A Spicy Chip, Dirt-cheap AIs, & RSI

TLDR;

Welcome back! This week is exciting. The news we’re about to cover is genuinely incredible. From new chips and models with genuinely new research, to what could mark the first signs of real evidence that we might be in the era of recursive self-improvement, or RSI.

Enjoy!

THEWHITEBOX

OpenAI’s Jalapeño marvels everyone

I’ve written a complete rundown of this chip in my Medium account, in case you want to read the long-form version of my coverage of this OpenAI chip.

OpenAI has presented results for its new inference chip, Jalapeño, and they're pretty extraordinary: they went from idea to reality in under two years while beating companies that have been playing this game for decades.

As you can see above, it even delivers better inference performance than NVIDIA’s new chip, Rubin, with great throughput (tokens/watt) and the same interactivity (tokens/user), even though NVIDIA’s models were running an MTP (multi-token prediction) setting.

In other words, although Rubin generates more tokens per prediction (which will be available on Jalapeño eventually too), Jalapeño still performs better.

This is the first time we’ve seen NVIDIA displaced from first place, and it didn’t come from AMD or Google. It’s also much faster with a single user, reaching 700 tokens/second, while Rubin maxes out at 400.

How did they pull this off?

The answer is a mixture of incredible engineering and many decisions that can only be explained in one way:

OpenAI has just proven that competitive advantages in chip design are less about decades-old experience and more about who has the best AI models, which play a vital role in design.

TheWhiteBox’s takeaway:

As I explain in the article, this will be portrayed as an NVIDIA killer, but only by those who misunderstand NVIDIA’s key advantage: supply chains and balance sheets.

Let’s not forget that OpenAI needs to tap into the same manufacturing supply chains as NVIDIA, mainly the same bottleneck: TSMC (the guys that actually manufacture the chips).

It’s irrelevant that you can produce this at scale if NVIDIA, AMD, and Google secure 80% of TSMC’s supply and DRAM (we’ll later see the extent to which NVIDIA dominates supply chains below).

OpenAI’s achievement is massive. But now comes the real hard part: real life, the constraints you can’t fight against just using your brain.

MODELS

Qwen 3.8-Flash Next, first production-grade Engram model

I’m very excited about this one. Alibaba has released Qwen3.8-Flash-Next, an open-weight model priced at just $0.16 per million input tokens and $0.47 per million output tokens.

The main reason it is so cheap is that it’s smaller than comparable models and can therefore produce far more tokens. The more tokens each server produces per second, the more it can spread its hardware and electricity costs. Crucially, although Qwen3.8-Flash-Next has 125 billion parameters, it activates only six billion per token.

For comparison, Qwen3.7-Plus activates 17 billion, and DeepSeek V4 Flash activates 13 billion. Qwen’s new model therefore performs much less computation per token, allowing each accelerator to generate considerably more of…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.