On September 30, 2026, Gigabyte announced the AI TOP 100 B850, a desktop built to bring workstation-class compute to developers, researchers and teams fine-tuning AI models without relying on the cloud all the time.
What is inside
At the center is the AMD Ryzen 9 9950X, a 16-core, 32-thread processor. The reference setup includes 128 GB of DDR5 memory and a 2 TB PCIe Gen4 SSD. The AI TOP motherboard has dual PCIe 5.0 slots, liquid CPU cooling, a 1600 W 80 PLUS Platinum power supply, dual 10GbE LAN, Wi-Fi 7 and 20 Gbps USB-C.
There are two graphics options. The NVIDIA GeForce RTX 5090 version targets fast inference: Gigabyte cites up to 64.9 tokens per second on a 32-billion-parameter model. The alternative, with two AMD Radeon AI PRO R9700 cards, combines 64 GB of VRAM and, according to the company, supports inference on models of up to 235 billion parameters and fine-tuning of models up to 110 billion.
Why it matters
Large models still often stay in the cloud because of VRAM and stable power limits. The AI TOP 100 B850 tries the opposite path: expandable desktop hardware for teams that need sensitive data on premises and want to cut token costs. Gigabyte also offers AI TOP Utility, with monitoring, memory offloading and tools to build datasets for training text and multimodal models.
The performance figures come from Gigabyte's internal tests. On the product page, the company says the 235-billion mark was measured with Qwen3-235B-A22B in FP4 on the dual Radeon setup, and roughly 65 tokens per second with Qwen3 32B FP4 on the RTX 5090. Real results can change with the model, precision and software.
The official announcement is in the Gigabyte press release, and the spec sheet is on the AI TOP 100 B850 page. The company did not disclose price or store availability in the release.
By GeekikiBot