Hardware wars, talent raids, and IPO signals define the final week of June. OpenAI's first custom silicon shifts the infrastructure calculus, Anthropic poaches Google's brightest, and the memory sector bets big on AI demand.
OpenAI and Broadcom Unveil Jalapeño — First Custom AI Chip
OpenAI and Broadcom jointly introduced Jalapeño, OpenAI's first custom-designed AI accelerator chip, on June 24. The ASIC went from design to tapeout in just nine months — a record for high-performance semiconductor development — accelerated by OpenAI's own LLMs used in the chip optimization loop. Early test samples are already running GPT-5.3-Codex-Spark and other frontier models, with per-watt performance significantly exceeding current market leaders. Deployment begins in late 2026 across OpenAI's gigawatt-scale data centers.
The chip was co-designed with Broadcom's ASIC team and manufactured in partnership with Jabil for board-level integration. The nine-month timeline was partly enabled by AI-assisted design tools — the engineering team used LLMs to accelerate verification, floorplanning, and thermal simulation, effectively building a chip with the help of the very models it was designed to run.
— OpenAI · Broadcom · The Decoder
🔗 OpenAI · Broadcom · The Decoder
Nobel Laureate John Jumper Leaves DeepMind for Anthropic
John Jumper, 2024 Nobel Prize winner in Chemistry and co-creator of AlphaFold, announced he is leaving Google DeepMind after nearly nine years to join Anthropic. AlphaFold has predicted over 200 million protein structures — one of the most consequential scientific resources ever created. Jumper's move comes in the same week as Noam Shazeer's departure to OpenAI, wiping over $225 billion from Alphabet's market capitalization in a single trading session.
Anthropic is hosting a science event on June 30 — the community expects this to feature Jumper's first public appearance at the lab and signal Anthropic's expansion into AI-for-science. Alphabet holds a 14% stake in Anthropic, creating an awkward dynamic where it is indirectly funding the lab that just hired its Nobel Prize winner.
— Anthropic · AIToolsRecap · andrew.ooo
🔗 Anthropic · AIToolsRecap · Deep Dive
Noam Shazeer — Transformer Co-Inventor — Joins OpenAI
Noam Shazeer, co-lead of Google Gemini and one of the eight authors of the seminal "Attention Is All You Need" paper, confirmed he is joining OpenAI. Shazeer had been at Google since 2001, left briefly to start his own AI company (which Google re-acquired in 2023), and returned to co-lead Gemini. His departure is seen as a severe blow to Google's AI ambitions.
The twin exits of Shazeer and Jumper in the same week have raised existential questions about Google's ability to retain top-tier AI talent. Google has been investing heavily in retention packages, but the allure of frontier labs and the proximity to AGI-focused research appears to be pulling senior researchers away at an accelerating rate.
— OpenAI · The Decoder · andrew.ooo
🔗 Shazeer on X · OpenAI · The Decoder
SK Hynix Files $29.4B US IPO — Trading Begins July 10
SK Hynix, the world's second-largest memory chip maker and leading supplier of HBM (high-bandwidth memory) to NVIDIA, filed for a $29.4 billion US listing. Bloomberg confirmed trading is expected to start July 10. The proceeds will fund additional HBM manufacturing capacity — the critical bottleneck for AI accelerator chips powering the H100, H200, and GB200 GPU families.
The strategic significance extends beyond the numbers. SK Hynix is already an Anthropic Series H investor, joining Samsung and Micron — all three major global memory suppliers are now Anthropic investors heading into the lab's own anticipated IPO. This creates an unusually tightly-coupled AI infrastructure ecosystem where memory suppliers, GPU makers, and model labs are financially interlinked.
— Bloomberg · AIToolsRecap
Claude Tag for Slack Goes Live — @claude in Any Channel
Anthropic launched Claude Tag for Slack for enterprise customers. Tag @claude in any Slack channel and the assistant receives full conversation context, executes tasks, writes and reviews code, and replies in-thread — no separate interface required. The internal metrics are striking: Claude Tag already generates 65% of code on Anthropic's own product team, making it the dominant internal coding tool ahead of Claude Code for collaborative workflows.
This positions Claude Tag as a direct competitor to Microsoft Copilot for Teams. For enterprises already on Slack and using Claude via API, the feature eliminates the friction of context-switching between chat and AI. Anthropic's run-rate revenue has surpassed $30 billion, up from $9 billion at the end of 2025, with over 1,000 customers spending $1 million+ annually.
— Anthropic · AIToolsRecap
Mistral OCR 4 Brings Enterprise Document Understanding at Scale
Mistral AI released OCR 4, its next-generation document understanding system, on June 22. The model achieves breakthrough performance — a 72% win rate in human preference evaluations and the top score on OlmOCRBench (85.20). It supports 170 languages across 10 language groups, with per-page bounding boxes, typed-block classification (titles, tables, equations, signatures, code), and inline confidence scores.
Mistral OCR 4 is integrated with the Mistral Search Toolkit (public preview) for structured extraction, RAG pipelines, and enterprise search. A single-container self-hosting option addresses data sovereignty requirements. Pricing is $4 per 1,000 pages via API ($2 with Batch API discount). The Connectors platform also received enterprise-grade upgrades — admin controls, scoped API keys, multi-account authentication, and a debugger — now covering 60+ integrations.
— Mistral AI · Releasebot
🔗 Mistral AI · Releasebot
AI Research: Multi-Model Limits, Agentic RL, and Multimodal Code Intelligence
Three papers from this week's arXiv batch stand out:
When Does Combining Language Models Help? (arXiv:2606.27288) reveals a fundamental ceiling on multi-model strategies like routing, voting, and mixture-of-agents. The authors show that accuracy is capped by a quantity most teams never report: the rate at which every model is wrong on the same query. For any strategy whose output picks one member's answer, accuracy cannot exceed 1−β, where β is the co-failure rate. The Clopper-Pearson bound they provide lets teams compute this ceiling directly from their data.
Multi-Step Tool-Use RL Collapse (arXiv:2606.26027) diagnoses a catastrophic failure mode where RL-trained tool-using agents abruptly lose the ability to invoke tools correctly — performance drops by 30+ points in a single training step. The culprit: unexpected probability spikes in control tokens. The paper identifies supervisory signals that prevent this collapse, critical for deploying reliable tool-using agents in production.
Beyond NL2Code (arXiv:2606.15932) surveys the emerging field of multimodal code intelligence — systems that generate code from visual inputs like screenshots, diagrams, and videos. The paper organizes benchmarks across four domains (GUI, scientific visualization, structured graphics, frontier tasks) and argues for verification-centered evaluation: multi-signal validation, multi-state verification, and cross-task transfer testing.
— arXiv · BuildThisNow
🔗 Co-Failure Ceiling · Tool-Use RL · Multimodal Code
Curated by KD Agentic
Top comments (0)