Nvidia Blackwell: Unbeatable in AI; AMD Is a Landslide Behind!

General news News Tech Tech News 08/17/2025

0 369 Views

TECH NEWS – The GB200 chip offers high profit margins for companies that use it. However, those looking to AMD will notice the opposite.

Morgan Stanley Research has released new data comparing the operating costs and profit margins of AI inference workloads. Most AI “factories,” or enterprises deploying multiple chips for inference, achieve margins above 50%, with Nvidia clearly leading the pack.

The evaluation focused on 100 MW AI factories running server racks from Nvidia, Google, AMD, AWS, and Huawei. Among them, Nvidia’s GB200 NVL72 Blackwell GPU platform posted the highest margin at 77.6%, translating into an estimated $3.5 billion profit. Google’s TPU v6e pod ranked second at 74.9%, while AWS Trn2 Ultraserver came in third at 62.5%. Most other solutions hovered around 40–50%. AMD, however, showed worrying results.

The AMD MI355X platform recorded -28.2% profit in inference, while the older MI300X showed -64.0%. Based on average rental prices of $10.50/hour, Nvidia’s GB200 NVL72 brought in $7.50/hour, followed by the HGX H200 at $3.70/hour. AMD’s MI355X managed just $1.70/hour, with most competitors in the $0.50–$2 range.

Nvidia’s dominance comes from FP4 support and ongoing CUDA AI stack optimization. Even older GPUs like Hopper and Blackwell are consistently improving with quarterly updates. AMD’s MI300 and MI350 platforms are excellent hardware-wise, and its software has improved, but AI inference remains a weak spot.

The total cost of ownership (TCO) of AMD’s MI300X reached $744 million—comparable to Nvidia’s GB200 platform at ~$800 million—offering no cost advantage. Newer MI355X servers have a TCO of $588 million, similar to Huawei CloudMatrix 384. Nvidia’s higher initial costs are offset by its superior inference performance, which will account for 85% of the AI market in the coming years.

Nvidia will launch the Blackwell Ultra this year, boasting 50% performance gains over the GB200. It will be followed by Rubin (2026), Rubin Ultra, and Feynman. AMD plans to counter with the MI400 next year, optimized for inference, setting the stage for an intense battle in the AI segment.

Source: WCCFTech, Morgan Stanley

Angyal Anikó

Anikó, our news editor and communication manager, is more interested in the business side of the gaming industry. She worked at banks, and she has a vast knowledge of business life. Still, she likes puzzle and story-oriented games, like Sherlock Holmes: Crimes & Punishments, which is her favourite title. She also played The Sims 3, but after accidentally killing a whole sim family, swore not to play it again. (For our office address, email and phone number check out our IMPRESSUM)