At this week’s GTC, Nvidia’s most important task is not to unveil one more fast chip but to make the post-Blackwell roadmap feel concrete through 2028.
- At this week’s GTC, the real story is unlikely to be a single hero chip; it should be a sharper roadmap from Blackwell into Rubin, Rubin Ultra and Feynman. Nvidia has already positioned the March 16 keynote as a full-stack AI event, and it formally launched Rubin in January as the post-Blackwell platform; public reporting from last year’s GTC placed Rubin Ultra in 2027 and Feynman in 2028. What matters now is whether Nvidia turns those names into a visible cadence of racks, fabrics and workload-specific products, rather than a sequence of distant chip labels.
- Expect the keynote to emphasize that Rubin is not one chip but a six-part system architecture: Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4 and Spectrum-6. Nvidia says Rubin is designed to cut inference token cost by up to 10x versus Blackwell and that partner systems begin shipping in 2H26; its product materials also stress software continuity, so the migration path matters as much as raw performance. The deeper point is that GTC should focus on system-level continuity—how customers move from GB200/GB300 into Rubin, and then into Rubin Ultra / Kyber-scale deployments, without breaking the stack.
- From there, the likeliest expansion is a broader, disaggregated inference family. Nvidia has already introduced Rubin CPX, a new class of GPU for million-token coding and generative video, with dedicated compute trays and NVL144 CPX racks; Dynamo already supports disaggregated serving with vLLM and SGLang; and BlueField-4 now underpins the new Inference Context Memory Storage Platform for KV-cache and long-horizon agent memory. Put together, that points to a GTC message that inference is being broken into separate jobs—training, prefill, decode, context storage and orchestration—each with its own optimized silicon or software layer.
- The most important unannounced product to watch is a Groq-derived low-latency inference chip—effectively an Nvidia answer to the LPU idea. Groq and Nvidia disclosed a non-exclusive inference-technology licensing agreement in December, along with the move of founder Jonathan Ross and senior Groq engineers to Nvidia, and Reuters reported on February 28 that Nvidia plans to unveil a new AI inference system at GTC incorporating a chip designed by Groq. If it appears, expect it to be framed not as a GPU replacement but as a decode coprocessor: an SRAM-heavy, deterministic engine for fast response generation that sits beside Rubin or CPX inside larger racks.
- A second underappreciated thread is Intel. Nvidia and Intel have already announced a formal collaboration under which Intel will build Nvidia-custom x86 CPUs for AI infrastructure, and Nvidia says those processors will integrate into its platforms via NVLink / NVLink Fusion; the same agreement also covers x86 SoCs with Nvidia RTX GPU chiplets for PCs. That makes GTC a plausible venue for more detail on a co-designed enterprise CPU path, especially for customers that want Nvidia racks without abandoning the x86 universe. In practice, this would extend Nvidia’s role from accelerator supplier to defining the entire control plane of the rack, even when the host CPU bears someone else’s logo.
- Networking and optics may end up being the most consequential announcements after compute. Nvidia has already launched Spectrum-X Photonics and Quantum-X silicon-photonics switching, while recent March announcements with Coherent and Lumentum show the company is locking up laser, packaging and optical-component capacity well ahead of the next scaling wave. The key watchpoint is whether GTC treats co-packaged optics as a premium option or as a structural requirement for Rubin Ultra / Kyber-era clusters, where Nvidia’s own OCP materials point to 800 VDC infrastructure and megawatt-class racks. If that tone shifts from “better networking” to “necessary fabric for future scale,” it will be one of the conference’s most important signals.
- Finally, quantum looks set to graduate from curiosity to architecture. Nvidia’s inaugural Quantum Day at GTC 2025 has become a full Quantum Computing / Quantum Computing and HPC track for 2026, and the company has already launched NVQLink to connect GPUs with quantum systems across 17 quantum builders and nine scientific labs. That does not make quantum a near-term revenue event next week; it means Nvidia is likely to present CUDA-Q, NVQLink and hybrid quantum-classical infrastructure as the next adjacency to the AI factory. The cleanest reading of GTC, then, is that Nvidia will try to show one continuous machine: Rubin for 2026, Rubin Ultra/Kyber for 2027, Feynman beyond that, Groq-derived decode silicon, Intel-based host options, photonic switching, and quantum at the edge of the roadmap.