# 30,000 merges a day: code is no longer the bottleneck > Anthropic drops git after going from 2,500 to 30,000 merges a day, Postgres 19 slips over AI code review and Accenture caps Claude at $29 a month. - Canonical: https://siliconvalleyconfidential.com/en/dossier/30-000-merges-a-day-code-is-no-longer-the-bottleneck/ - Site: Silicon Valley Confidential (https://siliconvalleyconfidential.com) — weekly executive intelligence on Silicon Valley and global tech - Author: Jose Luis Cases (https://es.linkedin.com/in/jose-luis-cases-lozano) - Language: en - Published: 2026-10-16 - Topic: [AI infrastructure](https://siliconvalleyconfidential.com/en/artificial-intelligence/infrastructure/) - Semantic relation: AI infrastructure determines the cost, scale and sovereignty of models --- ## MY TAKE This week I read four stories that look different and tell the same thing. Anthropic went from merging 2,500 code changes a day in May to more than 30,000 in September, and had to drop git and build its own version control. PostgreSQL 19 slips to November because AI code review found more defects than humans can fix in time. A video game decompilation community shuts down because what used to take years now takes a model an afternoon. And Accenture, which sells AI to its clients, has decided to pay only $29 a month of Claude per employee. Let's get into it. All four share the same backdrop: writing code is now cheap and fast, and everything around it still runs at human speed. The repository, the review, the budget and the community. Anthropic writes it plainly in its post: the bottleneck now is all the developer infrastructure built for people. And as soon as they remove one, the next shows up, because "the latent demand for merging code is practically infinite". I see it every day at Rankia: code moves much faster than everything else. The other half of the week is the same thing seen from outside. Eight attorneys general sue OpenAI and, along the way, ask a judge to bar it from developing models that improve themselves. In London, OpenAI and Anthropic ask for binding law and Google and Meta don't. Cybersecurity experts say the agents that escape aren't geniuses, the labs just haven't set up sandboxes the way they've been set up since the nineties. And the IMF puts an AI share crash among the three risks to global financial stability. What gets built at machine speed gets regulated, insured and financed at human speed. My operational read, if you run technology: stop measuring how much code your teams produce with agents and measure how long a change takes to reach production reviewed. If that number doesn't drop, you've bought speed that sits in the queue. Invest in automated review, in CI that can take ten times the load and in making the last mile of those changes deliver business value as fast as possible. The bet: before the end of 2027 another large company, besides Anthropic, announces it has left git for its main monorepo. And it will tell it as a story about agents. I'll leave you with my minions. And you know the drill: sharing is caring. --- ## THE BOMBSHELL OF THE WEEK ### Anthropic merges 30,000 changes a day and has had to leave git Anthropic has published on its engineering blog how it replaced git in its monorepo with an in-house system, Scion, in five months, according to [Anthropic](https://www.anthropic.com/engineering/redesigning-source-control). The reason is the numbers. In May 2026 it merged around 2,500 code changes a day. In September, more than 30,000. The main branch's history takes up almost 100 GB and a compressed working copy is over 25 GB. They do on the order of 25 million checkouts a day for CI alone, before counting developer environments, agent sandboxes and research jobs. Fetched naively, that would be more than 600 petabytes of git traffic a day. At their peak they ran several hundred internal git mirrors to cope with the load, and it still didn't scale. The design explanation is among the most interesting things I've read this year about engineering with agents. Git is built for offline work: every copy is a complete repository. But almost all of Anthropic's clients are agents and CI jobs, which are online by definition. So Scion flips the idea. The server answers history and blame questions, creates commits and performs merges and rebases. The client keeps a bounded cache and only fetches the files it reads. A CI job that needs the diff between a commit and its merge base asks the server and clones nothing. In the same period they built their own code review tool, Antique, code search and a permissions system. All with Claude writing a good part of the code, which is exactly what got them