OpenAI unveils its first chip; Qualcomm takes on Nvidia

OpenAI and Broadcom unveiled Jalapeño, OpenAI's first custom-built processor — a reticle-sized accelerator designed specifically for running large language models rather than training them. Broadcom will manufacture it, and OpenAI will use it to serve ChatGPT and its other products starting late this year. Early testing points to roughly 50 percent lower cost than today's top AI GPUs and substantially better performance per watt. OpenAI used its own models to speed parts of the design, taping out in about nine months — among the fastest chip-development cycles on record, and a direct bid to loosen its dependence on Nvidia.

Qualcomm made its most serious move yet beyond smartphones, launching a line of data-center processors — including the Dragonfly C1000 CPU, already signed by Microsoft and Meta — plus new high-bandwidth memory aimed at cutting the cost of running AI. The same day, it agreed to buy software startup Modular for about $3.9 billion in stock. Modular's software lets programs written for Nvidia's CUDA platform run on other chips, so the deal hands Qualcomm a neutral software layer meant to pry customers away from Nvidia's deepest lock-in.

Immunologist Derya Unutmaz at The Jackson Laboratory handed GPT-5 Pro an unpublished dataset his team had puzzled over since 2022 — how glucose shapes the way T cells develop and specialize. Within minutes the model pinpointed the likely mechanism, disrupted N-linked glycosylation, and proposed follow-up experiments that later confirmed it. OpenAI published the case on June 23 as evidence that frontier models can now generate original, testable scientific hypotheses, not just summarize what is already known.

Meta launched its first self-branded smart glasses — the Fury, the Adventurer, and a Starfire edition designed with Kylie Jenner — stepping out from under the Ray-Ban label it has shared with EssilorLuxottica. The two everyday models start at $299 and the Kylie pair at $399, all running Meta AI for hands-free questions, translation, and capture; the Kylie model even answers in a voice meant to sound like her. The move signals Meta wants the AI-glasses category to be its own brand, not a fashion-house collaboration.

Mistral released OCR 4, a document-intelligence model that extracts structured text, tables, and page layout from scanned files across 170 languages — and runs in a single container on a company's own servers. Independent annotators preferred it to every leading OCR system tested, with an average 72 percent win rate, and it leads the OlmOCRBench benchmark. At $4 per 1,000 pages, it folds a fragmented tooling category into a single enterprise building block for retrieval and search pipelines, with no data leaving the customer's infrastructure.

Assort Health raised a $120 million Series C led by Menlo Ventures at a $1.2 billion valuation, becoming a unicorn on the strength of AI voice agents that run the patient journey — scheduling, intake, referrals, medication refills, and payments. The company says it has now handled more than 190 million patient interactions and grown revenue 20-fold over the past 15 months. The round is a marker of how quickly health systems are handing routine front-desk work to AI rather than hiring for it.

Tokyo lab Sakana AI launched Fugu, an orchestration model that behaves like one AI but routes each request across a swappable pool of frontier models — GPT-5.5, Claude Opus, and Gemini 3.1 Pro — handling selection, delegation, and synthesis behind a single OpenAI-compatible endpoint. A heavier Fugu Ultra tier targets harder, multi-step problems. The pitch is that the best answers come not from any one model but from coordinating many, a direct challenge to the single-model assumption behind ChatGPT and Claude.