In just over 2 years, Positron AI has introduced innovations that put it in the race with top competitors, including NVidea, and has gained early investor support.
As AI moves from model training to widespread deployment, power consumption has become a key challenge.
Positron AI, a Nevada-based startup specializing in AI inference hardware, has announced an oversubscribed $230 million Series B round. With that funding, the company has a post-money valuation above $1 billion, establishing it as a new unicorn in AI infrastructure.
Thomas Sohmers, Barrett Woodside, and Edward Kmett founded the AI startup in 2023. Mitesh Agrawal became CEO in January 2025, and Barrett has since joined Opal Security as Growth and Strategy Lead.
The funding round was co-led by ARENA Private Wealth, Jump Trading, and Unless, with strategic participation from Qatar Investment Authority (QIA), Arm, and Helena.
Existing investors, including Valor Equity Partners and DFJ Growth, also participated.
The announcement coincided with Web Summit Qatar, highlighting the global interest in next-generation compute platforms.
Why this matters: inference is becoming the real cost center
While the AI industry has focused on training clusters in recent years, most enterprises face cost and capacity constraints during inference—the routine tasks of generating tokens, answering queries, summarizing documents, running agents, and serving media models.
Improving performance-per-watt and prioritizing memory-centric system design, rather than simply increasing compute power, is key to making AI widely accessible.
CEO Agrawal sees energy availability as a critical factor for AI deployment and states that the company’s next-generation roadmap aims for significant efficiency improvements over Nvidia’s upcoming Rubin platform in targeted workloads.
The company also highlights memory as a key limitation for modern inference, particularly in video, trading, large-scale models, and applications with extensive context windows.
Atlas today, Asimov next
Positron reports that it is already shipping Atlas, an inference system designed for rapid deployment and scalability.
The company notes that Atlas is developed and manufactured with a focus on the American supply chain, which may appeal to buyers seeking predictable capacity and reduced procurement risk.
Positron’s larger goal is Asimov, its next-generation custom silicon. The company plans to complete tape-out in late 2026 and begin production in early 2027. They have set an ambitious schedule in a field where delays are common.
Positron promotes a “memory-first” approach to inference. Asimov is designed to deliver terabyte-scale memory per accelerator and multi-terabyte memory per system, with Titan referenced as the next-generation system concept.
This aims to support long-context LLMs, agentic workflows, and advanced video and media models constrained by current memory limitations.
A customer becomes a co-lead investor
A key development in this round is Jump Trading’s decision to co-lead after initially deploying Atlas as a customer.
According to the release, Jump’s CTO notes that inference bottlenecks are increasingly tied to “memory and power”. He reports that testing showed significantly lower end-to-end latency than a similar Nvidia H100-based system across the evaluated workloads.
Positron views this customer-to-investor transition as strong validation of market demand.
The competitive backdrop
Positron is entering a market dominated by Nvidia. But it is also seeing increased competition from startups and alternatives focused on inference specialization.
With funding in place, the company can pursue its core mission. That is to challenge Nvidia’s AI chip dominance, which is driven by rising demand and growing interest in non-GPU solutions for production inference.
The leadership team believes that memory capacity, bandwidth, and energy efficiency will determine leadership in the next phase of AI infrastructure.
