OpenAI's New Chip Beats Nvidia. It Still Can't Quit Nvidia.

Today's stories all point at the same question: who actually owns the AI supply chain. OpenAI built a chip that beats Nvidia at its own game, Anthropic is telling IPO investors its market is roughly the size of the entire US economy, and a three-year-old chip startup just got valued at $21 billion. Let's get into it.

Skydive agents take on work for you, using the same tools your team already uses.

Talk to your agent on the web, Slack, email, iMessage, or right from your terminal. Wherever you pick up the conversation, your agent keeps the context.

Put agents to work across customer support, sales, marketing, engineering, ops, and more.

Give your agent a job. They’ll take it from there.

Make your first agent

What happened: OpenAI released benchmark results this week for its first custom chip, called Jalapeño, and independently verified numbers show it beating Nvidia's flagship GB300 systems on real workloads, according to Tom's Hardware. The chip, built with Broadcom, did up to 1.9 times more work per kilowatt and cut latency by as much as 3.6 times compared to Nvidia's best inference hardware, while using roughly half the power. OpenAI plans to run Jalapeño in its own data centers later this year, with two more chip generations already in development. One catch: Jalapeño only handles inference, meaning running already-trained models, not training new ones, so Nvidia keeps that part of the business for now.

Why it matters to you: This is not about your business buying a chip. It's about the biggest AI company on earth telling you, with its own money, that running AI tools is about to get cheaper and faster. That usually shows up downstream as your AI subscriptions getting more capable without getting more expensive, at least for a while.

What to do about it: Nothing to do today. Just don't panic if a tool you use suddenly gets faster or cheaper. That's the AI chip war working in your favor, for once.

What happened: Anthropic is preparing paperwork for investors that pegs its total addressable market above $30 trillion, roughly the size of the entire US economy, according to Yahoo Finance, citing the Wall Street Journal. That number would beat SpaceX's own $28.5 trillion claim from its IPO filing earlier this year. Anthropic's actual 2028 revenue projection, still enormous, sits at $190 to $200 billion, a small slice of that $30 trillion story. The company is reportedly aiming for a valuation around $2 trillion when it goes public.

Why it matters to you: A $30 trillion market claim is a sales pitch, not a forecast, and it tells you these companies expect investors to keep funding AI at almost any price. That flood of capital is part of why AI tools keep getting cheaper and more capable for you right now. It's also exactly the kind of number that turns into a very loud bubble conversation the moment growth slows down even a little.

What happened: OpenAI rolled out a feature that lets ChatGPT Work log into websites on your behalf and finish tasks like filling out forms, booking things, or searching internal tools, according to OpenAI's own announcement. You still enter your own password, so OpenAI says ChatGPT never sees or stores it. The tool can also work inside your existing Chrome tabs and cookies if you want it to use accounts you're already logged into.

Why it matters to you: This is the first real version of an AI that can go do your errands online instead of just answering questions about them. If you've ever wished you could hand off logging into a vendor portal or a benefits site to someone else, this is aimed straight at that.

What to do about it: Before you let any AI agent log into a business account, check what permissions it's asking for and turn on two-factor authentication anywhere it might have access.

What happened: Anthropic merged the memory systems behind Claude's chat app and Claude Cowork, so anything you tell Claude in one place is now remembered in the other, according to TechCrunch. The feature is on by default, and Claude can start saving useful details about your business or preferences while you're still mid-conversation, not just afterward. Everything remembered lives in an editable list under Claude's memory settings, so you can see, edit, or delete it anytime.

Why it matters to you: If you use Claude for your business, you'll stop having to re-explain who you are, what your business does, and how you like things done every single time you open a new chat. That's a real time saver, but it also means Claude is quietly building a running file on your business, which is worth knowing exists.

What happened: AI chip startup Etched more than doubled its valuation to $21 billion in under a month, raising $700 million in a round led by Jane Street, according to Reuters. The company, founded by three Harvard dropouts all around age 24, builds specialized chips meant to run AI models faster and cheaper than general-purpose Nvidia hardware. Jane Street is also Etched's first customer and has already deployed the chips in its own trading systems.

