U.S. funds $300 million to speed AI data centers with GLOBALFOUNDRIES photonics - CHOSUNBIZ - Chosunbiz

Skyler: Hey, welcome to Line Rate — the show about datacenter fabric and A I high-speed interconnect. I'm Skyler Monroe, and if it moves bits between chips, racks, or datacenters at ridiculous speeds, this is where we talk about it. Quick thanks to our sponsors: Ada, who helps companies actually integrate A I into their real workflows, and Zen Semiconductor, an umbrella company with a portfolio of businesses across the silicon industry. Really appreciate both of them keeping the lights on here. Alright, we've got a packed episode today — government money flowing into photonics, a massive new datacenter campus filing in West Texas, co-packaged optics reshaping the switch market, Cisco and Intel teaming up on Ethernet-native A I clusters, and TrendForce's read on where datacenter interconnect is headed in 2025. Let's get into it. Skyler: First up — GlobalFoundries is getting 300 million dollars in U.S. government funding to accelerate photonics for A I datacenters. And look, when federal money shows up at this scale for a specific technology area, that's a signal worth paying attention to. So what's actually happening here? The funding is tied to silicon photonics manufacturing — specifically building out GlobalFoundries' capability to produce photonic integrated circuits at scale. This isn't basic research money. This is about domestic manufacturing capacity for a technology that's becoming genuinely critical to the next generation of A I infrastructure. Skyler: So why GlobalFoundries and why now? GF has been quietly building a credible silicon photonics platform for years — their 300mm photonics process at their Malta, New York fab is designed to co-integrate photonic components with electronics on the same wafer. That matters enormously because one of the central problems in moving to co-packaged optics and eventually on-package optical I/O is that you need photonic devices that can be manufactured with the same repeatability and yield discipline as the electronics they're sitting next to. You can't ship a hundred thousand ports at scale if your photonic components are boutique parts. Skyler: The government angle here is also interesting strategically. The CHIPS Act funding framework has been very focused on leading-edge logic — your advanced nodes, your memory — but photonics is increasingly being recognized as a separate critical capability. You need photonics to move data between the leading-edge chips. What good is a cluster of next-generation A I accelerators if the interconnect fabric bottlenecks the whole thing? The administration is clearly connecting those dots. Skyler: For the industry, this has a few ripple effects. First, it validates the co-packaged optics roadmap in a very public way — government doesn't write 300 million dollar checks for technologies it thinks are speculative. Second, it puts GF in a stronger competitive position against TSMC and others in the photonics space, which has been somewhat dominated by foreign fabs. Third, and maybe most importantly for people building systems, it suggests that domestic supply for photonic components might actually materialize in a meaningful timeframe — which changes the risk calculus for U.S. hyperscalers who've been nervous about supply chain concentration. This one's worth watching closely as the fab buildout progresses. Skyler: Next story — Crusoe Energy, now rebranding more broadly as just Crusoe, has filed for a 4.8 billion dollar datacenter campus in Jayton, Texas. That's Jayton — J-A-Y-T-O-N — which is in Kent County, way out in West Texas. And they're also connected to a project in Childress, Texas, not too far away. So what's going on here and why does it matter for the interconnect world? Skyler: First, some context on Crusoe. They started as a company that was using stranded natural gas — basically gas that would otherwise be flared at oil wells — to power compute. It was a clever arbitrage play on cheap, stranded energy and they built a name running GPU clusters in unusual places. But they've been pivoting hard toward being a full-stack A I cloud infrastructure company, and this filing is a major statement of that ambition. Skyler: 4.8 billion dollars for a campus in rural West Texas is a big number, and the two-datacenter structure is interesting. When you file for a campus with two facilities that are geographically close, you're almost certainly thinking about them as a single logical fabric. The connection between the Jayton campus and the Childress project suggests Crusoe is building toward a multi-site cluster — which immediately raises the question of how you tie those together at the fabric level. Skyler: This is where it gets relevant for our audience. When you're talking about connecting two datacenters that are potentially tens of miles apart, you're in the territory of long-haul and metro optical interconnect — dark fiber, DWDM, maybe direct detect at 400G or 800G per wavelength depending on the distance and budget. But you're also thinking about how that inter-site fabric interacts with the intra-cluster fabric. If Crusoe is running GPU or accelerator clusters, they need low-latency, high-bandwidth east-west traffic within a site. Between sites, latency goes up, so you have to architect your training jobs and inference workloads accordingly. Skyler: The West Texas location is also not random. The region has been attracting datacenter investment because of land availability, proximity to wind and solar resources, and