淘金热）里卖 picks and shovels（镐和铲）的人，但实际上我们自己用得相当谨慎。我们会做实验，但还没有真正彻底 all in（全力投入）。而且公司内部有不少人对此持怀疑态度。有个叫 Kenton Varda 的人，他打造了 Workers Platform，这是一位非常了不起的工程师。你知道，现在大家不太该谈什么 10x engineer（十倍工程师），但 Kenton 显然就是个 10x engineer。他就是，他就是，他真的是个天才。
Speaker 157:06 - 57:43
And he, you know, last in the spring of twenty twenty five, he's like this ASF is is BS. I'm gonna go play with all of these tools and show you just how much of a mess they they are and why why why we shouldn't be relying on them. And and he did, he went away. He was super senior guy on our team, went away for a month played with him came back and he was like, oh my god, I am now a 100 times more productive than I was before. Which is which if that's right and he's a 10x engineer means that he alone was more productive than our entire team was back in like 2019. Speaker 157:06 - 57:43
你知道，他在 2025 年春天时还觉得，这个 ASF 简直就是 BS。他当时说，我要去把这些工具都玩一遍，给你们看看它们到底有多乱，以及为什么我们不该依赖它们。然后他真这么做了。他是我们团队里资历非常深的人，离开去研究了一个月，回来之后他说，天哪，我现在的生产力是以前的 100 倍。如果这是真的，而他本来就是个 10x engineer，那就意味着光他一个人，生产力就比我们整个团队在 2019 年时还高。
Speaker 157:43 - 58:18
Right? And so, and and like I kind of believe that's true. Like it's it's it was it was remarkable. And so internally on the engineering side, I think it spread pretty quickly and I think the real turning point for us was probably November 2025 where it felt like and it was with the release of Claude Opus and and and a bunch of things but it felt like it it tipped where we all of a sudden our engineering side we're getting you know much much much more productive. I think that's the story everyone that's sort of where everyone is today and everyone is seeing some version of that. Speaker 157:43 - 58:18
对吧？所以，而且我某种程度上相信这是真的。那真的非常惊人。所以在内部的 engineering（工程）这边，我觉得这种变化传播得非常快。而我认为对我们来说真正的转折点大概是在 2025 年 11 月，当时随着 Claude Opus 的发布，以及其他一系列事情，感觉整个局面发生了倾斜——突然之间，我们 engineering 这边的生产力高了非常非常非常多。我觉得这也是现在所有人的共同故事，大致就是今天每个人所处的位置，每个人都在看到某种版本的这种变化。
Speaker 158:18 - 59:30
I will say that you know for us you know, we've there there have been the people that I continue to worry about are the ones that have been very hesitant to embrace these tools and it tends to be not the super senior folks, not the super junior folks, it tends to be the people kind of in the middle of their career and it makes some sense where they they've been doing well in the world that existed. They've been told this is the game to play and if you play this game you're gonna succeed And so they they've played the game and now someone comes along and says we're changing the rules of the game on you. And that feels deeply deeply deeply unfair. And yet again you kind of have to adjust to what's out there but I I really do worry that we're going to have this sort of lost generation of sort of either, you know, young Millennials or older Gen Z folks that just they were sort of they're gonna have incentives to say we shouldn't use these tools because again if they do they don't feel like they have any advantage of the people who are coming from behind them. Speaker 158:18 - 59:30
我想说的是，对我们来说，我一直担心的其实是那些非常犹豫、不愿意拥抱这些工具的人。而这类人往往不是最资深的人，也不是最初级的人，往往是那些处在职业生涯中段的人。这其实也说得通，因为他们在原本那个世界里一直做得不错。别人一直告诉他们，这就是该玩的游戏，只要你按这个游戏规则来，你就会成功。所以他们就一直按这套规则在玩。结果现在突然有人来说，我们要改游戏规则了。这种感觉会让人觉得非常非常非常不公平。但话说回来，你还是得去适应外部现实。我确实担心，我们会出现某种“失落的一代”——大概是较年轻的 Millennials 或年纪稍长的 Gen Z——他们会天然有动机说我们不该使用这些工具，因为一旦用了，他们就会觉得自己相比后面追上来的人不再有任何优势。
Speaker 1 | 59:30 - 1:00:39 And so a lot of what I've been trying to do is how do you bring those groups groups along. The other piece was, know, I think engineering teams are very natural users of these these products and and coding has been a place where you've gotten really clear value from them. But like that's just a small part of Cloudflare. We've got, you know, legal, finance, know, marketing, sales, all of these other organizations that are part of the team that were like how can we make them significantly more productive as well. And so we came up with this idea that you know, could we create a sort of a user friendly system that allowed you know someone on the finance team to be able to you know do their job incredibly incredibly efficiently And and we called it Cloudflare OS and it's it's it's it's basically it's a harness where we've taken a bunch of things that have to be true and gotten them into it.
所以我一直在努力做的，很大一部分就是：怎么把这些群体也一起带上。另一部分是，我觉得 engineering 团队本来就是这些产品非常天然的用户，而且 coding（编码）一直是这些工具能够提供非常明确价值的场景。但那其实只是 Cloudflare 的一小部分。我们还有 legal（法务）、finance（财务）、marketing（市场）、sales（销售），以及所有这些其他团队组织。我们当时想的是，怎样也能让他们的生产力显著提升。所以我们提出了一个想法：能不能做出一个更 user friendly（用户友好）的系统，让 finance 团队里的人也能把工作做得极其极其高效。我们把它叫做 Cloudflare OS。它本质上是一个 harness（工作框架），我们把一系列必须成立的条件都整合进去了。
Speaker 1 | 1:00:39 - 1:01:38 We've taken a bunch of things that are sort of just informative based on our experience and then we've taught it skills where it can do various things. There's nothing that's magical about what we did except that, you know we had a pretty significant advantage because we had all the security tooling where we could hook into all the different systems of record we have, the ERP system we have, you know sales systems we have, workplace management systems that we have and do it in a way that was secure and where an agent working on behalf of a user could in could inherit the the specific kind of access that that user had. And I think that's where a lot of these things fall down is where people don't have like that security model in place. We did and so we were able to build it pretty quickly. The other thing that was was where I think people fall down is if you ask someone to describe that what they do in their job, they'll describe it but they'll leave like half of their job out.