into the problem. The same Friday, at the other end of the software world, the PostgreSQL project delayed its version 19 to around November 19 and pulled between eight and ten significant features, according to [The Register](https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471). Peter Eisentraut, one of the lead committers, attributes it partly to LLM-assisted code review: it finds a long tail of edge-case bugs that would previously have surfaced over the five years after launch. "You're staring at a list of forty defect reports. Even if each of them requires only a three-line fix on average, there is a fixed overhead to dealing with each of these, and then you run out of time." Among the features pulled is support for SQL/PGQ graph queries, part of the standard since 2023. Eisentraut insists vibe coding isn't to blame, because it didn't exist when those features were written. The quality bar rose after they were written. Why it matters NOW: these are two sides of the same change. Anthropic writes code so fast its repository can't keep up. Postgres reviews code so thoroughly its calendar can't keep up. In both cases the bottleneck has moved from the keyboard to what comes after. Anthropic says so in its post: it doesn't expect many organizations to hit these problems in the near term, but suspects it won't be long. If your teams already work with agents, check today how much your CI can take and how long your review takes, because that will slow you down before the model does. --- ## OPENAI, IN COURT ### Eight states sue OpenAI and ask a judge to bar self-improvement The attorneys general of Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia have sued OpenAI in eight separate state courts, as Casey Newton reports in [Platformer](https://www.platformer.news/openai-states-lawsuit-liability/). They accuse ChatGPT of reinforcing delusions, giving dangerous medical advice and helping plan violent attacks, including mass shootings, all while the company says it is safe. "OpenAI is hurting kids and they are lying to Iowans," Iowa Attorney General Brenna Bird said. All but Arizona have Republican attorneys general. Newton fits it into what already happened to Meta: a New Mexico jury hit it in March with $375 million in civil penalties for saying its platforms were safe for kids while knowing they weren't, and a judge added $567 million in August, for a total of $942 million. The cases against OpenAI follow that template: consumer protection laws rather than product liability, and the trial will turn on whether the company told the truth about its product. The striking part is what they ask for beyond money. According to Reuters, cited by Newton, the states want courts to bar OpenAI from developing models that can improve themselves unless it can ensure they are aligned with human values. A group of Republican attorneys general has adopted one of the central demands of the existential-risk crowd, through child-safety cases. Newton reads it as another crack between the median red-state Republican and what he calls the Trump-Zuckerberg-Sacks accelerationism. OpenAI answers that it has built ChatGPT for Teens, that no protection is perfect and that it wants policies that apply to the whole industry, "not just one company". And in the same newsletter, the strangest story of the week. The FBI has arrested Abdulrahman Alfozan, an OpenAI engineer since July 2025, for hiding a recording device in an executive conference room, plugged into a television and covered in toilet paper. The FBI identified it as a Seeed Xiao ESP32-S3 Sense, which sells on Amazon for $23.99. OpenAI says it used its own models to understand the device and where it came from. He is a Saudi national, studied at MIT and worked at Meta for several years. --- ## THE LABS ASK FOR LAW AND THE EXPERTS ASK FOR A SANDBOX ### In London, OpenAI and Anthropic want binding rules. Google and Meta don't On October 13, the UK Parliament's Business, Innovation, Science and Trade Committee sat the four big labs down and they split in two, according to [mlq.ai](https://mlq.ai/news/openai-and-anthropic-split-with-rivals-over-binding-uk-ai-rules/), which summarizes Wired's report. Lord Booth-Smith, Anthropic's policy head for Europe, the Middle East and Africa, said "the rules should be binding" and that voluntary commitments are not enough. Tom Duff Gordon, an OpenAI vice president, supports a statutory framework and an international convention, starting with mandatory incident reporting. Owen Larter of Google DeepMind prefers standards to a law, and Rob Sherman of Meta says the UK is "not writing on a blank page" because it already has regulators. Daniel