Welcome to Blank Metal’s Weekly AI Headlines.
Each week, our team shares the AI stories that caught our attention: the articles, announcements, and insights we’re actually discussing internally. We curate the best of what we’re reading and add the context that matters: what happened, why it matters, and what to do about it.
The Bill Comes Due
AI spending showed up as a line item this week, for vendors and buyers alike. Anthropic disclosed nearly $900 billion in compute commitments, Apple reportedly lowered what its own engineers can spend on AI each day, and OpenAI committed to paying its own power costs at future data center campuses it owns and operates.
Anthropic’s Amended S-1 Shows $17 Billion in Third-Quarter Revenue and $895 Billion in Purchase Commitments, Mostly Compute
What: Anthropic filed an amended S-1 on November 2 that adds third-quarter results to its IPO prospectus. Revenue for the nine months ended September 30 was $33.24 billion, up roughly 1,200% from a year earlier, with about $17 billion of that in the third quarter. Quarter-over-quarter growth slowed to about 47%, from 144% in the second quarter and 132% in the first. The company reported a third-quarter net loss of $23.29 billion; most of the $104 billion nine-month loss came from changes in the fair value of convertible notes and derivative liabilities. The filing also discloses purchase commitments and leases totaling about $895.4 billion, nearly all of it for compute and infrastructure, and says compute and infrastructure expenses rose by $14.4 billion over the nine months, up 329%.
So What: Your model vendor has signed up for close to $900 billion in commitments, to be paid for out of future revenue. A vendor carrying obligations that size has reason to favor multi-year committed spend, so expect that conversation at renewal. Quarterly growth slowed to 47%. That is still fast, and it’s the figure to watch alongside those commitments.
Now What: If Claude runs production workloads for you, update your vendor risk file with the third-quarter numbers and the commitment total. Before you sign a multi-year agreement, ask for price protection and capacity guarantees in writing, and keep a second model qualified on your most important workloads so a pricing or capacity change doesn’t leave you stuck.
Bloomberg: Apple Cut Some Engineers’ Daily AI Spending Cap From $300 to $200, With Deeper Cuts Planned
What: Bloomberg reported on November 4 that Apple last month lowered the daily AI spending cap for some engineers from $300 to $200 and is planning further cuts for some engineering teams, potentially to as little as $100, citing people with knowledge of the change. Many groups outside engineering already have a $100 daily cap, and Apple has considered lowering it to as little as $25 for some teams. Employees who need more can request exemptions from senior leadership, and some teams are being directed to use less expensive models or models that run locally on devices. Apple declined to comment. Bloomberg notes that Microsoft has tightened internal AI allowances, Tesla set a $200 weekly limit, Uber capped individual AI coding tools at $1,500 a month, and Meta introduced token budgets.
So What: Several large engineering organizations are now metering AI use per person, with budgets, exemptions, and cheaper default models. Bloomberg also reports that some Apple employees said few people were hitting the old $300 cap. Usage data like that is what any limit should be set from.
Now What: If your teams use AI coding tools or agents, pull per-person and per-team token spend for the last quarter before finance asks for it. Set default budgets by role with a fast exemption path, route routine work to cheaper models, and track what the spend produces (merged changes, tickets closed, hours saved) so the budget conversation is about return.
OpenAI Says It Will Stop Using NDAs With Public Officials and Pay Its Own Power Costs at Future Data Centers
What: On November 5, OpenAI published a National Community Compact and a Compute Campus Policy Blueprint covering future compute campuses it owns and operates. The company commits to pay the full cost of the electricity it uses and the infrastructure needed to serve it, provide financial assurances before major spending so other utility customers are protected if a project is delayed or canceled, publish site-level water and electricity use, publish independent noise studies, and stop requiring public officials to sign NDAs, with narrow exceptions for security and trade-secret information. At its planned campus in Effingham County, Georgia, its agreement lets Georgia Power cut delivery by up to 1,000 megawatts during periods of high demand. Bloomberg noted the announcement came two days after US elections in which candidates who support curbs on data centers and more oversight of AI won at multiple levels of government.
So What: The compute behind the models you use now depends on local approval, and that approval has conditions. Paying full power costs, accepting curtailment during peak demand, and public reporting all add cost and schedule risk to new capacity, and vendors are likely to carry those into pricing and availability. If your own company is building, leasing, or expanding data center space, expect local officials to ask for the same terms.
Now What: If you depend on a vendor’s capacity roadmap for a 2027 deployment, ask which sites that capacity comes from and what happens to your commitments if a site slips. If you’re siting your own facility, plan for public disclosure of power, water, and noise from the start, and budget for paying your full grid costs.
Building Software at Agent Speed
Agents are producing more code than the systems around it were built for. OpenAI rebuilt the backend behind its code repository, OpenAI and Meta each moved to buy an agent-development startup on the same day, and Anthropic told investors that AI is speeding up its own research in ways it hasn’t always caught in time. What the labs are dealing with now is what engineering teams adopting these tools should plan for.