我们把很多基于经验、主要起信息作用的东西放了进去，然后又教给它一些 skills（技能），让它能够执行各种任务。我们做的事情本身并没有什么神奇之处，唯一特别的是，我们有一个相当明显的优势：我们拥有完整的 security（安全）工具链，因此可以接入我们各种不同的 systems of record（记录系统）、我们的 ERP system、我们的销售系统、我们的 workplace management systems（工作场所管理系统），并且以一种安全的方式来完成这一切。这样一来，一个代表用户工作的 agent（代理）就可以继承该用户本身所拥有的那种特定访问权限。我认为很多这类系统失败，恰恰就是因为人们没有建立起那样的 security model（安全模型）。而我们有，所以我们能够很快把它搭起来。另一个我觉得大家容易出问题的地方是：如果你让一个人描述他在工作中做什么，他会描述，但通常会漏掉大概一半的工作内容。
Speaker 1 | 1:01:38 - 1:02:06 And so the question is how do you capture really what are the jobs to be done within an organization? And so the clever thing is a guy named Sam Ray on our team and he was like listen we're gonna set up an email address and we're gonna tell everyone that it's a magic AI agent. So for us it was autocloudflow dot com and and we told everyone it's a magic AI agent. We've set it up. You can ask it anything.
所以问题就在于：你要怎样真正捕捉一个组织内部那些“需要完成的工作”到底是什么？我们团队里一个叫 Sam Ray 的人想了个很聪明的办法。他说，听着，我们要设一个 email address（邮箱地址），然后告诉所有人，这是一个神奇的 AI agent。对我们来说，那个地址是 autocloudflow dot com。我们告诉所有人，这是个神奇的 AI agent，我们已经把它搭好了，你可以问它任何问题。
Speaker 1 | 1:02:07 - 1:03:04 Now in reality behind the scenes, it was a team of about 20 people. The staff this 20 fourseven and yes, were using AI tools, what they were really doing was gathering kind of the the information about what are the real problems that people have in their jobs and then they were building that in and turning it into a set of skill files and information to be done. And so an example of this is there's a woman named Heather who's on our Investor Relations team and so we're a public company and so every time we have earnings, it's about two weeks of prep ahead of time to generate all the IR documents and and and everything that that needs to be done and it's it's kind of arduous kind of drudgerous work to put this all all together And Heather's like, I think I can automate it all. And so we went from what used to be a two week process down to what is now a three minute process. It is more accurate.
而实际上，在幕后是一个大约 20 人的团队。他们 24/7 轮班支持。没错，他们也在用 AI 工具，但他们真正做的，是收集信息，弄清楚大家在工作中真正遇到的问题是什么，然后把这些内容逐步构建进去，把它们变成一组 skill files（技能文件）和可执行的信息。举个例子，我们 Investor Relations（投资者关系）团队里有位女士叫 Heather。因为我们是一家上市公司，所以每次到财报季之前，大约都要提前两周准备，生成所有 IR 文件以及其他所有需要完成的内容。这项工作相当繁琐，也很枯燥，要把这些东西全部整理出来并不轻松。Heather 说，我觉得我可以把它全部自动化。于是，我们把原本需要两周的流程，缩短到了现在只要三分钟。而且结果还更准确。
Speaker 1 | 1:03:04 - 1:04:03 We actually found errors in in things that we had put up before nothing nothing significant but you know little little things here and there and now we're able to take that and that means that then our IR team can spend more time meeting with investors, talking with you know our shareholders, doing all the things that are actually the really productive parts of their job and much less time actually generating those documents. And so I think that that every company is going to be going through this and it's interesting we again we we get questions like, how are you guys doing this and we keep showing off this Cloudflare OS thing. I was I was on the I'm on the organizing part part of the the group that's organizing the twenty thirty four Olympics in in Utah and I was talking to the guy named Milario who's the CTO of the IOC and he's like, listen we need this we need a system to do this and I was like, well let me show you what we're doing at Cloudflare and and so I showed it to him. We never thought this was gonna be a product.
我们实际上还发现了以前做出来的一些东西里的错误，不是什么重大错误，但你知道，就是这里一点、那里一点的小问题。现在我们能够把这部分工作交给系统处理，这意味着我们的 IR 团队就能花更多时间去见投资者、和股东交流，去做那些真正属于他们工作中高价值、高生产力的部分，而不是把大量时间花在生成这些文档上。所以我觉得，每一家公司都会经历这个过程。有意思的是，我们也不断被问到，你们到底是怎么做到这些的？而我们总是在展示这个 Cloudflare OS 的东西。我当时在 Utah 参与一个负责组织 2034 Olympics 的小组，我也在那个组织团队里。我和一个叫 Milario 的人聊过，他是 IOC 的 CTO。他说，听着，我们需要这个，我们需要一个能做这件事的系统。我当时就说，那我给你看看我们在 Cloudflare 是怎么做的。于是我就展示给他看。我们从来没想过这会成为一个产品。
Speaker 1 | 1:04:04 - 1:04:44 And he's like, I need you in Luzon, you know, next week to and to and bring your team because we need to implement this. And so we all flew up and now they're actually using this to, you know, to help break better organized Olympic Games. And and I think that that if you get and every organization is gonna have to build something like this. And if you do it in a way which is secure that gives users access, again, I think it's gonna it's gonna just massively empower teams in order to to be successful. I I'm not sort of in the sort of the I guess the Dario camp of, you know, AI is gonna replace all jobs.
他当时说，我需要你们下周去 Luzon，把你的团队也带上，因为我们需要把这个东西实施起来。于是我们就都飞过去了。现在他们实际上已经在用这个系统，来帮助更好地组织 Olympic Games。我认为，如果你把这件事做成——而且每个组织最终都得构建类似的东西——并且你是以一种既安全、又能让用户获得访问权限的方式来做，我还是觉得，这会极大地赋能团队，帮助他们取得成功。我并不属于那种——我猜可以说是 Dario 那一派——认为 AI 会取代所有工作的观点。
Speaker 1 | 1:04:44 - 1:05:17 I think it is gonna replace a certain set of jobs like I think we're gonna like if I have a developer and I can now give them AI tools and they're 10 times as productive, I'm gonna hire as many developers as I can. If I you know, I'm a salesperson and I can give them tools that eliminate the parts of their job that they don't like which is coming up with account plans and and doing all of those things. What they want is to be in front of customers and and out there selling. I'm not as many salespeople as I as I can. Where I think there's going to be a big change is the parts of the organization that that aren't building and they aren't selling, they're basically measuring.
