Analysis
Nvidia has announced new customers for its Vera CPU and Groq-derived LPX inference racks, The Information reported, giving the market its first named commercial evidence behind the roughly $20 billion bet Nvidia made absorbing Groq's inference technology.
Vera is Nvidia's next-generation custom CPU, designed to sit alongside its GPUs in a full server rack rather than relying entirely on third-party x86 processors from Intel or AMD. Pairing a custom CPU with custom inference silicon lets Nvidia sell a more complete, more margin-dense rack to hyperscalers and neoclouds, rather than ceding any part of the server bill of materials to a competitor.
The LPX racks build directly on Groq's architecture, covered earlier this week alongside the first published Groq 3 LPU benchmarks. Where those benchmarks offered a performance read, named customers offer a commercial one -- someone is paying for the hardware, which is a different and more reliable signal than a vendor-published chart.
“Custom CPU and inference silicon both move in that direction.”
- Vera CPU -- Nvidia's custom processor competing with Grace, and indirectly with ARM-based server chips from AWS's Graviton and Google's Axion lines
- Groq LPX racks -- Nvidia's inference-optimized hardware, addressing the latency-sensitive workloads Groq's original architecture was built for
- AMD, Intel -- the incumbent CPU suppliers whose share of AI server sockets narrows every time a hyperscaler adopts a custom alternative
The strategic logic mirrors what Nvidia has done with networking and storage: rather than being just the GPU vendor inside someone else's server design, Nvidia increasingly wants to sell the entire rack, capturing margin on every component rather than just the accelerator. Custom CPU and inference silicon both move in that direction.
The open question remains adoption breadth. New customers announced alongside a product launch are frequently early adopters or reference accounts secured with favorable terms, not yet a representative sample of the broader market. Whether Vera and LPX racks displace meaningful share from AMD's Instinct-adjacent CPU offerings or from hyperscaler-designed silicon will take several more quarters of disclosed deployment data to assess.
