Positron AI Raises $875M for Memory-First Inference

Positron AI has raised $875M in Series C financing at a $5B post-money valuation, giving the Reno, Nevada company enough capital to take its Asimov inference chip through tapeout and begin turning its Titan system from a technical plan into manufactured racks. The announcement was made on September 10, 2026, seven months after Positron raised a $230M Series B at a valuation above $1B.

The financing arrived in two parts. A $375M Series C at a $3.5B pre-money valuation was co-led by NEA, Andra Capital, Atreides Management, Valor Equity Partners, and Dylan Patel's SemiAnalysis Capital. A Series C-1 of up to $500M was led by NEA and Jim Clark, the founder of Silicon Graphics and Netscape.

The amount gets attention, but the use of it explains the wager. Positron is financing the expensive distance between a chip architecture and a customer-ready system: tapeout, emulation, power infrastructure, memory supply, manufacturing capacity, integration, and go-to-market work. Its first-generation Atlas product is already shipping, and Positron says more than 50 racks are being deployed at Oracle Cloud Infrastructure.

What Positron AI Is Financing

Positron says the round will fully fund the tapeout of Asimov, its next-generation custom silicon, on TSMC's N3P process by the end of 2026. Production is targeted for the second half of 2027, a later and more specific schedule than the early-2027 target published with the company's Series B in February.

The company also plans to build a 2 MW-plus engineering data center and emulation platform. That facility will support the work between design and deployment, where hardware teams have to validate silicon behavior, software, power, thermals, system integration, and manufacturing assumptions before customers inherit the consequences.

Asimov is designed with 288 GB to 2.304 TB of LPDDR5X memory per chip. Titan will combine four to eight Asimov chips into a system intended for multi-trillion-parameter models, context windows beyond 10M tokens, video, and persistent agent workloads. Those capacity and efficiency figures are company targets for products expected in 2027, not independent production results.

Why Positron Built Around Memory

AI training made raw compute the celebrity. Inference, the work of running trained models for every assistant, agent, search, and application request, has a different set of operating pressures. As models grow and applications hold more context, memory capacity, usable bandwidth, latency, power, and serving cost become part of the customer experience.

Positron's architecture is built around that claim. The company says Asimov can realize more than 90% of available memory bandwidth on transformer workloads and use commodity LPDDR5X rather than depending on scarce HBM and advanced CoWoS packaging. That design could offer more memory per chip and fit both air-cooled and liquid-cooled data centers, but its performance and cost comparisons remain Positron's projections until production systems face broader customer testing.

The business logic is easier to understand than the chip diagram. If a customer can keep larger models, longer contexts, and more persistent agent state close to the accelerator, the system may spend less time moving data and more time serving tokens. For cloud providers and enterprises, that can affect the number of users supported per rack, the power required per workload, and the revenue or productivity each system can produce.

Atlas Gave the Roadmap Operating Proof

Positron was co-founded in 2023 by Thomas Sohmers and Edward Kmett. Sohmers serves as co-founder and CTO, while Kmett is founder and chief scientist. Mitesh Agrawal joined from Lambda as CEO and now leads the company's commercial scale-up.

The team brought Atlas, an FPGA-based transformer inference server, to market before committing its next generation to custom silicon. Positron publishes benchmark claims of more than 4x performance per watt and more than 3x performance per dollar versus selected Nvidia Hopper configurations. Those comparisons are company-reported and workload-specific, but Atlas has supplied something every new chip company needs: a product customers can deploy while the custom architecture is still being built.

Positron says Parasail uses its Oracle Cloud capacity to power an inference service, while Jump Trading and i3d.net are additional production customers. Jump Trading also co-led the company's Series B after testing Atlas, turning a customer relationship into investment evidence. By September, those deployments had expanded to more than 50 racks at Oracle, a milestone Reuters independently reported.

That record matters because the market is crowded with claims about replacing or outperforming GPUs. Positron is making a narrower commercial argument: purpose-built inference hardware can win workloads where memory, power, latency, and system economics matter more than general-purpose flexibility. Atlas does not prove Asimov will work, but it gives the company deployment data, customer feedback, and a software path before the new chip arrives.

The Investor Group Is Buying a Manufacturing Schedule

The financing expands Positron's investor base while bringing several investors directly into governance. Forest Baskett of NEA, Gavin Baker of Atreides Management, Thomas Jermoluk from Jim Clark Office, and Dylan Patel are joining the board. Their backgrounds span semiconductors, growth investing, operating companies, and detailed analysis of AI infrastructure.

Additional investors include DFJ Growth, Qatar Investment Authority, Resilience Reserve, Arena Private Wealth, Natural Capital, Helena, 1517 Fund, Flume Ventures, Unless, Boardman Bay Capital Management, Fincadia Advisors, Banyan Ventures, U First Capital, VentureTech Alliance, Hudson River Trading, Cisco Investments, and Naver Ventures. QIA separately confirmed its participation and noted that it also invested in the Series B.

This syndicate is pricing demand for AI chips while financing supply commitments, a foundry schedule, a 2 MW-plus engineering environment, a production ramp, and the organization required to support customers. Specialized hardware does not scale because a benchmark looks good; it scales when design, software, manufacturing, power, service, and customer procurement arrive in the same window.

What the $875M Round Has to Prove

Positron's valuation rose from more than $1B in February 2026 to $5B in September, according to the company and Reuters. The new price reflects confidence in a shift from AI training toward inference, but it also advances the proof burden. Asimov must tape out on schedule, Titan must move from simulations into production, and the claimed economics must survive workloads selected by customers rather than by a marketing page.

The competitive field includes Nvidia's broad GPU platform, hyperscaler chips such as Google's TPUs and AWS Trainium, and other specialized inference systems. It also includes software and serving companies trying to improve economics above the silicon layer, a market movement visible in Baseten's reported infrastructure financing. Positron does not need to replace every GPU to build a substantial business; it needs to win enough memory-heavy inference workloads to justify a new system architecture and the operational change that comes with it.

The capital gives Positron room to make that case with physical systems instead of projections. By the second half of 2027, the company expects Asimov chips to be entering production and Titan racks to begin carrying the lessons learned from Atlas at Oracle. Between now and then, the financing will be consumed by the unglamorous work that decides whether a semiconductor roadmap becomes infrastructure customers reorder.