Why it matters to you: This is another sign that serious money is betting Nvidia won't be the only option for AI hardware much longer. More competition among chip makers is generally good news for anyone paying for AI tools downstream, since it puts pressure on prices across the whole supply chain.

What happened: ByteDance signed a copyright agreement with Hollywood's Motion Picture Association to add safeguards to its Seedance and Seedream AI video and image tools, the same models that power features inside TikTok, CapCut, and Dreamina, according to NBC News. The deal follows a cease and desist letter the MPA sent in February after Disney and other studios said the tools could generate copyrighted characters and celebrity likenesses without permission. ByteDance says newer versions of its models include stronger protections, though neither side has detailed exactly what those guardrails are, and studios still aren't being paid for any of their content that may have trained the models in the first place.

Why it matters to you: If your business uses AI video or image tools for marketing, this is a preview of the legal fights coming for the whole industry over what these tools were trained on and what they can generate. Assume any AI-generated content featuring a recognizable character, logo, or celebrity likeness is a legal risk for your brand, pact or no pact.

What happened: A new Stanford study using payroll data from ADP found that employment for workers aged 22 to 25 in AI-exposed occupations is now 19 percent lower than in less-exposed jobs, up from a 13 percent gap just a year ago, according to Computerworld. The drop shows up almost entirely through reduced hiring of young workers, not layoffs of people already in those roles. Older, more experienced workers in the same AI-exposed jobs haven't seen comparable declines, and jobs where AI complements rather than replaces tasks show a much more mixed picture.

Why it matters to you: If you hire entry-level or junior staff, this is the clearest data yet that AI is changing who gets that first job, not by replacing people already working, but by making companies hire fewer new ones. If you're relying on AI to absorb junior-level tasks in your own business, this study is basically describing your hiring decisions in aggregate.

What to do about it: Before you let AI absorb every entry-level task in your business, think about where your next experienced hire is actually going to come from in five years.

What happened: Apple introduced new M6 and M5 Ultra chips built for running AI models locally, launching in a refreshed Mac mini and Mac Studio on September 22, according to Apple's own announcement. The M6 Mac mini starts at $899 and claims up to 4 times the AI performance of the previous M4 chip. The M5 Ultra Mac Studio can be configured with up to 512GB of memory, enough to run very large AI models entirely on the machine instead of over the internet.

Why it matters to you: If you've been nervous about sending sensitive business data to a cloud AI tool, this is Apple explicitly building hardware so you, or your IT person, can run powerful AI models locally instead. It's not cheap, but it's a real option now, not just a theory.

Here's the throughline today: everyone in AI is racing to control the actual plumbing, the chips, the compute, the hardware, because that's where the real money sits long after the hype cycle around any one chatbot fades. OpenAI built a chip to get out from under Nvidia. Anthropic is telling investors its market is basically the whole economy. A three-year-old startup got valued at $21 billion for making chips almost nobody outside one trading firm has actually used yet. Meanwhile the actual humans this is all supposed to serve, especially the ones under 25 trying to get their first job, are already feeling it in ways a chip benchmark will never capture. Watch the plumbing, not the demos. That's the whole rollercoaster in one sentence.

If this issue saved you from reading eight different AI newsletters yourself, forward it to one person who'd get something out of it. New here? Subscribe below so tomorrow's issue lands in your own inbox too.

Talk tomorrow,
Mark Shilensky

Follow me personally on Social Media:
Facebook: https://www.facebook.com/MarkShilenskyPage
Instagram: https://www.instagram.com/markashilensky/
Threads: https://www.threads.com/@markashilensky
LinkedIn: https://www.linkedin.com/in/shilensky/
X: https://x.com/markshilensky
TikTok: https://www.tiktok.com/@mark.shilensky
BlueSky: https://bsky.app/profile/markshilensky.bsky.social