NVIDIA pledges up to $3.2 billion in Corning for AI optical fiber

NVIDIA and Corning build a US optical backbone for AI

NVIDIA and Corning Incorporated announced a multi-year supply and equity agreement on May 6, 2026. NVIDIA paid an upfront $500 million for warrants to acquire up to 18 million Corning shares and can invest up to $3.2 billion in equity over three years. In return, Corning will build three new US plants dedicated entirely to optical components for NVIDIA's AI systems.

What the deal actually covers

According to consistent reporting by CNBC and Tech Startups on May 6 and 7, 2026, the three new plants will multiply Corning's US optical manufacturing capacity roughly tenfold and create at least 3,000 jobs. The exercise price for the first warrant tranche is 180 dollars per share according to CNBC. On the day of the announcement, Corning shares rose about 12 percent, while NVIDIA gained about 6 percent.

Co-packaged optics at the heart of the strategy

The deal centers on shifting from copper to optical interconnects directly at the chip, known as co-packaged optics (CPO). In today's GPU racks for AI training, copper traces hit speed and energy limits once bandwidth crosses a few hundred gigabits per second. Optical fiber can carry those signals over the same distances at significantly lower power loss.

Part of a larger NVIDIA investment program

According to Benzinga and CNBC on May 9, 2026, NVIDIA has now committed more than $40 billion in equity stakes and warrants across partners, including roughly $30 billion in OpenAI alone. The Corning deal sits alongside the agreement also disclosed this week with data center operator IREN, worth up to $2.1 billion.

Why it matters

Optical interconnects are the often overlooked link between GPU counts and actual usable AI compute. When hyperscalers scale their clusters to hundreds of thousands of accelerators, the quality of the fiber backplane helps determine whether training runs stay stable for weeks or fail with bit errors. A domestic, predictable supply chain for these components reduces dependence on imports from East Asia and aligns with the industrial policy laid out in the CHIPS and Science Act.

For Corning, the agreement means three years of demand visibility plus an anchor customer. For NVIDIA, it secures exclusive access to manufacturing capacity that no other supplier currently offers at this scale. CEO Jensen Huang said, according to CNBC, that the partnership would "revitalize American manufacturing".

In plain language

Picture data inside an AI data center as parcels in a giant sorting hub. Until now, most parcels rolled along copper rails that get hot and trip up when they go too fast. Optical fiber is like a new rail network made of glass, where parcels race as flashes of light with almost no loss. NVIDIA and Corning are now jointly building three new factories that produce exactly this glass rail network.

A practical example

A German cloud provider in the Frankfurt area is building an AI training cluster with around 8,000 NVIDIA GPUs in 2027. With an old design based on copper interconnects, the team would spend about 18 percent of the cluster's energy on internal links and live with bit error rates that make longer training runs on 200-billion-parameter models unreliable. With co-packaged optics racks from the Corning-NVIDIA plants, industry studies suggest the share of energy used for interconnects drops to 6 to 8 percent, and training time for a language model falls by an estimated 11 percent. In the first year of operation, that saves more than 4 million euros in electricity and several weeks of time to production.

Scope and limits

First, the $3.2 billion figure is an upper limit over three years, not an immediately paid amount. How much is actually drawn down depends on share prices, delivery targets, and order volume.

Second, optical fiber only helps where bandwidth and power are the bottleneck. Smaller inference clusters or edge deployments only benefit at the margins and will keep running on copper through the end of the decade.

Third, the supply chain for laser drivers, module housings, and specialty glass remains globally distributed. A US fiber-drawing plant does not guarantee full sovereignty over co-packaged optics modules.

SEO and GEO keywords

NVIDIA, Corning, Co-Packaged Optics, AI optical fiber, Optical Networking, AI Data Center, GPU Interconnect, CHIPS Act, Jensen Huang, Hyperscaler, AI Infrastructure, US Manufacturing 2026.