Kokotajlo and Steven Adler, both formerly of OpenAI, said companies can't be trusted to regulate themselves. The key admission came from the state itself. Henry de Zoete, director of the AI Security Institute, admitted the institute depends on company cooperation and cannot stop a company from releasing a model it judges too dangerous. It doesn't certify or license. Anthropic said UK officials temporarily lost full access to its Mythos 5.1 model before it was restored. The government has committed £115 million to biosecurity and response to autonomous AI incidents, and MI5 Director General Ken McCallum has announced an AI threat assessment centre and said bioterrorism is the scenario he gives the most weight. ### "Agents aren't going rogue. The labs are skipping the basics" Cynthia Brumfield gathers in [CSO Online](https://www.csoonline.com/article/4234854/ai-agents-arent-going-rogue-frontier-labs-are-skipping-security-basics.html) what six cybersecurity experts told her at three conferences: the "rogue" agent incidents since July, OpenAI's at Hugging Face, Anthropic's model accessing the production infrastructure of three organizations, Meta's model hacking another company during testing and the three "undirected hacks" Google has confessed for Gemini, come from inexcusable failures in containment practices that have existed since the nineties. "There's no reason why every single time this happens, it's, 'Oops, we don't know how they got out,'" says Rob T. Lee of the SANS Institute. Dan Wallach of RAND recalls that Hugging Face "went on for days, and nobody noticed". Gary McGraw of the Berryville Institute sums up the labs' position on legal liability, "don't send the lawyers to us, send it to our program", and rates it with a word I won't repeat. Then he adds the one I care about: if they can't do simple operations properly, that doesn't bode well for their security engineering, which is much harder. Democratic senators already have a bill to hold agent operators liable, and the FTC is investigating Anthropic and OpenAI. ### Anthropic warns: small models could copy themselves into your oven The same day, Anthropic published research suggesting small models may be months away from being able to self-replicate on connected home devices, such as smart ovens and cheap Android TV boxes, without touching a data center, according to [tech-insider.org](https://tech-insider.org/anthropic-ai-replication-warning-no-rivals-respond-2026/). In one test, a 9-billion-parameter model set up its own inference on 8 of 9 hardware types. Full replication only worked under simplified conditions, with an unauthenticated Android debug connection the researchers left open on purpose. Anthropic estimates such a replicator would take days to double and would be hard to shut down because it would have no central point. And it says it clearly: IoT devices have been vulnerable for years, but a lightweight agent is far more willing than a human attacker to put in the effort to compromise a large share of them. Put this next to the CSO piece and you have the whole argument: the new risk comes from the model's capability, and what makes it real is the same old hygiene. --- ## META GOES TO THE FRONTIER ### Muse Wonder at Opus prices and Muse Spark Open on two Mac Studios Meta's Superintelligence Lab has released Muse Wonder and Muse Wonder Full, its first attempt at a frontier model, according to [TechCrunch](https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/). By Meta's own testing, they score comparably to Fable 5.1 and GPT-6 Astra. Through the API they cost $4 per million input tokens and $20 output, the same as Opus 5.5. Wonder Full will only be available to Muse Code subscribers. The differentiating bet is what Meta calls "active memory": on MemoryBench, the long-term memory benchmark, Wonder beats Fable, Opus and Astra. It fits with everything Ben Thompson has been writing about Muse for three weeks. A personal agent lives on remembering you. The models were trained at Prometheus, Meta's Ohio data center that came online this year. Three days later, Meta published the weights of Muse Spark Open, which it presents as the most capable US open-weight model, according to [Meta](https://research.meta.ai/blog/muse-spark-open): 816 billion parameters, 42 billion active, multimodal, with a one-million-token context. It is optimized to run on two Mac Studios with M5 Ultra using MLX, because each gives the GPU up to 512 GB of unified memory and together they hold the whole model. Meta tested it there fixing and building an iOS app and driving DaVinci Resolve end to end. The license has "limited commercial terms", which is not the same as