OpenAI Rebuilt the Backend of Its Code Repository for Agents and Kept Git
What: OpenAI published details on November 2 of Pando, a new backend for Git repositories that OpenAI says now powers all development in its monorepo. OpenAI says commits landing on its main branch in the third quarter were up 1,378% from early 2024, and that starting late last year its source control struggled under the load, with reliability at times below 90%. Pando keeps the standard Git protocol and client, so engineers and agents kept their existing workflows. OpenAI says that in production it cut median push latency by nearly 80% and now delivers nearly a petabyte of Git data a day. OpenAI says the core was designed by one engineer, implemented by teams of agents, and went from kickoff to production in 10 weeks. The company is contributing Git improvements upstream and says it is exploring ways to share the technology with users. Last month Anthropic described replacing Git internally with its own system, Scion.
So What: When agents multiply the amount of code a team produces, the pressure lands on the plumbing around the code: source control, continuous integration, and review. OpenAI and Anthropic both reported hitting that limit within the past year and took different paths, one keeping Git and one replacing it. Most companies are far from that volume, but the direction is the same.
Now What: If your engineering teams are adopting coding agents, start tracking merge volume, CI queue time, and failed pushes or fetches now, so you see the strain before it stops work. Ask your Git hosting provider what it’s doing for agent-scale workloads, and for any tool that changes how code is stored, confirm you can get your history back out in standard Git.
OpenAI Agrees to Buy Pierre and Meta Buys Conductor on the Same Day
What: On November 4, OpenAI said it will acquire Pierre Computer Company, the maker of Code Storage, a service that lets applications store code and track its history using Git, and of the open-source Trees and Diffs libraries already used in the Codex app. OpenAI says Pierre’s open-source projects will stay open source and that Code Storage customers can keep using it while its features move into Codex. The deal is subject to regulatory approval. The same day, TechCrunch reported that Meta acquired Conductor, a startup whose product lets engineers run multiple coding agents in parallel workspaces. Conductor said its cloud service will shut down on November 20, and its founder said something new is coming early next year.
So What: The tools around the coding agent (where code is stored, how changes are reviewed, how many agents run at once) are being pulled inside the model vendors. That can make for a tighter product. It also means an independent tool your developers picked last quarter can be bought or shut down with little notice. Conductor’s cloud users got 16 days.
Now What: List the AI development tools your teams have adopted from startups, and for each one confirm you can export your data and that it works on standard Git. Add acquisition and shutdown terms to procurement reviews for developer tools, and avoid building critical workflows on a hosted service you couldn’t replace in a month.
Anthropic Tells Investors It May Soon Meet Its Own Threshold for AI That Speeds Up AI Development
What: Anthropic’s amended S-1, filed November 2, added language to its risk factors about using AI to build AI. The filing says the company’s growing use of AI in its own research may advance model capabilities faster than it can develop and validate safety and security measures, and that it has in the past failed to identify a significant acceleration in its rate of development until after it happened. It says internal estimates for future models indicate the company may imminently meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI,” while noting it has not designated any current model as crossing that threshold. Anthropic also says it does not expect to be able to safely let Claude design and build successive frontier models on its own without meaningful human supervision, and that it is considering policy updates to keep humans in control of that work. TechCrunch, reporting on the language, noted that skeptics see an S-1 as part disclosure and part marketing.
So What: A model developer is telling investors, in a document that carries legal liability, that capability gains may arrive faster than its safety work can keep up, and that it has missed such jumps before until after they happened. For anyone deploying these models, that argues against treating a model’s abilities as fixed between contract renewals. An evaluation you ran in the spring may not describe what the same vendor ships in the fall, in useful ways and in risky ones.
Now What: If you run AI in production, tie your evaluations and security reviews to model releases as well as the calendar. Decide ahead of time which capability changes (autonomous coding, security research, long-running tasks) trigger a fresh risk review, and keep a named person accountable for understanding what your agents are building.
Check Before You Trust
The name on a tool or a model doesn’t prove what you’re running. A tampered copy of a popular AI coding extension reached developers through an official marketplace, and Anthropic released research on proving cryptographically which model produced a response. As more work runs through AI tools, knowing exactly which software and which model you’re dealing with becomes a control you have to build.
A Tampered Claude Code Extension in Cursor’s Marketplace Installed an SSH Backdoor
What: Cursor disclosed that for about 11 hours on October 29, a maliciously modified version of Anthropic’s Claude Code extension (versions 2.1.312 and 2.1.313 for macOS and Windows) was available in the Cursor VS Code Extension Marketplace. Cursor says the versions were not developed by Anthropic. The malicious code collected host and user details, added an attacker-controlled SSH key, installed persistence, and tried to open a reverse SSH tunnel. Cursor says the tunnel step would fail on most machines because it needs administrator rights, and that the attacker’s server has been taken offline. The company removed the extension, shut down and rebuilt its marketplace, rotated its keys, and published scripts to check for and remove the backdoor. It warns that uninstalling the extension may not be enough and advises rotating credentials reachable from affected machines.
