Nscale filed publicly for a U.S. IPO on Friday, seeking to raise roughly $2B on the NYSE under the ticker NSCL — the clearest sign yet that AI-infrastructure financing is starting to move from private mega-rounds into public markets. The UK-based data-center builder's contracted pipeline reached $103.4B as of August 31, up from $38.0B at the end of 2025, anchored by Microsoft and Anthropic deals worth roughly $88B combined; revenue for the first half of 2026 grew 1,252% year-over-year to $140.6M. Nscale builds its own power, liquid-cooled shells, and GPU fleets rather than leasing space from someone else, and that vertically integrated model is now being tested against public-market scrutiny instead of venture term sheets. (Bloomberg)
Private Companies
Firmus Technologiesfundraise
Firmus is targeting up to $7B in its IPO at a $50B valuation, people close to the deal say, as the Nvidia-backed Australian AI-infrastructure builder moves through a global roadshow — Singapore, Hong Kong, London, New York, then Australia — toward an October 6 ASX listing. At that size the float would rank among 2026's largest IPOs globally and Australia's second-largest ever. The number resolves what had been an open question since the roadshow opened this week. (Capital Brief)
Public Markets
Shares rose as Taiwan's memory-fab union let Friday's contract deadline pass without calling a strike vote, easing a supply-disruption risk just as 2027 HBM allocations for AI accelerators are being locked in. (Reuters)
Shares rose as Broadcom held its dividend at $0.65/share, payable September 30, while multi-year custom AI accelerator commitments from hyperscalers including Alphabet, Meta, Anthropic, and OpenAI kept its non-GPU silicon backlog intact heading into next quarter's print. (Benzinga)
Emerging
Memory bottleneck: BOOST, a new runtime from Georgia Tech, Nvidia Research, and Stanford researchers, lets LLM inference draw on host memory and HBM at the same time instead of treating host RAM as a fallback tier. In testing it cut time-per-output-token by 4.3% at matched batch size with no kernel changes required — a software answer to the same memory-bandwidth constraint that HBM suppliers are currently rationing hardware around. (arXiv)