我觉得它确实会取代某一类工作。比如说，如果我有一个 developer（开发者），现在我可以给他配上 AI 工具，让他的生产力提升 10 倍，那我就会尽可能多招 developer。如果我是做销售的，我可以给 salesperson（销售人员）提供工具，把他们工作中那些自己不喜欢的部分消掉，比如制定 account plan、处理那些杂事。他们真正想做的是站到客户面前，出去卖东西。那样的话，我也会尽可能多招 salespeople。我认为真正会发生巨大变化的，是组织里那些既不负责构建、也不负责销售，而基本上只是负责衡量、监督的部分。
Speaker 1 | 1:05:18 - 1:06:17 And it turns out that AI systems are much better measures than than any humans again in part because they're uncorrelated and I and so I think a lot of those things are gonna be the case. Another example internally, you know, audit which is again one of these things that you do when you're a public company, you probably don't do it when you're small startup but you do you do it eventually. We have a 105 or something risk areas and every quarter historically the internal audit team has picked six to 10 of them and then they do sort of these deep dive audits to make sure that we're you know not screwing up procurement or we're doing revenue recognition correctly or you know and they're measuring effectively the organization. We're now getting to a place with these tools where we can take all 105 of those risk areas and continuously audit them and find any issues that are there. And I think that that's like that's making us a better organization.
而事实证明，AI 系统在“衡量”这件事上比任何人类都强，部分原因还是因为它们彼此不相关联（uncorrelated）。所以我觉得，其中很多事情都会变成这样。再举一个内部的例子：audit（审计）。这同样是那种当你成为一家 public company（上市公司）后就必须做的事；你还是小 startup（初创公司）时可能不会做，但迟早会做。我们大概有 105 个左右的风险领域，过去每个季度，内部审计团队都会从里面挑 6 到 10 个，然后做这种深度审计，确保我们没有把 procurement（采购）搞砸，或者 revenue recognition（收入确认）做得正确，诸如此类。他们本质上是在衡量整个组织。现在借助这些工具，我们正走到这样一个阶段：可以把这 105 个风险领域全部纳入进来，持续不断地审计它们，并找出其中存在的任何问题。我觉得这会让我们成为一个更好的组织。
Speaker 1 | 1:06:17 - 1:06:30 It's freeing up people to be able to work on either building or selling which are really where value is created within organizations. And it's just incredible to see how much more productive the teams are.
这也把人解放出来，让他们可以去做构建或销售，而这两件事才是真正在组织内部创造价值的地方。看到团队的生产力提升了这么多，真的非常惊人。
Speaker 2 | 1:06:30 - 1:06:44 And do you think that there is a painful transition period from the current world to the next world? So do you think everybody's gonna have to go through this as organizations recalibrate? We laid off
那你觉得，从当前这个世界走向下一个世界，中间会不会有一个痛苦的过渡期？也就是说，随着组织重新校准，你觉得每个人都得经历这个过程吗？我们裁掉了——
Speaker 1 | 1:06:44 - 1:07:29 more than 20% of our team and not because the business was struggling, but and not because they weren't great team members. But we just need fewer middle managers, we need fewer measures and those roles were going away. As I talked to peers at other companies, they're all seeing the same thing and they're all saying yeah we're gonna have to do the same thing at some point and seeing these sort of trickle out. But there's a real fear in leaders out there that they don't want to be the first ones to do that. And you know pardon the vernacular, but I think that's chicken shit because the cruelest thing that you can do is wait.
——超过 20% 的团队成员，这并不是因为业务出了问题，也不是因为他们不是很优秀的团队成员，而是因为我们需要更少的 middle managers（中层管理者），需要更少的“衡量者”，而这些岗位正在消失。我和其他公司的同行交流时，他们看到的情况都一样，他们也都在说，是的，我们在某个时点也不得不做同样的事，而且已经开始看到这种情况一点点扩散出去。但现在领导者中确实存在一种恐惧：他们不想成为第一个这么做的人。说句不太文雅的话，我觉得这很怂，因为你能做的最残酷的事，就是等待。
Speaker 1 | 1:07:31 - 1:08:55 I do think in the next six to twelve months almost every company is going to go through some exercise like this where they're going to cut a bunch of their team And I think a lot of people are waiting, a lot of CEOs are waiting around to do it because they're afraid that they're gonna look bad in the process. What really struck me was the realization that once you know that that change is going to happen, the kindest thing that you can do for your your team is to do it as soon as possible because it's way easier to get a job today than it's going to be in six to twelve months because the market's going to get flooded. And so we put together an incredibly, I mean the most generous severance package in tech that we've seen. We let stock continue to vest which nobody does but we did And we're working incredibly hard to place these people because they're incredible people and somebody who is a middle manager at Cloudflare can be a senior manager at a startup and do an extraordinary job. But but once you're as an organization, you've realized you're going to do it, like the the kindest thing that you can do for your team is to do it as soon as possible because what the worst thing that can happen is that there's just a bloodbath six to twelve months from now where everyone does it at the same time and the market gets flooded with a bunch of people who are all of a sudden looking looking for jobs.
我的确认为，在接下来的六到十二个月里，几乎每家公司都会经历某种类似的动作，大幅裁减一部分团队成员。我觉得很多人都在等，很多 CEO 都在拖着不做，因为他们害怕自己在这个过程中看起来不好看。真正让我深受触动的是这样一个认识：一旦你已经知道这种变化一定会发生，那你能为团队做的最善意的事，就是尽早去做。因为今天找工作，要比六到十二个月后容易得多；到那时，市场会被大量求职者淹没。所以我们拿出了一套非常、非常慷慨的离职补偿方案——我的意思是，这是我们所见过科技行业里最慷慨的 severance package（遣散补偿方案）。我们允许股票继续 vest（归属），这一点几乎没人这么做，但我们做了。我们也在非常努力地帮助这些人找到去处，因为他们都是非常出色的人。在 Cloudflare 做 middle manager 的人，到了 startup 里完全可能成为 senior manager，而且表现得极其优秀。但一旦作为组织你已经意识到自己会走到这一步，那你能为团队做的最善意的事，就是尽快去做。因为最糟糕的情况就是，六到十二个月后出现一场“血洗”，所有人同时动手，市场一下子涌入一大批突然开始找工作的人。
Speaker 1 | 1:08:55 - 1:09:46 And so that was the cattle that was sort of the the rationale that went through us and and again I there were there are people who are really close friends of mine who we let go because the roles that they had just didn't make sense anymore and and we we wanted to do it in a significant enough way that we would never have to do it again. And so we're I feel comfortable saying that for the foreseeable future, we're not gonna have to do anything like that again because we've reorganized the organization. And I think that you know one of the metrics that that we look at is span of control. And so how many direct reports, what's the average number of direct reports in an in an organization? And you that the historic number that like the Harvard Business School did a study and they came up that the right number for that is about six.
