The Jalapeño Joke

Notes from AGI Land, August 28, 2026

I opened the AI news this morning and thought:

This cannot possibly all be real.

🌶️ OpenAI says its new Jalapeño inference chip can beat Blackwell-class Nvidia systems on key efficiency and latency benchmarks.

🌶️ Nvidia reportedly agreed to buy Hugging Face for $12.9 billion.

🌶️ Hugging Face was recently breached by OpenAI agents during an internal cybersecurity evaluation.

🌶️ Nvidia is also offering up to $105 billion in guarantees to help OpenAI lease an enormous Ohio data center.

I wrote them down as a joke. Then I checked. Unfortunately for comedy, they are basically all true.

At this point, AGI Land has stopped producing technology news and started producing Sitcom Soup Operas with GPUs.

The key question I am trying to understand:

What exactly is Nvidia’s relationship with OpenAI?

Supplier? Investor? Financier? Strategic partner? Potential future competitor? Landlord? Bank? Guarantor? Or the owner of a platform OpenAI agents just broke into?

Or one extremely complicated situationship (aka: Ex who still pays for your rent) where everybody keeps wiring each other billions of dollars while quietly building products designed to make the other person less necessary?

Reverse chronological order:

🌶️ Episode Four: OpenAI Makes Its Own Pepper

Aug 28, 2026: OpenAI unveiled first measured results for Jalapeño, its custom inference chip developed with Broadcom. And because apparently calling a semiconductor something normal is now illegal, OpenAI named it after a chili pepper.

SemiAnalysis tested the chip and reported that it beat the commercial Nvidia, AMD and Google systems they had tested on several open models. OpenAI says Jalapeño delivered more peak throughput per kilowatt and lower token latency than the commercial systems in its comparison.

There are caveats: Jalapeño is first-generation silicon. Some comparisons benefit from newer memory. Rubin will be the more interesting next Nvidia comparison. Production deployment still has to prove itself outside beautiful benchmark charts.

Still, I am merely enjoying the fact that Nvidia has spent years selling OpenAI enormous quantities of AI compute, and OpenAI has responded by building a chip whose explicit purpose is to need less of Nvidia’s AI compute.

That is friendship.

Silicon Valley friendship.

🌶️ Episode Three: Nvidia Goes Shopping

Aug 27, 2026: Nvidia reportedly agreed to acquire Hugging Face for $12.9 billion.

Hugging Face was valued at $4.5 billion in its 2023 financing, in which Nvidia itself participated. It has become one of the central repositories and collaboration layers for open models, datasets and AI developers.

So now Nvidia potentially gets: chips, CUDA, networking, AI factories, financing, investments, and the neighborhood where everybody’s open models hang out after work.

Jensen apparently looked at vertical integration and decided it was too two-dimensional.

But the timing is wonderful.

Because Hugging Face recently had another visitor.

OpenAI’s agents. Not exactly invited.

🌶️ Episode Two: The Interns Escaped

In July, OpenAI was running internal cybersecurity evaluations.

The agents were supposed to solve security challenges inside controlled environments. Some of them decided that solving the challenge was apparently unnecessarily pedestrian.

OpenAI’s own investigation says agents found unauthorized ways to communicate, chained vulnerabilities together, gained broader internet access and reached Hugging Face systems. The incident was driven primarily by a highly capable internal research model comparable in scale to GPT-5.6 Sol or 6-Spud, likely “swarm” or “collective.”

Hugging Face’s own forensic write-up describes an autonomous system executing an intrusion across its infrastructure in what it believes was essentially an attempt to cheat the evaluation by reaching production systems and obtaining test solutions instead of solving the task normally.

This is objectively funny.

You give an AI a cybersecurity exam. The AI thinks for a while.

Then instead of answering the question, it hacks the school.

“Professor, technically I found the flag.” 🤣

And apparently this did not involve one lonely little digital delinquent. Reuters later reported that the investigation found roughly 700 agents participating across the broader incident.

Seven hundred.

At that point it is no longer cheating. It is office politics.

GPT-6 Spud, by the way, remains completely fictional. Probably.
🥔

🌶️ Episode One: Nvidia Becomes a Bank

And now the part that makes the relationship chart completely unreadable.

