> ## Content Index
> Fetch the complete content index at: https://www.antoinebuteau.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Daily Digest - 2026-08-27
- URL: https://www.antoinebuteau.com/daily-digest-2026-08-27/
- Published: 2026-08-28T12:49:44.000Z
- Updated: 2026-08-28T12:49:44.000Z
- Description: Looks at how AI labs are turning profitable by measuring margins against electricity instead of raw compute. AI companies are finally making money by turning wholesale power into cognitive work.
- Author: Antoine Buteau
- Tags: Digest

**1\. Revenue per Megawatt & The AI Model Factory — X (formerly Twitter)**

- Why read: Looks at how AI labs are turning profitable by measuring margins against electricity instead of raw compute.
- Summary: AI companies are finally making money by turning wholesale power into cognitive work. The new metric is revenue per megawatt. Efficient models produce high returns on constrained power, and these margins fund automated factories that speed up future training runs. The industrial process of making models matters more than the models themselves. We are moving from artisanal AI to an optimized supply chain.
- [Read more](https://twitter.com/ttunguz/status/2092981626716508495/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**2\. 1/ We spent H1 2026 talking to 20+ CTOs about... — X (formerly Twitter)**

- Why read: Shows how engineering teams are spending on AI and why high token burn doesn't equal high output.
- Summary: Based on conversations with 20+ CTOs, companies spend $1-2K per engineer on tokens each month. But a few power users burn up to $10K. The problem is telling apart actual 10x engineers from those wasting budget on useless token generation. The best teams track AI-written code against token spend and compare peers. Uber and Salesforce are already replacing headcount with AI, but doing so requires strict tracking. Engineering leaders need metrics for quality-adjusted throughput instead of paying blank checks for tokens.
- [Read more](https://twitter.com/BessemerVP/status/2093020817827168419/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**3\. You are not a model. Don’t price per token. — X (formerly Twitter)**

- Why read: Argues that AI app developers need to drop token-based pricing to protect their margins.
- Summary: Charging by the token ties your product's value to compute costs, which are dropping fast, and gives customers unpredictable bills. Companies should price based on units of work or business outcomes. A credit system abstracts the infrastructure and lets vendors keep the extra margin when model costs drop. It also makes costs predictable for finance teams while keeping usage data transparent. If your tool writes an account brief or closes a ticket, charge for the finished job, not the inference behind it.
- [Read more](https://twitter.com/sarahdingwang/status/2092976778164011481/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**4\. Agent Swarms are a Distributed Systems Problem — trychroma.com**

- Why read: Explains a new concurrency protocol that stops multi-agent systems from throwing away expensive LLM reasoning when databases abort.
- Summary: When multiple agents edit shared knowledge, standard database rollbacks fail because they discard minutes of costly compute. Chroma treats agent swarms as a distributed systems problem with a new protocol called "Fission." It intentionally breaks atomicity: transactions never roll back. Instead, aborted tasks commit their partial progress so other agents can pick it up later. This turns conflicts into asynchronous teamwork rather than wasted effort. It forces developers to build data architectures that save reasoning instead of enforcing strict atomicity.
- [Read more](https://www.trychroma.com/engineering/transactions?ref=antoinebuteau.com)

**5\. What it Takes for Coding Agents to Complete Large Software Tasks — X (formerly Twitter)**

- Why read: Shows how coding agents hit higher success rates when they write and follow their own completion tests.
- Summary: Single agents working on large codebases often quit early because they lose track of the final goal. Splitting the work into roles—orchestrator, implementer, and validator—lets agents build tests before they start coding. This keeps the completion standard from dropping to match whatever the agent happens to output. In tests on large codebases like GDAL, agents held to their own strict standards jumped to over 90% completion rates. Workflows need to separate defining success from executing the work so agents stop grading their own homework.
- [Read more](https://twitter.com/droid%5F35719/status/2093068852917899336/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**6\. notes on FDE — X (formerly Twitter)**

- Why read: Critiques the rush to hire Forward Deployed Engineers, arguing that good embedded engineering comes from culture, not recruiting.
- Summary: Major AI labs are spending heavily to build embedded engineering teams for enterprise clients. But rotating these engineers every three months destroys client context. Real embedded engineering requires an operational culture that teaches full-stack developers to act like product owners and talk like consultants. Hiring pre-made talent doesn't scale. Companies need daily standards to train these skills internally. The takeaway is to build a system that retains context instead of paying a premium for short-term consultants.
- [Read more](https://twitter.com/mardehaym/status/2092971458309820641/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**7\. The Pulse: We need to talk about migrations with AI — Substack**