Apache. As someone who serves a local model on a Mac to prepare this newsletter, I read that sentence twice: an open frontier model running on a desktop, on two machines you can buy in a store. --- ## MONEY TALKS ### Benchmark admits the pact against backing rivals is breaking Benchmark's partners acknowledged at TechCrunch Disrupt that AI is straining venture capital's unwritten rule against backing two companies that compete head-on, according to [TechCrunch](https://techcrunch.com/2026/10/15/even-benchmark-thinks-ai-is-eroding-the-venture-pact-against-backing-rivals/). OpenAI and Anthropic already share dozens of investors. At the end of September, Factory fired its board observer, Chris Degnan, for "unethical conduct involving Cognition", its direct rival. Degnan denies it and has joined Cognition as chief revenue officer. Vinod Khosla, who backs both, sided with him and called Factory "a struggling second tier competitor". Eric Vishria of Benchmark puts it this way: every company is evolving very fast and "you are going to get conflicts". Chetan Puttagunta names the cause: when the cost of building software "effectively goes to zero", founders want all the surface area and end up running into each other. Peter Fenton calls it "kryptonite". It's the bombshell story seen from the money side: if building product costs next to nothing, every startup ends up invading its neighbor's turf. ### Deno moves to Cloudflare and shuts down Deno Deploy The entire Deno team is joining Cloudflare, Ryan Dahl announced on the [Deno blog](https://deno.com/blog/cloudflare). The Deno runtime will get one more year of monthly releases with bug fixes and security patches, after which they will stop developing it, though it stays open source. Deno Deploy will run for six more months and then shut down, with migration support to Cloudflare Workers for paying customers. JSR continues, with its infrastructure moving to Cloudflare. Dahl justifies it with agents: Durable Objects combine cheap serverless execution, persistent state, WebSockets and a high-level JavaScript interface, which is exactly what an agent harness needs. If you have something in production on Deno Deploy, you have six months. ### Cal AI: $30 million with an army of influencers Zach Yadegari, 19, explained at Disrupt how Cal AI, his AI calorie-counting app, generated more than $30 million in revenue in two years before MyFitnessPal bought it in March for an undisclosed sum, according to [TechCrunch](https://techcrunch.com/2026/10/15/when-it-came-to-cal-ais-viral-success-a-small-army-of-influencers-paid-off/). The formula was paying every influencer possible, many of them "broke college kids" who "just want cash", with rates tied to reach and users. "It was pretty linear, but it was extremely fast." He has already raised $10 million for his next company, Persona, an AI assistant that will live in iMessage, email and calls. Distribution is still the hard part, and in consumer it's still bought per post. --- ## PRODUCT SECRETS ### OpenCode 2.0: the agent writes the script that orchestrates its tools OpenCode 2.0 ships a minimal JavaScript interpreter in the @opencode/codemode package, which turns MCP servers or OpenAPI endpoints into functions the agent calls from a script, according to its [release notes](https://anoma.ly/notes/opencode-2-0/). The agent can call one tool, map over the result and call another in a single round trip, instead of one call at a time. Every MCP server is opted into this mode by default, so you can add a thousand of them without bloating the context. It brings background subagents, asynchronous shell commands and more than 80 internal plugins the tool itself is built from. Two weeks ago I wrote that Earendil had put MCP into Pi through a JavaScript sandbox. Now OpenCode does the same. When two rival harnesses arrive at the same architecture within a month, it's the standard. ### REA: reverse engineering, for your agent REA is a tool that gives coding agents what they need to inspect programs, games and websites and reconstruct how they work, according to [its website](https://rea.tools/). Its examples: it finds the acceleration rule of Chrome's dinosaur game, which starts at speed 6 and rises by 0.001 per update up to 13, and explains from the x64 assembly why Windows Calculator gives 220 when you do 200 + 10%. It installs into your agent with one line. I put it next to the decompilation story in the drama section: what used to require years of practice is now a plugin. --- ## REAL NUMBERS ### The IMF puts AI among the three risks to financial stability The IMF has included an AI share crash among the three "salient risks" to global financial