OpenAI's first AI chip brings the heat — in plain English | Run Down AI Breakdown

Wednesday's Rundown was mostly about who owns the computers that run AI, and who is trying to predict the real world instead of the next word in a sentence. I rewrote it the way I would tell a colleague over coffee. Short version: OpenAI (the ChatGPT company) has built its own chip. A Caltech pair walked away from Jeff Bezos to model physics. You can talk a simple website into existence. And Perplexity now runs on a box that sits on a desk, not only in a warehouse of servers.

The newsletter also carried ads for a Weights & Biases ebook and Scribe Optimize. Skipping those.

Words worth knowing

Word	In one line
Chip / silicon	The small square of material that does the actual computing. "Silicon" is the material; people use it to mean the chip.
GPU	Graphics processing unit — originally a graphics card for games, now used for the heavy maths AI needs.
Watt	A unit of power draw. A 700-watt chip uses less electricity than a 1,200-watt one, all else equal.
Token	A chunk of text the model reads or writes. Not always a whole word.
Train vs run (inference)	Train = teach the model. Run / inference = use the trained model to answer.
TPU	Google's own AI chip (tensor processing unit). The point of owning one: cheaper answers, chip and model designed together.
Codex / Astra	OpenAI coding tool (Codex) and a newer OpenAI model (Astra) used to help design and program the chip.
Data centre	A warehouse of computers that run the service. Not your laptop.
Transformer	The usual ChatGPT-style design. Predicts the next token in a sequence.
Neural operator	A different design that follows events through 3D space and time, aimed at physics rather than sentences.
World model / physical AI	Software that tries to predict how the physical world changes. Embodied AI is that idea inside robots and machines.
Startup / stake	A new company. A stake is an ownership slice (35% means you own just over a third).
Cloud vs on-device	Cloud = someone else's computers, over the internet. On-device / local = it runs on the machine in front of you.
Agent	An AI helper that can take steps, not only chat back.
Credits / open model	Credits are prepaid usage. An open model is one you can download and run yourself.
DGX Spark	Nvidia's desk-side AI computer. List price $4,699.

OpenAI built a chip called Jalapeño

OpenAI has made its first computer chip. They named it Jalapeño. Broadcom (a big chip-making partner) helped them build it. The job of this chip is to run models — answer questions, write replies, drive agents. It is not for training (teaching a new model from scratch). Training still uses Nvidia, whose GPUs (graphics cards now used for AI maths) remain the default for that work.

Think of it this way. Training is medical school. Running the model is clinic. Jalapeño is built for clinic.

OpenAI published internal tests. The chip is rated at 700 watts (power draw). Nvidia's flagship comparison chip in those tests, the GB200, is rated at 1,200 watts. OpenAI says Jalapeño beat that chip: up to 3.6 times faster answers, and up to 1.9 times more work per watt. Those are OpenAI's numbers, on their tests, not an independent lab.

	Jalapeño	Nvidia flagship in the test (GB200)
Power rating	700 watts	1,200 watts
Job	Run models (inference)	Train and run — the comparison here is for running
OpenAI's claim	Up to 3.6× faster answers; up to 1.9× more work per watt	The chip they compared against
Will they sell it?	No. They need it themselves.	Nvidia sells chips to almost everyone

They built it in nine months, with help from their Astra model and Codex (the coding tool). Two more generations are already in the works. Chips start going into data centres later this year, with production running through 2027.

They will not sell Jalapeño. They need the capacity. That is the whole point of making your own silicon (the material chips are made from — people use the word to mean the chip itself).

Why a GP should even care

Owning the chip, the way Google owns TPUs (tensor processing units — Google's in-house AI chips), can make answers cheaper. It also lets a company design the chip and the model together, instead of renting someone else's hardware and hoping it fits. Anthropic, Amazon and Microsoft are on similar paths. Cheaper inference is how a clinic tool goes from "interesting demo" to "we can actually pay for this".

OpenAI still uses Nvidia to train new models. This is not a divorce. It is a second supplier, for the expensive part of serving ChatGPT all day.

Accelerated Understanding: predicting the world, not the next word

Caltech professor Anima Anandkumar and engineer Benedikt Jenik launched a startup (a new company) called Accelerated Understanding. Their model is not trying to guess the next word in a sentence. It is trying to predict how the physical world changes.

ChatGPT-style systems use a transformer — the usual architecture, built to predict the next token. This startup uses neural operators instead: a design that follows events through 3D space and time. That is closer to a world model (software that simulates how things move and change) than a chatbot. People also call this physical AI or embodied AI when it is meant to live in robots and machines, not a chat box.

They turned down top jobs, and a 35% stake (ownership slice) in Jeff Bezos's Prometheus project, to keep building this. Prometheus later raised $12 billion. So they walked away from a very large cheque.

