Micky LI

hello,world with AI~~ practice makes perfect. Skill thrives on diligence, but is lost to idleness.

[Negligible Capital @negligible_cap]

Bernstein’s 2030 CPU TAM buildup, after raising the base case from $137B to $223B this week:

[QuikInsightz @QuikInsightz]

🚨 #BREAKING: Michael Dell said $DELL currently holds a 31.2% share of the enterprise storage market.

➜ According to the figures he shared, the closest competitor is $NTAP with a 9.9% market share.

➜ $P follows with an 8.9% share.

[NYSE 🏛 @NYSE]

The NYSE welcomes @Cloudflare to celebrate its investor day! $NET https://t.co/iLPcW18udF

[DD @DD29397053]

周末在家仔细思考：AI 投资到底还能不能持续？回到最朴素的生意逻辑——投入多少，产出多少，能持续多久，我脑子似乎清晰了一些。

如果一门生意每投入一单位资本，都能持续产生很高收入和现金流，那它就会继续吸引资本；如果投入越来越大，但产出跟不上，泡沫就会开始破。这个道理大家都懂。

问题是，AI 最头部的两家公司 OpenAI 和 Anthropic 都不是上市公司，财务不透明，真实 token 量、GPU/TPU 利用率、训练和推理占比都不公开，那怎么判断？

我觉得可以从一个很简单的角度看：算力吃紧本身就是一个非常重要的信号。算力吃紧说明可售产能的利用率已经接近极限。这个时候，只要估算出它在这种产能紧张状态下的年化收入，再用独立的算力线索估算支撑这笔收入所对应的有效算力，就可以大概看出这个行业最顶级玩家的收入/算力比，也就是 AI 这门生意当前的投入产出效率。AI这门生意到底有多好。

以 Anthropic 为例，它自己在 5 月底披露 run-rate revenue 已经超过 470 亿美元。算力端，Anthropic 官方披露目前使用超过 100 万颗 AWS Trainium2 芯片训练和服务 Claude，并且与 Amazon 扩大到最高 5GW 的合作；Reuters 还报道 Anthropic 获得 SpaceX Colossus 约 300MW、22 万颗以上 Nvidia 芯片的算力容量。用这些独立产能线索粗略估算，Anthropic 当前有效可调用算力大概率是 约 1–1.6GW 这个区间。

这样算下来，Anthropic 当前大概是：

年化收入：约 470 亿美元
当前有效算力：约 1–1.6GW
收入/GW：约 290–470 亿美元/GW/年
如果 1GW AI 数据中心年化总成本大约 85–120 亿美元，那么 Anthropic 的收入/年化算力成本大约是 2.5–5.5 倍。

也就是说，粗略看，Anthropic 现在每投入 1 美元年化算力成本，可以产生约 2.5–5.5 美元收入。即使扣掉云厂商分成、训练、研发、销售、免费用户补贴，它仍然说明一件事：Anthropic 当前的边际算力非常值钱。它更像一个满产、高收入密度的 AI 工厂。

OpenAI 也有一个很好的公开锚点。OpenAI 官方说，2025 年 ARR 超过 200 亿美元，对应可用算力约 1.9GW；3 月底又说现在每月收入约 20 亿美元，也就是年化约 240 亿美元。考虑到 2026 年有部分新增算力上线，但并非所有已宣布算力都已经 online，我粗略估算 OpenAI 当前有效可调用算力大约 2.2–3GW。

这样算下来，OpenAI 当前大概是：

年化收入：约 240–250 亿美元
当前有效算力：约 2.2–3GW
收入/GW：约 80–115 亿美元/GW/年
如果 1GW 年化算力成本约 85–120 亿美元，那么 OpenAI 的收入/年化算力成本大约是 0.7–1.3 倍。

这个数比 Anthropic 低很多。这不一定代表 OpenAI 生意差，而是因为 OpenAI 承担了大量免费用户、低价订阅、多模态、图像、语音、搜索、训练、冗余和战略性入口流量。OpenAI 更像 AI 超级入口，Anthropic 更像高价值企业/API/Claude Code 产能被打满的工厂。

这俩到底是啥水平呢？可以拿著名公司的 ROIC 做参照。苹果这类顶级公司 ROIC 大概 40–50%；Visa 大概 30% 左右；Meta 大概 25–35%；微软大概 20% 出头。这些都是世界上最好的生意之一。