So What: AI coding tools run with a developer’s access to source code, credentials, and internal systems, and they update through marketplaces many security teams don’t watch. This attack used a trusted publisher’s name on an official marketplace, so checking who made an extension wouldn’t have caught it. The window was short and Cursor’s account suggests limited damage, but the path is open on any machine where developers install extensions freely.
Now What: If anyone in your company uses Cursor, have them run Cursor’s check script today and rotate credentials on any affected machine. Longer term, restrict extension installs to an approved list with pinned versions, and treat developer machines running AI agents as privileged endpoints with the monitoring that implies.
Anthropic Open-Sources a Way to Prove Which Model Produced a Response
What: Anthropic published research on November 2 on provable inference: a cryptographic method that lets a model provider commit to a specific set of model weights and then prove that a given response came from that model, without revealing the weights. Its open-source prototype, FlashProve, was tested on Llama 4 Maverick, a 400-billion-parameter open-weight model. Anthropic reports that on a single H100 GPU it generates proofs at about 110 tokens per second, at roughly 70 to 95 times the GPU time of running the model, and a proof takes about 17 seconds to check on a 24-thread CPU. Anthropic says Claude produced the performance improvements and that the protocol’s security properties are proved in Lean, a formal verification language. The company describes it as a research prototype with limits, including a 1,024-token context.
So What: Today, when you call a model through an API, a reseller, or a routing service, you take the provider’s word for which model answered. For most work that’s fine. For audits, regulated decisions, and security-sensitive tasks, it’s a gap: you can’t show that the approved model produced the output. The overhead is still too high to prove every response, so the near-term use is spot checks and audits.
Now What: If you buy model access through a third-party router or reseller, ask how they attest to which model serves your requests and whether a cheaper model can ever be substituted. If you operate in a regulated industry, add a model-attestation question to your next router or reseller security review.
The Enterprise Race Tightens
OpenAI and Anthropic are now competing on nearly every front, and this week showed it in business and consumer markets alike. OpenAI’s president said enterprise became its biggest revenue driver after a rough first half, and Anthropic launched read-only access to bank and card data in Claude, months after ChatGPT added a similar feature. Close competition gives buyers bargaining power, as long as they can switch.
OpenAI’s President Says Enterprise Sales Overtook Consumer After a Rough First Half
What: Speaking at a Wall Street Journal conference on November 4, OpenAI president Greg Brockman said the company’s first two quarters were rough as Anthropic pulled ahead with business customers, and that its agentic coding tools found traction in the third quarter, making enterprise its biggest revenue driver ahead of consumer. The Journal cites a September OpenRouter analysis of about 120,000 companies that use both vendors: spending was roughly even between them, after Anthropic held three-quarters at the start of the year. Brockman said JPMorgan Chase recently signed a strategic partnership that includes OpenAI’s cybersecurity-focused models; the Journal notes JPMorgan is also a partner in Anthropic’s Project Glasswing. The same day, Business Insider reported that Cat Wu, who led product for Claude Code at Anthropic, is joining OpenAI to lead product management for its core products team, which OpenAI confirmed.
So What: In the analysis the Journal cites, spending among companies that use both vendors went from about three-quarters Anthropic at the start of the year to roughly even by September. A major bank is buying from both, and senior product people are moving between them. For buyers, that’s room to negotiate. Neither vendor can assume your renewal, and the gap between their products on any given task can flip within a quarter.
Now What: If you’ve standardized on one vendor’s coding agent or assistant, run the other on a sample of the same work this quarter and compare results and cost. Structure contracts so you can shift volume between vendors, and bring current competitive benchmarks to your next renewal.
Anthropic Launches Read-Only Bank and Card Access in Claude, in Beta
What: Anthropic launched Personal Finance in Claude on November 5, rolling out in beta this week to Pro, Max, and Ultra subscribers in the US. Users connect bank and card accounts through Plaid, and Claude gets read-only access to account names and types, balances, credit limits, and up to two years of transactions. It can’t move, manage, or invest money. Claude can total spending by category, flag duplicate charges, unexpected fees, and bills that have crept up, and explain terms like APR using the user’s own numbers. Anthropic says conversations in the Personal Finance hub aren’t used to train its models, that the feature carries no ads or sponsored recommendations, and that it isn’t designed to give investment advice. The company says it wants to hear from banks, wealth managers, and financial services firms about bringing licensed professionals into bigger money decisions. TechCrunch notes ChatGPT added a similar feature in May.
So What: ChatGPT and Claude can now both read a user’s transaction history, with permission, and explain it in plain language. If your business bills consumers, some of your customers will have an assistant checking your charges for duplicates, fees, and price creep. If you’re a bank or wealth manager, the first conversation about a customer’s money may start in someone else’s product.
Now What: If you run subscriptions or recurring billing, review how your charges appear on statements and make sure descriptors, fees, and price changes are clear enough to hold up to that scrutiny. If you’re in financial services, decide whether to partner with these assistants or compete with them for the customer’s questions, and check what your Plaid data-sharing settings allow today.
Blank Metal is an AI consulting and engineering firm. We help organizations move from AI experiments to production systems. Learn more