Nscale raised $3.36B in pre-IPO convertible notes led by hedge fund Third Point, with Nvidia adding a separate $1B tranche due in mid-November, about a week after the British AI-cloud operator filed paperwork for a roughly $35B NYSE listing. The financing funds data-center buildout across Nscale's Norway and West Virginia campuses and lets investors including Apollo, Citadel and the Abu Dhabi Investment Council convert their notes into equity once the IPO prices, rather than wait to buy in at the offering. Institutional demand for AI-infrastructure exposure is outrunning what any single financing vehicle — private round, convertible note, or the IPO itself — can absorb. (Bloomberg)
Private Companies
DensityAIfundraise
DensityAI is in talks for a round that would value the year-old chip startup at $10B, with Andreessen Horowitz set to lead a raise of several hundred million dollars. Founded by Ganesh Venkataramanan and roughly 20 engineers who built Tesla's Dojo supercomputer, the company has also reportedly secured a conditional AWS purchase agreement committing Amazon to buy its data-center chips — built around 3D DRAM stacking — if they hit specified performance benchmarks. (The Information)
Crusoeother
Crusoe walked away from a $1.25B agreement to buy 29 of Boom Supersonic's 42-megawatt Superpower gas turbines for its AI data centers, with first deliveries originally set for 2027. Boom CEO Blake Scholl said turbines are no longer part of Crusoe's near-term primary power mix; Crusoe said its energy plans haven't changed but that Boom isn't the right fit today. (TechCrunch)
Public Markets
Credo's fiscal Q1 revenue rose 115% year-over-year to $479M — its seventh straight quarter of triple-digit growth — with guidance reiterating more than $600M of optical revenue for fiscal 2027, evidence AI-network optical demand keeps outrunning the company's own targets. (Credo)
Emerging
Agentic chip design: Siemens and TSMC rolled out an AI-powered agent, built on Siemens' Calibre software, that automatically fixes design-rule-check violations across TSMC's digital and custom IC design flows, part of a broader push toward autonomous, self-verifying EDA workflows under Siemens' Fuse EDA AI System. As chip designs grow too complex for manual verification at each node, the tooling vendors themselves are racing to make agentic AI standard in the tapeout pipeline, not just a productivity add-on. (Siemens)
Split-brain reasoning: Imperial College London researchers propose HeteroReason, offloading a reasoning model's draft generation to an FPGA while GPUs handle verification and refinement, with a backtracking step so the system can recover from a bad early reasoning trajectory instead of compounding the error forward. Accepted at MICRO 2026, the design delivers a 4.2% accuracy gain alongside 1.01–1.42x latency speedups and up to 1.57x better energy efficiency than GPU-only baselines — evidence reasoning inference has distinct enough compute phases to justify splitting work across chip types, not just scaling more GPUs. (arXiv)