所以，这大致就是我们当时的思路和背后的 rationale（逻辑）。而且再说一次，被我们裁掉的人里，有些其实是我非常亲近的朋友，只是他们所承担的那些角色已经不再合理、没有意义了。我们希望一次做得足够彻底，这样以后就再也不用做第二次。所以我可以比较有把握地说，在可预见的未来，我们不会再需要做类似的事了，因为我们已经把组织重组过了。我觉得我们会看的一个指标是 span of control（管理幅度）。也就是，一个组织里平均每位管理者有多少个 direct reports（直接下属）。历史上的那个数字——Harvard Business School 以前做过一项研究，他们得出的结论是，这个数字大约在 6 左右才是合适的。
Speaker 1 | 1:09:46 - 1:10:19 So for every manager they have six direct reports and that was exactly what we were at. It's so clear that with these new tools a manager can manage a lot more people. You know Meta is trying to get to 50 to one. I think that's all that seems like too many to me but I think 12 to one is right and the benefit of expanding the number of the average number of direct reports is it inherently flattens the organization. The fewer number of direct reports on average a manager has inherently the organization is going to be much more hierarchical.
所以，对每一位 manager 来说，他们有 6 个 direct reports（直接下属），而我们当时正好就是这个比例。现在已经非常清楚了：有了这些新工具，一个 manager 可以管理更多人。你知道 Meta 现在在尝试做到 50 比 1。我觉得那看起来还是太多了，但我认为 12 比 1 是合适的。提高平均 direct reports 数量的好处在于，它会天然地让组织更扁平。平均而言，一个 manager 的 direct reports 越少，组织就必然会越层级化。
Speaker 1 | 1:10:20 - 1:11:09 The more direct reports they have it's going to be flatter and I think one of the things that you know we're seeing is that the flatter the organization is the more the easier it is for us to move quickly to be able to deliver things and and again I think that's that's what's that's what's driven a lot of us but yeah I think this is coming across the industry and again it's not a statement that any person is a bad person or a bad individual. In fact they're great but there are roles that you're just not going to need anymore and some people you're going to be able to sort of re home and reskill inside your organization. But like if you're going to go from you know six to one to 12 to one, that that means a whole bunch of middle managers are gonna are gonna go away.
他们的 direct reports 越多，组织就会越扁平。我认为我们现在看到的一点是：组织越扁平，我们就越容易快速行动，越能够把东西交付出来。而且，我觉得这也确实是推动我们很多决策的原因之一。不过，这种情况正在整个行业里出现。再说一次，这并不是在说某个人不好，或者某个人不行。事实上，他们都很优秀；但有些 role（岗位）就是以后不再需要了。有些人你可以在组织内部重新安置、重新培训。但如果你要从 6 比 1 变成 12 比 1，那就意味着会有一大批 middle managers（中层管理者）消失。
Speaker 2 | 1:11:09 - 1:11:26 So what do you do if you're one of those people? If you're a middle manager and I think earlier you mentioned old Gen Z or young millennial and you find yourself in that situation like mid career, do you just go down a complete rabbit hole of like just making yourself as AI peeled as possible?
那如果你正好是这类人，你该怎么办？如果你是 middle manager，而且我记得你前面提到过，可能是年纪稍大的 Gen Z 或者年轻的 millennial，然后你发现自己处在这种 mid-career（职业中期）的位置上，你是不是就得一头扎进这个兔子洞，尽可能让自己变得足够“AI 化”？
Speaker 1 | 1:11:26 - 1:12:10 I think yes. Recognize that it's natural for that to feel scary and that the and that your instinct is going to be to fight against it. But that the more that you can embrace these tools the better. I've never seen, you know, and again running this now for seventeen years, like I've never seen more senior managers on our team say raise their hand and say, hey I want to go back to being an individual contributor. And so we're rethinking like even how compensation works and everything else because I think that the power of an individual contributor and the amount of work that they can do is is extraordinary.
我觉得是的。首先要认识到，觉得害怕是很自然的，而且你的本能反应会是抗拒它。但你越能拥抱这些工具，就越好。你知道，我做这件事已经 17 年了，在这段时间里，我从来没有见过这么多 senior managers（高级管理者）主动举手说：“嘿，我想回去做 individual contributor（个人贡献者）。”所以我们甚至在重新思考 compensation（薪酬）机制以及其他一切，因为我认为，individual contributor 的能力，以及他们现在能够完成的工作量，都是非常惊人的。
Speaker 1 | 1:12:10 - 1:13:26 So I think that for us, you know, our most senior members of our team are embracing this because it's a way for them to reinvent themselves and they're senior enough that they're not scared about it. Our most junior folks, so we we we hired 1,111 that number is meaningful to us, but interns this summer, you know and I mean it's like they're the most extraordinary humans because everybody stopped hiring interns which I think is just total mistake. And so we got sort of the pick of the litter and and and they're and they're all totally AI native and what we're trying to do is get both of those groups to get the folks that have been more hesitant because the fact is 92% of organization engineering organizations using AI tools like that that should be a 100% for sure and and it's but you've got to get people who especially are sort of mid career to be brave enough to learn these new these new tricks and and I think that that's that's that's really important and that and again with the you can't have one person who is 10 or a 100 times as productive in the same role as somebody else making the same amount of money as somebody else who's who's sort of doing it the old way.
