OpenAI Announces Its First Custom Chip "Jalapeño" — Co-Developed with Broadcom

On June 24, OpenAI announced "**Jalapeño**", an LLM inference-dedicated chip co-developed with Broadcom. It is the company's first in-house silicon, designed specifically for the inference workloads of large language models.

Jalapeño is optimized for the unique demands of OpenAI's inference systems, aiming to deliver improvements in performance, power efficiency, and scalability. According to a report by The Decoder, the company plans to bring it into full operation by late 2026.

This move sits squarely in the hyperscaler custom-silicon trend, following Google (TPU) and Amazon (Trainium). Given how heavily OpenAI has depended on NVIDIA GPUs, this is a significant step forward in terms of cutting inference costs and diversifying supply.

"Tokenpocalypse" — A Growing Number of Companies Are Struggling with AI Spending

A phenomenon that 404media has dubbed the "**Tokenpocalypse**" is spreading: companies have begun scrambling to cut their spending on AI services.

InformationWeek reports a similar trend, noting that CIOs are starting to map out "exit strategies" for their AI projects. In practice, this means cancelling AI projects with no visible ROI (return on investment) and optimizing token usage.

Rather than the end of a temporary boom, this should be read as a transitional phenomenon as AI adoption shifts from the "experimentation phase" to the "profitability phase." Budgets that were once poured in with early enthusiasm are increasingly becoming the target of pragmatic cost management.

Swiss Federal Supreme Court Evaluates Adopting the Censorship-Evasion Model "Heretic"

An interesting development has been reported out of Switzerland: the **Swiss Federal Supreme Court** is reportedly considering adopting "Heretic", an abliterated (censorship-removed) model.

According to information shared in the LocalLLaMA community on Reddit, the court had been struggling with "excessive refusals" from commercial LLMs. The models were frequently rejecting legitimate business questions as inappropriate content.

The fact that a government institution is looking to use a decensored open-source model for its own operations is a symbolic development — evidence that the balance between AI model safety and practicality is now being debated at the working level.

Local LLM Optimization: A 20× Speedup Takes GLM5.2 from 2.5 tok/s to 50+ tok/s

A case of speeding up the GLM5.2 model by **roughly 20×** has been reported in the LocalLLaMA community.

On a GH200 system (equipped with two Hopper H100 96GB GPUs), the developer combined model hacks with optimizations to push the initial speed of about 2.5 tok/s to over 50 tok/s. Specific techniques included inference engine tuning, memory allocation optimization, and kernel-level adjustments.

It's a textbook example of how dramatically performance can change on identical hardware depending on software optimization — an important benchmark demonstrating the practicality of local LLMs.

AI Designs Radio Chips Beyond Human Imagination

According to IEEE Spectrum, AI is producing **radio chip designs** that human designers would never think of.

Conventionally, radio frequency (RF) chip design has been carried out by human engineers drawing on experience and intuition, but AI generates structures that are difficult for humans to interpret — and they reportedly deliver superior performance in practice. This shows AI's range of applications expanding from the software domain into physical hardware design.

Wrap-Up

What emerges from today's news is a picture of the AI industry simultaneously facing two currents: "bringing hardware in-house" and "the reality of costs." OpenAI's custom chip symbolizes long-term infrastructure investment, while the Tokenpocalypse is a manifestation of short-term profitability pressure. How these two vectors find their equilibrium looks set to be a key point to watch over the coming quarters.

---