L&T Deploys India’s First 10,000-GPU Unified AI Cluster for Together AI in Chennai - Tech Times

Skyler: Welcome to Line Rate — the show tracking the chips, standards, and infrastructure moving data at the speed of modern A I. I'm Skyler Monroe, and if you care about what's happening inside the datacenter fabric, rack-to-rack interconnect, and the optical and silicon layers tying it all together, you're in the right place. Big thanks to our sponsors Ada and Zen Semiconductor — Ada helps enterprises actually integrate A I into their workflows, and Zen Semiconductor is the umbrella behind some really interesting fabric work including their Loom A I fabric, their DC-to-DC scale-across interconnect, and Ada for A I deployment at scale. Alright, we've got a packed show today — India's first ten thousand GPU cluster just landed, Corning is making a serious move on inside-the-box optics, Celestial A I is pitching photonics as the post-NVLink future, and Marvell's custom silicon play is getting some serious analyst attention. Let's get into it. Skyler: First up — L&T, that's Larsen and Toubro, the Indian engineering and technology conglomerate, has deployed what's being called India's first ten thousand GPU unified A I cluster, and they've done it in Chennai for Together A I. Now, Together A I is a cloud platform focused on open-source model inference and fine-tuning, so this isn't a hyperscaler vanity project — this is a serious production infrastructure play. Ten thousand GPUs in a single unified cluster is not trivial. The word "unified" is doing a lot of work in that headline, and it's worth unpacking. A unified cluster means these GPUs aren't siloed into smaller logical pools — they're networked together in a way that allows jobs to span the full scale, which means your fabric has to support that. At ten thousand GPUs, you're almost certainly looking at a multi-rail InfiniBand or RoCE deployment, possibly both depending on whether they're separating storage traffic from compute traffic. L&T has been quietly building out their digital and cloud infrastructure business, and this is a significant milestone for them. For India's A I ecosystem, this matters because local compute availability has been a real bottleneck for Indian AI startups and researchers. Having ten thousand GPUs in Chennai — that's on-shore, lower latency to Indian users, and it potentially brings Together A I's open model ecosystem to a region that's been underserved by domestic GPU capacity. From a fabric standpoint, what I'd want to know is what the spine looks like. At this scale, you're talking about a fat-tree or a dragonfly topology, probably with 400G or 800G ports on the leaf-spine interconnect. If they're running NVLink for intra-node and then InfiniBand or Ethernet for inter-node, that's a pretty standard hyperscale playbook. But together A I's workload — open model inference at scale — puts different demands on the fabric than, say, training. Inference is often more east-west bursty with tighter latency sensitivity on individual request paths. So the quality of the fabric here matters operationally, not just for benchmarks. This is a big deal for India's A I infrastructure story, and it puts L&T on the map as a serious GPU cluster integrator. Skyler: Alright, let's pivot to optics. MarketsandMarkets just dropped a photonic packaging market report, and while market research reports can sometimes feel like a lot of numbers in search of a thesis, the photonic packaging space right now is genuinely worth the attention. The core story here is that photonic packaging — meaning how you physically integrate photonic components with electronic chips — is becoming a critical bottleneck and a critical opportunity in A I datacenter buildouts. For a long time, optics sat outside the package. You had your ASIC, your SerDes, your pluggable optic module, and everything talked over copper traces on a PCB before light happened. That model works fine up to a point, but as bandwidth demands scale — and with A I training clusters, they are scaling hard — the electrical reach from chip to pluggable module starts burning power and eating signal integrity. Co-packaged optics, or CPO, is the answer the industry is converging on, and photonic packaging is the discipline that makes CPO real. MarketsandMarkets is projecting strong growth in this market, which honestly isn't surprising given the tailwinds. Every major hyperscaler is pushing toward 800G and 1.6T connectivity. Every new switch ASIC generation — Broadcom's Tomahawk, Cisco's Silicon One, Marvell's Teralynx — is designed with higher port density that stresses the electrical-optical interface more. Photonic packaging covers a wide range of integration approaches — from silicon photonics in a separate package co-located with the ASIC on a substrate, to fully integrated photonic chiplets in the same package as the compute die. The latter is where things get really interesting from a performance standpoint, and also really hard from a manufacturing standpoint. The yield and testing challenges for co-packaged photonic devices are non-trivial. You're dealing with thermal management, optical alignment tolerances, and testing optical ports mid-package — none of which is straightforward. But the market is moving this direction, and the packaging ecosystem — OSATs, foundries, substrate suppliers — is racing to catch up. This is one of those markets where the infrastructure enabling the infrastructure is where a lot of the value gets created. Skyler: Staying on optics — Corning just made a pretty bold statement. They're targeting what they're calling a ten billion dollar "inside-the-box" opportunity driven by A I optics demand. Now, Corning is a company most people associate with fiber optic cable — the outside-the-box stuff, the long-haul and campus fiber runs. But what they're signaling here is a deliberate pivot toward the optical components that live inside the datacenter chassis — the short-reach, high-density, co-packaged and near-packaged optical interconnect that A I clusters are driving demand for. Ten billion dollars is not a rounding error, even for Corning. And it reflects a real structural shift in where fiber and optical component demand is concentrating. Historically, the big optical capex in a datacenter was the fiber plant — the trunk cables, the patch panels, the outside plant connecting buildings. That's still a business, but the explosive growth is now inside the box, inside the