increasingly, proximity to the Permian Basin energy grid. For a company like Crusoe with roots in energy arbitrage, that's a natural fit. But for the interconnect ecosystem, what this really signals is that the buildout of A I infrastructure is spreading beyond the traditional Northern Virginia, Silicon Valley, and Phoenix clusters — and that creates demand for interconnect and fabric solutions that work in more distributed, sometimes more constrained environments. Skyler: Alright, story three — LightCounting put out research arguing that co-packaged optics are actually growing the total market for scale-out switches, not just disrupting existing pluggable optics revenue. This is a nuanced but really important market dynamics story, so let me break it down. Skyler: The conventional worry in the optics industry has been that co-packaged optics — where you integrate the optical engines directly into the switch package rather than using pluggable modules in the faceplate — would cannibalize pluggable transceiver revenue. The logic was: fewer pluggables sold equals less revenue for the optics guys. And that's not wrong as far as it goes. But LightCounting is making a more interesting argument. Skyler: What they're saying is that C P O — co-packaged optics — is enabling switch radix and bandwidth that simply isn't achievable with pluggable form factors. When you eliminate the electrical serializer-deserializer links that connect a traditional switch ASIC to pluggable optics in the faceplate, you remove a big power and signal integrity constraint. That means you can build switches with dramatically more ports at dramatically higher speeds. We're talking about moving from 51.2 Terabits per second switch ASICs toward 102.4T and beyond, with port counts that would be physically impossible to cable up with pluggables given faceplate space and power density limits. Skyler: And here's the key insight: if you can build bigger, higher-bandwidth switches, you can build bigger, flatter network topologies for A I clusters. Instead of a three-tier fat tree, you start to get viable two-tier or even single-tier designs at scales that previously required multiple layers. That means more total switch bandwidth deployed per cluster — not less. So the market for switching silicon and optical engines actually expands, even if the unit economics per port shift. Skyler: For the competitive landscape, this is significant. Broadcom is the dominant scale-out switch ASIC vendor, and their Tomahawk and Jericho families are the backbone of most hyperscaler fabrics today. But C P O requires very close co-design between the switch ASIC and the optical engine — you're literally putting them in the same package or at minimum the same substrate. That opens the door for new entrants who can do that co-design natively, and it puts pressure on traditional optical module vendors to partner or pivot. Companies like Ayar Labs, Lightmatter, and others have been positioning here, and this LightCounting analysis suggests the market they're going after is bigger than the bear case assumed. Skyler: There's also a standards dimension here. The CPO consortium has been working on interoperability specs — because if every switch vendor has a proprietary optical engine interface, you lose the pluggable ecosystem's key advantage, which is multi-vendor interoperability. Watching how those specs develop will tell you a lot about whether C P O ends up being an open market or a set of closed vertical stacks. That's a story that's still very much in motion. Skyler: Story four — Cisco published a blog post detailing how they've been working with Intel to run Ethernet-native A I clusters using Intel Gaudi 3 A I accelerators connected through Cisco Nexus 9000 switches. On the surface this looks like vendor co-marketing, and yes, it partly is — but there's real technical substance here worth unpacking, especially in the context of the Ultra Ethernet Consortium and the U E C ecosystem. Skyler: So the setup: Intel's Gaudi 3 is their latest A I accelerator, and one of its differentiating claims has always been that it uses standard Ethernet for both its scale-up and scale-out fabric — as opposed to NVLink for scale-up or InfiniBand for scale-out, which is the NVIDIA approach. The Cisco Nexus 9000 is a well-established data center switch platform. What the blog is really documenting is a reference architecture for running distributed A I training workloads — specifically large model training — entirely over Ethernet. Skyler: Why does this matter? Because the longstanding knock on Ethernet for A I fabric has been congestion and latency. InfiniBand has had a significant lead in RDMA performance and congestion management, which matters a lot for the all-to-all communication patterns in distributed training — things like AllReduce across hundreds or thousands of accelerators. If congestion spikes during that collective communication, your GPU utilization tanks, your job slows down, and you're paying for expensive accelerators to sit idle. Skyler: The Ultra Ethernet Consortium — U E C — has been developing a set of enhancements to Ethernet specifically to close this gap. Things like improved transport protocols designed for A I workloads, better multipathing, and congestion control mechanisms tuned for the bursty, synchronized traffic patterns of distributed training. U E C specs are still evolving — the transport layer spec in particular is still being finalized as of early 2025 — so deployments right now are using existing Ethernet capabilities plus software-layer