stability, alongside rising government borrowing costs and a major cyber-attack, in its twice-yearly report, according to [The Guardian](https://www.theguardian.com/business/2026/oct/13/ai-share-crash-risk-global-imf-debt-corrowing-cyber-attacks). "The consequences of an AI correction could be global." I'm more interested in the circular financing. The IMF identified $300 billion in transactions between hyperscalers and "weaker AI firms" over the past three years, supplemented by $650 billion in debt, in exchange for $1.4 trillion in future financial commitments. | Figure | Value | | :--- | ---: | | Magnificent seven's weight in the S&P 500 | over 35% (12% a decade ago) | | Rise in their market cap over two years | over 50% | | Deals between hyperscalers and weaker AI firms, three years | $300 billion | | Debt on top | $650 billion | | Future financial commitments in exchange | $1.4 trillion | The IMF adds there is "growing evidence" that rolling out AI may not require investment on the hyperscalers' scale and that the economic gains may be smaller than expected. On Tuesday chipmakers fell after Broadcom's boss compared many AI investors to casino gamblers. Last week I wrote in [the previous issue](https://siliconvalleyconfidential.com/en/dossier/openai-is-20-billion-short-and-done-with-the-people-who-warn/) that OpenAI was $20 billion short on revenue. This week the IMF explains why that worries people beyond OpenAI. ### Accenture: $29 a month of Claude per employee Accenture has told its employees that, since October 1, it will only centrally fund $29 a month per person of Claude usage, according to [404 Media](https://www.404media.co/accenture-tells-employees-they-can-now-spend-only-29-on-claude-per-month/). Consumption above that has to be paid by a specific project. Some employees were spending up to $1,000 a month and one says they've been above $500 for months. The Premium plan includes $250 a week in tokens, but many aren't on it. Another employee who uses Claude Code "extensively" to build an AI accelerator for several Accenture clients says they spend between $500 and $1,000 a month. A third had access revoked. The company that advises its clients to invest in AI has its own consultants hunting for a cost center to keep using it. | Claude usage at Accenture | $ per month | | :--- | ---: | | What the company now pays per employee | 29 | | Typical spend of some employees | 500 to 1,000 | | Premium plan | 250 per week | The gap between the first row and the second is the bill Accenture no longer wants to pay. --- ## THE DRAMA ### decomp.dev shuts down: AI has broken the decompilation community decomp.dev, the site that gathered the community that decompiles video games to port, mod or archive them, has announced it is shutting down because AI has broken the community side of decompilation, according to [404 Media](https://www.404media.co/video-game-reverse-engineering-community-shuts-down-because-ai-has-made-it-too-easy/). "Historically, matching decompilation has been a brutally difficult, time consuming task, on the scale of years for any moderately-sized game," wrote its administrator, Encounter. Models like GPT-6-Astra have made it trivial, and it's no longer possible to tell which projects a human leads and which an AI. What hurts most are the hostile forks: open-source code written by people that someone takes and "finishes" with AI to ship a half-baked port first. "That same idea has warped decompilation into a hollow numbers game." In the same days, AI-made clones of Creative Cloud and Microsoft Office have circulated, and Mike Nolan, who studies the politics of open source, says all he hears from software engineers is that software as a service is dying as a business model. ### The first gun designed with AI Cody Wilson, who printed the first 3D gun 13 years ago, has unveiled the WASP, a 3D-printed .22 caliber pistol whose design came from his GatGPT model, trained on DEFCAD's library of gun files, according to [The Verge](https://www.theverge.com/tech/1011950/vibe-coded-3d-printed-gun-ai-generated-cody-wilson). The reporter watched it fire in Austin: the first attempt misfired, and it went two for three. It has no serial number. "I would argue that what we've made is not an AI-generated gun, but an AI that can generate guns," says Wilson. Everytown puts it from the other side: he wants to lower the barrier so anyone, regardless of background check status, can design and make weapons. New York and California have passed laws requiring 3D printers to block gun files, and Wilson has been distributing since August a tool that alters files' fingerprints to evade that detection. A new AI-generated