Positron AI Raises $875M for Chips That Cut AI Inference Costs
Positron AI announced on September 10, 2026 that it raised $875 million in a Series C financing round at a $5 billion post-money valuation. The round was co-led by NEA, Atreides Management, Valor Equity Partners, Andra Capital, Dylan Patel’s SemiAnalysis Capital, and Jim Clark, the founder of Silicon Graphics and Netscape. The funding is earmarked for one goal: taping out Asimov, a custom AI inference chip that Positron claims can deliver 26 times more tokens per dollar than Nvidia’s Blackwell GB300 NVL72.
That claim has not been validated in silicon yet. But the round size, the investors, and the architecture bet behind it signal that a serious alternative to Nvidia-centric AI inference is closer to reality than most enterprise teams realize.
What Positron Actually Builds
Positron AI is a 2023-founded company based in Reno, Nevada. It sits in a specific and strategically important position in the AI stack: it builds purpose-built hardware for AI inference, not training.
This distinction matters. As enterprises move from piloting AI to deploying it at production scale, inference costs, not training costs, become the constraint. Every API call, every agentic loop, every customer-facing AI interaction runs on inference. The rise of enterprise AI pricing pressure has made per-token cost a boardroom metric.
Positron’s current product, Atlas, is already shipping. More than 50 racks of Atlas are deployed at Oracle Cloud Infrastructure, with Parasail using that capacity to power its inference service. Jump Trading and i3d.net are also listed as production customers. This is not a paper product. The team has operating experience at hyperscaler-adjacent scale, which informed Asimov’s design.
The LPDDR5X Bet
The architectural thesis at Positron is that HBM, the high-bandwidth memory inside Nvidia H100 and Blackwell GPUs, has become a bottleneck rather than an advantage for inference workloads. HBM is expensive, constrained by CoWoS packaging supply chains, and increasingly difficult to procure in volume. Etched’s similar $700M raise earlier this year for transformer-specific ASIC silicon highlighted the same dynamic.
Positron’s answer is LPDDR5X, the memory technology used in high-end smartphones. The tradeoffs look counterintuitive until you run the math:
- LPDDR5X is commodity supply with no CoWoS dependency
- Positron’s architecture achieves more than 90% of available memory bandwidth
- Asimov configs span 288 GB to 2,304 GB of memory per chip
- A Titan node with eight Asimov chips provides up to 18.4 terabytes of memory and 23.68 terabits per second of bandwidth
That memory capacity is the key number. Large frontier models, and especially the next generation beyond current-generation 16 trillion parameter scale, cannot run on systems with insufficient memory. Positron is betting that memory capacity and bandwidth will matter more than raw compute throughput for serving the coming generation of frontier models, and that LPDDR5X delivers both without the supply chain risk.
The Titan System: Numbers That Matter
| Specification | Titan (Asimov) | Nvidia GB300 NVL72 |
|---|---|---|
| Memory per node | Up to 18.4 TB LPDDR5X | ~13.5 TB HBM3e (72 GPUs) |
| Memory bandwidth | 23.68 Tbps | ~67 Tbps (72 GPUs total) |
| Max model size (single node) | 32 trillion parameters | Approx. 13 trillion parameters |
| Max context window (single node) | 10 billion tokens | Approx. 1 million tokens |
| Claimed tokens per dollar | 26x better (simulation) | Baseline |
| Memory type | LPDDR5X (commodity) | HBM3e (constrained supply) |
| Cooling | Air or liquid | Liquid required |
| Tapeout process | TSMC N3P (3nm) | TSMC CoWoS-L (5nm) |
Positron’s 26x figure is from internal simulation data; production benchmarks against real workloads will be the acid test when Asimov enters production in 2H 2027.
Who Is Backing This and Why It Signals Real Conviction
The investor list is not generic venture capital. Dylan Patel of SemiAnalysis is among the most credible independent analysts of AI chip architecture. His capital in this round means he reviewed the architectural claims and found them credible enough to back with money, not just analysis. Jim Clark, who founded both Silicon Graphics and Netscape, brings direct experience building hardware and platform companies that reshaped their respective eras.
Strategic investors include Cisco Investments, Hudson River Trading (a quantitative trading firm that runs enormous compute workloads), and Naver Ventures. These are operating businesses with real inference cost pressures, not passive financial players.
The valuation progression reinforces the signal. Positron raised $230 million in February 2026 at a $1 billion valuation. Seven months later it has raised at $5 billion, a 5x increase. The capital market is pricing a real probability that purpose-built inference silicon reaches commercial viability before the next wave of frontier model releases.
Three Enterprise AI Implications
1. Model size constraints will loosen. The memory capacity of Titan enables serving 32 trillion parameter models in a single node. As frontier labs release next-generation models (post-GPT-6 Astra, post-Claude Fable 5.x), enterprise teams today constrained by memory capacity will have options beyond hyperscaler APIs. This changes the self-hosted vs API cost calculus.
2. Long-context applications become economically viable. Context windows of 10 million tokens in a single node matter for enterprise workloads: ingesting full codebases, legal document repositories, customer history archives. Today these workloads require chunking, retrieval pipelines, and quality tradeoffs. Titan’s native capacity eliminates the chunking tax.
3. The Nvidia dependency story now has real competition. Etched, Cerebras, and now Positron represent three distinct architectural approaches to inference silicon. Etched targets transformer-specific throughput. Cerebras has wafer-scale advantage on single-chip model serving. Positron targets memory-first, energy-efficient token economics. Enterprise procurement teams building 2027-2028 AI infrastructure plans now have multiple credible alternatives to evaluate.
What Comes Next
Asimov tapes out on TSMC’s N3P node at the end of 2026. If the tapeout is clean and performance matches simulations, production silicon ships in the second half of 2027. That timeline means enterprise customers cannot rely on Titan for current deployments but should factor it into procurement planning that extends into 2027 and beyond.
The first-generation Atlas platform is available now at Oracle Cloud Infrastructure. Teams that want to evaluate Positron’s memory-bandwidth approach on real inference workloads can do so through Parasail’s inference service, which runs on Atlas.
The broader pattern here is a structural shift in where AI capex flows. Training was the first wave. Inference optimization, including custom silicon, software, and system architecture, is the second wave. For enterprise teams managing AI costs at scale, that wave is arriving faster than most procurement timelines anticipated.
If your team is modeling AI infrastructure costs through 2027, talking to us about architecture before locking in vendor commitments is worth the conversation. The inference silicon landscape is moving fast, and the cost implications of a wrong call compound quickly at agentic AI volumes.
Sources: Positron AI press release (primary); Reuters; SiliconANGLE; Converge Digest; AWS Bedrock announcement