Supermicro CBO Vik Malyala and SemiAnalysis' Dylan Patel discussed a 7,000-pound double-wide rack at Computex 2026. The rack targets the B300 HGX memory squeeze by enabling denser GPU packing and liquid cooling.
Key facts
- Double-wide rack weighs 7,000 pounds
- B300 HGX faces memory squeeze from HBM3e supply
- Helios is launch partner for AMD MI450X
- Vera Rubin expected smoother ramp than GB
- PCIe Gen 6 and CXL 3.0 targeted for petascale storage
At Computex 2026, Supermicro CBO Vik Malyala and SemiAnalysis' Dylan Patel outlined a hardware strategy centered on a double-wide rack weighing 7,000 pounds. According to @SemiAnalysis_, the rack addresses the B300 HGX platform's memory squeeze, where HBM3e supply tightens as GPU demand grows.
The Double-Wide Rack Rationale
Malyala argued that a double-wide rack works now because of improved power and cooling density, enabling denser GPU packing without sacrificing thermal performance. The rack supports liquid loops and bus bars for cooling, a necessity given the 7,000-pound weight and high power draw. This contrasts with prior rack designs that struggled with weight distribution and cooling at scale.
Helios and AMD MI450X
Supermicro's Helios platform was announced as a launch partner for AMD MI450X, targeting AI inference workloads. Malyala noted that Helios leverages the double-wide rack to pack more MI450X GPUs per node, aiming to compete with Nvidia's B300 HGX in inference efficiency. The partnership signals Supermicro's bet on AMD's inference stack as a viable alternative to Nvidia's dominance.
Vera Rubin vs. GB Ramp
Malyala predicted that Vera Rubin, Nvidia's next-generation architecture, would have a smoother ramp than the current GB platform. He attributed this to lessons learned from GB's supply chain and software stack challenges. The statement suggests that Nvidia's ecosystem maturity could accelerate Vera Rubin adoption, potentially pressuring AMD's MI450X launch timeline.
Storage and Connectivity
Storage took center stage in the interview, with discussions on petascale systems, PCIe Gen 6, and CXL 3.0. Malyala emphasized that storage bottlenecks are becoming as critical as compute and memory, especially for large-scale AI training clusters. Connectivity improvements via PCIe switches, Tomahawk 6, and UALink were highlighted as key enablers for scaling beyond single-rack configurations.
Memory Price Dynamics
Memory prices were a recurring theme, with Malyala noting that HBM3e costs remain elevated due to supply constraints. The double-wide rack's ability to maximize memory bandwidth per GPU is a direct response to this cost pressure, allowing operators to extract more value from each HBM3e module. The exact memory price figures were not disclosed.
What to watch
Watch for Nvidia's Vera Rubin launch in late 2026 and whether its ramp indeed outpaces GB's. Also monitor AMD MI450X inference benchmarks against B300 HGX, and HBM3e pricing trends through Q3 2026 earnings calls.
