Excited to announce Volantis's $88M Series A. We are solving Al's memory bottleneck by using optics, enabling chips with huge amounts of fast & cheap memory. By boosting both the memory bandwidth and capacity per chip by orders of magnitude, we enable ultra-fast inference (up to 10,000 tps/user) for large models (>10T) - with low $/tok to boot. Initially, this will enable insanely fast agents - think coding agents that finish in minutes or even seconds instead of hours. More excitingly, optics is a fundamentally scalable way to increase memory systems. Not 2X/year, but by orders of magnitude across new generations. This will enable a structurally new Al industry, including restarting scaling laws, holding entire repos in context windows & more. Our team has pioneered many core semiconductor technologies: the 1st CoWoS product, early HBM, the 1st silicon photonics CPO systems, the 1st high volume tunable VCSELs, the 1st processors to directly communicate using light & more. We’ve already sent data >10× farther than equally tiny electrical wires inside a chip package. Our next iteration is already taped out and targets world-record bandwidth density over relevant distances, read more: volantissemi.ai/news-insight…

dare i say fidget hardware is the best piece of hardware i’ve ever seen

We've been following @semiDL and team for over a year and are thrilled to be a part of their Series A. Tapa's assembled a world-class team tackling the core problems of memory capacity and bandwidth that hold back model intelligence. Excited to see this team build!

Proud to be an angel investor in @semiDL and Volantis! We as an industry have a lot of work to do to keep scaling intelligence, and we need all the creative minds we can throw at the problem! Moving data around is a central problem and these guys are working on something truly breakthrough.

Well will soon look back on the days of 50 tok/s as dial-up speeds

Insane place to live

SF is so gorgeous Undoubtedly one of the best cities in the world

Someone should design a scalable hardware system that brings SRAM-system speeds to large, future models! The world can do much, much better than 300 tok/s/user

Slower than TPU?

Be different, stand out: use Grok to make your decks instead

Today this is unusable in practice because it runs 24X slower than real time New computing hardware will fix that, enabling real time frontier model inference

The impact of ultrafast inference for big models at a low cost is going to be titanic

Transformers can learn to predict sequences of natural data without access to any data because they can be trained to approximate Solomonoff induction, the ideal next-token predictor. This is the basis of our project described in x.lingyaoai.com/michaelyli_/status/210… . Here,I explain why you shouldn’t be shocked that this is possible. A really fun project co-led with @michaelyli_ and @KfirDolev and great co-authors @gbruno_dl @ANourya @noahdgoodman, and @YoavLevine

Very cool work

Can an LM, starting from random init (!!), learn to generate all of its pretraining data? Introducing Self-Play Pretraining with Zero Data. Two models start from random initialization: a generator proposes programs for a universal Turing machine and a learner trains on their outputs. We never train on any real data, but see predictable scaling on natural datasets: zero-shot val loss on images, text, audio, and melodies decreases predictably with self-play compute. And the learner develops in-context learning capabilities. A fun proof-of-concept, co-led with @AdityaCowsik and @KfirDolev and co-authors @gbruno_dl, @ANourya @noahdgoodman, and @YoavLevine.

If you’d rather have 200+ HBM per package in 2027, consider Volantis :)

If you’d rather reach 10,000 GB per chip by 2027, join us at Volantis

If only we could use light to solve the memory bottleneck for the entire AI industry :)

Prediction: As the bottleneck moves to the physical layer for AI, 1st principles simulators will become more and more important, increasing demand for HPC-like compute systems