optimizations. Skyler: What Cisco and Intel are doing here is essentially a proof point that you don't have to wait for U E C to be fully baked to run serious A I workloads on Ethernet. They're showing workable performance on existing infrastructure. That's strategically important for both companies — it keeps Gaudi 3 relevant for customers who have Ethernet infrastructure and don't want to invest in an InfiniBand overlay, and it gives Cisco a story in the A I datacenter market where they've been somewhat on the outside relative to Arista and the InfiniBand vendors. Skyler: For the broader ecosystem, this is a data point in the long-running debate about whether Ethernet can truly displace InfiniBand for A I fabric. The honest answer right now is: for many workloads, yes, especially with good software stacks. For the most latency-sensitive, tightly-coupled large-scale training jobs, InfiniBand still has an edge. But that edge is narrowing, and every reference architecture like this one chips away at the conventional wisdom that InfiniBand is mandatory. Keep an eye on U E C progress — when those transport specs land and silicon starts shipping with native U E C support, the calculus could shift pretty quickly. Skyler: And finally — TrendForce put out their 2025 Global A I Datacenter Interconnect Trends report, and it's worth going through the key themes because TrendForce does solid supply chain and market research and this gives us a useful macro view of where the industry is heading. Skyler: The headline finding is pretty much what you'd expect if you've been following this space: 2025 is shaping up to be a critical inflection year for optical interconnect in A I datacenters, driven by the scaling requirements of next-generation accelerator clusters. But the interesting stuff is in the specifics. TrendForce is seeing a clear bifurcation in the interconnect market — on one side, you've got the intra-cluster, scale-up fabric, which for the hyperscalers and the leading A I labs is increasingly proprietary and tightly integrated. Think NVLink Switch systems, Google's TPU interconnect, custom silicon from the big players. On the other side, you've got scale-out fabric connecting clusters, pods, and ultimately datacenters — and that market is much more open and contested. Skyler: On the optical side, TrendForce is tracking a rapid migration toward 800G per-lane coherent and direct detect solutions for inter-datacenter and intra-datacenter spine layers. The 400G generation isn't going away, but the growth is at 800G, and the next wave of planning is already at 1.6T. That's a faster cadence than traditional telecom optical cycles, and it's creating real pressure on component suppliers to ramp new generations before the previous one is fully depreciated. Skyler: The report also highlights co-packaged optics as a key 2025 theme — consistent with the LightCounting piece we talked about earlier. TrendForce sees the first real commercial deployments of C P O in hyperscaler environments happening this year, which would mark a transition from lab demos and limited trials to actual production infrastructure. The bottleneck they flag is the same one the industry has been wrestling with: yield and supply chain maturity for the photonic components. The designs exist, the performance is there, but getting consistent, high-volume supply is still a work in progress. Which, incidentally, is exactly why the GlobalFoundries funding announcement we led with today is relevant. Skyler: TrendForce also calls out the geographic diversification of datacenter buildout — more capacity going to Southeast Asia, Middle East, and yes, rural domestic U.S. locations like what we're seeing with Crusoe — and notes that this creates new requirements for long-haul interconnect between distributed A I campuses. You can't run a globally distributed training cluster without serious thought about how you manage the latency and bandwidth budgets across those links. That's driving demand for high-capacity subsea cables, terrestrial DWDM buildouts, and increasingly, a new generation of metro-scale optical networking that bridges the gap between the intra-datacenter fabric and the long-haul backbone. Skyler: The overall picture TrendForce paints is one of an industry in rapid transition — the optical speeds are going up, the integration is going deeper, the geography is spreading out, and the standards are still catching up to where the deployments need to go. Which honestly is a pretty accurate description of where we are. There's a lot of infrastructure being built right now on specs that aren't fully standardized, timelines that are aggressive, and supply chains that are still maturing. That's the nature of building at the frontier. Skyler: Alright, that's a wrap on today's episode of Line Rate. We covered a lot of ground — GlobalFoundries photonics funding, Crusoe's massive West Texas campus filing, co-packaged optics growing the switch market, Cisco and Intel making the Ethernet-native A I cluster case, and TrendForce's macro read on 2025 interconnect trends. Big thanks again to our sponsors — Line Rate, brought to you by Ada and Zen Semiconductor. Ada helps businesses actually integrate A I into their real workflows, and Zen Semiconductor is an umbrella company supporting a portfolio across the silicon industry. If you got value from this episode, please subscribe and share it with someone in your orbit who's building or investing in this space. Show notes with all the source links are available wherever you're listening. I'll see you next time on Line Rate.