Artificial intelligence workloads are simply overwhelming traditional datacenter infrastructure, and the only way forward for the next generation of AI datacenters is through new tech like Coherent’s optical solutions. We’re seeing a fundamental change in how data gets moved and processed in these insane compute environments, well beyond just incremental upgrades.
Key Takeaways
- Coherent’s indium phosphide (InP) photonics platform is the backbone for building the high-speed, power-sipping 800G+ optical transceivers that AI datacenters need.
- Switching from copper to optical interconnects, especially with linear-drive pluggable optics (LPO), slashes power draw and latency inside AI clusters.
- By controlling everything from materials to modules (vertical integration), Coherent maintains tight control over performance and can guarantee a stable supply of its optical components.
- To handle the ballooning power and bandwidth density of new AI processors, Coherent is developing next-gen architectures like direct-drive optics and co-packaged optics (CPO).
- With its work on silicon carbide (SiC) substrates, the company also makes power electronics that are needed to manage the massive energy demands of AI infrastructure.
The Imperative for Optical Innovation in AI
AI datacenters are a different beast entirely from your standard cloud or enterprise setup. The insane data volumes for model training and the need for thousands of GPUs to chat with near-zero latency creates a ridiculous appetite for bandwidth and power. Copper wiring works fine for short hops, but in these massive GPU clusters, it just falls apart, the signal degrades, and the retimers and digital signal processors (DSPs) needed to clean it up eat way too much power. At this scale, optical solutions are the only real answer.
This is where a company like Coherent comes in, with its background in materials science and photonics. They have a serious leg up because they’ve been making the core components, laser diodes, photodetectors, optical fibers, for ages. Engineering these parts from scratch and controlling the whole manufacturing line is a world away from just buying and assembling them. That vertical integration means they can squeeze out performance and power optimizations that are impossible with off-the-shelf parts. It all comes down to the physics of light and electrons, where every nanometer and every picosecond matters when you’re building for AI.
Advanced Photonics Platforms: Indium Phosphide and Silicon Photonics
Coherent’s whole strategy for AI datacenters hinges on its indium phosphide (InP) photonics platform. The material’s direct bandgap properties are what make it so good for high-speed, high-power optical parts. It’s just really efficient at generating and detecting light. That efficiency is exactly what you need to build the 800G and 1.6T optical transceivers that are now the go-to for linking GPUs in AI clusters. And it’s not a small market, a report from LightCounting Market Research projects that demand for these 800G+ transceivers in AI will blow past $5 billion by 2028 LightCounting Market Research, which tells you everything you need to know about where the industry is heading.
But it’s not just about InP. Silicon photonics is also a huge piece of the puzzle. The main draw there is that it uses the same manufacturing processes as regular computer chips, so it’s cheap to scale up, especially for things like putting optics right on the silicon. Coherent’s real smart play is using a hybrid approach, they use InP for the parts that need its high-power performance (like the lasers and detectors) and then use silicon photonics for the more passive stuff like waveguiding. Having expertise in both gives them the freedom to pick the right tool for the job in any given datacenter design, like pairing an InP light source with a silicon modulator to get the best of both worlds in performance and cost.
Being able to actually make these complex optical integrated circuits in-house is what really sets a company like Coherent apart. It’s about having deep control over the yield, reliability, and performance from the ground up. And when your components are the things determining the efficiency and uptime of a multi-billion dollar AI supercomputer, that kind of control is non-negotiable.
The Rise of Linear-Drive Pluggable Optics (LPO)
A huge step forward for AI datacenters has been the move to linear-drive pluggable optics (LPO). Your standard optical transceiver has a digital signal processor (DSP) inside to clean up signal noise. That DSP works, but it’s a power hog and adds latency, two things you absolutely can’t afford in an AI cluster where every milliwatt and nanosecond is being counted. The DSP just becomes a major bottleneck.
LPO modules get around this by just ripping the DSP out of the transceiver entirely. They rely on really clean analog circuits to keep the signal strong, which can cut power consumption by up to 50% on an 800G transceiver compared to one with a DSP, based on what we’re seeing in the industry. That’s a massive power saving when you’re talking about thousands of these things in a single AI cluster. And with no DSP, you also cut latency, which means your AI models train faster. Coherent saw this coming and was one of the early companies pushing and developing LPO because they knew the old architecture wouldn’t cut it for AI.
