This week's material points to one thing. Competition at the model layer is starting to commoditize, while the ability to put AI reliably into enterprise workflows — deployment, process knowledge, operational memory — is being repriced by capital, the giants, and founders all at once. Indigo's calls on X this week line up with more than a dozen primary sources, and together they sketch a map of value changing layers.
2026.05.10 — 2026.05.17 · Once a week: spot the signals, recalibrate.
This Week's Signals
Start with the facts. On May 14, Cerebras went public, 20x oversubscribed, with its valuation stepping up to 48 billion. The same week, Anthropic CFO Krishna Rao publicly confirmed that annualized revenue grew from $9 billion to $30 billion in a single quarter, up 333%. OpenAI acquired Tomoro AI, gaining 150 FDEs in one move (Forward Deployed Engineers — engineers stationed at customer sites who wire AI into business workflows). And on May 13 alone, Richard Socher's Recursive announced a $650 million round, and David Silver's Ineffable Intelligence closed a $1.1 billion seed — the largest in European history.
Most people will read this string of news as a continuation of the model race: who raised money, who went public. But put the items side by side and Indigo's reading this week is different — what is really changing is that the value-capture structure itself is shifting layers. Vercel's production data shows that agents running in real businesses use 35 models on average, and switching model vendors is just a one-line config change. Meanwhile, OpenAI, Anthropic, Sequoia, and YC all pushed resources toward deployment and services in the same week. This is not five news stories. It is one: models are getting cheaper, and the ability to put models into real businesses is getting more expensive.
The real competition is no longer whose model is smarter. It is who can reliably put a model inside someone else's business.
Where the Wind Is Blowing
#01 FDE: The New Key Role of the Agent Era
Indigo recorded a full podcast episode on this role this week. On X he said: 和 Cresta 的 Head of FDE 同学 Jove 聊最近很火的 FDE（Forward Deployed Engineer），这应该是未来两三年很重要的工程岗位，Agent Native 的顾问公司会大量出现
— "Talked with Jove, Cresta's Head of FDE, about FDE (Forward Deployed Engineer), which is getting a lot of buzz lately. This should be a very important engineering role over the next two or three years, and Agent Native consulting firms will appear in large numbers" (original post). Agent Native means companies organized around agents from day one. The industry has already voted with real money: OpenAI acquired Tomoro AI and picked up 150 FDEs in one stroke, replaying Palantir's old playbook — station engineers at the customer site and load enterprise knowledge, piece by piece, into an Ontology (an enterprise knowledge graph). Long Lake, with an architecture of 80% horizontal infrastructure plus 20% vertical deployment, has compressed the cycle for entering a new industry from over a year to a few days. Canadian telecom TELUS used this playbook to build more than 13,000 custom AI solutions and save 500,000 hours. Every's interview puts it even more bluntly: infrastructure is the wall where most agent projects die; the hard part was never the model — it is harness engineering (the software layer that connects the model to enterprise workflows). An FDE is not a systems integrator. An FDE is someone who turns enterprise process knowledge into a reusable asset.
#02 Continual Learning: A Three-Layer Architecture and a 50x Collision
Indigo's weightiest original call this week: LLM 卡在《记忆碎片》式的永恒当下，需要持续学习才能真正进化！…… 一个成熟的系统应该分层使用：ICL 是第一道防线，模块层做个性化，参数层做真正的发现。…… 这个架构跟 Anthropic Code with Claude 发布的 Routines / Dreams / Outcomes 三个 Managed Agents 特性在抽象层级是同构的
— LLMs are stuck in a Memento-style eternal present; they need continual learning to truly evolve! … A mature system should work in layers: ICL as the first line of defense, the module layer for personalization, the parameter layer for real discovery. … This architecture is isomorphic, at the level of abstraction, to the three Managed Agents features — Routines / Dreams / Outcomes — that Anthropic released at Code with Claude
(original post). Here ICL means in-context learning — things a model picks up temporarily inside a conversation's context; Managed Agents is a product form where agents are hosted on a platform and run automatically. Interestingly, this three-layer framework was independently confirmed by several parties the same week: a16z research shows an 8B model with the right module layer can perform like an 800B model — a 100x gap in parameter efficiency. Anthropic's CFO also publicly acknowledged that more than 90% of Claude Code's code is written by Claude Code itself — recursive self-improvement (AI improving AI itself) just went from a probabilistic prediction to a production fact. And the scales of capital tilt hard: on the acceleration side, Recursive plus Ineffable total $1.75 billion; on the warning side, Bengio's LawZero has only $35 million in philanthropic funding — a 50x gap in scale. Continual learning is not an academic topic. It is where the heaviest bets landed this week.
