AI Builders Digest — 2026-09-26

X / TWITTER

Vercel CEO Guillermo Rauch published AI model spend data from the Vercel AI Gateway (last 2 months): Anthropic remains #1 in spend but dropped from 69% to 40%; OpenAI climbed from 10% to 24% as GPT-6 Astra and GPT 5.6 Sol gain traction, and now leads in token volume; Kimi K3 and DeepSeek captured roughly half of Anthropic's lost share; Opus 5.5 hit 10% of spend within 2 days; OpenAI handles 62% of image generations.
https://x.com/rauchg/status/2103216656747262419

Vercel CEO Guillermo Rauch 公布了 Vercel AI Gateway 过去两个月的模型支出数据：Anthropic 仍居支出第一，但份额从 69% 跌至 40%；OpenAI 从 10% 升至 24%，GPT-6 Astra 和 GPT 5.6 Sol 势头凶猛，token 调用量已是第一；Kimi K3 和 DeepSeek 拿走了 Anthropic 流失份额的约一半；Opus 5.5 上线 2 天就占到 10% 支出；图像生成有 62% 跑在 OpenAI 上。
https://x.com/rauchg/status/2103216656747262419

Thariq (Claude Code engineer at Anthropic) outlined how Claude Code plan mode is being reworked into a built-in "mod": users will be able to customize the plan mode prompt, create and share their own modes, or ignore plan mode entirely and rebind shift+tab to something else. The change responds to feedback that many users plan for themselves, while others prefer a mode where Claude just thinks and brainstorms with you.
https://x.com/trq212/status/2103212052391354794

Thariq（Anthropic Claude Code 工程师） 披露 Claude Code 的 plan mode 将重构为内置 "mod"：用户可以自定义 plan mode 的 prompt、创建并分享自己的 mode，或者干脆无视它，把 shift+tab 重新绑定到别的功能。这个改动来自用户反馈——很多人自己就能规划，不需要 plan mode；另一些人则想要一个 Claude 只负责思考和头脑风暴的模式。
https://x.com/trq212/status/2103212052391354794

Replit CEO Amjad Masad announced that Muse can now make apps on Replit, with a demo of the agent building an app directly on the platform.
https://x.com/amasad/status/2103129037011120525

Replit CEO Amjad Masad 宣布 Muse 现在可以直接在 Replit 上做应用了，并附上了 agent 在平台内构建应用的演示。
https://x.com/amasad/status/2103129037011120525

Peter Steinberger (OpenClaw core, working with OpenAI) shared two field notes on agents: after running Daybreak on the codebase he found 8 more long-standing leaks — "care for your OSS dependencies". And on directing agents: never just tell an agent to "clean up" — it stops far too early; give it an ambitious, measurable goal like "remove 20% of the least useful tests while maintaining code coverage within 2%".
https://x.com/steipete/status/2103200311641076100
https://x.com/steipete/status/2103148444701610233

Peter Steinberger（OpenClaw 核心开发者） 分享了两条 agent 实战笔记：用 Daybreak 扫了一遍代码库，又找出 8 个存在已久的泄漏——开源依赖要用心维护。指挥 agent 时别只说"清理一下"，它会过早收工；要给有野心的可量化目标，比如"删掉 20% 最没用的测试，同时代码覆盖率变化不超过 2%"。
https://x.com/steipete/status/2103200311641076100
https://x.com/steipete/status/2103148444701610233

Swyx (Latent Space) reports his January "Scaling without Slop" content strategy is paying off: the first 100k YouTube subscribers took 3 years, the next 100k took 1.2 months, with similar traction across AEO/SEO and newsletter subscribers. He's officially teasing the next phase of Latent Space, AINews, and the rest of swyx inc.
https://x.com/swyx/status/2103361254433993165

Swyx（Latent Space 主理人） 宣布他年初提出的 "Scaling without Slop" 内容策略开始奏效：YouTube 前 10 万订阅花了 3 年，下一个 10 万只用了 1.2 个月，AEO/SEO 和邮件订阅的增长同样亮眼。他正式预告 Latent Space、AINews 以及 swyx inc 其余项目的下一阶段。
https://x.com/swyx/status/2103361254433993165

Josh Woodward (VP at Google Labs / Gemini App) highlighted Dreambeans, a newer Google Labs experiment with "a growing cult following": a fixed number of "beans" brew every morning, each pointing you at something in the real world, with real people, doing things you care about together.
https://x.com/joshwoodward/status/2103182635992514569

Josh Woodward（Google Labs / Gemini App VP） 推荐了 Google Labs 的新实验 Dreambeans，已有一批忠实拥趸：每天早上"酿造"固定数量的"豆子"，每一颗都把你指向真实世界里、与真实的人一起做你在乎的事。
https://x.com/joshwoodward/status/2103182635992514569

Peter Yang on the blistering pace of AI media generation: "Astra blew up 3D models. Then Opus blew up videos. I don't even know what's next anymore."
https://x.com/petergyang/status/2103318850641260959

Peter Yang 感叹 AI 媒体生成的进化速度："Astra 把 3D 模型卷疯了，Opus 又把视频卷疯了，我已经不知道下一个会是什么了。"
https://x.com/petergyang/status/2103318850641260959

Nikunj Kothari (partner at FPV Ventures) shared a vertical-AI thesis from a founder: every small business owner — in the trades specifically — will have bespoke software, and that last mile is where differentiation gets built, through unique experiences for customers and employees.
https://x.com/nikunj/status/2103360292973633770

