Phison unveils solution to reduce $3m cost of upcoming 1T parameter AI model to $100K

NVIDIA's GTC conference has officially kicked off, and Phison, one of the world's leading companies in controllers for NAND flash memory chips, has unveiled what it's describing as an affordable path to supporting a 1 trillion parameter AI model.

2

The race to the first 1 trillion parameter AI model is nearing completion, at least according to Phison, which predicts we will see the first 1 trillion parameter AI model before 2026. Phison has briefly outlined the progression of AI models, with a 69 billion parameter model unveiled in 2023 with Llama 2, then a big jump to the 405 billion parameter Llama 3.1 model in 2024, and then the 671 billion parameter DeepSeek R3 model in 2025. Keeping on this same trajectory, Phison expects the world's first 1 trillion parameter model will be unveiled before the end of 2025.

With the increase in the size of these AI models, the hardware needed to train and support them will also need to be increased, which is why Phison is looking at ways to reduce operational costs, and the company has a solution. Phison expects the operational costs of a 1 trillion parameter model will be approximately $3 million in raw GPU power, but the company has outlined a plan to alleviate the cost down to just $100,000. How? A combination of SSDs, Phison's aiDAPTIV+ software, and an NVIDIA GH200 Superchip.

By offloading the processing from the GPUs onto SSDs through the use of Phison's intuitive aiDAPTIV+ software, the company is able to significantly reduce the number of GPUs required to successfully run a large AI model. Phison anticipates it will be able to reduce the cost of training the largest-sized model by 1/30th of the price, or less than 4% of the currently estimated operating cost.