#03 Cerebras's Ceiling, Anthropic's Acceleration
Two IPO narratives crossed paths this week. Indigo warned a day in advance: Cerebras 即将于 5月14日 IPO，股票代号 $CBRS …… 大家无脑冲的时候，注意天花板
— Cerebras is about to IPO on May 14, ticker $CBRS … While everyone is piling in without thinking, watch the ceiling
(original post) — his most widely shared post of the week. Where exactly is the ceiling? Cerebras's revenue depends heavily on a $24.6 billion backlog (orders signed but not yet delivered) from OpenAI alone, plus a $1 billion interest-free working-capital loan, and OpenAI will end up holding about 12% of the company. SemiAnalysis also points out that the SRAM (on-chip cache) its wafer-scale chip depends on gains almost nothing on the new N3E process, while Nvidia's interconnect density is 130x higher. In the end, Cerebras is not an era-defining giant; it is a niche winner financially locked in by a single large customer. On the other side, Anthropic's IPO prelude advanced on several fronts this week: the 333% single-quarter growth went public, a 100B+ portfolio of compute commitments landed, flagship customer Mercado Libre put all 23,000 of its engineers on Claude Code, and the market's valuation talk stretched from 900B all the way to $1.5 to $2 trillion. Both are IPO stories. One turns on customer concentration; the other turns on the slope of growth.
On the Ground
#04 The Model Is a Config Setting
Agents in production carry 35 models on average, and the share of tokens going to tool-calls (the model calling external tools) rose from 31.6% to 58.9% in six months. A model is not an asset. It is one line of config you can swap at any time. That is the most tangible shape of commoditization, and the starting point of the layer shift.
#05 Organizational Memory Beats Model Upgrades
At Shopify, 5,938 employees use an internal agent, submitting 1,870 PRs (pull requests) a week, with the merge rate rising from 36% to 77% in two months — without switching models. The gains came from observing what should be observed and writing down what should be written down. The improvement comes from organizational memory, not model capability. This is the most granular evidence of value in the deployment layer.
#06 From Writing Code to Directing Agents
After trying multi-agent orchestration, Indigo said: 终于你可以在 Claude Code 里一览正在工作的牛马 Agent 的全貌了 …… 我们现在需要一个真正的 Agent IDE 而不是 Coding IDE
— Finally you can see the full picture of the workhorse agents laboring away in Claude Code … What we need now is a true Agent IDE, not a Coding IDE
(original post). IDE means development environment — competition at the tool layer is shifting from helping people write code to helping people direct a fleet of agents.
#07 The Big Three's Differentiated Steady State
我们一家人用三款AI模型 - 小孩喜欢 Gemini（学习最佳）/ 老婆用 ChatGPT（日常问答）/ 我用 Claude（工作流自动化）
— "Our family uses three AI models - the kids like Gemini (best for learning) / my wife uses ChatGPT (everyday Q&A) / I use Claude (workflow automation)" (original post). Indigo used his own family's real usage to confirm one thing: commoditization is not homogenization — it is a division of labor, each model in its own niche.
#08 A New Unit of Measure for Compute
Coatue argues that the atomic unit of AI investment has shifted from the GPU to the gigawatt (billions of watts of electricity, the new yardstick for compute clusters), and the supply-scarce links — memory, optical modules, power equipment — have seen profitability jump 3 to 7x, with contracts locked out to 2029 through 2030. Below the deployment layer, the power layer is being repriced too.
#09 Services Are the New Software
Sequoia and YC offered the same framework in the same week: behind every $1 of software sits $6 of service-delivery labor, and AI agents are folding that $6 into software pricing, pointing to a $10 trillion services market. What AI companies deliver is no longer a tool. It is an outcome. This is the deployment-layer story pushed to its limit.
