Vera Rubin is the product. Feynman is the roadmap. The $10 billion investment spree is the strategy. GTC 2026 is where all three converge.
NVIDIA's GPU Technology Conference opens March 16 at SAP Center in San Jose. Jensen Huang's keynote is Monday at 11 a.m. Pacific. He has promised a chip that will "surprise the world" and technology "never unveiled before."
The known story is the Vera Rubin platform — NVIDIA's successor to the Blackwell generation, pairing an 88-core Vera CPU with Rubin GPUs carrying 288 GB of HBM4 memory each. NVIDIA has already begun shipping early samples to select customers. On paper, Vera Rubin is the product GTC was built to showcase.
But multiple data points from the past two weeks suggest GTC 2026 is about something much larger than a single chip.
Multiple sources, including Korean outlet Chosun Biz (cited by Fudzilla and WCCFTech), report that Huang's "surprise" is Feynman — the architecture generation after Vera Rubin. The key details as currently reported:
Feynman will be the first chip built on TSMC's A16 (1.6nm) process node, which uses Super Power Rail backside power delivery. NVIDIA is reportedly TSMC's first and, during initial high-volume manufacturing, likely only A16 customer. Other chipmakers may require architectural revamps to adopt the node — giving NVIDIA a meaningful head start.
A more speculative but potentially seismic detail: Feynman may integrate Groq LPU (Language Processing Unit) technology via hybrid bonding, analogous to AMD's X3D stacking. If confirmed, this would mean NVIDIA is no longer content to iterate on GPU architecture alone — it would be incorporating fundamentally different compute paradigms.
Mass production is not expected until 2028, with customer shipments in 2029–2030. So GTC will be a reveal-and-roadmap event, not a product launch. That distinction matters: Feynman at GTC is less a product announcement than a subscription to NVIDIA's future, designed to lock in customer loyalty and deter investment in alternatives years before the silicon arrives.
Even the nearer-term Vera Rubin story has an underappreciated subplot. NVIDIA reportedly had to revise Vera Rubin's HBM4 memory bandwidth specifications downward after SK Hynix and Samsung struggled to hit the original 22 TB/s bandwidth target. NVIDIA subsequently bumped specs back up by roughly 10% to maintain distance from AMD's Instinct MI455X, which runs at approximately 1.7 kW versus Vera Rubin's 2.3 kW power envelope.
This supply-chain tension is genuinely significant. It reveals that NVIDIA's roadmap is now bottlenecked not by its own silicon design capability — which remains industry-leading — but by its memory suppliers. The pace of AI compute is being set by HBM4 yield rates in South Korean fabs. That's a strategic vulnerability NVIDIA has never faced at this scale before, and it's worth watching whether Huang addresses it at GTC or papers over it.
The chip story alone would justify GTC coverage. But NVIDIA has simultaneously been deploying capital at a pace that would make a sovereign wealth fund blush:
The pattern is unmistakable: NVIDIA is investing across every layer of the AI infrastructure stack — cloud compute (CoreWeave, Nebius), chip design tools (Synopsys), optical interconnects (Lumentum, Coherent), and model development (OpenAI, Thinking Machines Lab). This isn't diversification. It's vertical integration by checkbook.
NVIDIA is also preparing to unveil NemoClaw at GTC — an open-source enterprise AI agent platform. CNBC reports it's being pitched to Salesforce, Cisco, Google, Adobe, and CrowdStrike. If the investments represent the hardware side of NVIDIA's stack ambitions, NemoClaw represents the software side: a play to become the default orchestration layer for enterprise AI agents running on NVIDIA hardware.
The real GTC story isn't any single product. It's that NVIDIA is simultaneously showing up as chipmaker (Vera Rubin samples shipping), roadmap architect (Feynman reveal), infrastructure financier ($10B+ in strategic investments), and software platform (NemoClaw). No company in tech history has tried to own this many layers of the AI stack at once.
Jensen Huang has spent a decade turning NVIDIA from a GPU company into the computing backbone of modern AI. GTC 2026 is where that thesis gets its most ambitious statement yet — and where its risks become most visible. The HBM4 supply chain can't keep up. The power budgets are climbing. The capital deployments assume a future where every enterprise runs AI workloads at scale. If that future arrives on schedule, NVIDIA will be the most important technology company in the world. If it doesn't, the company will have built an empire on a timeline that slipped.
Monday's keynote will tell us which version of the future Jensen Huang is betting on. Based on everything we've seen this month, the answer is: all of them, simultaneously.
Sources: Tom's Hardware (Vera Rubin sample delivery), TechPowerUp (HBM4 spec revisions), Fudzilla & WCCFTech (Feynman/A16 details via Chosun Biz), CNBC (NemoClaw, March 10; Nebius, March 11), NVIDIA GTC official site.