OpenAI and Broadcom Unveil Jalapeño, a Custom Inference Chip
OpenAI took a major step toward controlling its own destiny on Tuesday, unveiling Jalapeño — its first custom-designed AI inference processor, built in partnership with Broadcom. The chip reached tape-out in a remarkably fast nine months and is slated for deployment beginning in late 2026.
Jalapeño is a reticle-sized ASIC designed specifically for running large language models at scale. Early lab testing shows performance-per-watt substantially better than current state-of-the-art hardware, with engineering samples already running ML workloads — including GPT-5.3-Codex-Spark — at production target frequency and power.
What makes the announcement particularly notable is that OpenAI used its own models to accelerate parts of the chip design process. The company is signaling that it intends to compete across the full stack — from model architecture and training to the silicon that powers inference. This positions OpenAI alongside Google (which has its TPU line) and Amazon (with Trainium/Inferentia) as companies that design both the AI and the chips that run it.
Google's AI Talent Exodus Wipes $250 Billion Off Alphabet
It has been the worst week for Google's AI talent base in the company's history. The departures started on June 18, when Noam Shazeer — co-author of the landmark "Attention Is All You Need" paper that introduced the Transformer architecture and co-lead of Google's Gemini models — announced he was leaving for OpenAI. The move came less than two years after Google paid $2.7 billion to bring Shazeer back through its acquisition of Character.AI.
Two days later, John Jumper, the computational chemist who led AlphaFold and won the 2024 Nobel Prize in Chemistry, announced his departure from Google DeepMind to join Anthropic. Then on June 24, two more DeepMind researchers — Jonas Adler and Alexander Pritzel — were confirmed to also be heading to Anthropic.
The market reaction was swift: Alphabet shares fell as much as 7.2% intraday, the steepest single-day drop since February 2026, erasing roughly $250 billion in market capitalization. While Alphabet maintains zero analyst sell ratings and delivered 82% earnings growth in Q1, the concentration of foundational AI talent flowing to competitors is raising questions about Google's long-term position in the AI race.
GPT-5.5-Cyber: OpenAI's Gated Cybersecurity Model Goes Live
OpenAI launched the full version of GPT-5.5-Cyber on June 22, a specialized model engineered for automated vulnerability detection, patch generation, and remediation. The model achieves 85.6% on CyberGym (vs. 81.8% for standard GPT-5.5), 39.5% on ExploitGym (vs. 25.95%), and 69.8% on SEC-bench Pro (vs. 63.1%).
Unlike most OpenAI releases, GPT-5.5-Cyber is not publicly available. Access is restricted to vetted organizations through the Trusted Access for Cyber program, which includes Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler. The model can navigate large codebases, trace attack paths, validate exploitability, generate targeted patches, and produce remediation evidence — all within a single automated workflow.
Alongside the model, OpenAI is launching the Daybreak Cyber Partner Program and co-funding Patch the Planet, an initiative with Trail of Bits and HackerOne to help open-source maintainers remediate vulnerabilities found by AI tools.
Fable 5 Export Ban Enters Second Week — No Resolution in Sight
It has now been twelve days since the U.S. Commerce Department ordered Anthropic to suspend access to its newest models, Claude Fable 5 and Claude Mythos 5, under the Export Administration Regulations. The directive, issued on June 12, cited a claimed jailbreak with national security implications.
The scope of the ban is unusually broad: it covers not just users outside the U.S. but also any foreign national within the United States, including Anthropic's own non-citizen employees. Given these constraints, Anthropic disabled the models for all users globally. Access to other Claude models, including Claude Opus 4.8, remains unaffected.
Anthropic has publicly disagreed with the decision, arguing that "if this standard was applied across the industry, it would essentially halt all new model deployments for all frontier model providers." As of June 25, there is no indication the ban will be lifted soon — a Polymarket prediction market on the question shows low confidence in a near-term resolution.
The Great American AI Act Takes Shape in Congress
On the regulatory front, the bipartisan Great American AI Act discussion draft, released on June 4 by Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA), continues to generate debate. The 269-page bill would create the first comprehensive federal framework for AI governance in the United States.
Key provisions include requiring frontier AI developers with over $500 million in annual revenue to establish public safety frameworks, codifying the Center for AI Standards and Innovation within the Commerce Department with $100 million in annual funding, and a three-year preemption of state-level AI regulation — a provision that has drawn both praise from industry and pushback from state legislators.
The bill's four titles cover Frontier AI Governance, Workforce, Cybersecurity, and Research & International Cooperation. It arrives alongside a separate Trump administration executive order from June 2 that asks AI developers to voluntarily share new models with the federal government up to 30 days before broader release. Together, these moves signal that Washington is moving more aggressively to establish federal AI rules — even as the Fable 5 episode shows the tension between security concerns and industry innovation.