Slow Thinking
Indigo did change his framework this week. Before, he viewed AI value capture mainly through two layers: the model layer and the application layer. Now his map has at least five — a supply-scarcity layer (power, memory, optical modules), the model layer, a process-knowledge and operational-memory layer, a deployment and FDE layer, and a services layer run directly by AI — with the thickest layer sitting at the bottom. The trigger was the convergence of signals on May 13: Anthropic's CFO disclosing 333% single-quarter growth, Recursive and Ineffable closing large rounds the same day, Sequoia proposing that services are the new software, and Skills (written-down process knowledge fed to agents) being named as an architectural layer of its own. There is also a steeper update: recursive self-improvement moved from future tense to present perfect — it used to be a probabilistic prediction, and now Anthropic's CFO has confirmed in his own words that more than 90% of the code is written by Claude Code. The warning camp's concerns are no longer just a hypothesis.
But on the other side, the strongest rebuttal to this week's narrative is this: the model-commoditization story is being told precisely by the beneficiaries of the deployment layer. OpenAI absorbing FDEs, Anthropic pushing Managed Agents, Sequoia and YC pricing service companies — every party profits directly from the narrative. Indigo himself has called out this kind of interest structure before: Anthropic 的最新 AI 政策研究，真是为封堵来自中国的竞争操碎了心 …… 但研究最严重的问题是：里面充满了利益相关者偏见，Anthropic 同时是政策诉求者、Mythos 能力的拥有者、和 Project Glasswing 的发起者。它推的每一条政策都对它直接有利
— "Anthropic's latest AI policy research really goes to great lengths to block competition from China … But the most serious problem with the research is that it is full of stakeholder bias: Anthropic is at once the policy advocate, the owner of Mythos capabilities, and the initiator of Project Glasswing. Every policy it pushes benefits it directly" (original post). There are two ways to falsify the framework. One: if parameter-level continual learning in the next generation of models lets companies deploy without hiring FDEs, the value the deployment layer has built up over these years will depreciate fast. The other: if frontier models' pricing power keeps expanding instead of converging — Anthropic's 333% single-quarter growth itself looks more like a scarce good than a commodity — then the commoditization premise was wrong from the start.
Indigo on X
Cerebras 即将于 5月14日 IPO，股票代号 $CBRS …… 大家无脑冲的时候，注意天花板
— Cerebras is about to IPO on May 14, ticker $CBRS … While everyone is piling in without thinking, watch the ceiling
From @indigox, 76 likes
计划明天约 Cresta 负责 FDE 的同学录一期 Indigo Talk，来详解一下什么是 FDE？以及为什么在 Agent 时代它很重要
— Planning to sit down tomorrow with the person who runs FDE at Cresta and record an episode of Indigo Talk, to break down what FDE is, and why it matters in the Agent era
From @indigox, 56 likes
和 Cresta 的 Head of FDE 同学 Jove 聊最近很火的 FDE（Forward Deployed Engineer），这应该是未来两三年很重要的工程岗位，Agent Native 的顾问公司会大量出现
— "Talked with Jove, Cresta's Head of FDE, about FDE (Forward Deployed Engineer), which is getting a lot of buzz lately. This should be a very important engineering role over the next two or three years, and Agent Native consulting firms will appear in large numbers"
From @indigox, 46 likes
Closing
One Thought
What Palantir did in the last cycle is exactly what everyone is doing today: station engineers at customer sites, load the enterprise's tacit knowledge into an Ontology piece by piece, and grind through it project by project. For a long stretch, that got it treated as a labor-heavy consulting firm that could not scale. This week, OpenAI absorbed 150 FDEs in one move, and Anthropic is co-building deployments with a flagship customer — the two largest model makers publicly paying tribute, at the same time, to a model once looked down on. The heavy-services model mocked in one cycle is often the very moat of the next. To judge whether a model is worth anything, don't ask whether it looks like software today. Ask whether what it accumulates can be reused.
One Thing to Try
Indigo wrote this week: AI 已经把分析能力压缩成商品，真正稀缺的是判断力、品味、和提出好问题的能力！这些只能向内寻找
— AI has already compressed analytical ability into a commodity. What is truly scarce is judgment, taste, and the ability to ask good questions! These can only be found by looking inward
(original post). Spend 30 minutes testing that sentence. Pick one process from your own work that repeats every week and write it up as a one-page spec — what the inputs are, what the steps are, what counts as an exception, and how to tell when it was done well. Odds are you will discover two things: 90% of the steps could be handed to an agent tomorrow, and the remaining 10% you cannot write down is exactly where judgment lives. That one page is a minimal skill — and a personal-scale sample of deployment-layer value.