Chinese AI startup DeepSeek has released its open-source reasoning model, R1; it claims the tech matches or outperforms OpenAI's ChatGPT on benchmarks for a fraction of the cost
The company said its approach included using heterogeneous infrastructure, reducing numerical precision and applying a technique called an intelligent mixture of experts, which it said had cut training costs from $100m to $5m and reduced hardware needs from 100,000 graphics processing units to 2,000
DeepSeek is part of a group of Chinese companies developing their own graphics processing units that can rival Nvidia and other international chip firms
OpenAI, Oracle, Softbank and MGX's Stargate Project, which President Trump has called 'the largest AI infrastructure project in history', aims to invest $500bn in US AI infrastructure before 2029
Nvidia currently holds a market lead in graphics processing units, though rival AMD has launched some promising products which use application-specific integrated circuits
Edge and embedded AI represent a significant area of opportunity for either disrupting or partnering with leading chip firms to create specialized products, as different requirements and specifications arise, according to Chris Kachris, founder and CEO of InAccel
Kachris added there are still chances for upstarts to find innovation in areas too small for larger companies
In-memory computing, which combines storage and processing, and neuromorphic computing, which enables an AI system to adapt to new situations, have been put forward as areas to watch
Kachris also noted that although FPGAs may not be the most powerful solution for AI training, they may be the most energy-efficient solution for inference
The software layer for AI solutions is also important in terms of driving adoption: Nvidia's CUDA stack provides an advantage to the leading provider on the software front, although AMD's ROCm software competes with it