$11 billion AI chip startup SambaNova positioned as potential next Cerebras; raises stakes in custom-silicon competition.
Artificial intelligence chip startup SambaNova just raised $1 billion from investors, giving it a total valuation of $11 billion. The private company makes chips for AI inference, and it seems to be following in the footsteps of Cerebras Systems, which conducted the largest semiconductor initial public offering ever in May. That IPO was one of the most popular in years—will SambaNova be one of the next big IPOs?
SambaNova bills itself as "the world's fastest AI inference platform," and like Cerebras, it builds AI chips that process inference workloads. Inference is the process of generating AI chatbot outputs based on user inputs, and it's becoming a larger part of overall AI workloads.
SambaNova's latest chip is called the SN50, and it's used in servers installed in AI data centers. The company says its chip runs the largest AI models with trillions of parameters and fits them on a single server rack, helping models run faster. According to SambaNova, the SN50 delivers better performance than Nvidia's B200 chips.
The firm has signed up well-known clients for the SN50, which is expected to roll out by the end of the year. JPMorgan Chase signed a deal to deploy SN40L and SN50 systems for on-premises AI inference. Tech investor SoftBank will be the first customer to put the SN50 into use. SambaNova also identified Saudi Aramco, Meta Platforms, and Intel as clients.
Intel is both an investor in SambaNova and home to its executive chairman, CEO Lip-Bu Tan, who has held that position since 2017. Intel participated in the latest fundraising round as well as in SambaNova's $350 million round in February. The two companies announced a multiyear deal earlier this year to develop AI inference based on Intel's Xeon semiconductor.
SambaNova said its recent capital raise would help grow its business and secure its supply chain, as waves of chip companies are driving demand higher. Chipmakers are rushing to build out alternatives to challenge Nvidia's dominance and capture a piece of its hefty margins.
Nvidia seems to be the target of every AI chip company these days. A range of startups has Nvidia—and its incredibly wide 65.6% operating margin—in their crosshairs, as the AI industry increasingly shifts workflows from training models to inference.
AI workloads are quickly moving to inference. Deloitte estimates that inference workloads will make up about two-thirds of all AI workloads in 2026, up from just half in 2025. That number has climbed from a mere 20% a few years ago, and it could rise to 80% by 2027, according to McKinsey.
AI chip companies such as SambaNova are looking to deploy solutions that deliver inference quickly and cost-effectively, so users aren't waiting around for models to process their tasks. Part of SambaNova's solution to improving efficiency is to bring inference in-house, as shown in its recent deal with JPMorgan Chase. The company is deploying AI servers at data centers owned by specific business clients, a move that also improves data security, since the data is controlled by the company itself rather than by a third-party AI provider or AI lab.
Other AI chip companies, such as Cerebras Systems, have taken different approaches to compete in the AI inference race. Cerebras produces wafer-scale chips about the size of a plate that process inference workloads faster than Nvidia's graphics processing units. Cerebras bills itself as the "maker of the world's fastest AI infrastructure."
Cerebras' WSE-3 chip packs 4 million transistors divided across 900,000 cores and 44 gigabytes of memory. The large supply of memory means it can run AI models on the chip, sidestepping the slower process of moving them off it.
Billions of dollars continue to flow into the AI chip sector as well as into memory chips, with South Korea's SK Hynix recently completing a secondary listing on Nasdaq.
SambaNova is considering an IPO in 2027 and is leaning toward listing on the U.S. market, according to co-founder and CEO Rodrigo Liang. Of course, any IPO depends on factors beyond the company's control, such as the state of the market itself. If the market turns cold or the AI bubble bursts, it will be tough for private companies to debut.
But in a hot sector such as AI chips, a company can reach the public markets much faster than expected. Cerebras conducted its own IPO in May. In a period of eight months, Cerebras went from a lower valuation than SambaNova's ($8.1 billion versus $11 billion) to an IPO that multiplied its market capitalization roughly seven times. The market moved quickly on Cerebras, and it could do the same for SambaNova.
It's also possible that another semiconductor company makes a bid for the firm. SambaNova's CEO has been noncommittal on whether the company will actually hold an IPO. Intel was already in talks last year to acquire the company for $1.6 billion, including debt, according to Bloomberg. The recent capital raise appears to put that figure well below what the company can fetch in private markets—and maybe soon in the public market.