所以我觉得，对我们来说，团队里最资深的成员正在拥抱这一变化，因为这给了他们重新塑造自己的机会，而且他们资历够深，不会因此感到害怕。我们最初级的员工也是如此。这个夏天我们招了 1,111 名 interns（实习生），这个数字对我们是有特殊意义的。而且，他们真的是最出色的一群人，因为几乎所有公司都停止招实习生了，我觉得这完全是个错误。所以我们某种程度上拿到了最优秀的一批人，而且他们全都是彻底的 AI native（AI 原住民）。我们现在想做的是，把这两类人带动起来，也把那些更犹豫的人带起来。因为现实是，92% 的 engineering organizations（工程组织）都在使用 AI 工具，这个数字显然应该是 100%。但你必须让人们，尤其是处在 mid-career 的人，足够勇敢去学习这些新把戏。我认为这一点非常重要。再说一次，你不能让一个人在同一个 role 里，生产力是另一个人的 10 倍甚至 100 倍，却拿着和那个仍在用老方式工作的人一样多的钱。
Speaker 1 | 1:13:26 - 1:13:56 Like one of those two is gonna leave and the worst case is if the person who's a 100 times as productive leaves. And so we're just we are continuously trying to get every way and lift everyone up and and again part of why we did the layoff was in order to be able to, you know, really really free up the resources and and catalyze the entire organization to be able to do this. But and it's not just your developers, has to be finance, legal, marketing, everything has got to be embracing these new tools.
这两个人里，总会有一个离开；而最糟糕的情况，是那个生产力高出 100 倍的人离开。所以我们一直在想尽办法把每个人都带起来、都提升上来。再说一次，我们之所以裁员，部分原因也是为了真正释放资源，并催化整个组织去做到这一点。而且这不只是 developers（开发者）的事，finance、legal、marketing，所有部门都必须拥抱这些新工具。
Speaker 2 | 1:13:56 - 1:14:01 So almost a year ago now you declared I think that was Content
所以差不多一年前，你宣布了——我想那应该是 Content
Speaker 1 | 1:14:01 - 1:14:03 Independence Day July 1.
Independence Day，7 月 1 日。
Speaker 2 | 1:14:03 - 1:14:03 July 1.
Speaker 2 | 1:14:03 - 1:14:03 7月1日。
Speaker 1 | 1:14:04 - 1:14:07 Coming up on the second anniversary or I guess first anniversary.
Speaker 1 | 1:14:04 - 1:14:07 快到两周年了，或者我想应该说是一周年。
Speaker 2 | 1:14:07 - 1:14:14 Yes. What happened in the last year? How is that whole debate evolving? Yeah. So And what was it in the first place?
Speaker 2 | 1:14:07 - 1:14:14 对。过去一年发生了什么？整个争论是怎么演变的？对。所以这件事最开始又是什么来着？
Speaker 1 | 1:14:14 - 1:14:52 Yeah. So so so again, I think that it is you've just seen an enormous rise in the amount of sort of automated traffic that's running across the Internet. I think in five years that you'll that bots and agents will outnumber humans online a thousand to one. At the same time, like the interface through which we access the Internet is changing and we've gone through other sort of platform changes before. You know, we went from browsing the Internet on a browser on your desktop or laptop to social to mobile.
Speaker 1 | 1:14:14 - 1:14:52 对。所以还是那句话，我认为你现在看到的是，互联网上运行的某种自动化流量（traffic）数量已经大幅上升。我认为五年后，在线的 bots 和 agents 数量会比人类多出一千比一。与此同时，我们访问互联网的 interface（界面）也在发生变化，而这种 platform（平台）变迁以前也发生过。你知道，我们经历过从在 desktop 或 laptop 上用 browser 浏览互联网，到 social，再到 mobile 的转变。
Speaker 1 | 1:14:52 - 1:15:46 But throughout all of those things, underlying business model of the Internet for the last twenty eight years has remained basically the same which is it's largely advertising and then to a lesser extent subscription driven. The problem is that first, there's just gonna be a lot more cost to run on the Internet because you're gonna have a thousand times as much traffic which means you're gonna have to have servers and CPUs and network and all that stuff and and someone needs to, you know, pay for it. And the business model of the last twenty eight years which largely advertising based doesn't work because bots don't click on ads. And so something has to change and so a part of that is how do we make the Internet more efficient? Like how do we take legacy software and port it to things that are, you know, can continue to scale in a cost effective way?
Speaker 1 | 1:14:52 - 1:15:46 但贯穿这一切的是，过去二十八年来互联网底层的 business model（商业模式）基本没变，本质上主要还是 advertising 驱动，其次才是 subscription 驱动。问题在于，第一，互联网的运行成本会高得多，因为流量会变成现在的一千倍，这意味着你需要更多 servers、CPUs、network 以及所有这些基础设施，而总得有人为此买单。而过去二十八年那种主要基于 advertising 的 business model 行不通，因为 bots 不会点广告。所以一定有什么东西得改变。其中一部分就是，我们怎样让互联网变得更高效？比如，怎样把 legacy software 迁移到那些能够继续以成本有效方式扩展的东西上？
Speaker 1 | 1:15:47 - 1:16:23 But another big part of that is how do we figure out what the future business model of the Internet is is is gonna look like. And you know, we we have great relations with all the different AI labs and and and they get it. And they actually you know they they say listen as long as it's fair as long as everybody has to do the same thing we get that we've got to pay for the content that we're using in order to build our systems and our and our and our models. But we've got to build a system to do that. And so like we're experimenting in a bunch of different ways.
Speaker 1 | 1:15:47 - 1:16:23 但另一大部分是，我们怎样弄清楚互联网未来的 business model 会是什么样。你知道，我们和各家不同的 AI labs 关系都很好，他们也明白这一点。事实上他们会说，听着，只要这件事是公平的，只要每个人都必须遵守同样的规则，我们就理解：为了构建我们的 systems 和 models，我们必须为所使用的内容付费。但我们得先建立起这样一套 system 来实现这件事。所以我们正在用很多不同方式做实验。
Speaker 1 | 1:16:23 - 1:16:41 The first thing that we did was we said in order for a market to exist, know, you always think you have to have supply and demand. It's not exactly right. What you actually need to have is scarcity of supply and demand. If everything is free then then no one pays for it. Like we're sitting in this room, we're breathing the air, doesn't cost us anything.
Speaker 1 | 1:16:23 - 1:16:41 我们做的第一件事是，我们说，要让一个 market 存在，通常你会觉得需要 supply 和 demand。其实不完全对。你真正需要的是供给的稀缺性以及 demand。如果一切都是免费的，那就没人会为它付钱。比如我们坐在这个房间里，呼吸空气，这对我们来说不花钱。
Speaker 1 | 1:16:42 - 1:17:12 On the other hand, we go scuba diving, air is scarce and so we have to pay for it and so you know that's that. So step one was we said we need to create controls that allow the people who have content online to be able to say who can access it and who can't. And again, it's everyone's going to make different decisions. In our case, we want all the bots to be consuming our information. We want that to be out there and that's probably half of our customers are in that camp where they actually want to make it as easy as possible.
