Oxmiq Labs: Raja Koduri’s new silicon symphony | Festina Lente - Your leading source of AI news | Turtles AI

Raja Koduri returns to the field with Oxmiq Labs: a startup rewriting GPU architecture "from atoms to agents," built on licensable RISC-V IP and a "software-first" software ecosystem (OXCapsule, OXPython). He discusses chiplets, AI multimodality, memory modalities, and partnerships—all without any actual silicon yet.

Key Points:

• Illustrious founder with 500 years of combined team experience
• OXCORE: Modular RISC-V GPU core for maximum scalability
• OXPython: CUDA-Python compatibility on non-Nvidia hardware
• IP and software-only business model, funded at $20 million

It’s as if an old orchestra conductor—Raja Koduri, with dusty symphonies of Polaris, Vega, Navi, and Arc behind him—found himself once again at the podium, this time orchestrating sonatas not for gaming PCs but for ocean-scale data centers and multimodal AI payloads. His baton: Oxmiq Labs, a hybrid of visionary architectural thinking and pragmatic financial efficiency. The stage has been silent for two years, but now the curtain rises: we have OXCORE, a RISC-V computational heart that blends scalar, vector, and tensor computing in a stackable module, coupled with DRAM or SRAM memory under cryptic labels like MCB/CCB—leave the reader to guess, because so far the only diagram is more suggestive than detailed.

The direction? Not toward desktop enthusiasts, but toward data center brains, where scientific and multimodal AI rules the roost. And the soundtrack is all software-first: OXCapsule as the connective tissue between heterogeneous hardware, and OXPython as the expert translator capable of running CUDA-Python applications as-is, on non-Nvidia hardware without a hitch—via TenStorrent first and foremost.

The business model is more "licensing" than "manufacturing": without silicon in its pocket, Oxmiq avoids costly tape-outs and multi-million dollar EDA tools, relying instead on seed capital—€20 million raised, with MediaTek among its backers—and initial software revenue. Not only that, but it’s also opening up to emerging markets like India, where access to IP could be a springboard to sovereign, zettascale AI infrastructures.

And as long as the cocktail of "nano-agents," in-memory computing, "light transport," and modular chiplets generates more curiosity than clarity, it remains to be seen whether this collection of buzzwords will transform into true innovation or an elegant, instrument-less melody on stage. In any case, the score has been delivered, even if the musicians have yet to play it.

It’s a debut that leaves unanswered questions, but the conductor is one of those who knows how to keep time.