The Brief

NVIDIA’s GTC 2026 opens March 16 in San Jose with over 30,000 attendees from 190 countries, as CEO Jensen Huang prepares a keynote expected to debut the Vera Rubin architecture and unveil what the company calls “several new chips the world has never seen.” The conference follows NVIDIA’s announcement of $4 billion in combined strategic investments in optical interconnect firms Coherent and Lumentum, signalling that the next constraint on AI scaling is no longer silicon — it is light.

The Report

NVIDIA’s annual GPU Technology Conference returns to San Jose from March 16–19, spanning ten venues including the SAP Center, where Jensen Huang will deliver a keynote at 11 a.m. PT on the opening day. More than 30,000 attendees from over 190 countries are expected — developers, researchers, enterprise leaders, and representatives from over 240 NVIDIA Inception startups.

The conference programme runs to over 1,000 sessions covering AI factories, robotics, digital twins, quantum computing, and enterprise AI deployment, alongside nine full-day workshops, 60 hands-on labs, and 150 poster presentations. Scheduled presenters include representatives from Google DeepMind, Meta, Microsoft, OpenAI, Tesla, and Adobe. A pregame broadcast featuring the CEOs of Perplexity, Mistral AI, and LangChain will precede the keynote. An investor Q&A follows on March 17.

Huang has indicated the keynote will include “several new chips the world has never seen.” Industry analysts widely expect the formal debut of NVIDIA’s Vera Rubin architecture, which pairs a proprietary Vera CPU with sixth-generation HBM4 memory and is reported to deliver up to five times the inference performance and 3.5 times the training performance of the current Blackwell generation. There is additional speculation about an early showing of the Feynman architecture, scheduled for 2028, which would be built on TSMC’s 1.6nm process and incorporate silicon photonics directly into the chip design.

The photonics thread connects to NVIDIA’s most consequential pre-conference move. On March 2, the company announced parallel multiyear strategic partnerships with Coherent Corp. and Lumentum Holdings, investing $2 billion in each. The agreements include multibillion-dollar purchase commitments for advanced laser and optical networking products, with both companies expanding U.S.-based manufacturing capacity. The investments address what analysts have identified as a structural bottleneck: global production capacity for indium phosphide lasers — the specialised material required for high-speed optical transceivers — currently falls short of demand by a factor of two.

Optical interconnects serve two distinct functions in AI infrastructure. Scale-up optics handle chip-to-chip communication within GPU clusters, where copper signalling is approaching its physical bandwidth limits. Scale-out optics manage rack-to-rack connectivity across data centres requiring hundreds of thousands of optical links. NVIDIA’s dual-vendor approach secures complementary capabilities — Coherent’s strength in co-packaged optics integration alongside Lumentum’s expertise in high-power laser components.

The $4 billion combined investment extends a relationship between NVIDIA and Coherent that spans two decades. Jensen Huang framed the context in characteristically expansive terms: “In the age of AI, software runs on intelligence with tokens generated in real time by AI factories for every interaction.”

GTC 2026 will also be available virtually, with the keynote livestreamed without registration.

The Angle

The $4 billion in optics investment is the more revealing story than whatever Huang unveils on stage on Monday — because it tells you where the actual constraint is. NVIDIA has spent the past three years as the company that builds the engines. Now it is moving backwards down the supply chain to secure the roads between them.

The pattern is familiar. In 2023 and 2024, advanced packaging bottlenecks at TSMC — specifically CoWoS capacity — constrained GPU production more effectively than any competitor or regulator managed to. NVIDIA learned the lesson. When the next chokepoint emerged — indium phosphide laser fabrication, a niche process with a handful of global producers and roughly half the capacity the market requires — the company did not wait for the market to solve it. It wrote cheques large enough to reshape the supply chain directly.

What makes this structurally interesting is the implication for where AI infrastructure is heading. Copper interconnects work at current scale. They do not work at the scale NVIDIA is building toward — clusters of tens of thousands of GPUs drawing gigawatts of power, generating data volumes that require optical signalling not as an upgrade but as a physical prerequisite. The Feynman architecture, expected in 2028, reportedly integrates silicon photonics at the chip level. NVIDIA is not investing in optics because it is fashionable. It is investing because without it, the next generation of hardware has no nervous system.

There is a quieter dimension worth noting. Both Coherent and Lumentum are being funded specifically to expand U.S.-based manufacturing. At a moment when semiconductor supply chain geography is the subject of intense geopolitical negotiation, NVIDIA is building domestic optical capacity with the urgency of a company that has run the numbers on what concentration risk looks like when your product is the substrate of an entire industrial transition.

Thirty thousand people will gather in San Jose next week to see new chips. The chips will be impressive. But the thing that determines whether those chips can actually be connected into systems that function at the scale the next decade requires was announced two weeks earlier, with considerably less fanfare, in the form of purchase orders for lasers.

The infrastructure that will carry the weight of the next generation of intelligence is being built not in silicon but in light. The companies that control the photons will matter as much as the companies that design the transistors.