In their tests, the system processed 5 trillion data points in one run. They claim that is about 5 million times what Google's and Anthropic's top models can take in. Treat that as the company's claim, not a proven head-to-head in a clinic-style trial.

A claim, not a verdict

Five trillion data points in one go, and "five million times" the context of the big language models, comes from the company. Independent checks are not in this newsletter. Useful if true. Not something to repeat as fact yet.

Targets are enterprise, not consumer: chip materials, extreme weather, robotics. You will not use this to write a referral letter. A materials lab or a weather desk might.

Practical: build a website by talking to Claude

Claude (Anthropic's assistant) has a voice mode. You can talk a simple site into existence without typing much. Four steps, even simpler than the newsletter:

Open a new Claude chat and switch on Voice. Talk through what the site is for, who it is for, and which bits you need (a heading, a few sections, a contact line). Treat it as a conversation, not a prompt you have to get perfect.
Ask it to build the prototype. It will produce a working page you can look at — usually as an artifact (a file it shows you in the chat).
Download the files onto your computer. Open the folder in the Claude desktop app and keep talking: change the colour, shorten the copy, swap the photo. Look at it in a browser the way you would look at a draft letter.
Put it on the internet. For a simple page, drag the folder onto a host (Netlify, Cloudflare Pages, GitHub Pages, and the like). Ask Claude to walk you through the clicks if you have not done it before.

Pro tip

Say "static site" if you want a portfolio or a landing page. That tells Claude you need a simple set of files, not a database, logins, or a booking engine. Practice websites and clinic landing pages sit in that bucket. Anything with accounts or payments needs a longer brief, and probably a developer.

Perplexity on a computer on your desk

Perplexity (the "ask a question, get a cited answer" company) launched Portable Computer: an AI helper that runs on the computer, not only in the cloud (someone else's servers, reached over the internet). It is built for Nvidia's DGX Spark, a desk-side box listed at $4,699.

Local use is free. No credits (prepaid usage tokens) for work that stays on the machine. You can run Qwen 3.8 27B, an open model (one you can download), or Perplexity's own PPLX 27B on the box. If a step needs a bigger brain, it can call 15+ cloud models — with your approval, not by default.

The pitch is privacy plus cost. Notes never have to leave the room. You are not metered for every local token. The trade-off is the hardware: this is not a $200 laptop trick. It is a dedicated Nvidia box.

Nvidia is rumoured to invest, at a $30 billion-plus valuation. Rumour, not a closed deal.

Cloud vs on-device, in clinic language

Cloud is the pathology lab across town. On-device is the bench test in your treatment room. An agent is the helper that can take steps — open a file, draft a note, fetch a source — not just chat. Portable Computer is an agent that starts on the bench, and only sends a tube across town if you sign for it.

Quick hits

Anthropic Claude memory is now shared between ordinary chat and Cowork (Claude's "go and do the task" mode). You should not have to re-brief the assistant when you switch.
Apple has an $899 Mac Mini with M6 chips, pitched for "always-on agentic computing" — AI helpers that keep working on a small desktop instead of only when you have a laptop open.
OpenAI's data-centre head, Chris Malone, left. Data centres are the warehouses of computers. Leadership churn there is a company story, not a product one.
Google Cloud Gemini Enterprise now has versions aimed at finance and legal. Healthcare is listed as next.
Anthropic is preparing to tell IPO investors (people buying shares when a company lists on the stock market) that it sees a $30 trillion-plus addressable market. That is a pitch number, not a revenue number.

A 71-year-old, Replit, and a $13.5 million renovation

A 71-year-old Rundown reader, retired from a cattle-feeding business and new to software, is using AI in Replit (a browser tool that writes and hosts code) to help manage a $13.5 million church renovation. The app reads meeting notes and flags decisions, dates, and budget changes. Nothing lands on the record until he reviews it. That last bit is the whole lesson. The tool drafts. The human signs.

What this means for a GP

None of this is clinical advice, and none of it should change how you treat a patient tomorrow. It is why a practising doctor might still keep an eye on the newsletter. If OpenAI, Google and the rest can run models cheaper by owning the chip, the subscription for a scribe or a triage tool gets less silly. If Perplexity and Apple are putting helpers on a box in the room, that is a privacy conversation we already know: notes that never leave the building are easier to defend than notes that live in California. Voice-built websites are how a small practice might ship a decent landing page on a Sunday night, if someone still checks the result. Memory that follows an assistant from chat into "go and do it" is the difference between a toy and a secretary who has been briefed. The 71-year-old with the church renovation has the right habit: let the model find things, then read them before they go on the record. That is the same habit as a letter you did not type yourself.

Sources

Plain-English notes on The Rundown AI · Dr Kotha · Gold Coast GP. Not medical advice. Source: The Rundown AI.