Before we dive into today’s newsletter… check out our latest episode of podcast with Jascha Achterberg, CTO of Callosum!
Jascha Achterberg started Callosum with his Cambridge PhD collaborator Danyal Akarca. Callosum is building infrastructure that matches complex AI workloads to specialized models and chips. The company raised $100 million in their seed round led by Atomico, with significant participation from Plural, DCVC, and the UK Sovereign AI Fund.
Jascha’s bet is that the future of compute is heterogeneous: different chips working together on different parts of a problem. We’ll dig into the commercialization opportunity: discover where each architecture excels, help chipmakers find production workloads, and let enterprises adopt new compute without replacing the stack they already trust. It is all in pursuit of faster, cheaper, and more capable computing systems, which is the biggest prize in infrastructure today. Give the episode a listen below!
d-Matrix unveiled Squadron, clusters of up to 16 racks and 1,024 Corsair accelerators built with Arista, Broadcom and Supermicro, and said the systems are shipping to priority customers, with Corsair servers and SquadRacks now generally available. At the OCP Global Summit, CEO Sid Sheth showed attention-FFN disaggregation running on real silicon, with GPU servers handling attention and Corsair racks handling the feed-forward and expert layers, at 845 tokens per second per user and 20,415 tokens per second across 64 concurrent users. Splitting inference by layer type gives a memory-centric chip a defined job beside GPUs instead of a contest for the whole workload; the figures are d-Matrix's own and no customer was named. (d-Matrix)
Private Companies
Lightmatterlaunch
Lightmatter showed a 204.8 Tbps co-packaged optics switch tray built with Flex and an XPU tray carrying four accelerators and two CPUs at 102.4 Tbps of scale-up bandwidth, each in one rack unit to Open Rack and DC-MHS specifications. The XPU tray replaces copper backplanes, which Lightmatter says limit scale-up systems to one or two racks, with blind-mate optical links. Both follow a white paper it contributed to OCP; the figures are Lightmatter's own. (Lightmatter)
Frore Systemslaunch
Frore Systems introduced LiquidJet Diamond, a GPU coldplate with diamond wafers over hotspots, rated for 770 W/cm² against about 600 W/cm² for its original LiquidJet. Frore claims GPU temperatures 16C to 22C lower than conventional coldplates and 30% to 35% more tokens per watt on parts such as Nvidia's 2,300W Rubin. The figures come from Frore's own models, and it has not said how the diamond is made or bonded. (Tom's Hardware)
Public Markets
Micron fell after CXMT said it will ship server DDR5 modules on a 4F² cell by year-end, about 33% smaller per cell. CXMT gave no yield or capacity figures, and Samsung and SK hynix each fell more than 7% in Seoul Tuesday. (Korea JoongAng Daily)
Cisco said Microsoft has moved into production on systems powered by its Silicon One P200, and Alibaba has deployed its Quantum Cat router, also built on the P200, at scale in its eCORE backbone at 51.2 Tbps of routing capacity. Both customer statements come from Cisco's release. (Cisco)
AMD introduced Helion, a 21-inch OCP rack reference design holding up to 32,768 Zen 6 cores in a CPU-only build, and moved its ship date to November 2026 from the first half of 2027. The rack targets agentic AI servers. (Tom's Hardware)
Emerging
MoE serving: A paper from Chang Chen, Andrew Yang, Tiancheng Chen, Jiangfei Duan, Xinwei Qiang, Zhongkai Yu, Xiang Fang and Yufei Ding, Zepp: Accelerating Distributed MoE Serving under Relaxed Balance Constraints, treats load balance as a constraint instead of the objective and optimizes inter-node communication directly. It places expert replicas to cut token traffic under GPU limits, reshapes flows under NIC limits, and overlaps expert compute with the transfers. Against seven MoE serving systems it reaches up to 6.68x MoE layer speedup and a 1.86x geometric mean over the fastest baseline, which locates the bottleneck of expert-parallel serving in inter-node traffic rather than GPU balance. (arXiv)
Meta networking: Meta introduced IcePack, its first liquid-cooled network rack on the OCP ORV3-HPRv1 standard, holding IceCube 102.4 Tbps switch blades, double its 51.2T platforms. IceCube comes in a Broadcom Tomahawk6C variant and an Nvidia Spectrum-6 variant, both running Meta's FBOSS, so the switch operating system is the constant across vendors. Meta also described Flattened Planar Fabric, which connects each GPU over eight independent 100G links to eight switches, one per network plane. (Engineering at Meta)