Qualcomm says it has validated its first-generation High Bandwidth Compute (HBC) silicon for AI inference. The milestone is tied to the company’s Dragonfly AI250 rack-scale platform, whose product page lists 768 GB of memory and 133 TB/s of effective memory bandwidth per card. DiarioBitcoin reports that Qualcomm expects to distribute AI250 samples in the first half of 2027.
What Qualcomm says it validated
HBC places compute close to memory, an approach Qualcomm says is intended to reduce data movement in AI inference. The company describes its first-generation HBC silicon as validated in its newsroom announcement.
DiarioBitcoin reported that Qualcomm’s laboratory tests covered memory reads, writes and computation near memory, using automated test equipment and randomized data. The outlet also described HBC as DRAM layers bonded directly above a logic chip.
What Qualcomm lists for the AI250
Qualcomm describes the AI250 as a second-generation rack-scale inference platform featuring HBC Gen 1. Its AI250 product page lists specifications for individual cards, the rack and its infrastructure:
| System element | Qualcomm-listed specification |
| AI250 card | 768 GB of memory; 133 TB/s effective memory bandwidth |
| Model and context support | Models up to 10 trillion parameters; context lengths up to 1 million tokens |
| Rack configuration | 56 cards; 43 TB of memory |
| Rack design and connectivity | Open Rack v3 (ORv3) compliant; PCIe 6.0 scale-up; Ethernet with RoCE scale-out |
| Cooling and power | Air or direct-liquid cooling; 140 kW rack thermal design power (TDP) |
The per-card figures and rack specifications describe different levels of the system: the card’s memory and bandwidth, then the rack’s configuration and infrastructure. Qualcomm’s page also characterizes the platform by the model sizes and context lengths it supports.
When AI250 samples are expected
DiarioBitcoin reported that Qualcomm’s target is to distribute Dragonfly AI250 samples in the first half of 2027.
