We're all fully aware of what's going on with the supply chain shortages on RAM and SSD. [Kingston](https://www.pcgamer.com/hardware/memory/kingston-sounds-the-ssd-pricing-alarm-as-the-company-has-seen-a-246-percent-increase-in-nand-wafer-prices-with-the-biggest-increase-within-the-last-60-days/) sent out a warning stating that NAND costs are up 246% since early 2025 and says RAM/SSD prices are likely to keep rising through 2026 rather than come back down. [SK Hynix](https://www.techpowerup.com/344063/sk-hynix-forecasts-tight-memory-supply-lasting-through-2028) is telling investors that the global memory shortage could drag on until 2028, even as they and Samsung print multi year high profits on AI memory. NAND suppliers are already reporting that their 2026 enterprise SSD allocations are basically sold out, with double-digit price hikes locked in. On top of this, hyperscalers are only getting about [70% of the server DRAM they order](https://www.tomshardware.com/pc-components/storage/server-dram-prices-surge-50-percent), even after accepting contract price increases of up to 50% in a single quarter. Then you got Sam Altman doing his best impression of Dr. Evil prebooking [40% of the world's DRAM](https://www.mooreslawisdead.com/post/sam-altman-s-dirty-dram-deal) wafers with money that has yet to be seen.
I say all this to say, HBM/DRAM/flash has become the new oil for AI, and everyone is fighting over the same barrels. What I have come to notice is that real choke point isn’t just how much memory you can buy, it’s how efficiently you can use the memory you already have, and how fast you can shove bits between GPUs, DRAM pools, and SSDs without the copper plumbing melting down and draining power grids of every last drop of electricity they can possibly produce.
In light of the RAM/SSD shortage, this has become a conversation about **finding solutions that can maximize the compute potential of the currently existing infrastructure to meet current AI demands**
This is where POET’s photonic engines come in: they don’t manufacture a single RAM chip or SSD, but they act like adding extra lanes to the AI data highway, letting data centers squeeze more training and inference out of every scarce byte of HBM/DRAM/flash.
**So POET has two main products:**
[The Optical Interposer](https://www.poet-technologies.com/technology) \- think of this as an “optical motherboard tile/chiplet.”
[Starlight](https://www.poet-technologies.com/products/starlight) \- an external laser farm that feeds those tiles with light.
Today’s optics usually look like this:
* You’ve got a discrete transceiver module in a metal “gold box.”
* Inside that box: lasers, modulators, drivers, receivers, lenses, fiber alignment, etc.
* Every module is handheld, robot aligned, and tuned. Slow, expensive, power-hungry.
POET’s interposer basically says fuck the gold box
* They take silicon and III-V lasers and build everything into a flat wafer-scale platform.
* Instead of doing micron-level alignment one module at a time, they pattern the waveguides in lithography and drop the chips on with cheap pick-and-place tools.
* Fibers line up passively, so there's really no heavy and expensive manufacturing process involved
* The parts on the interposer are configureable, so this gives hyperscalers flexibility to choose which components and tier of components they would like on the interposer. think of the interposer like a complex lego block and you can piece together whichever legos you want to it.
To visualize, the interposer sits between the hot compute chips (GPUs/ASICs/NPUs) and the outside world (network, memory pools, storage) and replaces a bunch of copper with light paths.
**Starlight is more of a accessory to the interposer.**
* Instead of putting hot, fragile lasers on the GPU/package, they park the lasers off to the side in a dedicated module.
* That Starlight box pumps clean laser light over fiber into a bunch of POET interposer engines sitting near the compute.
**The benefits are:**
* Keeps lasers away from 700W space heaters (GB200, H100, etc.).
* One laser farm can feed many engines instead of every module lugging its own lasers.
* Better reliability, easier maintenance, and lower $/bit because you share the light source across multiple links.
**So together these two components work like this:**
Starlight = shared light source
Interposer = cheap, configurable, mass-produced optical engine tile that is easily scalable. This is probably the biggest sell for me on this product, as this gives hyperscalers more flexibility to customize the interposer
**So how is different from the other companies that offer a photonics solution?**
Most of the big incumbents (Broadcom, Coherent, Lumentum, etc.) are still:
* Shipping discrete pluggable modules (400G/800G, etc.).
* Using a lot of active alignment and labor-intensive packaging.
* Typically putting lasers inside every module, which:
* Wastes power
* Adds heat right where you don’t want it
* Doesn’t scale nicely to 1.6T / 3.2T and giant AI fabrics.
**POET’s edge vs “generic” silicon photonics:**
* Wafer-scale, passive alignment which lowers capex, lowers assembly cost, and is easier to ramp volume.
* Allows hybrid integration: can mix and match best-of-breed lasers with silicon in one platform.
* Remote lasers (Starlight) which results in fewer total lasers and less heat on the package.
And if you're wondering if any of this technology has even been validated, look no further than Marvell Technologies, [who purchased one of POET's largest customers](https://investor.marvell.com/news-events/press-releases/detail/1000/marvell-to-acquire-celestial-ai-accelerating-scale-up-connectivity-for-next-generation-data-centers), Celestial AI, for up to $5+ billion. Marvell's main reason for acquiring Celestial AI was for its Photonic Fabric, an optical interconnect architecture that is built atop POET's Optical Interposer and uses POET’s Starlight as a remote laser source.