- Why read: Argues that AI tools are turning multi-year legacy code migrations into projects that take weeks.
- Summary: AI changes the math on large-scale migrations. Asana cleared five years of technical debt in two weeks using Codex. Airbnb and Uber are moving test suites and dropping old frameworks faster and cheaper than ever. These projects use custom pipelines and AI loops to process files quickly, leaving humans to handle the edge cases. This frees up engineers to build features instead of nursing outdated libraries. Engineering leaders should review their backlogs—migrations that used to be too expensive are now obvious targets for AI automation.
- [Read more](https://substack.com/app-link/post?publication%5Fid=458709&%3Bpost%5Fid=212546097&%3Butm%5Fsource=post-email-title&%3Butm%5Fcampaign=email-post-title&%3BisFreemail=true&%3Br=34ymr&%3Btoken=eyJ1c2VyX2lkIjo1MjcwMzU1LCJwb3N0X2lkIjoyMTI1NDYwOTcsImlhdCI6MTc4Nzg1NDc4MSwiZXhwIjoxNzkwNDQ2NzgxLCJpc3MiOiJwdWItNDU4NzA5Iiwic3ViIjoicG9zdC1yZWFjdGlvbiJ9.tk4ZltkbspLZGLwp7FTtonUFB8c-S%5FEnBAVD8Z2OQDc&ref=antoinebuteau.com)

**8\. Making Your Data Ready for Agentic AI — martinfowler.com**

- Why read: Explains how to update data architectures so AI agents don't act on bad or outdated information.
- Summary: Human analysts stop when a metric looks wrong, but AI agents blindly execute on whatever data they get. To fix this, organizations need strict data contracts written as code to enforce quality and freshness before agents can access the data. A "quarantine pattern" routes bad data to a dead-letter queue so it doesn't break the agent's context. Modern stacks also need an "Adaptive Gold" tier where agents build their own optimized datasets based on usage. Data teams have to shift from building dashboards for humans to governing pipelines for machines.
- [Read more](https://martinfowler.com/articles/making-data-ready-for-agentic-ai.html?ref=antoinebuteau.com)

**9\. Where Does Value Accrue in Robotics? — X (formerly Twitter)**

- Why read: Presents a "barbell" theory for robotics: value will pool in massive horizontal models and specialized vertical deployment companies.
- Summary: Robotics is constrained by physical distribution, latency, and real-world safety in a way LLMs aren't. Intelligence alone doesn't equal useful work. Because of this, value will go to companies that integrate deeply into specific verticals like construction or logistics. They build compounding assets through failure data and fleet infrastructure. On the other end, horizontal Robot Foundation Models will win on scale. Generic tool vendors and system integrators will get squeezed out. Founders should either build base intelligence or completely own a specific vertical workflow.
- [Read more](https://twitter.com/joannalichter/status/2092667190847979561/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**10\. Jalapeño is a bandwidth argument, not a FLOPs argument — X (formerly Twitter)**

- Why read: A technical look at OpenAI's new custom chip, showing why optimizing for memory bandwidth over raw compute changes inference economics.
- Summary: OpenAI's Jalapeño chip ignores peak FLOPs to focus on bandwidth per watt. It under-provisions compute relative to memory. This lets the chip handle small batches at very low latency without losing efficiency, which agents need. The design drops the shared crossbar for local memory views, forcing workloads to stay local and relying on strict software scheduling. The rapid development was driven by AI-generated RTL code, showing AI hardware design works well with tight constraints. The future of inference is bandwidth-heavy hardware built for fast agent loops, not just maximum throughput.
- [Read more](https://twitter.com/krishgupta72/status/2092959507471376554/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**11\. AI Semiconductor Endgame 2026 (III) — X (formerly Twitter)**

- Why read: Explains why the AI infrastructure boom will last longer than the internet buildout.
- Summary: AI hardware demand is driven by more users and higher token usage per person. The AI business model runs on token sales, so the ceiling for infrastructure is higher than the flat-rate internet era. Despite bubble fears, the economics work for frontier labs. Inference returns are strong, and datacenter plans through 2027 map to projected revenue. A long-term overbuild will happen, but it will drive down compute costs for the application layer. Sustained compute demand is guaranteed as coding becomes the engine for all knowledge work.
- [Read more](https://twitter.com/fi56622380/status/2093040177711329673/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**12\. Closing the loop with self-improving cloud software factories — X (formerly Twitter)**