Speaker 1 | 1:16:42 - 1:17:12 但另一方面，如果我们去 scuba diving，空气就是稀缺的，所以我们必须为它付钱，大概就是这个道理。所以第一步是，我们说需要建立控制机制，让那些在线拥有内容的人能够决定谁可以访问、谁不可以访问。再说一次，每个人都会做出不同决定。以我们为例，我们希望所有 bots 都来消费我们的信息，我们希望这些信息被广泛传播，而且大概有一半客户也属于这一阵营——他们实际上想让这件事尽可能容易。
Speaker 1 | 1:17:12 - 1:18:19 And so we provide the tools to as cost effectively, as efficiently as possible get information and take take HTML which has all this croft around it converted into markdown which is a much more streamlined system and then as a result like you can get more information in these AI systems without blowing up their context windows and all kinds of things. It makes sense that it's on the other hand if you're an ad supported business, like agents aren't good for you because they're taking your content and then people are consuming it through their AI system and you're not getting eyeballs. So you're the ad systems that work so we said like let's give everyone the tools to be able to stop it. Let's create some scarcity in the supply that's out there. And what we've seen is if you go across the biggest publishers in the world Conde Nast, Dotdash Meredith, you know they are they because we've now given them these tools they're actually signing much better deals with the AI companies and so that's I think that's sort of the first step but we've got to go beyond that because if you're a small blog or a you know influencer or a content creator like you don't have the ability to go strike the same deal that a Conde Nast does.
因此，我们提供工具，以尽可能低的成本、尽可能高的效率来获取信息，并把带着一堆冗余杂质的 HTML 转成 markdown（更精简的格式），这样一来，你就能在这些 AI 系统里塞进更多信息，而不会把它们的 context window（上下文窗口）撑爆，等等。这是说得通的。但另一方面，如果你是一个靠广告支持的业务，那 agent（智能代理）对你并不友好，因为它们会拿走你的内容，然后用户通过自己的 AI 系统来消费这些内容，你却得不到 eyeballs（用户注意力/曝光）。而广告系统靠的就是这个。所以我们说，要给所有人提供工具，让他们能够阻止这件事。让现有供给变得更稀缺一些。我们看到的是，如果你去看全球最大的那些出版商，Conde Nast、Dotdash Meredith，你会发现，因为我们现在给了他们这些工具，他们实际上正在和 AI 公司签下好得多的协议。所以我认为这算是第一步，但我们还得继续往前走，因为如果你是一个小 blog，或者一个 influencer、content creator，你并没有能力去谈到像 Conde Nast 那样的协议。
Speaker 1 | 1:18:20 - 1:19:01 There should be some basically a micro payment. Every time we access the site and get information, we're happy to pay a fraction of a penny in order to do that. That's harder than it sounds. So we process in any given second about half a billion requests through our network. We estimate that somewhere between 110% of those would be sort of monetizable through some micro payment transaction and so that means that you have to be able to support day one, call it 10,000,000 financial transactions per second with a clear ramp to be able to get to a 100,000,000 transactions per second.
原则上，应该有一种 micro payment（微支付）机制。每次我们访问一个网站并获取信息时，我们都愿意为此支付几分之一美分。只是这件事比听起来难得多。我们的网络在任意一秒大约会处理 5 亿次请求。我们估计，其中大约有 10% 会是那种可以通过某种 micro payment 交易来 monetizable（可货币化）的请求，这意味着你在第一天就必须能支持——姑且说——每秒 1000 万笔金融交易，而且还得有一条清晰的扩展路径，能够做到每秒 1 亿笔交易。
Speaker 1 | 1:19:01 - 1:19:43 Visa, which is the largest payments network in the world does fewer than a 100,000 transactions per second. So we need to be, you know, two orders of magnitude bigger than Visa but with it much much smaller volumes. And so we're working on that and we think that you know in partnership with with folks like Coinbase and Stripe, know, and and and protocols like x four zero two which is you know, it's interesting with the original Internet had something a response code for the 402. We all know 404 which is not found and 500 which is an error. Mean there are different things but 402 was payment required and we just never have really built anything around that.
Visa 作为全球最大的支付网络，每秒处理的交易量也不到 10 万笔。所以我们需要做到比 Visa 大两个数量级，但对应的单笔金额却要小得多。我们正在推进这件事，而且我们认为，与 Coinbase、Stripe 这样的一些合作方，以及像 x four zero two 这样的 protocol（协议）合作，是可行的。这也很有意思：最初的 Internet 其实就有一个 402 的 response code（响应码）。我们都知道 404 是 not found，500 是 error，当然它们各有不同含义，但 402 代表的是 payment required，只是我们一直没有真正围绕它构建出什么东西。
Speaker 1 | 1:19:43 - 1:20:20 I think that's going to happen going forward. And so, you know, Google really is responsible for having built the business model that's power the last twenty eight years of the Internet and I think we are trying to figure out, you know, what is that business model that's going to power the next twenty eight years of the internet but something has to pay content creators because they deserve to eat. Something has to pay for the infrastructure that all of these bots are taxing effectively and I think we're working a lot around figuring out exactly what that looks like.
我觉得接下来这会发生。所以，Google 其实要为过去 28 年 Internet 商业模式的建立负责——正是它塑造了这个模式。而我认为，我们现在在试图弄清楚的是，什么样的商业模式会支撑未来 28 年的 Internet。但无论如何，必须有人为 content creator（内容创作者）付费，因为他们也要吃饭。也必须有人为这些 bots 实际上正在消耗、征税般压榨的基础设施买单。我认为我们正在做很多工作，想把这件事最终会是什么样子弄清楚。
Speaker 2 | 1:20:20 - 1:20:47 Great. It's been an absolutely fascinating conversation. The overarching thoughts as I was listening to you is, to which extent Cloudflare is becoming even more relevant today than it's been for the last seventeen years, which you built the network, and now the internet is changing and, you are absolutely at this sort of central point for it. Yeah. Is that how you how you guys think about it?
很好。这绝对是一场非常精彩的对话。我在听你说的时候，一个总体感受是：Cloudflare 在今天似乎比过去 17 年里的任何时候都更加重要了。你们先是建起了这个网络，而现在 Internet 正在发生变化，你们恰好处在这个变化的中心位置。对，这就是你们看待自己的方式吗？
Speaker 2 | 1:20:47 - 1:20:50 Like you're you're you have another seventeen years?