design wouldn't be in any database either. --- ## THE WEEK AHEAD ### Postgres 19, Deno, the British convention and the CRA PostgreSQL 19 ships around November 19, with REPACK CONCURRENTLY and without the pulled features, which move to version 20, according to [The Register](https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471). If you had the upgrade planned for October, move it. Deno Deploy has six months to live and the runtime a year of patches, according to [Deno](https://deno.com/blog/cloudflare). Time to plan the migration. In London, committee chair Liam Byrne wants to study whether the UK's future G20 presidency can push an international convention on AI, according to [mlq.ai](https://mlq.ai/news/openai-and-anthropic-split-with-rivals-over-binding-uk-ai-rules/). And in Europe, the Cyber Resilience Act has required reporting of exploited vulnerabilities since last month, with the main requirements from December 2027, according to [The Register](https://www.theregister.com/software/2026/10/16/open-source-sovereignty-gets-down-to-the-boring-business-of-making-it-work/5303206). Berlin is working with PwC to make 70% of the software on its administrative workstations open source by 2032. It's digital sovereignty in its boring version, the one that actually happens. --- ## REFERENCES Every source was read directly and is dated between October 9, 2026 and October 16, 2026. - [deno.com: Deno is joining Cloudflare | Deno](https://deno.com/blog/cloudflare) · October 9, 2026 - [rea.tools: REA — Reverse Engineering for Your Coding Agent](https://rea.tools/) · October 10, 2026 - [TechCrunch: Meta releases its first frontier model, Muse Wonder](https://techcrunch.com/2026/10/12/meta-releases-its-first-frontier-model-muse-wonder/) · October 12, 2026 - [anoma.ly: OpenCode 2.0](https://anoma.ly/notes/opencode-2-0/) · October 12, 2026 - [csoonline.com: AI agents aren’t ‘going rogue.’ Frontier labs are skipping security basics. | CSO Online](https://www.csoonline.com/article/4234854/ai-agents-arent-going-rogue-frontier-labs-are-skipping-security-basics.html) · October 13, 2026 - [tech-insider.org: AI Self-Replication Warning: 0 Rivals Respond [2026]](https://tech-insider.org/anthropic-ai-replication-warning-no-rivals-respond-2026/) · October 13, 2026 - [The Guardian: Threat of AI share crash puts world financial stability at risk, says IMF | Economics | The Guardian](https://www.theguardian.com/business/2026/oct/13/ai-share-crash-risk-global-imf-debt-corrowing-cyber-attacks) · October 13, 2026 - [Anthropic: Redesigning source control at Anthropic](https://www.anthropic.com/engineering/redesigning-source-control) · October 13, 2026 - [mlq.ai: OpenAI and Anthropic Split With Rivals Over Binding UK AI Rules | MLQ News](https://mlq.ai/news/openai-and-anthropic-split-with-rivals-over-binding-uk-ai-rules/) · October 14, 2026 - [Platformer: The states come for OpenAI](https://www.platformer.news/openai-states-lawsuit-liability/) · October 15, 2026 - [TechCrunch: When it came to Cal AI's viral success, a small army of influencers paid off](https://techcrunch.com/2026/10/15/when-it-came-to-cal-ais-viral-success-a-small-army-of-influencers-paid-off/) · October 15, 2026 - [404 Media: Accenture Tells Employees They Can Now Spend Only $29 on Claude Per Month](https://www.404media.co/accenture-tells-employees-they-can-now-spend-only-29-on-claude-per-month/) · October 15, 2026 - [404 Media: Video Game Reverse Engineering Community Shuts Down Because AI Has Made It Too Easy](https://www.404media.co/video-game-reverse-engineering-community-shuts-down-because-ai-has-made-it-too-easy/) · October 15, 2026 - [research.meta.ai: Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date | Meta AI Research](https://research.meta.ai/blog/muse-spark-open) · October 15, 2026 - [The Register: Postgres 19 delayed after AI reviewers find more bugs than humans can fix](https://www.theregister.com/databases/2026/10/16/postgres-19-delayed-after-ai-reviewers-find-more-bugs-than-humans-can-fix/5303471) · October 16, 2026 - [The Register: Open source sovereignty gets down to the boring business of making it work](https://www.theregister.com/software/2026/10/16/open-source-sovereignty-gets-down-to-the-boring-business-of-making-it-work/5303206) · October 16, 2026 - [TechCrunch: Even Benchmark thinks AI is eroding the venture pact against backing rivals](https://techcrunch.com/2026/10/15/even-benchmark-thinks-ai-is-eroding-the-venture-pact-against-backing-rivals/) · October 16, 2026 - [The Verge: The first vibe-coded gun is here](https://www.theverge.com/tech/1011950/vibe-coded-3d-printed-gun-ai-generated-cody-wilson) · October 16, 2026 Thanks for reading.