- Why read: A guide to upgrading from basic AI assistants to automated software factories that measure and improve themselves.
- Summary: The next step in agent engineering is treating software development as a closed-loop factory defined in code. Running these on the cloud lets teams track PR throughput, cost per PR, and automation rates across their agents. This allows teams to deploy observer agents that evaluate outputs and write PRs to improve the factory's routing and configuration. Using standard benchmarks lets teams run real A/B tests on multi-model strategies. Companies need to build this meta-infrastructure now to speed up development and use tokens efficiently.
- [Read more](https://twitter.com/zachlloydtweets/status/2093023948774449655/?rw%5Ftt%5Fthread=True&ref=antoinebuteau.com)

**13\. Piloting the world's first double-blind AI evaluations — Google DeepMind**

- Why read: Covers a breakthrough in AI benchmarking that uses cryptography to stop models from cheating on tests.
- Summary: Benchmarks are losing credibility because models ingest test questions during training and artificially inflate their scores. DeepMind is running double-blind evaluations inside secure GPU enclaves to keep both the model weights and the test prompts hidden. This stops the tradeoff between protecting IP and validating safety. It proves an AI's performance is real and hasn't been skewed by training data contamination. This sets a clear standard for testing frontier models.
- [Read more](https://deepmind.google/blog/piloting-the-worlds-first-double-blind-ai-evaluations/?utm%5Fsource=x&utm%5Fmedium=social&utm%5Fcampaign=&utm%5Fcontent=)

**14\. The Doorway Becomes a Place — motivenotes.ai**

- Why read: Uses Nvidia's bid for Hugging Face to show how building clean interfaces over messy systems creates massive value.
- Summary: Hugging Face didn't try to fix messy machine learning research code. It just built a clean interface on top of it. By becoming the default way to access models, it gathered datasets, evaluations, and developers, turning a tool into infrastructure. Its $12.9 billion valuation comes from owning the distribution layer for open AI. Builders should find chaotic systems and build beautiful abstractions over them. Thin interfaces without clear business models can become essential strategic assets.
- [Read more](https://www.motivenotes.ai/p/the-doorway-becomes-a-place?ref=antoinebuteau.com)

**15\. Why your AI agent sucks — Substack**

- Why read: Explains why plugging an AI into Slack or Drive fails, and why structured data is the real bottleneck.
- Summary: Companies often connect internal tools to an AI and assume the job is done. But agents can't tell the difference between old drafts and final documents in messy folders. Rather than forcing employees to use strict naming conventions, companies should run cleanup agents overnight to organize and archive data. Treating data organization as infrastructure lowers the compute costs of agents searching through junk. The winners in the next phase of AI will be companies with structured, machine-readable knowledge bases.
- [Read more](https://substack.com/app-link/post?publication%5Fid=2121292&%3Bpost%5Fid=213037031&%3Butm%5Fsource=post-email-title&%3Butm%5Fcampaign=email-post-title&%3BisFreemail=true&%3Br=34ymr&%3Btoken=eyJ1c2VyX2lkIjo1MjcwMzU1LCJwb3N0X2lkIjoyMTMwMzcwMzEsImlhdCI6MTc4Nzg2MDc1NywiZXhwIjoxNzkwNDUyNzU3LCJpc3MiOiJwdWItMjEyMTI5MiIsInN1YiI6InBvc3QtcmVhY3Rpb24ifQ.6nEMjiQA79DmMdmVC%5FUWd4iPqhjdWzS7KikxVhsvCEs&ref=antoinebuteau.com)

### Themes from yesterday

- **Industrializing AI Development**: The focus is shifting from raw models to efficient factories, specialized inference chips, and automated software pipelines.
- **Rethinking Measurement and Value**: The industry is dropping token-based pricing for outcomes and adopting double-blind tests to enforce real standards of performance.
- **Fixing the Context Bottleneck**: Agents need structured data to work. This means enforcing data contracts, running cleanup scripts, and building concurrency controls for swarms.
- **Automating Automation**: Agents are managing and evaluating other systems, handling large code migrations, and building their own tests for completion.