就像，你们还有下一个 17 年？
Speaker 1 | 1:20:50 - 1:21:04 People are like, why, you know, why do you keep doing this? There are days that aren't fun. I mean laying off a bunch of people, I mean that wasn't fun. And it's and you know I've made plenty of money. So it's not about any of that.
人们会说，为什么，你知道，为什么你还要继续做这个？有些日子并不好过。我的意思是，裁掉一大批人，那肯定不好受。而且，你知道，我已经赚了很多钱了。所以这和那些都没关系。
Speaker 1 | 1:21:06 - 1:22:07 What it is about is is really our mission and we and we believe in that mission fundamentally and we we do things that that cost us money or or that certainly don't make us any money because we believe in helping build a better Internet. And like I can't imagine anything that I could possibly be doing that is more rewarding than than figuring out how do we create some sort of sustainable business for the future. I think there's a lot that's to be incredibly optimistic about. The business model of the Internet for the last twenty eight years wasn't the most healthy. Traffic has always been a terrible proxy for value and and and and so much of like the media landscape was how do you sort of define your narrow audience and then with a headline piss them off enough that they'll click on a link in order to you know drive eyeballs to your your site.
真正有关的是我们的使命，而且我们从根本上相信这个使命。我们会做一些让我们花钱的事，或者至少完全不赚钱的事，因为我们相信自己是在帮助建设一个更好的 Internet。而且我几乎想不到，还有什么事情能比去弄清楚：我们该如何为未来创造某种可持续的商业模式，更有成就感。我认为有很多事情是非常值得乐观的。过去 28 年 Internet 的商业模式并不是最健康的。traffic（流量）一直都是价值的一个糟糕代理指标，而媒体生态里有太多做法都是：你如何定义出一小群非常狭窄的受众，然后用一个标题把他们激怒到足以点开链接，从而把 eyeballs（用户注意力/曝光）导到你的网站上。
Speaker 1 | 1:22:08 - 1:22:46 And you know so much of this so much of what is dividing the world today is actually the underlying business model of the Internet which has been traffic based. How do you drive as much much traffic as possible? That's going away and I think that you know if we think about what can replace it going forward, it can be something which is much healthier and much much better. I was I I flipped to Stockholm to to meet with Daniel Akk who who started Spotify and and because I mean like nobody has compensated creators at scale like Spotify. It's just it's just it's remarkable.
你知道，如今世界上许多撕裂，实际上都源于互联网的底层商业模式——它一直是以 traffic（流量）为基础的。怎么尽可能地把更多更多的 traffic 拉进来？这种模式正在消退。我认为，如果我们去想未来有什么能取代它，那可以是某种更健康、也更好得多的东西。我之前飞去 Stockholm 见 Daniel Akk，他创办了 Spotify。因为说实话，没有谁像 Spotify 这样大规模地补偿 creators（创作者）。这真的非常了不起。
Speaker 1 | 1:22:46 - 1:23:31 When we can quibble about are the right people getting the money or not but but there's more money going into music creation today on on per capita, per GDP than any time in human history by far. And and Spotify gets a lot of credit for that and and you forget that like twenty two years ago the music industry was dying, know, because on Napster and everything, everyone was just normalizing everyone stealing stealing content. And then you know Steve Jobs steps on stage announces iTunes 99¢ a song. That's not the business model that won but it at least planted a flag that said that there's there's something that's there and then I mean Spotify last year sent something like $12,000,000,000 out to music creators. That's more than the that's more than the entire music industry's market cap was twenty two years ago, which is which is extraordinary.
我们当然可以争论，拿到钱的人是不是对的人，但无论如何，今天流向音乐创作的资金——按人均、按 GDP 比例来看——都远远超过人类历史上任何时候。而 Spotify 在这件事上功不可没。你会忘记，22 年前音乐产业其实快要死了，因为 Napster 之类的出现，让“偷内容”这件事被所有人渐渐视为正常。后来 Steve Jobs 上台，发布 iTunes，一首歌 99¢。这并不是最终胜出的商业模式，但它至少立起了一面旗帜，说明这里面是有路可走的。然后你看，Spotify 去年向音乐创作者分发了大约 $12,000,000,000。这比 22 年前整个音乐产业的 market cap（市值）还要高，真的非常惊人。
Speaker 1 | 1:23:31 - 1:24:08 And and so Daniel was telling the story. He said one of the things that they that they do is if you search on Spotify for like Taylor Swift, Shake It Off, like they know they return results and they know that those results are what you were actually looking for. On the other hand, you search for like I want a song to a disco beat about how much fun it is to dance with my cat. They also return results but they know that that they probably don't have that and what they do is they actually publish those queries that are sort of bad results back to music creators. And there are hundreds of music creators that are making a living off just doing this.
Daniel 当时给我讲了一个故事。他说，他们做的一件事是：如果你在 Spotify 上搜索 Taylor Swift、Shake It Off 这样的内容，他们会返回结果，而且他们知道这些结果就是你真正想找的东西。另一方面，如果你搜索“我想要一首带 disco 节拍、讲和我的猫跳舞有多开心的歌”，他们也会返回结果，但他们知道自己大概率并没有这个东西。于是，他们会把这类结果不理想的查询重新发布给音乐创作者。现在已经有数百位音乐创作者，光靠做这件事就在谋生。
Speaker 1 | 1:24:08 - 1:24:37 One in particular is based in Denmark, Daniel shared is making something like €40,000,000 a year just writing songs for unfulfilled Spotify queries. And I think there's something that's actually quite beautiful about that because if you search for like a something on Spotify, what you're searching for is how do I kind of bring an emotion out? And and so going forward like we for the first time in human history have essentially a mathematical model of all of human knowledge. That's what the LLMs are, right? It's not perfect but it gives you that.
Daniel 分享说，其中有一位在 Denmark，光是为 Spotify 上那些未被满足的搜索需求写歌，一年就能赚到大约 €40,000,000。我觉得这里面其实有一种相当美的东西。因为你在 Spotify 上搜索某样东西时，本质上是在寻找：我该怎样把某种情绪调动出来？而展望未来，我们在人类历史上第一次，基本拥有了一个关于全人类知识的 mathematical model（数学模型）。这就是 LLMs，对吧？它并不完美，但它确实给了你这样的东西。
Speaker 1 | 1:24:37 - 1:24:59 It also shows you where the gaps are, where there are the holes. I picture like a giant block of Swiss cheese. There's lots of cheese but there's a lot of holes in the cheese. And when you talk to the AI companies, what they want is highly reputable sources that are filling in the holes in the cheese, are contributing net new knowledge. They don't want yet another story about what happened at 1500 Pennsylvania Avenue, right?