rack, and inside the package. Corning is specifically calling out A I-driven demand as the catalyst, which makes sense. A single large A I training cluster might have more optical connections within the rack than a traditional hyperscale compute cluster has in the entire facility. The density is just orders of magnitude higher. What Corning brings to this market is deep materials expertise — they're one of the best in the world at specialty glass and optical fiber fabrication, and that expertise translates to the waveguides and optical components needed for high-density photonic packaging. The question is how they compete with companies like II-VI, Coherent, and the silicon photonics players who've been in the co-packaged space longer. Corning's angle is likely around manufacturing scale and materials — they can bring volume and consistency that newer entrants struggle to match. If they can execute on that ten billion dollar target, it tells you something important about where the overall photonic packaging market is heading. This is not a niche story. This is a major materials and optical components player making a generational bet on A I datacenter architecture. Skyler: Now let's talk about Celestial A I, because this is one of the more technically ambitious startups in the interconnect space, and there's a good analytical piece out from SemiVision that digs into their photonic interconnect thesis and their capital strategy. The headline framing is "Beyond NVLink," which is intentionally provocative, and worth taking seriously. NVLink is NVIDIA's proprietary chip-to-chip and GPU-to-GPU interconnect — it's fast, it's high bandwidth, and it's one of the key reasons NVIDIA's multi-GPU systems dominate A I training. NVLink 4.0 runs at 900 gigabytes per second of total bandwidth per GPU in the H100 configuration. That's a lot. But NVLink has a fundamental architectural constraint — it's designed for scale-up within a node or a tightly coupled NVLink domain. Once you go beyond that domain boundary, you're back to InfiniBand or Ethernet, and bandwidth drops off significantly. Celestial A I's bet is that as models scale into the trillion-parameter regime — and we're already seeing models pushing that direction — the interconnect bottleneck moves from compute to communication, and the only way to address it at the required bandwidth and power envelope is photonics. Their core technology is what they call "Photonic Fabric" — a disaggregated optical interconnect that can provide very high bandwidth at low latency between compute elements at distances that NVLink can't cover. Think connecting accelerators across a rack, across multiple racks, potentially across a row — with bandwidth that's competitive with what you'd get on-package today, but at greater distance. The capital strategy piece is also interesting — Celestial A I has been raising aggressively, positioning themselves as a platform play rather than just a component vendor. The idea being that if the photonic fabric becomes the backbone of how you compose large A I compute pods, the company controlling that fabric has significant architectural leverage. The competitive risks are real — Ayar Labs is doing optical I/O chiplets, Intel has its silicon photonics work, and even some of the GPU vendors are thinking hard about optical integration. But Celestial A I's focus on the fabric layer specifically, rather than just optical I/O, is a differentiated angle. This is definitely one to watch. Skyler: Last story — Marvell. There's a piece on Seeking Alpha titled "Custom Is The Dark Horse Catalyst," and while it's an investor-focused analysis, there's real technical substance here worth discussing. Marvell has been on an interesting trajectory over the last few years. They divested a bunch of non-core businesses, they doubled down on datacenter, and they've been building out a custom A I silicon capability — meaning they design custom ASICs for hyperscaler customers who want something more tailored than off-the-shelf merchant silicon. The "dark horse" framing is about the market not fully pricing in the upside from this custom silicon business. And honestly, there's something to that. The hyperscalers — Google, Amazon, Microsoft, Meta — have all been investing heavily in custom A I accelerators. Google has its TPU. Amazon has Trainium and Inferentia. Microsoft has the Maia A I accelerator. And they don't design all of this in-house — they work with merchant silicon design houses, and Marvell has been winning that work. Marvell's position is interesting because they bring both the custom compute capability and a strong portfolio of connectivity IP — Ethernet PHYs, SerDes, DSPs — that are essential for the interconnect fabric around those custom accelerators. So a hyperscaler partner isn't just getting a compute chip, they're getting a partner who can also design the networking silicon that connects it. That vertical integration across compute and connectivity is genuinely differentiated. The competitive landscape here includes Broadcom — who is also deep in custom A I silicon, especially for Google — and increasingly startups like Enfabrica, Tenstorrent, and others chasing the custom silicon opportunity. But Marvell's scale, their process technology access, and their existing hyperscaler relationships make them a serious player. The interconnect angle matters here too — as custom A I chips proliferate, the fabric that stitches them together becomes critical infrastructure, and that's squarely in Marvell's wheelhouse. Watch their custom silicon revenue line — if that breaks out meaningfully over the next couple of quarters, the dark horse narrative becomes the main story. Skyler: Alright, that's a wrap on today's Line Rate. We covered India's ten thousand GPU cluster milestone, the photonic packaging market heating up, Corning's ten billion dollar inside-the-box bet, Celestial A I's case for photonics beyond NVLink, and Marvell's custom silicon opportunity. Line Rate, brought to you by Ada and Zen Semiconductor — Ada helping enterprises get real value from A I integration, and Zen Semiconductor pushing the frontier on A I fabric and scale-across interconnect. If you found this useful, please subscribe and share it with someone who needs to stay sharp on datacenter fabric — there are a lot of those people right now. Show notes and links to all the source material are in the episode description. I'm Skyler Monroe, and I'll see you next time on Line Rate.