Nvidia Prepares to Strike the Inference Market

On February 28, 2026, insiders leaked information about a new chip from Nvidia, scheduled to be announced at the upcoming GTC conference. The company is integrating architectural approaches similar to the LPU (Language Processing Unit) from startup Groq.

Nvidia absolutely dominates in training neural networks, but the inference stage — generating responses in real time — requires different math. Traditional GPUs are overkill here and too expensive in terms of energy costs (TCO). By introducing "deterministic inference," Nvidia intends to kill two birds with one stone: destroy emerging competitors like Groq on their own turf and radically reduce the cost of API requests for cloud providers. This is the beginning of a new phase of price wars in hardware.

Source: WSJ / Reuters