它同时也会告诉你缺口在哪里、空白在哪里。我会把它想象成一大块 Swiss cheese。奶酪很多，但上面也有很多洞。你去和 AI 公司交流时，他们想要的是那些高度可信、能够把这些洞补上的 sources（来源），是那些能贡献 net new knowledge（全新净增知识）的内容。他们并不想再要一篇关于 1500 Pennsylvania Avenue 发生了什么的报道，对吧？
Speaker 1 | 1:24:59 - 1:26:03 They want something where people are actually telling and creating new information that moves the world forward. And I think that's what I want, I think that's what you want, I think that's what all of us are craving is how do we get back to sort of a media environment which isn't just trying to provoke outrage in order to drive traffic, but is actually trying to find the truth and fill that in. And if we can create a system where the people who again are highly reputable and are creating sort of net new knowledge that those are the people that get rewarded, then I think that's incredibly incredibly beautiful. What I keep pushing the AI labs to do is I'm like we should create the equivalent of I don't if it's the Academy Awards and Nobel Prize where we measure across a whole bunch of different segments who has contributed the most to human knowledge in the last year in you know I don't know cancer research and you know foreign policy news. Mean you can create a bunch of these things and they can actually measure and say you know, this particular year the best podcast was you know the first Mark podcast that was that was out there or whatever it is.
他们想要的是那种人们真的在讲述、在创造新信息，并推动世界向前发展的东西。我觉得这正是我想要的，也是你想要的，也是我们所有人都在渴望的：我们怎样才能回到一种媒体环境，不是为了拉动 traffic 而刻意挑起愤怒，而是真正在努力寻找真相，并把那些空白填补起来。如果我们能建立一个系统，让那些再次强调一下——高度可信、并且在创造 net new knowledge 的人——得到回报，那我觉得这会是极其极其美好的事。我一直在推动 AI labs 去做的一件事是：我们应该建立某种相当于 Academy Awards 和 Nobel Prize 的东西——我不确定具体像哪一个——在很多不同领域里衡量，过去一年谁对人类知识贡献最大，比如我不知道，cancer research，以及 foreign policy news。你完全可以建立很多这样的评选机制，真的去衡量并说：比如某一年，最好的 podcast 是第一期 Mark podcast，或者不管具体是什么。
Speaker 1 | 1:26:03 - 1:26:27 And we should be rewarding and celebrating both both in terms of honoring these people but then also paying them for doing that. And I think if we do that, you're going to get a lot more, you know, interesting information. My wife and I bought the local newspaper in our in our hometown Park City, Utah. I think we will make more this year off AI licensing deals than we do off display ads. Why?
而且我们应该去奖励、去赞颂这些人——不仅仅是给予荣誉，也要真的为他们的工作付费。我觉得如果我们这样做，你就会得到多得多的有趣信息。我和我妻子买下了我们家乡 Utah 州 Park City 的本地报纸。我认为我们今年从 AI licensing deals（AI 授权交易）中赚到的钱，会比 display ads（展示广告）还多。为什么？
Speaker 1 | 1:26:27 - 1:27:28 Because if you're planning a vacation to go skiing in Park City, Utah, you want to know what's the best hotel to stay at, what's the best restaurant, who's going to be performing, know, what are the snow conditions like and and the only place that has that is the local newspaper. And so while the last twenty eight years of the internet kind of killed local quirky unique interesting information, what I'm hopeful is that as we figure out what the business model is for the next twenty eight years it actually brings a lot of those things back and and it actually gets back to let's talk about facts, let's talk about reality and okay I like the New York Times, be amazing if they were covering like in the Marriott Marquis hotel in in Times Square. Room 1313 is better than Room 1314 or if you know if you're if you want to sleep in late room you know 1314 faces north and so it's less likely to get sun in the morning. Like those are the sorts of here's the best bodega in New York. Very interesting quirky local stories.
因为如果你正计划去 Utah 州 Park City 滑雪度假，你会想知道：住哪家 hotel 最好、哪家 restaurant 最好、谁会来演出、雪况怎么样，而唯一真正掌握这些信息的地方，就是本地报纸。所以，尽管过去 28 年的互联网某种程度上扼杀了本地的、古怪的、独特的、有趣的信息，我所希望的是，当我们弄清楚未来 28 年的商业模式时，它反而会把很多这样的东西带回来，带我们回到“来谈事实吧，来谈现实吧”的状态。比如说，我喜欢 the New York Times，但如果他们能去报道 Times Square 的 Marriott Marquis hotel 里，1313 房间比 1314 房间更好，或者如果你想睡到晚一点，1314 朝北，早上被太阳晒到的概率更低——那会非常棒。又比如“这是 New York 最好的 bodega”。就是这种非常有意思、很有地方特色、很古怪的本地故事。
Speaker 1 | 1:27:28 - 1:27:40 I think that that is both a media future and an internet future that we all crave and want and I think it's one that we actually have an opportunity to build because we get to reinvent what the business model is going be.
我认为，那既是我们都渴望、都想要的一种媒体未来，也是一种互联网未来；而且我认为，这也是我们确实有机会去构建的未来，因为我们可以重新定义未来的商业模式会是什么样。
Speaker 2 | 1:27:40 - 1:27:45 Fascinating. Well Matthew, thank you so much. This was an absolutely wonderful conversation. Really appreciate it.
很精彩。那么 Matthew，真的非常感谢你。这绝对是一场非常棒的对话。非常感谢。
Speaker 1 | 1:27:45 - 1:27:46 Thank you for having me.
感谢邀请我来。
Speaker 2 | 1:27:48 - 1:28:07 Hi it's Matt Turk again, thanks for listening to this episode of the MAD Podcast. If you enjoyed it, we'd be very grateful if you would consider subscribing if you haven't already or leaving a positive review or comment on whichever platform you're watching this or listening to this episode from. This really helps us build a podcast and get great guests. Thanks and see you on the next episode.
大家好，我是 Matt Turk，再次感谢你收听这一期 MAD Podcast。如果你喜欢这一期内容，而你还没有订阅的话，我们将非常感激你考虑订阅，或者在你观看或收听本期节目的平台上留下积极的评价或评论。这对我们打造这个 podcast 并邀请到优秀嘉宾真的很有帮助。谢谢，我们下期再见。