当然A和O两家不是严格会计意义上的 ROIC，而是“AI 算力投入产出比”，也就是每年收入相对于年化算力成本的倍数。它更像是在问：这个 AI 工厂每年烧掉 1 美元算力成本，能产生多少美元收入。但 Anthropic 当前这个“收入/年化算力成本”不是 20%、30%、50% 这种级别，而是 250%–550% 的收入覆盖倍数。当然这不是净利润率，也不是严格 ROIC，但它说明：只要收入和算力估算没有大错，Anthropic 当前这门生意的收入密度非常夸张。OpenAI 则更像是 70%–130% 的收入覆盖倍数，说明它的 AI 工厂在经济性上能跑通，但还不是 Anthropic 这种满产高收入密度状态。

所以到目前为止，我的定性判断是：AI 这门生意在头部公司身上已经证明了很强的收入产出能力，尤其是 Anthropic 这种企业/API/代码代理驱动的模式，当前更像一个满产高 ROI 生意，而不是低 ROI 泡沫。

但这是现状。未来回怎么样呢？

未来最关键的是：年化收入的增速来自哪里。

之前 Anthropic 年化收入增速爆炸，有几个因素叠加：第一，算力增加；第二，产能利用率快速提高；第三，Claude Code、企业 API、高端模型这些高价值场景占比提高。到了 3 月底、4 月以后，利用率提升这个变量大概率已经接近极限了，后面收入增速主要就看新增算力上线速度，以及新增算力能否继续被需求吃掉。

这就分成两种情况。

第一种情况：收入增速放缓，是因为算力不够。状态页仍然紧，限流仍然存在，高端模型不降价，新算力一上线就被吃掉。这是好的放缓，是良性的。说明需求仍然强，只是产能交付是阶梯式的，股价如果因为“ARR 增速没以前爆炸”而下跌，可能反而是买点。

第二种情况：收入增速放缓，是因为需求变弱。状态页变松，限流取消后没人吃，高端模型开始降价，新增算力上线后收入/GW下降。这就是危险信号。说明 AI ROI 开始恶化，硬件链估值就要重估。继续👇

[Micky LI @compundtrend]

@DD29397053 nb 是 Anthropic 现在联合 spacexai 降本 就是降这个本 ，openai 虽然 codex 还可以 但是每一块钱背后的成本都在滴血

[Matthew Prince 🌥 @eastdakota]

The #1 feature CIOs are requesting from us…

[SemiAnalysis @SemiAnalysis_]

BREAKING NEWS: according to CloudFlare Radar Data, Agentic traffic has SURPASSED human traffic across the worldwide internet for HTML webpages.

[Cloudflare @Cloudflare]

VoidZero, the team behind Vite, Vitest, Rolldown, Oxc, and Vite+, is joining Cloudflare. Vite stays open source, vendor-agnostic, and built for everyone. https://t.co/DJTpX4Q9Xt

[Elon Musk @elonmusk]

Grok on Cloudflare

[dominik kundel @dkundel]

Codex can now deploy and host websites for you using Sites! 🎉

This includes storage for data and files using D1 and R2 💥

We started rolling out to Business and Enterprise customers before rolling out more broadly.

More details 👇

[RPCS3 @rpcs3]

PSA: @TencentGlobal is aggressively scraping the Internet to build yet another AI slop chatbot, DDoSing many websites in the process.

We've found that, as of last week, their scraping bots can now solve Cloudflare challenges and behave like real users while ignoring robots.txt. In the last 24 hours alone, our website received more than 3 million successful requests from Tencent bot IP addresses, plus another 1 million that were blocked by Cloudflare challenges.

These recurring DDoS attacks from Tencent have been going on for over a year, and we have been constantly adjusting our firewall rules to filter them while trying not to impact Tencent's real users. Because that is no longer possible, we're now fully blocking Tencent IP addresses, starting with ASN 132203. We recommend other sysadmins do the same.

Other ASNs displaying similar abusive behaviour will also be fully blocked from our services.

We'd also like to thank @Cloudflare for sponsoring us with Project Alexandria as of 2025, giving our sysadmin the tools to keep RPCS3's online services running without service disruptions.

[Matthew Prince 🌥 @eastdakota]

Welp, that happened faster than I predicted. Thought it would be end of 2027, then early 2027, but agentic traffic growing so fast that bots have now passed human traffic online for the first time in the Internet's history. https://t.co/2zX5bHdhsa

[Micky LI @compundtrend]

@mingchikuo 也有一种可能，由于AI本身能力的外遇用户对于图形界面的操作系统的需求并没有之前那么的粘性。ubuntu之类大行其道。

[Michael Dell 🇺🇸 @MichaelDell]

We have the first @DellTech + @nvidia Vera Rubin NVL72 @CoreWeave. Here we go! 🚀

[Brandon Carl @brandonjcarl]

More GPU price normalization via @OrnnExchange.

Easing is across H100, H200, B200 and A100.

