Hyun-joong Kim, co-founder and CEO of SqueezeBits.
Courtesy of SqueezeBits
that Running AI models costs a lot of money. The cost of a single Nvidia AI chip can exceed $30,000. According to one estimate, it costs OpenAI $700,000 per day to operate ChatGPT. With such costs, only large companies have the funds to develop and run AI. SqueezeBits, a two-year-old South Korean startup, claims it can reduce that cost and democratize the hot technology. To help with that, SqueezeBits recently received support from one of the country’s top Internet companies.
In January, SqueezeBits raised 2.5 billion won (approximately $2 million) in pre-Series A funding. The Seoul-based startup declined to disclose its valuation, but the round values SqueezeBits at about $15 million, according to people familiar with the matter. Investors in the round include Kakao Ventures, the venture capital arm of billionaire Kim Bum-soo’s Kakao, one of South Korea’s top two internet companies.
“We are not only focused on how companies utilize AI, but also how it is democratized and pervasive to create more AI applications,” Kakao Ventures said Justin Singh, senior associate at , in a video interview. Kakao Ventures was also an early backer of South Korea’s Rebellions, which is developing cheaper AI chips (about half the price of Nvidia’s H100 chips) and was recently valued at $124 million at a $650 million valuation. , becoming the most funded AI chip startup in Japan. .
Other investors in the round include Samsung Electronics’ Samsung Next. POSCO Venture Capital, the venture capital arm of South Korean steel giant POSCO. And Postech Holdings is an accelerator supported by Pohang University of Science and Technology (Postech), one of South Korea’s top universities of engineering.
SqueezeBits previously raised 1 billion won in seed funding in 2022 from D2 Startup Factory and Postec Holdings, which is backed by South Korean billionaire Lee Hye-jin’s Naver (a South Korean internet giant other than Kakao). The startup has raised a total of 3.5 billion won in venture funding.
“Many of these AI applications aim to use AI models to minimize cost and maximize performance,” Shin says. “Cost is definitely the biggest issue. It takes a lot of money to build a scalable AI model and turn it into a working product. It costs a lot of money.”
SqueezeBits says it can reduce costs by increasing efficiency. “AI models are very over-parameterized,” Hyungjun Kim, co-founder and CEO of SqueezeBits, said in a separate video interview. “Many companies don’t actually use their models fully optimized.” The number of parameters is an important indicator of the size of an AI model and is usually, but not always, correlated with performance. To do. For example, OpenAI’s GPT-3 language model has 175 billion parameters, and his GPT-4, which powers the latest version of his ChatGPT, reportedly has 1.7 trillion parameters.
“There are some useless parameters and data in the model. We basically remove such useless or unimportant data in the model and calculation process so that we can reduce the calculation cost and memory usage. We’re deleting data,” explains Kim, 29, who has a PhD in electrical engineering and computers. Science of Postec. “This enables cheaper and faster AI inference.”
According to Kim, SqueezeBits can speed up models by a factor of 3 to 5 and reduce memory usage by a factor of 4. Last month, the startup announced a Software-as-a-Service toolkit to help companies optimize open source AI models, as well as their own large-scale language models, for cloud services.
“AI models are far too parameterized. Many companies don’t actually fully optimize their use of the models.”
Of course, SqueezeBits isn’t the only company optimizing its AI models. Other companies include OmniML, founded in San Jose in 2021, and Seattle-based OmniML, which spun off in 2017 from the Allen Institute for AI, a nonprofit dedicated to AI research founded by the late Microsoft co-founder Paul Allen. and Xnor.ai, which is based there.
According to the tech news site, OmniML is owned by GGV Capital, Qualcomm Ventures, and IMO Ventures, whose portfolio includes billionaire Peter Thiel’s data mining company Palantir Technologies and Google-backed AI drug discovery startup Xtalpi. ) is an investor, and it was acquired by Nvidia in February. information. Xnor.ai was acquired by Apple for approximately $200 million in 2020 with backing from Madrona Venture Group (Seattle’s largest VC firm and early investor in Amazon) and Nokia-backed NGP Capital, Technology News This was reported by the site GeekWire.
Other independent startups include Israel’s Deci, which is backed by Insight Partners, one of the world’s largest technology investors. Square Peg Capital is an Australian venture capital whose portfolio includes Airwallex and Canva, and investors in Massachusetts-based Neural Magic include Andreessen Horowitz and early Robinhood investor NEA, cloud-based networking and They include cybersecurity services provider Cloudflare and data analytics company Databricks. .
SqueezeBits’ domestic competitors include Nota, which raised $14.7 million in Series B in 2021. Investors in the round include Stonebridge Ventures (an early backer of Seoul and Seattle-based chip design startup MangoBoost) and Company K Partners (whose portfolio companies include autonomous drones) (Nearthlab), one of Asia’s 100 startups to watch.
Kim is unfazed by the competition. Hardware development will always be slower than software development, creating a demand for his AI optimization providers like SqueezeBits. “There’s a lot of new stuff coming out when it comes to AI, but the hardware can’t keep up with that kind of speed,” Kim says. “That means there is a huge gap between super-rapid advances in AI algorithms and model sizes and hardware that doesn’t support it.”
“And that’s where we come in,” he says.
