Jensen Huang doesn’t do things halfway. Fresh off the NVIDIA GTC 2026 spectacle where he redefined what a GPU keynote can look like, NVIDIA is now writing $4 billion in checks for a technology most people haven’t heard of: photonics.
The recipients? Lumentum and Coherent — two companies that make components for moving data with light instead of electrons. NVIDIA is putting $2 billion into each, in the form of multi-year purchase agreements. Not investments. Purchase agreements. Jensen is buying product, not stock — and that tells you everything about where this is going.
Wait, What Exactly Is Photonics?
Here’s the quick version for those of us who didn’t major in optics: photonics is the science of transmitting data using light rather than electrical signals through copper wire. You know how fiber optic cables can carry a staggering amount of internet data across continents? That’s photonics.
Now imagine applying that same principle inside a data center — between the chips, between the racks, between the servers. Instead of copper cables that heat up, degrade signal over distance, and cap out at certain bandwidths, you get optical interconnects that are faster, cooler, and can carry far more data per second.
For AI workloads — which are basically continuous, massive data shuffling operations between thousands of GPUs — this is kind of a big deal. Training a frontier model like GPT-5 or whatever comes next isn’t bottlenecked by raw compute anymore. It’s bottlenecked by how fast data can move between GPUs. Fix that, and everything gets faster.
Why NVIDIA Is Doing This Now
NVIDIA already makes the GPUs. They make the networking (InfiniBand, via Mellanox). They make the software stack (CUDA). They’re increasingly involved in the full-stack AI infrastructure story — from chip design to software frameworks to cloud services. This photonics move is the next logical step.
By locking in multi-year purchase agreements with Lumentum (lumentum.com) and Coherent (coherent.com), NVIDIA is effectively securing the supply chain for a component that every serious AI data center will need. It’s not just about getting components for themselves — it’s about making sure those components exist at scale when the market demands them.
And the market will demand them. A lot. According to reporting from Bloomberg and Reuters, the AI infrastructure buildout isn’t slowing down — if anything, 2026 is doubling down on 2025’s already absurd investment pace.
The Context: Everyone Is Building. Fast.
This photonics announcement doesn’t exist in a vacuum. A few other things happened around the same time:
- Amazon just committed ~$40 billion to new data center infrastructure in Spain. Not the US. Spain. That’s how global this AI buildout is getting.
- The US government is weighing new export caps on NVIDIA H200 chips to China — potentially capping purchases per customer. This is the ongoing geopolitical chess match over who controls AI compute, and it’s reshaping supply chain strategies across the industry.
- Hyperscalers are spending hundreds of billions collectively in 2026 on AI infrastructure. Microsoft, Google, Meta, Amazon — they’re all building at a pace that would have sounded like sci-fi five years ago.
In this context, NVIDIA’s photonics bet makes perfect sense. The bottleneck isn’t GPUs anymore — NVIDIA has that covered with Blackwell and whatever comes after it. The bottleneck is interconnects. The wiring. How do you get data from one GPU cluster to another fast enough that the whole system doesn’t waste cycles waiting?
You replace copper with light.
What This Means for the Photonics Industry
For Lumentum and Coherent, this is a massive validation. These are not startup companies — they’ve been in the optical components business for years — but they were somewhat niche players in the telecom supply chain. AI is turning them into critical infrastructure vendors almost overnight.
When NVIDIA signs a $2 billion purchase agreement with your company, every other hyperscaler suddenly wants to know if you have capacity for them too. This is how industrial ecosystems form: one dominant player de-risks the technology, creates supply chain certainty, and suddenly everyone follows.
We saw this playbook with TSMC and advanced chip manufacturing. We saw it with Samsung and memory. We’re watching it happen right now with photonic interconnects.
The Bigger Picture: NVIDIA’s Full-Stack Ambition
There’s a pattern here that’s worth stepping back to appreciate. NVIDIA started as a GPU company. Then they got into networking (Mellanox acquisition). Then software (CUDA, cuDNN, the whole NVIDIA AI Enterprise stack). Then cloud services (DGX Cloud). Now optical interconnects.
At every level of the AI infrastructure stack, NVIDIA is showing up. Not just as a chip vendor, but as an architect of the entire ecosystem. Jensen Huang has talked about this explicitly — the idea of NVIDIA as an “AI factory” builder, not just a chip maker.
This photonics move fits that narrative perfectly. If you want to sell complete AI data center solutions — the computing equivalent of a turnkey factory — you need to have an answer for every bottleneck. Right now, optical interconnects are one of the biggest remaining bottlenecks. So NVIDIA is fixing it.
Should You Care?
If you’re building AI applications, running inference at scale, or just watching where the industry is heading: yes.
Faster interconnects mean larger models can be trained more efficiently. They mean inference clusters can be bigger and more coordinated. They mean the AI applications of 2027-2028 will have capabilities that would be physically impossible today — not because the algorithms got smarter, but because the hardware got faster at moving data.
It’s one of those infrastructure improvements that’s invisible to end users but completely changes what’s possible. Like when we moved from hard drives to SSDs — same data, same software, completely different experience.
AI is about to get a similar upgrade. And NVIDIA is writing the check to make it happen.
Sources: NVIDIA Newsroom, Bloomberg, Reuters. This article is based on announcements from March 3, 2026.