Ricursive publishes Ripple-65B Preview on Hugging Face
Ricursive Intelligence published Ripple-65B Preview on 4 November, a 65-billion-parameter dense Transformer released under Apache 2.0 with weights on Hugging Face. The model card says it was pretrained on 12 trillion tokens and supports context up to 1 million tokens. Ricursive reports 80.5 on Terminal-Bench 2.1, an agentic coding benchmark, above the published figures of four larger US open-weight models, the closest being Reflection Beam 501B at 80.1. On Ricursive's own re-runs with the official Terminus 2 harness, Reflection Beam falls to 70.8 and Meta Muse Spark Open 816B comes closest at 78.7. Ripple's own scores are self-reported.
The company trained a dense model rather than a mixture-of-experts one, and says it held 40% model FLOPs utilisation through pretraining and post-training, with a team that grew from one person to five. An NVFP4 checkpoint ships alongside the BF16 weights, and the card says the model fits on consumer hardware such as an Nvidia DGX Spark. A 65B model small enough for a single machine is one a European team can host inside the EEA without a provider contract, so check which checkpoint fits your hardware before planning around it. Ricursive puts full R&D cost at about $32 million, and estimates that at a tenth of the cost of the previous open-weight leader on agentic coding.
On-premises HPC and AI-for-science server revenue reached $16.5bn, mostly on price
Hyperion Research said on 3 November that worldwide on-premises HPC and AI-for-science server revenue was $16.5 billion in the first half of 2026, up from $13.8 billion in the second half of 2025, which Hyperion puts at 19.9% growth. Unit sales over the same period grew 4.4%. Hyperion attributes the gap to rising system costs from growing GPU content and higher memory costs. Dell became the market share leader with $4.2 billion of server revenue in the half.
Earl Joseph, Hyperion's chief executive, said organisations are replacing ageing machines and adopting AI into their workflows, and that the growth reflects both a nominal increase in units bought and significant increases in system prices. For a European buyer sizing on-premises AI capacity for 2027, price per system is the variable that moved, not the number of systems. The Register reports memory costs have tripled or quadrupled since last year. Hyperion says HPE had led the market for more than 25 years, and that HPE booked revenue from several exascale-class machines in 2025, a scale of procurement it calls unrealistic to expect every year.
Quick Hits
- Broadcom's TPU compute shipments roughly doubled from its fiscal second to its third quarter of 2026 — Epoch AI estimates about 1.97 million H100-equivalents in Q3, on roughly 800,000 chips, a 65% rise in units.
- Supermicro will build inference racks pairing SambaNova's SN50 chips with Nvidia GPUs — announced 3 November and planned around an initial 10,000 SN50s. The companies will start with neoclouds and sovereign clouds, then expand to enterprises (TNW).
- Data Centers on the Ballot counted 43 of 46 municipalities approving tighter restrictions on server farms — Harwood, North Dakota, rejected them, and two had not posted results when The Register reported the tally on 4 November.
- Meta acquired Conductor, whose product runs several coding agents in parallel workspaces — terms were not disclosed. Conductor's cloud service shuts on 20 November, after which the company will begin deleting user cloud data (TechCrunch).