Of course, the trick with LPO is keeping the signal clean without a DSP to do digital cleanup. It demands top-notch analog parts and a very, very careful system design. The fact that LPO is not just working but becoming the preferred choice for new AI interconnects says a lot about how far optical and electrical engineering have come. There’s a clear trend emerging: for the short-reach, high-density connections inside AI racks, getting as close as possible to direct analog transmission is the best path for performance and power efficiency.
Beyond Pluggable: Co-Packaged Optics (CPO) and Direct-Drive Optics
LPO is solving today’s problems, but everyone’s looking at even tighter integration for the future. That’s where co-packaged optics (CPO) comes in. This is the idea of putting the optical transceiver right on the same package as the network switch ASIC or the AI processor. By doing that, you get rid of the electrical traces running between the chip and the optical module, which slashes power consumption and lets you cram way more ports into the same space. Coherent is pouring money into CPO R&D, bringing its wafer-scale manufacturing and packaging know-how to the table.
CPO comes with its own set of serious engineering headaches. When you pack hot electrical and optical parts so close together, managing the heat is a nightmare. On top of that, you need brand new ways to test these super-integrated modules to make sure they’re reliable. But the payoff in power, latency, and bandwidth density is so big that the industry is investing heavily anyway. Yole Group projects the CPO market, mostly for AI applications, will grow to over $1 billion by 2030 Yole Group. It’s a long-term bet that’s going to completely change how datacenters are designed.
Another path people are exploring is direct-drive optics. Think of it as LPO on steroids: you integrate the optical drivers right into the network ASIC or AI processor itself. This gets rid of the pluggable module entirely, which shortens the electrical path even more and saves more power. With its deep knowledge of lasers and detectors, Coherent is in a great spot to supply the optical engines for both CPO and direct-drive setups. These new architectures are forcing a total rethink of the line between the electrical and optical worlds to keep up with AI’s demands. It’s a tough but fascinating time to be an optical engineer.
Powering AI with Silicon Carbide (SiC)
The amount of energy AI datacenters consume is just off the charts. While moving data efficiently with optics is half the battle, you also have to get power to the whole setup without wasting tons of it. That’s where Coherent’s work in silicon carbide (SiC) substrates comes into play. SiC is a wide-bandgap semiconductor that blows traditional silicon out of the water for high-power, high-frequency jobs. This makes it perfect for the power electronics, the voltage regulators, power supplies, and inverters, that have to handle the insane electrical load of AI servers and GPUs.
Because SiC devices can run hotter, switch faster, and have lower conduction losses than silicon parts, they’re much more efficient. When a single AI datacenter is pulling down tens or hundreds of megawatts, saving even a few percent on power conversion adds up to massive energy savings and a smaller cooling bill. By making its own SiC wafers and devices, Coherent can supply components that improve the total energy efficiency of AI infrastructure. This power management side of things is often overlooked, but it has a direct effect on opex and the building’s environmental impact. Frankly, without this kind of progress in power management, the insane scaling we’re seeing in AI compute just wouldn’t be possible.
Conclusion
By working on everything from advanced photonics and new interconnects like LPO to high-performance SiC power electronics, Coherent is hitting the AI datacenter problem from all angles. Their deep roots in material science and their control over the manufacturing process give them a real edge in providing the speed, power efficiency, and scale that AI workloads demand.
Why use optical instead of copper in AI datacenters?
Optical connections provide way more bandwidth, have less latency, and use less power than copper. For the massive data transfers and nonstop processing in AI, copper just can’t keep up.
What’s special about Coherent’s indium phosphide (InP) platform for AI?
The InP platform is how Coherent builds its high-speed, high-power lasers and detectors. These are the key parts inside the 800G and 1.6T transceivers needed to move data quickly between GPUs and other AI chips.
What are Linear-Drive Pluggable Optics (LPO) and why do they matter for AI?
LPO modules are optical transceivers without the power-hungry DSP chip. This simple change can cut power use by 50% and reduces latency, both of which help AI models train faster and more cheaply.
What is co-packaged optics (CPO)?
CPO means putting the optical I/O right on the same package as the main processor or switch chip. This gets rid of short, power-hungry electrical connections, which saves a ton of power, frees up space for more ports, and cuts latency in future datacenter designs.
Where does silicon carbide (SiC) fit into the AI datacenter picture?
Coherent uses its SiC tech to build more efficient power electronics. For an AI datacenter, this means power supplies and voltage regulators that waste less energy as heat. That improves the building’s overall power efficiency and lowers the cooling bill.