Nikunj Kothari（FPV Ventures 合伙人） 分享了一个创业者的 vertical AI 判断：每个小企业主——尤其是水电网这类 trade 行业——都将拥有自己的定制软件，最后一公里才是差异化的来源，才能给客户和员工独一无二的体验。
https://x.com/nikunj/status/2103360292973633770

Matt Turck (FirstMark VC, MAD Podcast host) published his conversation with VAST Data CEO Renen Hallak: in Jensen's five-layer AI cake, the middle software-infrastructure layer is the least understood — and that's where VAST quietly became a $30B company few know about, powering xAI, CoreWeave, Nebius, Mistral and nscale.
https://x.com/mattturck/status/2103167531917721866

Matt Turck（FirstMark VC、MAD Podcast 主播） 发布了与 VAST Data CEO Renen Hallak 的对话：在黄仁勋的五层 AI 蛋糕里，中间的软件基础设施层最不被理解——而 VAST 正是在这里悄悄做成了 300 亿美元的公司，客户包括 xAI、CoreWeave、Nebius、Mistral、nscale。
https://x.com/mattturck/status/2103167531917721866

No notable posts this cycle: Dan Shipper (Every CEO), Aditya Agarwal (SPC General Partner).
本轮无实质内容：Dan Shipper（Every CEO）、Aditya Agarwal（SPC GP）。

PODCASTS

The MAD Podcast with Matt Turck — "Who Feeds the GPUs? Inside AI's Hidden $30B Layer | Renen Hallak, VAST Data"

The Takeaway: the most profitable company in AI infrastructure may be the one you've never heard of — the software layer that feeds data to the GPUs — and demand there is now outrunning even its own customers' projections.

Renen Hallak is founder and CEO of VAST Data, most recently valued at $30B, powering xAI and several of the biggest AI clouds from the exact middle of Jensen Huang's five-layer AI cake (power → chips → software infrastructure → models → applications). He calls that layer "the operating system for this new era": it manages storage, compute, networking and databases — and increasingly the models themselves, which he treats as just another resource like a GPU, needing routing, permissions and policy.

His demand signal is startling: a customer planned for 500 petabytes over three years, then came back a quarter later asking for an extra two exabytes on top. AI clouds are sold out 18 months ahead; the bottleneck is physical — land, power, chips. "Sometimes it scares me... I think in the next ten years, we'll see more difference than we did in the last thousand years."

The contrarian core of his worldview: enterprise IP will not live in documents — it will be distilled into model weights. "We infer during the day, we fine-tune at night." Agents will internalize an organization's experience through reinforcement learning, which is exactly why you can't hand your intelligence processes to someone else's cloud. Hence VAST's new DataEnclave, built with NVIDIA: end-to-end encrypted confidential computing that runs frontier-model inference inside the enterprise while model weights stay encrypted — no one sees what they're not allowed to see, opening regulated industries to frontier models.

Also notable: VAST is profitable in a money-burning category, because software margins ride AI's ~10x-every-2-years growth and "unsexy" plumbing meant almost no competition ("storage was where startups go to die"). Hyperscalers' lunch is being eaten by neo clouds that learn by actually building. And with OpenAI reportedly cracking Navier–Stokes by unleashing 10,000 agents on it, AI has proven it can do things we don't know how to do.

https://www.youtube.com/@DataDrivenNYC/videos

核心要点：AI 基础设施里最赚钱的公司，可能是你从没听说过的那家——给 GPU 供数的软件层——而那里的需求增速，现在连客户自己的预测都追不上。

Renen Hallak 是 VAST Data 的创始人兼 CEO，最新估值 300 亿美元，客户包括 xAI 和多家最大的 AI 云，正好处在黄仁勋五层 AI 蛋糕（电力 → 芯片 → 软件基础设施 → 模型 → 应用）的正中间。他把这一层称为"新时代的操作系统"：管理存储、计算、网络和数据库，以及越来越多的模型本身——在他看来，模型只是和 GPU 一样的资源，需要路由、权限和策略管理。

他的需求信号令人吃惊：有客户原本规划三年 500 PB，一个季度后回来追加了 2 个 EB。AI 云的容量已经卖到 18 个月后，瓶颈是物理性的——土地、电力、芯片。"Sometimes it scares me... 未来十年的变化，会超过过去一千年的总和。"

他最反共识的判断：企业的 IP 最终不会存在文档里，而是被蒸馏进模型权重。"We infer during the day, fine-tune at night"——agent 会通过强化学习把组织经验内化进自己拥有的模型，所以绝不能把智能流程交给别人的云。这也是 VAST 与 NVIDIA 推出 DataEnclave 的原因：端到端加密的 confidential computing，前沿模型在企业内部跑推理、模型厂商的权重全程加密，谁也看不到不该看的东西——把前沿模型带进强监管行业。

其他要点：VAST 在烧钱的赛道里保持盈利，因为软件毛利吃上了 AI 约两年 10 倍增长的红利，而且"不性感"的水管生意几乎没人竞争（"storage was where startups go to die"）；超大云厂商的午餐正在被真正下场干活、边干边学的新 AI 云吃掉；OpenAI 据传用一万个 agent 攻克 Navier–Stokes 方程——AI 第一次证明了它能做人类不知道怎么做的事。

https://www.youtube.com/@DataDrivenNYC/videos

Generated through the Follow Builders skill: https://github.com/zarazhangrui/follow-builders