Ayar Labs closed an additional $150M in Series E financing on Thursday, bringing its total primary capital raised in 2026 to $650M, alongside a separate $225M secondary sale led by Antero Peak Group at Artisan Partners that values the co-packaged-optics company above $5B, up from the $3.75B mark set at its March Series E close. New investors including MediaTek, Insight Partners and Qatar Investment Authority joined existing strategic backers AMD and Nvidia, both of which depend on Ayar's optical links to keep GPU clusters talking to each other as clusters scale past what copper can carry. CEO Mark Wade put it plainly: "Copper interconnect is becoming the limiting factor for AI scale-up." (SiliconANGLE)
Private Companies
Groqother
The Justice Department opened a formal antitrust inquiry into Nvidia's $20B license of Groq's inference-chip technology, Bloomberg reported, examining whether the December deal, under which Groq's CEO Jonathan Ross and other senior leaders moved to Nvidia while Groq kept operating independently under a new CEO, amounts to an unreported acquisition. The probe tests whether licensing technology paired with a mass leadership transfer can substitute for a merger filing, a template Nvidia could reuse against other compute rivals it would rather absorb than compete with. (Bloomberg)
d-Matrixpartnership
d-Matrix will connect its next-generation Raptor inference chips to Nvidia's MGX rack-scale architecture through NVLink Fusion, under a multi-year product-roadmap collaboration announced Thursday, with rack-integrated availability targeted for the fourth quarter of 2027. The deal gives d-Matrix's chiplet-based inference accelerators a direct path into Nvidia's dominant data-center reference architecture rather than requiring customers to build separate interconnect fabric around them. It is also a tell on Nvidia's own strategy: opening NVLink Fusion to a specialized inference rival suggests Nvidia would rather own the interconnect standard than the whole chip. (NVIDIA)
Public Markets
Micron led a broad chip-stock decline after Brent crude topped $108 a barrel and the 10-year Treasury yield neared 5% on renewed Middle East war fears, with investors cutting exposure to the sector's most richly valued AI winners regardless of company-specific news. (CNBC)
SK hynix's US-listed shares fell in the same macro selloff, giving back most of Wednesday's 7% gain as the memory maker's stock proved as sensitive to oil and bond yields as to its own record HBM demand. (CNBC)
Emerging
Inference chiplets: A new paper, CHIPSMORE, proposes a chiplet-based inference accelerator that pairs resistive-RAM analog compute-in-memory with SRAM digital compute-in-memory processing elements, linked by a programmable inter-chiplet network so the same silicon serves both full-parameter and LoRA-adapted models under shifting context lengths and request loads. Its composable key-value memory scheme reallocates scratchpad, on-chip SRAM and embedded DRAM on the fly, evidence that inference accelerators are increasingly designed around workload variability itself, not just raw throughput, as production LLM serving diversifies beyond a single fixed shape. (arXiv)