Fast inference could be a $550 billion market by 2030, Mizuho Securities says, giving specialized chip maker Cerebras (CBRS) a shot at one of AI computing's biggest segments. The bank kept its outperform rating and $300 price target in an Oct. 4 note reviewed by The Latent.

Vijay Rakesh, a managing director at Mizuho, puts the fast-inference market at about $1 billion in 2025. By 2030, he expects it to account for about 20% of all AI workloads. Cerebras' chips enable the faster processing.

His revenue estimates for Cerebras are about $885 million this year, $3 billion in 2027 and $13.5 billion in 2029. Most of that would come from the company's "expanding pipeline of customer engagements, anchored by key partnerships with OpenAI and AWS, with potential additional customers driving future upside into 2027/28E," the note says.

At $300, Cerebras would trade at about 12 times Mizuho's estimate of 2028 sales. The target is roughly 80% above Friday's $166.43 close. Rakesh also laid out a $465 bull case and a $165 bear case, writing that Cerebras "must continuously execute on its roadmap to maintain and widen its [specialized chip] technology moat."

Buyers pay more for speed

AI tokens delivered at high speed command prices about six to 10 times higher than standard tokens, Rakesh wrote, "making it a higher-revenue/margin segment."

Take Anthropic's fast mode for Claude: it runs about 2.5 times faster and costs six times as much. Rakesh also cited OpenAI's $500-per-month Pro plan, which includes access to the Ultrafast tier rolled out at the company's DevDay event last week.

Cerebras said in August that it powers Ultrafast for OpenAI's GPT-5.6 Sol. The company says the fast mode can generate up to 750 output tokens per second, or up to 14 times the speed of OpenAI's standard mode. The latest model offered at that tier is GPT-6.1 Sol, currently OpenAI's second-best model on The Latent's benchmark.

Early demand from trading firms and the Pentagon

Jane Street, Citadel, Jump and Hudson River Trading are among the market makers and high-frequency trading firms Mizuho expects to drive early demand. The bank also sees fast inference gaining traction in cybersecurity monitoring and defense systems that require real-time inference, pointing to defense technology companies Anduril and Shield AI.

One contract is already in hand: Cerebras and Ranovus won $45 million from DARPA to build an AI test bed for battlefield simulation, with a goal of sub-second inference, according to the note.

Rakesh sees Meta (META) as a possible next inference customer. He also says Amazon Web Services, or AWS, plans to offer its Nova models in Bedrock in the first quarter of 2027 using its Trainium chips alongside Cerebras CS-3 systems.

More revenue rests on OpenAI

OpenAI alone is expected to account for more than 80% of Cerebras revenue, by Mizuho's estimate. That follows a year in which Abu Dhabi-based G42 and the Mohamed bin Zayed University of Artificial Intelligence supplied about 86% of revenue. The bank expects those two customers' share to fall below 40% this year, from that 2025 level.

Cerebras went public in May at $185 a share. Last week, the stock slipped below its IPO price after SemiAnalysis reported that OpenAI's newer GPT-6.1 Sol Ultrafast tier runs on Nvidia GPUs. More shares also became available for sale as IPO lockup restrictions expired.

OpenAI CEO Sam Altman wrote on X that "Cerebras is a close partner, and we have a deep engagement pushing on the frontiers of speed."

According to Mizuho, Cerebras plans to ship CS-4 in the fourth quarter, doubling the current generation's token speed, and CS-5 in 2027. OpenAI has its own inference hardware in development, too. It said in August that its first inference chip beat Nvidia's GB300 in lab tests.