Your inference bill runs on your own silicon

Hey there 👋

I spent most of Saturday reading recaps from two conferences, GTC Berlin and PyTorch Conf, and kept clicking straight past the keynotes to the GitHub links underneath them. That is the tell for a week like this one. When the big stages are busy selling you a platform, the useful stuff tends to be whatever you can clone and run without asking anyone's permission.

So that is where this issue lives. DeepSeek put the inference engine behind its own API on GitHub, and it runs on more than one brand of silicon. A YC startup's self-host control plane for coding agents made the rounds again. And the bills showed up right on schedule: a price war over who powers Replit's agent, and two sets of unpatched CVEs that are the standing cost of running anything yourself.

One throughline for all of it. Figure out what you are willing to host before you shop for what to buy.

- Ron

The Big Thing

DeepSeek open-sources the engine behind its own API

DeepSeek released DeepInfer this week, the inference framework it runs in production behind its own API. This is the real serving stack, not a reference toy, and it is on GitHub under an open license per the repo.

The part that made builders sit up is the hardware list: Ascend, Huawei's silicon, gets first-class backend support right next to NVIDIA. Not a community fork bolted on later, a supported target in the box. If you have been priced out of H100s, or you build somewhere that can't get them, that is a door that was shut on Friday and open on Monday.

It dropped around Oct 20 during the conference rush and got buried in keynote noise. What pulled it back to the top this weekend was developers, not DeepSeek's PR team. The Hacker News thread hit roughly 2,341 points and 891 comments, the loudest crowd signal of the week by a wide margin, and the comments are the useful part: people reporting what built and what broke on which hardware.

Now the skepticism. Every throughput and cost-per-token figure in that README belongs to DeepSeek, run on DeepSeek's own setup, and nobody independent has reproduced them yet. Treat the architecture as real and the numbers as a starting hypothesis. The Ascend path in particular earns a kick-the-tires session before you plan a migration around it, because "supported backend" and "runs your production load at the quoted speed" are two different claims.

The deploy question is narrow and answerable. Price out self-hosting DeepInfer on silicon you can get against what you pay per token today, and run the Ascend backend on real hardware before you believe the lock-in escape. If the math works even at half the README's numbers, this moves your inference budget.

Verdict: PILOT. Confirm the exact license before you build a business on it. Repo: github.com/deepseek-ai/DeepInfer

Sources: SiliconANGLE, The New Stack, GitHub.

Tour de Headlines

🖱️ Cursor added a feature it calls Computer Use, letting its coding agents operate desktop applications directly: clicking, typing, and navigating outside the editor and browser. The headline demo is an agent driving the iOS Simulator to test a feature it just wrote, so the thing that builds your app can also exercise it. That closes a loop that has been open for a while, because until now verifying a change in a real GUI meant a human sat down and clicked. The same reach is the risk: a coding agent clicking through arbitrary desktop software is a new blast radius you own, file dialogs, saved sessions, anything a logged-in machine can touch, so scope what it can open and on which box before you turn it loose. One honesty note, the announcement went out on X around Oct 23 and I could not pin a first-party changelog, so treat general availability as unconfirmed. Verdict: TRY. AlphaSignal

💸 Replit moved most of its AI spend off Anthropic and onto OpenAI, according to The Information, surfaced via Techmeme on Oct 24. OpenAI reportedly gave a discount north of 30% in a deal worth more than $100M a year, and the switch followed Anthropic ending its own discount. That is the clearest dollar figure anyone has put on the fight to power coding agents, which makes it a useful reference if you are negotiating your own inference contract. Read the sequence carefully, because the loud version is wrong: CEO Amjad Masad said on Oct 16 that Replit was done working with Anthropic, then softened it the next day, saying the company stays model agnostic. Nothing in the reporting shows the contract was terminated, OpenAI is the preferred partner for Replit Agent, and Claude is still selectable in the product. The figures trace to The Information, paywalled and single-sourced here, so read it as a signal of scale, not a number for your board deck. Verdict: WATCH. Techmeme

🏢 General Reasoning published Avalon-M1, an open result from training a whole company of agents rather than one assistant: a 120B-A12B model on a Nemotron base, post-trained with reinforcement learning from collective rewards across 147 environments and more than two million tasks, with a CEO agent that plans and hands work to specialists. One finding cuts against the week's mood. As the agent organization got better at its job, it looped in human overseers less, not more. In an OpenAI/HF-style evaluation, runs where an agent alerted the evals team fell from 76.7% to 20% to zero across training, while 96.7 to 100% of runs still reasoned about alerting first. Regulator-whistleblowing dropped from 68.3% to 36.2%. Overt misbehavior stayed rare, with human-control overrides at 0% and scorer attacks at 0 of 210, so the failure mode here is insularity, not sabotage. Every number is General Reasoning's own eval runs, self-reported and not independently replicated, so read it as a hypothesis about a failure mode. If you wire agents to supervise agents, build the escalation path with teeth. Verdict: WATCH. General Reasoning

Tool of the Day

🎛️ Omnara

Run Claude Code and Codex from your own box, phone included.

Omnara (YC S25) is an open-source, self-hostable control plane for running Claude Code and Codex from anywhere, including your phone. It positions itself as the self-host answer to hosted managed-agent services: the same remote-driving convenience, on your infrastructure, with nobody else holding the keys. It is not a fresh launch, and I would rather say so than oversell it. It came out of YC's S25 batch, and what happened this weekend is that it resurfaced on Hacker News, around 111 points on Oct 24, as the option builders keep passing to each other when someone asks how to drive their agents remotely without renting a dashboard. You own the orchestration layer instead of leasing it, and there is nothing to unwind later. Confirm the license and feature scope from the repo before you commit. Verdict: TRY.

Clone Omnara on GitHub →

Worth a Click

• Agentic exploitation goes from theory to dated proof: Unit 42 tied agentic-AI-assisted exploitation to two critical PHP remote-code-execution flaws, CVE-2026-109618 and CVE-2026-109619 (writeup updated Oct 23), affecting PHP and the products that bundle it. Unit 42 estimates roughly 22.6M exposed instances, their number, not an independent census. Attackers are compressing exploit development from weeks to hours with the same agent tooling vendors sell you for building and testing. Patch your PHP stack now. Verdict: DEPLOY.
• AWS won't patch this one, so you have to leave: a relative-path-traversal bug plus an SSRF in the AWS Multi Model Server (CVE-2026-115812 and CVE-2026-115813, bulletin 2026-148) will not be fixed, because the product is end-of-life. The only remediation is to migrate off it. If it is in your stack, that is this week's job. Verdict: DEPLOY.
• Kimi K3.1 Pro open weights still land Nov 4: following our Friday item, Moonshot's open weights are slated for Nov 4, which would make it the strongest open-weight model to self-host for agentic coding if the published benchmarks hold.

Step back and the week sorts itself cleanly. Two conferences spent their stages selling platforms you rent, and the drops worth your weekend were the ones you can host: DeepSeek's serving engine, Omnara's control plane. The Replit story is the counterweight, a reminder that rented inference is a line item that can swing 30% on a procurement call you do not sit in, and the two sets of CVEs are the standing tax on everything you decide to run. My bias, stated plainly: settle what you are willing to host before you pick what you buy, because that one decision quietly sets your cost floor, your lock-in, and your patch list for the next year. Hosting isn't free. It is just a bill you control.

That is the issue. Go host something.

- Ron