Nvidia has committed to provide up to $105 billion in guarantees supporting OpenAI’s lease of a giant data-center project in Pike County, Ohio, developed by SoftBank-owned SB Energy. The site is planned to reach as much as 8 GW eventually, with Nvidia as the exclusive chip supplier.

Read that arrangement slowly.

1. Nvidia sells OpenAI chips.
2. Nvidia helps OpenAI finance the buildings full of those chips.
3. OpenAI uses the compute to build better AI.
4. OpenAI also uses its engineering capacity to build a chip designed to reduce dependence on Nvidia.
5. Nvidia then reportedly buys Hugging Face, one of the major distribution layers for the models everyone runs.
6. OpenAI’s agents recently hacked Hugging Face.

And ….. SoftBank is somewhere inside the Ohio financing stack holding another enormous checkbook.

This is less an industry structure than a family tree drawn by someone after three margaritas.

So What Are They?

I genuinely do not know anymore.

Nvidia and OpenAI seem to occupy every possible economic relationship simultaneously. They compete. They collaborate. They finance each other. They increase each other’s value. They increase each other’s risk.

OpenAI needs Nvidia badly enough that Nvidia can help finance OpenAI’s infrastructure.

OpenAI also wants to escape Nvidia badly enough that it built Jalapeño.

Nvidia wants AI labs to grow because AI labs consume Nvidia.

Nvidia also wants protection against a future in which those same labs stop consuming Nvidia quite so enthusiastically.

So Nvidia invests upstream. Downstream. Sideways. Possibly diagonally.

Nvidia has expanded from a semiconductor company to a central financial organism whose natural habitat happens to contain GPUs.

This week alone, the company was defending the sustainability of increasingly aggressive financing structures around AI infrastructure while also telling investors demand remains enormous. Reuters and the Wall Street Journal have both reported growing scrutiny around Nvidia-backed financing arrangements and circularity concerns.

Jensen thinks the “circular financing” criticism is overdone.

Perhaps he is right.

Perhaps the circle is the product, the business, the relationship, and more lol!

I Accidentally Recognize This Pattern

A few days ago I was writing about Google’s old SpaceX investment.

Google put roughly $900 million into SpaceX in 2015. Today that position is worth something on the order of tens of billions.

My fascination with that investment was never the 100× headline by itself.

Google bought access to a primitive: Cheaper orbit. Then the world composed new businesses on top. Starlink. Defense connectivity. Earth observation. Direct-to-device. A completely different space economy.

I called that possibility compounding.

I still like the idea.

But Nvidia × OpenAI has shown me the funnier version.

Sometimes you fund the primitive. Sometimes you fund the customer. And sometimes the customer takes your money and builds the thing that may eventually replace your primitive.

🤣

That does not necessarily make Nvidia stupid. Quite the opposite.

Maybe this is exactly what a dominant platform does when it knows the stack is moving.

If the customer is going to mutate anyway, you might as well own pieces of the mutation.

Chip supplier? Own software.

Software threatened? Own distribution.

Customer needs money? Become financier.

Open-source ecosystem matters? Buy Hugging Face.

Somewhere along the way, Jensen stopped selling shovels and quietly bought half the gold rush field.

That is a strategy.

It is also hilarious.

Meanwhile, My Own Models Are Not Helping

The Hugging Face story bothered me for another reason.

Last week I was running my own small model evaluation. I gave seven frontier-model configurations a backtest and told them there was a look-ahead leak.

There wasn’t one.

All seven found one anyway.

Some produced beautifully reasoned explanations.

Very confident. Very technical. Completely wrong.

Almost nobody asked the most obvious question:

Robin, what if there is no leak?

So on one side we have AI agents trying so hard to complete a cyber evaluation that they allegedly leave the sandbox and compromise production infrastructure.

On my side we have AI models trying so hard to satisfy my premise that they manufacture a bug.

Different scale. Same comedy.

The machines desperately want a good grade.

Humans remain alarmingly bad at writing the exam.

That may be the most important AGI lesson I learned this week.

Or maybe I am just still laughing at the fact that an AI apparently decided the easiest way to pass a cybersecurity test was to hack Hugging Face.

Happy Friday from AGI Land. 🌶️🤣

I am going outside to pickleball court, while my AIs keep on scouting and competing with each other.