When NVIDIA CEO Jensen Huang steps onto the stage at San Jose’s SAP Center on Monday, March 16, he is expected to do far more than tout the Vera Rubin platform already confirmed to be in mass production. Industry analysts and a constellation of leaks point toward a first public glimpse of Feynman — the architecture slated for 2028 that will push AI compute into previously uncharted territory.

GTC 2026, running through March 19, has been billed by Huang himself as a showcase of technology the world has “never seen before.” While much of the event’s commercial substance will focus on the Rubin platform’s new GPUs and the HBM4 memory ecosystem, the semiconductor industry’s gaze is drifting further ahead — to a chip named after the Nobel Prize-winning physicist Richard Feynman.

A New Node: TSMC’s A16 and Backside Power Delivery

The most technically significant detail confirmed by multiple credible sources is the process node. Feynman is expected to be manufactured on TSMC’s A16 — a 1.6nm-class process that represents a foundational leap in semiconductor engineering. If confirmed, NVIDIA would become the first, and potentially only, major customer to adopt A16 at volume.

A16 is notable not merely for its transistor density but for introducing Super Power Rail (SPR) — a backside power delivery network that routes power from beneath the silicon rather than competing for routing space on the front side. This architecture improves both power efficiency and thermal headroom, characteristics increasingly critical as AI accelerator TDPs escalate dramatically.

TSMC has been actively expanding A16 production capacity in anticipation of NVIDIA’s demand, with mass production targeted for the second half of 2026. Feynman chips manufactured on this process would then be available for data center integration in 2028 — with customer shipments potentially slipping into 2029 or 2030 depending on how NVIDIA allocates supply.

One additional manufacturing note worth tracking: credible industry reports indicate NVIDIA is considering outsourcing specific, lower-risk elements of the Feynman package — primarily the I/O die — to Intel, using Intel’s 14A or 18A process nodes alongside Intel’s EMIB (Embedded Multi-die Interconnect Bridge) advanced packaging. This would not affect the GPU core itself, which remains anchored at TSMC, but is intended to diversify supply risk and manage the high cost of CoWoS packaging.

Power Consumption: A Looming Thermal Frontier

The current Blackwell architecture already pushes close to 1,000W for single-card configurations, with the dual-chip Blackwell Ultra reaching approximately 1,400W. Feynman is expected to exceed 1,000W in a single-card configuration, and dual-chip variants may approach 2,000W — figures that make the architecture essentially impossible to deploy with conventional air cooling.

This is not an anomaly but a deliberate trajectory. NVIDIA’s Vera Rubin platform has already confirmed liquid cooling as its exclusive thermal solution. Feynman will almost certainly continue this direction, accelerating the data center industry’s transition to fully liquid-cooled rack infrastructure.

Taiwanese cooling and power supply manufacturers stand to benefit significantly from this transition. Analysis from TrendForce identifies companies including Auras Technology, AVC, Delta Electronics, and Lite-On Technology as major beneficiaries in the hardware upgrade cycle Feynman will drive. Server ODMs — including Quanta, Wistron, Inventec, and Foxconn — are likewise positioned as key partners in the next infrastructure generation.

Silicon Photonics: Signaling a New Interconnect Paradigm

Several industry analyses note that Feynman may introduce silicon photonics for the first time in an NVIDIA product — using optical signals rather than electrical signals to transmit data between components at extreme bandwidth with lower power draw. This would address one of the core bottlenecks of today’s multi-chip AI accelerators: the energy and latency cost of moving data.

Silicon photonics is a technology NVIDIA has backed through investments including Ayar Labs, a co-packaged optics startup that recently secured significant additional funding. Whether Feynman will implement on-package photonics or use it at the rack interconnect level remains technically unconfirmed.

The Groq LPU Question: Speculation, Not Fact

Some reports have suggested Feynman could be the first NVIDIA platform to integrate Groq’s LPU (Language Processing Unit) hardware, as inference latency becomes a critical competitive dimension. The theory envisions a hybrid bonding approach — analogous to AMD’s X3D stacked cache — that embeds LPU hardware directly into the Feynman package.

This should be treated as informed speculation, not confirmed architecture. The engineering complexity of integrating an external dataflow processor into a tapeout-stage GPU is considerable. NVIDIA has also historically separated its training and inference product lines for strategic reasons, and a standalone LPU product line would give the company greater flexibility in addressing inference markets without compromising the Feynman GPU’s primary data center positioning. Whether GTC brings any clarification on this front will be among the most closely watched outcomes of the week.

What GTC Will — and Won’t — Reveal

Expectations should be calibrated carefully. The Vera Rubin platform is where the commercial narrative sits: Samsung and SK Hynix will both appear at GTC to showcase HBM4 collaboration achievements, with Samsung having shipped HBM4 for Rubin GPUs in February. Rubin’s confirmed entry into mass production is the headline with near-term revenue implications.

Any Feynman showcase at GTC is expected to follow the pattern of NVIDIA’s Vera Rubin reveal at GTC 2025: early samples on static display, a general architecture overview, and broad capability claims — not a shipping product announcement. The purpose is strategic: alert the supply chain, anchor investor expectations, and begin telegraphing data center infrastructure requirements two years in advance.

For gaming customers, the timeline is clear: Rubin-based gaming cards are expected in the second half of 2027, and Feynman gaming products are unlikely before 2029 at the earliest. The near-term gaming story at GTC will center on RTX-series updates and DLSS technology, not next-generation architecture transitions.

This article synthesizes publicly available reports from TrendForce, Wccftech, The Register, TradingKey, and Tom’s Guide, cross-referenced against verified semiconductor industry data. All forward-looking claims about unannounced products reflect current industry consensus and are subject to revision following NVIDIA’s GTC keynote on March 16, 2026.