Positron AI

@mitesh711 nnbw.com/news/2026/apr/…

Our CEO, @mitesh711, spoked with Rob Sabo at NNBW about where Positron is heading in 2026. The point we keep coming back to: in this industry, product cadence matters as much as product quality. What customers teach us on Atlas is already shaping Titan, and that loop is how a team of under 50 competes with trillion dollar incumbents. Shipping today, going live with a hyperscaler next, and hiring fast to keep up. Full article in the comments. More to come.

NVIDIA has a monopoly. Thomas Sohmers is unbothered. @JordanNanos sits down with the Co-Founder & CTO of @Positron_AI to talk FPGA inference, LPDDR memory, and running 16T parameter models on a single box. @trsohmers Tune in: youtu.be/B8O3pLcX2w4

Most people think AI scaling is just about more compute. They’re wrong. ❌ The real bottleneck? Memory. In this exclusive clip from GTC 2026, Thomas Sohmers (CEO of @positron_ai) breaks down how their new "Asimov" chip is flipping the script: - 16 Trillion parameter models running on a single server. - Millions of tokens in context length. - Massive realized memory bandwidth that reduces the need for complex networking. If you want to know what the next phase of LLM infrastructure actually looks like, you need to watch this. 👇 Full conversation in the replies!

Check out the interview with our CEO @mitesh711 by @ndahad from @eetimes at the Arm AGI CPU Launch Event: youtube.com/watch?v=3yOnil…

Great to have @positron_ai be part of the @Arm AGI CPU launch today, with us shipping Positron Atlas powered by arm later this year. Positron’s next-generation Asimov accelerator is designed from the ground up with Arm IP Inside, and our Titan Server built specifically around ARM AGI.

Positron AI has been named one of Fast Company's Most Innovative Companies of 2026. As our CEO put it: "Inference hardware is going to get specialized, and the market will target the best perf/$ and perf/watt." That conviction built Atlas, and it's what's driving everything we're building next.

Some exciting announcements tomorrow along with @positron_ai!

@mitesh711 codestory.co/podcast/e9-mit…

Our CEO @mitesh711 sat down with Code Story to talk origins, architecture, and the road ahead. He walks through the journey from Lambda to Positron, the memory first thesis behind our architecture, and the philosophy of shipping early and staying capital efficient. Give it a listen.

Thrilled to have @positron_ai partnered with @Oracle, and honored to have Oracle's CEO @ClayMagouyrk feature us in their earnings call today. Together, we'll be delivering leading performance per dollar and per watt generative AI solutions to the world!

Being in SF is like being in Wuhan right before the pandemic Something is happening, it's gonna hit everywhere but so few people know it

We will be world’s first terabyte plus memory density silicon and will be in production in 2027. Another cool feature that weaver allows us is to have configurable silicon sku for amount of memory, so instead of only one set amount of memory per chip, we can have anywhere from 576GB to 2304GB per chip based on customer’s application and this can be done at system build out time.

Grateful to Brian Baumann and the New York Stock Exchange for celebrating our $230M Series B with an incredible banner, and to Kyt Dotson, John Furrier, and the SiliconANGLE team for the great coverage. Raising at a $1B+ valuation in just under two years is a team effort, and moments like this remind us why we're building: to deliver inference that's efficient, affordable, and finally makes GPUs optional. Thank you to our investors, our team, and every customer betting on purpose-built hardware. Onto the next milestone.

We're excited to announce our $230 million Series B at a valuation exceeding $1 billion. The round was co-led by @ARENA_pw , @jumptrading , and Unless, with strategic investment from Qatar Investment Authority (QIA), @Arm , and Helena. Existing investors, @Atreidesmgmt, Valor Equity Partners, @dfjgrowth, Resilience Reserve, and Flume Ventures, also participated. This funding accelerates our roadmap from shipping Atlas inference systems today to our next-generation Asimov custom silicon, targeting tape-out in late 2026 and production in early 2027. We're building the future of energy-efficient AI inference. More to come.

Joined Bloomberg TV today to discuss our $230M Series B. The key question from Caroline Hyde and Ed Ludlow: how do you take on NVIDIA? Our answer: memory architecture. We're building systems with 2.3TB of attached memory while NVIDIA's next-gen Rubin launches with 384GB. For decode-heavy workloads like video generation, code generation, and reasoning models, memory bandwidth and capacity is the bottleneck. The thesis is simple: as inference spending overtakes training, the architecture has to change. Memory capacity becomes the unlock. bloomberg.com/news/videos/20…

@mitesh711 eqvista.com/interview-with…

Our CEO, @mitesh711 , recently joined Eqvista to unpack the shift from GPU-dominated infrastructure to purpose-built inference hardware. The core thesis? Energy constraints and utilization rates will define who wins the next era of AI. Everyone's obsessed with training bigger models. But the real bottleneck? Getting them into production without going broke. Inference is the silent budget killer. It's the workload running 24/7: chatbots, code generation, agents. And the hardware powering it wasn't designed for the job. GPUs excel at training, but for inference, memory bandwidth becomes the constraint. Utilization hovers around 30%. Companies pay for compute they barely touch while power costs spiral. Positron took a different approach: hardware built specifically for transformer inference, achieving over 90% memory utilization and multi-model concurrency that lets enterprises consolidate entire racks without touching their code. The conversation covers why we proved the architecture on FPGA before taping out ASIC, how customers like Cloudflare validated performance in production, and what it actually takes to challenge Nvidia where they're weakest. Five years from now, the winners will be the ones solving for watts per token today. Full interview in comments.

Our CEO @mitesh711 joins @credosemi next Monday (Nov 10, 8AM PT/11AM ET) to discuss the real bottleneck in AI inference: memory. Building at scale? Don't miss this. Register: lnkd.in/g4SJzHBX