Microsoft has ended an agreement to rent AI infrastructure at Nscale's planned Loughton, Essex site, The Telegraph reported Sunday, after repeated construction delays and a warning from UK grid officials that they could not supply adequate power until the 2030s. The site was announced in September 2025 as 23,040 Nvidia GPUs and an initial 50 MW, expandable to 90 MW. Nscale says the Microsoft relationship is "strong and long-term," that its other Microsoft contracts are unaffected, and that it is looking at alternative power sources for Loughton. The company filed confidentially for a US IPO in September, so grid access at each contracted site now sits inside the neocloud story it is taking to investors (The Telegraph)
Private Companies
Tenstorrentfundraise
Tenstorrent closed a $1.5B Series E at a $6.2B post-money valuation, led by Trace Capital and Innovation X, with Fidelity, Morgan Stanley, ARK Invest, Sandisk and Tether Investments participating. The company will use the capital to deploy its Galaxy servers at scale and to advance the CPU and AI IP it licenses to chip designers, including Samsung and LG. It sells one architecture from single cards to racks, and the release gives no revenue or Galaxy performance figures (Tenstorrent)
Anthropicfundraise
Anthropic is offering Blackstone a yield above 10% on debt financing its latest chip deal, the Financial Times reported Friday. Blackstone funds would invest $9B, backed by Anthropic's five-year lease payments on the chips, and the terms have not closed. Anthropic is paying for accelerators through leases and private credit, so the yield on that debt becomes part of what each unit of compute costs it. Reuters said it could not immediately verify the report (Reuters)
Public Markets
AMD fell with the Philadelphia Semiconductor Index, which lost 3.8% for an eighth straight decline, its longest streak since 2008. The index is down about 16% from its October 6 close (24/7 Wall St.)
Emerging
Agentic system design: A paper from Murali Emani and Kashyap R Manjusha, Agentic AI for Heterogeneous System Design Across Chip, Scale-Up, and Scale-Out Domains, presents ArchAgent, in which an LLM planner splits system design into chip, scale-up and scale-out subproblems and a deterministic analytical evaluator, not the LLM, produces the performance estimates. For a mixed AI and HPC workload under a 1 MW power budget it reaches a 1.93x weighted geometric mean speedup over a homogeneous GPU baseline in 18 evaluations, against 1.65x in 100 evaluations for hill climbing. The gain comes from the authors' analytical models, so it shows what the search method can find rather than what built hardware delivers (arXiv)