OpenAI's first custom inference chip reaches engineering samples while SK Hynix moves to bring $29 billion in AI memory exposure directly to U.S. equity markets.
🗄️ OpenAI and Broadcom Unveil Jalapeño: First Custom LLM Inference Chip
Decoded: OpenAI and Broadcom (AVGO) jointly unveiled Jalapeño on June 24, OpenAI's first custom AI chip — an inference accelerator built from the ground up for large language model workloads. Engineering samples are already running production-target workloads in the lab, including GPT-5.3-Codex-Spark, and the chip is designed for deployment at gigawatt scale in partnership with Microsoft (MSFT) and other data center operators across multiple hardware generations. The project was completed from design to production in nine months, a timeline OpenAI said was accelerated by its own AI models. The Jalapeño architecture reduces data movement and balances compute, memory, and networking to achieve utilization closer to theoretical peak — Broadcom's silicon implementation handles the chip design while Celestica is responsible for board and rack integration. (OpenAI official announcement, openai.com, June 24, 2026)
Why it matters: Jalapeño puts OpenAI in the same hardware vertical integration tier as Google (TPUs), Amazon (Trainium), and Apple (Apple Silicon). For Broadcom (AVGO), this confirms a multi-generation custom silicon program with a hyperscaler-grade client at exactly the moment Broadcom's AI chip revenue crossed $10.8 billion in a single quarter — the Jalapeño relationship extends that revenue visibility further. The nine-month design-to-production timeline, achieved with AI-accelerated engineering, also signals a structural compression of custom chip development cycles: if AI labs can now build inference-optimized silicon at this pace, the dependency on general-purpose GPU architectures decreases over time, reshaping the competitive landscape for Nvidia and AMD in the inference workload category specifically.
📊 SK Hynix Files for $29B Nasdaq ADR Listing as AI Memory Leadership Deepens
Decoded: SK Hynix filed with the SEC on June 24 to raise approximately $29.65 billion through American depositary receipts on Nasdaq — what would rank among the largest share sales in history. The company plans to issue up to 17.79 million new shares as ADRs, with trading expected to begin July 10, 2026. SK Hynix shares surged 12% in Seoul trading on June 25 following the announcement. The South Korean company is the world's leading producer of high-bandwidth memory — the specialized stacked memory required for Nvidia H100 and B200 GPU clusters — and said the U.S. listing would allow "its true corporate value to be properly evaluated." SK Hynix is the primary HBM competitor to Micron (MU), and both companies report sold-out HBM order books through end-2026. (CNBC, SEC filing, June 24-25, 2026)
Why it matters: SK Hynix's Nasdaq listing converts the world's most critical AI memory supply chain into a directly tradable U.S. equity. For investors who understand that HBM is the binding constraint on AI infrastructure scale — every GPU cluster requires it, supply is fully allocated through end-2026, and only three companies globally produce it — the ADR listing provides direct exposure to that constraint without navigating Korean market access. The $29.65 billion raise signals that SK Hynix intends to fund next-generation HBM4 production capacity that will determine AI infrastructure pricing power through 2027 and 2028. For Micron (MU) investors, a better-capitalized SK Hynix is a stronger competitor, though it also validates the structural premium the entire HBM sector currently commands.
Stay decoded. See you tomorrow.
— The Get AI Decoded Team
Enjoyed this article?
Subscribe free — AI news decoded for investors, every morning.
No spam. Unsubscribe anytime. Privacy Policy