OpenAI and Broadcom unveiled Jalapeño on Wednesday — OpenAI's first custom AI inference chip, co-designed from scratch and taped out in nine months, which Broadcom describes as the fastest ASIC development cycle in the history of high-performance semiconductors. The chip is optimized specifically for LLM inference; Broadcom CEO Hock Tan cited approximately 50% cost savings relative to standard AI GPUs. Initial deployment is targeted by end of 2026 at gigawatt-scale data centers with Microsoft and other partners.
OpenAI has now disclosed several non-Nvidia inference commitments including Jalapeño, its owned-silicon ASIC layer built with Broadcom as well as the $20-billion and the 750-megawatt Cerebras commitment announced Tuesday for wafer-scale burst inference, where on-chip SRAM eliminates off-chip HBM latency.
For grounding in the architecture driving these decisions, check out Chris Zeoli's piece on The Inference Unbundling to understand why prefill and decode are separating onto different hardware.
Qualcomm announced Wednesday it has agreed to acquire Modular, the AI software infrastructure company co-founded by LLVM creator Chris Lattner, at a valuation of approximately $4 billion — up from the $1.6 billion at which Modular raised $250 million in September 2025. The transaction is expected to close in the second half of 2026, subject to regulatory approvals. Modular's MAX runtime and Mojo language let engineers deploy AI models across heterogeneous hardware — CPU, GPU, NPU, and custom ASIC — without rewriting for each chip's backend, a capability that maps directly onto CEO Cristiano Amon's push to extend Qualcomm's software stack from mobile NPU into edge and data center inference. The 2.5x step-up in valuation in under a year signals how quickly hardware-agnostic inference software has been repriced.
Q1 2026: $193.4 million revenue (+94% year-over-year), full-year guidance raised to $855–$865 million (+69% year-over-year). The after-hours move is driven by the margin path, not the top line. Cerebras guided FY2026 core gross margins to 38–41%, a step down from Q1's 47%, with Q2 specifically guided to 36–38%. The compression likely reflects mix shift: cloud and services revenue ($82.8 million, +178% year-over-year) is growing nearly three times faster than hardware ($110.6 million, +59%), and the cloud segment carries structurally lower margins. At 38–41% for the year, Cerebras core gross margins would sit materially below Nvidia's mid-70s and AMD's mid-50s.
SemiAnalysis published a detailed analysis of CXMT on June 23 that separates the Chinese DRAM maker's real progress from its limits. The numbers: $8.6 billion in DRAM revenue for 2025, up 156% year-over-year, crossing into profitability for the first time ($1 billion net income), with 2026 revenue potentially exceeding $50 billion and wafer capacity approaching Micron's — roughly 350,000 wafer starts per month by year-end. The structural threat is in commodity DRAM, not HBM: CXMT's yield on HBM3 8-hi is roughly 25%, and its cost-per-bit remains more than 30% above incumbents for DDR5. HP and Dell have begun qualifying CXMT parts for PC builds.
The transition to watch is from PC OEM qualification to AI server OEM qualification. If Chinese-made DRAM reaches that threshold, the margin compression for SK Hynix and Micron on standard DRAM would arrive faster than current forecasts assume — and would matter most in the period before HBM4 fully dominates the AI server bill of materials.
SK Hynix's board approved Wednesday a $29 billion Nasdaq ADR listing — 17.79 million new shares at a value of 45.45 trillion won — targeting a July 10 trading debut, subject to SEC approval. If completed, it would rank among the largest equity offerings of the AI hardware cycle. SK Hynix holds approximately 60% of the HBM market and has seen its market capitalization cross $1 trillion. Proceeds are directed toward the Yongin Cluster, a new campus of memory fabrication plants in South Korea coming online in 2027, and a $4 billion HBM packaging plant in Indiana. For US investors, the Nasdaq listing removes the KRW currency friction that has historically kept Korean semiconductor names underweight in US portfolios — at a moment when HBM supply allocation is the most consequential bet in the AI memory trade.
Micron Technology reports fiscal Q3 2026 earnings after market close. Consensus: revenue of approximately $35 billion (+276% year-over-year), gross margin ~81%, EPS ~$20. HBM capacity for all of 2026 and into early 2027 is reported fully booked. With Samsung and SK Hynix each down 12%-plus in Tuesday's session on AI infrastructure spending concerns, Micron's guidance on HBM4 allocation and early signals on 2027 platform demand — particularly Vera Rubin Ultra — are the most consequential data points for the AI memory trade this week. The report follows a session in which AI memory stocks were the primary target of forced selling, making any forward commentary on allocation and pricing especially load-bearing for sector sentiment.