In other words, a serious, blue-chip semi company and at least one hyperscaler ([AWS](https://mlq.ai/earnings/highlight/MRVL-marvell-partners-strategically-with-aws-dd4124/#google_vignette)) have already done the homework on this architecture and decided it’s real enough to bet billions and multi-year roadmaps on. You don’t build a next-gen AI fabric around external lasers and wafer-scale photonics if you’re not confident the vendors in that stack can actually meet the eye diagrams, BER targets, thermal budgets, and manufacturing yields you need at hyperscale. The Marvell/Celestial deal is basically the the big dawgs in the room saying: “Yes, this class of photonic solution is how we’re going to attack the I/O wall.”
Also, Celestial AI's CEO, [David Lazovsky](https://www.linkedin.com/in/david-lazovsky-734058a), was the Executive Chairman at POET
How does any of this actually help with the RAM/SSD shortage?
POET obviously doesn’t fab a single DRAM chip or SSD, so the way they solve the shortage is indirect, but very real: they change the economics of how much useful work you can squeeze out of every byte of memory and every GPU you already own.
Right now, hyperscalers are in a ridiculous situation. They’re paying insane prices for HBM, DRAM, and enterprise SSDs, but a lot of that capacity is effectively stranded. You have GPUs sitting on islands of local HBM, servers with their own DRAM, storage boxes full of flash, and the copper between them can’t move data fast enough. So to keep latency acceptable, the big guys end up replicating the same data all over the place, sharding models, caching datasets on every node, and overbuilding memory just so the network doesn’t become a brick wall. That’s fine when RAM is cheap. It’s a killer when HBM/DRAM/flash prices are up 200% over the last month.
On the power side, POET’s own FAQ for their 800G/1.6T engines spells out that by using directly modulated lasers and getting rid of extra modulators, wirebonds, and gearbox/DSP overhead, they can cut about **6–8 watts of power per module** versus conventional approaches at these speeds. Six to eight watts doesn’t sound like much until you remember you’re not deploying one module, you’re deploying thousands. At 10,000 modules across a big AI cluster, that’s **60–80 kW of power and cooling you’re not wasting just on the optics layer.** That “freed” power budget can either lower your overall TCO or be reallocated to more GPUs and more memory in the same rack/power envelope.
Then there’s cost. In their Starlight FAQ, POET flat-out says that for Celestial AI, the Starlight external light source is expected to be **up to 75% cheaper than comparable light sources on the market**, thanks to the Optical Interposer’s lower bill of materials and wafer level assembly. Cheap optics are what allow hyperscalers to overspec bandwidth relative to memory, which is exactly what you want when memory itself is the constrained resource.
There’s also a quiet supply angle baked in: their 1.6T transmitter is the first to efficiently use an electroabsorption modulated laser array in this way, and POET explicitly calls out that it gives a “dramatic capacity increase” for Indium Phosphide electroabsorption modulated lasers at a time when the industry is facing shortages of those lasers. In other words, they’re not just stretching HBM/DRAM/SSD; they’re also stretching scarce laser capacity on the photonics side by packing more usable throughput per laser device.
**Put that together in plain English: a single POET-based engine gives you terabit-class bandwidth in a tiny footprint, saves you several watts per port at 800G/1.6T, cuts your light source cost by as much as three-quarters, and increases the amount of traffic you can push per scarce laser. That’s what makes it a lever against the memory/SSD shortages. this results in:**
* **GPUs spend more time actually computing instead of waiting on I/O.**
* **You need fewer duplicate copies of data sprayed across the cluster just to hide latency.**
* **You can justify architecting shared memory pools and far-flash tiers because the fabric is fast and efficient enough to make them usable.**
So in a world where HBM/DRAM/SSD are the new oil and the barrels are sold out years in advance, POET’s pitch is basically: “Let us give you 1.6 Tbps-class optical engines that burn 6–8 W less per port and slash light-source cost by up to 75%, so you can stop wasting the obscenely expensive memory you already have.” This ultimately maximizes the compute potential of the present AI infrastructure to meet current AI demands and lessens the reliance on more compute power and SSD/RAM.
**TL;DR**
AI has turned HBM/DRAM/flash into the new oil. Prices are ripping, supply is pre-sold years out, and hyperscalers still can’t get all the memory they want – but the real choke point now is how efficiently they can use the insanely expensive RAM/SSD they already have and how fast they can move data between GPUs, DRAM pools, and storage. POET attacks that bottleneck: its Optical Interposer replaces bulky “gold box” optics with a wafer-scale, Lego-style optical tile, and its Starlight module moves the lasers off the hot GPU package into a shared external laser farm, making high-bandwidth links cheaper, cooler, and easier to scale. That lets photonic fabrics deliver terabit-class bandwidth, cut power and light-source cost, and reduce stranded/duplicated data, so you get more training and inference per GPU and per GB of HBM/DRAM/SSD instead of just brute-forcing with more hardware. In a world where memory is scarce and pre-booked, POET doesn’t make bytes – it makes the plumbing that squeezes more value out of every byte, and the Marvell–Celestial AI deal is the big-cap validation that this external-laser, photonic-fabric path is actually where hyperscalers are headed.
**Position:**
https://preview.redd.it/r5bsoo1s3v7g1.jpg?width=1290&format=pjpg&auto=webp&s=3f4ee644739ce9eac72992f7d84be909b9c8f6e8
Edit:
Interesting Statement:
“Dave Brown, Vice President of Compute and Machine Learning Services at AWS, commented on Marvell’s acquisition of Celestial AI in their official announcement, stating: “At AWS, we aim to be at the forefront of major technology inflections, and we believe optical interconnects will play an important role in the future of AI infrastructure. Building a scalable, high-performance, and power-efficient cloud starts with an approach built upon differentiated technologies. Celestial AI has made impressive progress, and we expect their combination with a large-scale semiconductor company like Marvell will help further accelerate optical scale-up innovation for next-generation AI deployments.””