[Cerebras @cerebras]

Cerebras is now running Kimi K2.6 – a trillion parameter model – in enterprise trials.

At ~1,000 tokens/s, this is the fastest frontier model performance ever measured by Artificial Analysis @ArtificialAnlys.

[Jared Sleeper @JaredSleeper]

Every day for the next long while, I'm going to tear down a new public software company and highlight the AI risks/opportunities around it- products launched to date, top startups, key quotes from earnings calls, etc.

Day twenty-five: Cloudflare $NET

Peak share price: $253.3 (Oct 31, 2025)
Share price today: $200.99 (-21%)
EV today: $70.2bn
ARR today: $2.46bn (+34% Y/y
NRR: 120%
EV/ARR: 28.5x
GAAP Operating Margin: 6.2%
EV/Run-rate GAAP EBIT: 453x
Headcount: 5,200 (+21% Y/y)

What Cloudflare does:

Cloudflare operates an edge network that it uses to provide an array of value-added connectivity, CDN, and security services to customers worldwide.

Its core business, application services, is built on upsells on top of its free CDN offering, which undercut the low-end of the paid CDN market and became an onramp to cross-selling higher value security services( WAF, DDoS mitigation, bot management, etc.

It then built and executed on an impressive second act, moving into enterprise access with a SASE business model that replaces corporate VPNs and firewalls with cloud native security.

Finally, Cloudflare uses the same infrastructure to run a developer platform, offering serverless compute (Wokers), object storage (R2), and more at the edge as a sort of distributed, developer-native alternative to traditional hyperscalers.

The company hasn't disclosed much on the relative sizes of these businesses, but its strength in enterprise customers and expanding net retention rate suggest that it is seeing significant success in the latter two businesses.

AI bear case:

This is less of an AI-specific bear case (those are hard to build), but Cloudflare's second-act businesses both face heavy competition from existing cybersecurity (Zscaler/Palo Alto) and hyperscaler (AWS, GCP, Azure) incumbents with substantially larger capex and go-to-market budgets.

On the AI side, perhaps most worrisome is the idea that provisioning of AI infrastructure might become the purview of the model vendors (OpenAI/Anthropic) who leverage their capabilities and the ubiquity of their APIs to ensure that inference compute happens on their backends, not Cloudflare's edge servers.

AI bull case:

The AI bull case for Cloudflare is pretty crisp (and arguably reflected in the multiple and metrics). Cloudflare occupies an increasingly strategic position between users and web content, and as agents proliferate, the types of security that Cloudflare specializes in (especially advanced bot detection) become increasingly critical.

Beyond that, there ARE AI workloads that likely belong at the edge, and in general enterprise security challenges will compound as agents come to the fore.

Finally, Cloudflare has done a surprisingly good job of maintaining its developer-first ethos (I use R2 for some of my projects!), giving it valuable exposure to the broader vibe-coding trend.

In short, Cloudflare has a history of executing well and beachfront real-estate from which to add value in an AI native world.

AI traction:

Cloudflare has seen a material acceleration in growth of late, with revenue growth increasing from 27% in Q1 to 34% Y/y in Q4, driven by increased net expansion, particularly with large customers.

The company doesn't disclose any AI-specific revenue (and indeed, is stingy on revenue breakouts for any segments), but it seems safe to assume that AI is showing up as a tailwind in the business one way or another. One of the most plausible candidates is heavy use of Cloudflare's Workers product by vibecoding customers (both developers and applications).

Strikingly, the company called out that the number of weekly agent-made requests on its platform doubled in the month of January alone (!!), highlighting the intensity of the usage tailwind.

Adjacent AI-native startup summary:

Though not fully AI native, @vercel is the most relevant competitor to the core CDN business (912 employees, +28% Y/y), and clearly executing well.

On the enterprise security side- ZScaler, Netskope and Palo Alto are the companies to watch, although all relatively mature. @Tailscale (315 employees, +68% Y/y) competes with Cloudflare's WARP and Tunnel products at the low end.

On the dev-tooling side, the competition is fierce, with @vercel showing up again along with a host of other inference platforms running from the low end to enterprise. It's hard to specify competitors given the breadth of Cloudflare's offering and the particular wedge it brings (a massive, highly optimized edge network).

Management Quotes:

"We blew away our previous record for new ACV in the quarter, with strong year-over-year and quarter-over-quarter acceleration. In Q4, new ACV book grew nearly 50% year-over-year, making it not only a record quarter in absolute ACV dollars but also the fastest growth rate we've delivered since 2021."

"A leading AI company expanded their relationship with Cloudflare, signing a 2-year $85 million pool of funds contract for our full platform, selecting Cloudflare as their single long-term infrastructure provider with 100% traffic allocation."

