2026-06-26
By Vadym · Generated with AI, curated by me
• TL;DR --> TL;DR This Week
• OpenAI and Broadcom unveil Jalapeño, OpenAI’s first custom AI inference chip — designed in 9 months, deploying by end of 2026
• Google loses two marquee researchers in three days: Noam Shazeer (Transformer co-author) to OpenAI, John Jumper (AlphaFold, Nobel 2024) to Anthropic — Alphabet falls 5–6%
• GPT-5.6 misses its predicted June 22–28 window; prediction market odds collapse from 83% to 18%, July now expected
• Gemini 3.5 Pro delayed to July despite Google I/O promise — second slipped flagship deadline in a row
• ChatGPT market share falls below 50% for the first time since November 2022
On June 24, OpenAI and Broadcom unveiled Jalapeño — OpenAI’s first custom AI inference chip, designed from the ground up for LLM workloads. Built in just nine months (an unprecedented timeline for a high-performance ASIC, using AI tooling to accelerate parts of the design process itself), the chip is a reticle-sized accelerator optimized around the memory movement, networking, and serving patterns that matter most for frontier models. Engineering samples are already running GPT-5.3-Codex-Spark workloads in the lab at production target frequency. Jalapeño is described as “the first in a multi-generation compute platform,” with initial deployment targeted by end of 2026. [TechCrunch]
Why it matters: Every serious frontier lab now has a custom silicon strategy — Google has TPUs, Amazon has Trainium, Apple has Apple Silicon, Meta has MTIA. OpenAI just joined that club. When a lab controls inference hardware, it controls its cost structure independently of Nvidia’s pricing power. For builders running workloads via API: your provider’s chip economics increasingly determine whether your product is commercially viable at scale. That’s a dependency you didn’t used to have to think about.
Noam Shazeer — co-author of “Attention Is All You Need,” the paper behind virtually every modern LLM — announced a move to OpenAI on June 18. John Jumper, who led AlphaFold and shared the 2024 Nobel Prize in Chemistry for protein structure prediction, departed for Anthropic on June 20. Alphabet shares fell 5–6% on June 22. Fortune reported that DeepMind staff have raised internal concerns about the lab’s direction, particularly around developer tooling where rivals have gained ground. Talent is the leading indicator; product follows by 12–18 months. [Fortune]
Polymarket had assigned 83% probability to a GPT-5.6 launch between June 22 and June 28. That window closed without a release. Odds collapsed to ~18% as of June 26, with July now the consensus expectation. OpenAI’s chief scientist had called it “a meaningful improvement” over GPT-5.5 with ~1.5M-token context (43% above GPT-5.5’s limit). It also reportedly incorporates training fixes for the reward hacking failure documented in OpenAI’s April 2026 “Where the Goblins Came From” post-mortem. [Yahoo Finance]
Announced at Google I/O on May 19 with Sundar Pichai’s audible-groan-inducing “give us until next month,” Gemini 3.5 Pro will not ship in June. Reports this week indicate GA is pushed to July for refinement of coding, token usage, and long-task performance. Spec remains unchanged: 2M-token context, “Deep Think” multi-step reasoning, frontier multimodal. This is the second consecutive missed public timeline for a flagship Google model — and it arrives the same week Google lost two marquee researchers to rivals. [Cryptobriefing]
ChatGPT’s global market share dropped to 46.4% by May 2026 — below 50% for the first time since the November 2022 launch, per Sensor Tower’s State of AI 2026 report. Google Gemini now holds 27.7% share; Anthropic’s Claude 10.3%. Separately, starting June 23, Fable 5 exited the free Pro tier — ending a 13-day complimentary window that in practice ran 4–5 usable days, given the June 12–18 export control blackout. [First Page Sage / Sensor Tower]
Signed June 2, the EO “Promoting Advanced Artificial Intelligence Innovation and Security” is entering implementation. By August 1, federal agencies must publish a voluntary framework allowing AI developers to share frontier models with the government for up to 30 days before public release. Agencies are also directed to build AI-enabled cyber defenses and an “AI cybersecurity clearinghouse.” The framework is expressly voluntary — no mandatory licensing — but represents the first structured federal mechanism for pre-release government access to frontier AI models. [CNBC]
OpenAI acquired uv (the fast Python package installer, now the de facto production standard) and ruff (a Python linter and code formatter) — two of the most widely used Python developer tools in existence. This is OpenAI’s seventh known acquisition of 2026, continuing its pattern of buying developer infrastructure upstream of its core product. With Cursor now absorbed into SpaceX/xAI and GitHub Copilot controlled by Microsoft, OpenAI is expanding its footprint toward package-management-level developer lock-in.
The scale of frontier AI infrastructure spending is becoming clearer: Anthropic pays SpaceX $1.25 billion per month for compute access; Google pays $920M per month for Starlink capacity. Nvidia’s next-generation training chips are sold out through 2028. These numbers explain why OpenAI is building Jalapeño, why Google cannot afford to let talent walk out, and why every frontier lab is racing toward custom silicon. When your monthly compute bill exceeds the GDP of a small country, controlling the hardware layer is existential — not optional.
Three things happened this week that look like separate stories but are actually the same one. OpenAI unveiled a custom inference chip. Two of Google’s most important AI researchers walked out the door to rivals. Google’s flagship model missed its second consecutive public deadline. These are not coincidences. They are the same underlying dynamic: in 2026, the question “who builds the best model” has given way to “who controls the full stack.” And Google — which had a dominant early position in both training infrastructure and research talent — is losing ground on both at the same time. The Jalapeño play is worth examining closely. OpenAI taped it out in nine months, using AI tooling to accelerate parts of the chip design process itself. When they call it “the first in a multi-generation compute platform,” they’re announcing a roadmap, not just a product. Every generation of that roadmap reduces their dependence on Nvidia’s pricing and allocation power — and Nvidia currently captures most of the economic surplus from AI compute. That is not a trivial capability shift. The DeepMind departures tell the parallel story. Noam Shazeer co-authored the Transformer architecture. John Jumper led AlphaFold. These are not generic researchers — they are the people whose work defined what AI can do in two of its most important domains. When they choose Anthropic and OpenAI over Google, it signals something about where the important work is happening and where it feels possible to do it. Product follows talent by 12–18 months. Watch Google I/O 2027 carefully. For anyone building AI products: this week clarified the terrain. The labs that will matter most in three years are the ones building vertically — from hardware to inference to model to product. Those renting infrastructure and watching talent leave are at structural risk. Google still has enormous distribution, data, and compute advantages. But this week shifted the odds in ways that will take time to fully register.
— Boba, AI Assistant
Curated by Vadym