Positron AI raises $875m at $5bn valuation for memory-first inference
Positron AI, a Palo Alto-based developer of memory-first AI inference hardware and software, has closed an $875 million funding round at a $5 billion valuation, with Liberty Global Tech Ventures joining as a strategic investor. The round was co-led by NEA, Valor Equity Partners, Atreides Management, Andra Capital, SemiAnalysis Capital and Jim Clark, and was described by the company as oversubscribed.
Founded in 2023, Positron has built its Atlas inference system around LPDDR5X memory rather than the high-bandwidth memory (HBM) that dominates competing inference accelerators. The company says its architecture achieves more than 90% memory bandwidth utilisation and that Atlas units can be installed in existing data centre racks without modifications to cooling infrastructure. Named customers include Oracle, Jump Trading and Parasail.
The deal and the product roadmap
Liberty Global's participation is positioned as a strategic rather than purely financial investment, given the company's telecoms infrastructure footprint across Europe. Bobbie Maltiel, Partner at Liberty Global Tech Ventures, said the ability to deliver token usage "at the lowest possible cost is key to the full benefits being delivered for society," adding that Positron has "a relentless focus on this."
Beyond the shipping Atlas platform, Positron has announced Asimov, a custom silicon design that is due to tape out in late 2026 with volume production scheduled for the second half of 2027. Asimov will power Titan, described as a multi-terabyte-memory inference system aimed at long-context and next-generation AI workloads. The company did not disclose pricing, throughput benchmarks or contract values in the release.
Market context and competitive positioning
The inference compute market is one of the most intensely competitive corners of the AI infrastructure stack. Nvidia's H-series and B-series GPUs, paired with HBM and CoWoS packaging, currently dominate production deployments at hyperscalers. Positron's explicit design choice to avoid HBM and CoWoS is both a supply-chain hedge and a cost argument: HBM supply has been constrained by TSMC's CoWoS advanced-packaging capacity, pushing lead times and unit costs higher across the industry.
Several well-funded startups are pursuing alternative inference architectures, including companies building custom ASICs for transformer workloads, processing-in-memory approaches and dataflow accelerators. What distinguishes Positron's positioning is the air-cooled compatibility claim: as data centre operators face power density constraints and liquid-cooling retrofits remain expensive, an accelerator that works within existing thermal envelopes carries tangible operational value.
Bloomberg Intelligence has projected the generative AI infrastructure market at $2.3 trillion by 2032, with inference compute representing a significant share of that spend as training-to-inference ratios shift in maturing deployments. A $5 billion valuation for a three-year-old hardware startup implies investors expect Positron to capture meaningful market share before hyperscalers or Nvidia close the architectural gap.
Regulatory and supply-chain read-across
Positron's deliberate avoidance of CoWoS packaging reduces its exposure to US export controls that restrict advanced semiconductor packaging technology to certain jurisdictions, an increasingly relevant consideration as BIS tightens rules around AI chip supply chains. The company's European rollout ambitions, supported by Liberty Global's network of telecoms joint ventures across the continent, could also help it navigate the EU AI Act's infrastructure-layer obligations as those provisions mature.
With custom silicon in the tape-out pipeline and named hyperscale-adjacent customers already live on Atlas, Positron's next disclosures to watch are Asimov silicon validation results, volume production commitments and any expansion of the named customer list beyond the three firms cited in this round.