"A U.S. media company signed a 3-year $3.1 million contract for AI Crawl Control, along with application services and Workers. This customer was facing a massive increase in AI scraping, which was crushing their network and driving up infrastructure costs. They chose Cloudflare to gain visibility into which AI models are consuming their data, allowing them to protect and eventually monetize their unique content."

"That means we win when AI applications are built on Cloudflare Workers, but we also win just from the increased usage of all of our products and agentic Internet drives."

"It's not a coincidence that most so-called vibe coding platforms are either built on Cloudflare Workers or have us as their preferred deployment target. We exited 2025 with more than 4.5 million human developers active on our platform. It's a lot more if we count their agents."

Commentary:

No comment on Cloudflare would be complete without reference to the company's sky-high valuation multiple. At 28.5x sales only Palantir is higher, and the next highest in SaaS in CrowdStrike at 19.7x with another cliff down to Palo Alto at 12.7x. This is a very, very expensive stock and while the metrics are truly impressive (and more importantly, trending the right direction) it's strike that they aren't yet standout on an absolute basis (other companies have similar/higher growth rates and similar net retention, yet trade at a massive discount). The market clearly believes that Cloudflare is a secular AI winner- and yes, it does have good reason to believe so.

The biggest challenge to writing about Cloudflare's AI tailwinds more specifically is the inscrutability of the business given the range of business lines and market segments served. The company's investor day on June 9th will hopefully provide some helpful context.

What seems true is that AI is a meaningful tailwind for Cloudflare's business, although there is no one hero SKU to point to and analyze. Still, taking the sky high, top-in-all-of-SaaS multiple aside, it is hard to argue against a well-oiled machine with a stellar management team that occupies beachfront AI real estate. It will be interesting to see if/how the story evolves into something more legible (and quantifiable) for investors over time.

[KawzInvests @KawzInvests]

The CPU is the bottleneck in agentic AI, not the GPU.

Up to 90.6% of an AI agent's runtime, the GPU is idle. Georgia Tech and $INTC hooked five production-style agents to real hardware and stopwatched every step. Where the latency actually goes.

> RAG retrieval. CPU document search consumes 83 to 89% of total latency.

> SWE-Agent coding. CPU bash and Python execution accounts for 38 to 65%.

> Web-augmented agents. CPU summarization runs 40 to 55% of total time.

> ChemCrow chemistry. CPU conformer generation hits 85 to 88% on heavy molecules.

Then they swapped in the H200, Nvidia's most expensive GPU. The CPU share got worse, not better. On SWE-Agent it climbed from 38% to 65%. Finishing the GPU part faster only makes the CPU wait look bigger.
Every Nvidia upgrade widens the gap.

This is why Intel's CFO told analysts unmet CPU demand "starts with a B." Billions of dollars in orders they cannot ship.
$AMD $INTC $ARM $QCOM $TSM

[Aakash Gupta @aakashgupta]

Microsoft sold every spare CPU it had to Anthropic and OpenAI. Amazon tripled its CPU buys year over year and still can't keep up. Two of AWS's biggest customers asked Andy Jassy if they could buy the entire 2026 production run of Graviton chips. He said no.

The ratio inside an AI datacenter used to be 100 megawatts of GPUs to 1 megawatt of CPUs. CPUs handled storage, checkpointing, pre-processing. Light work. GPUs did the actual training and inference.

Then OpenAI shipped o1-preview in September 2024. RL post-training went from "check the model output with a regex" to "run classifiers" to "compile the code and run the unit tests" to "spin up a sandbox, call three databases, run a physics simulation, verify the answer."

Every rollout now needs a CPU-backed environment to verify against.

Codex 5.4 runs agentically for 6-7 hours at a time. Each database call, each cron job, each scraped URL is CPU work. Coding agent revenue went from a couple billion to north of $10B in six months. That compute is sitting on CPUs.

The CPU to GPU ratio is now approaching 1:1. The entire global cloud was built for 1:8.

That's why GitHub has been unstable for weeks. Nvidia and Arm both announced they're entering the server CPU market in March. TSMC will only meet 80% of server CPU wafer demand this year. High-end server CPU prices are already up 50%.

When the GPU king and the IP licensor both pivot to CPUs in the same month, the boring chip isn't boring anymore.

[Brayden @BraydenWilmoth]

Cloudflare dashboard can now complete tasks for you.

- "Create a Worker and bind a new R2 bucket to it"
- "Change my DNS records to 1.1.1.1"
- "How many errors have happened this week"

Not only do we tell you, but we show you with generative UI.

PROTIP: Use full-screen mode.