Several years back, investing heavily in AI chip startups might have seemed a smart move. However, had that capital been funneled into Nvidia’s shares instead, the returns would have been significantly higher.
Consider Graphcore, which secured $200 million from backers including Microsoft and Samsung in December 2018, aiming to challenge Nvidia’s GPU dominance in AI workloads with its innovative chip designs and comprehensive software stack, Poplar.
Despite these efforts, Graphcore’s journey has been rocky, with regulatory filings last year raising ‘going concern’ doubts. Had the investors redirected their $200 million to Nvidia at that time, they would have witnessed a staggering 3,000% surge in stock value, turning their investment into $6.2 billion within 5.5 years—a substantial $6 billion gain.
An experienced venture capitalist, preferring anonymity, labeled betting against Nvidia as ‘a very stupid investment.’
Yet, Silicon Valley’s ethos often celebrates persistent technologists who defy skepticism to pursue their visions.
One such hopeful is Positron AI, led by CEO Thomas Sohmers, a former technical prodigy at an MIT research lab since age 13. Positron AI, which recently unveiled itself, is tackling Nvidia’s stronghold with a strategic and cash-efficient approach, initially focusing on transformer models that underpin OpenAI’s GPT frameworks.
Positron has designed the Atlas server, equipped with eight specialized AI chips boasting superior memory for handling transformer models. Sohmers revealed that the company is starting with inference tasks, betting on the growth of this market segment.
The startup’s hardware is built on field-programmable gate arrays (FPGAs), facilitating a swift product rollout. With about $12 million raised and half spent, Positron has kept its team lean at 20 employees.
Other contenders include Groq, a more established AI chip startup with $367 million in funding, and Rivos, led by ex-Googlers with a history in developing Chrome OS and hardware. Rivos is working on a combined CPU-GPU chip optimized for large language models and challenges Nvidia’s CUDA with its own software stack.
Then there’s Etched, co-founded by Harvard mathematicians specializing in AI compilers, and Rain AI, supported by OpenAI’s Sam Altman, both developing AI accelerator chips. Cerebras stands out with its colossal AI chip for model training, while OpenAI’s Altman himself is reportedly raising funds for a chip venture, Tigris.
Not to be outdone, Softbank’s Masayoshi Son—known for his audacious investments—is also rumored to be seeking a colossal sum for an AI chip startup codenamed Izanagi, aiming to enter the arena dominated by Nvidia.
