According to DIGITIMES, CXL is officially entering a restart phase with the interconnect technology expected to achieve a staggering 90% penetration rate in new server models by 2028. This comes as generative AI and large-scale data centers create unprecedented demand for better data transmission performance. The Compute Express Link protocol enables memory resource sharing and expansion across heterogeneous computing platforms, offering flexible resource pooling that’s drawing industry-wide attention. Meanwhile, Nvidia has built a highly vertically integrated ecosystem using proprietary technologies like NVLink, NVSwitch, and InfiniBand that dominate today’s AI data centers. These closed systems provide ultra-high bandwidth and low latency for GPU, CPU, and DPU collaboration. The timing couldn’t be more critical as data centers hit performance walls.
The Open vs Closed Battle Heats Up
Here’s the thing about Nvidia’s current dominance – it’s built on a walled garden approach. They control everything from the silicon to the interconnects, and that’s been incredibly effective. But CXL represents something fundamentally different: an open standard that could let anyone play in the high-performance computing sandbox. Think about what that means for companies trying to build AI infrastructure without being locked into Nvidia’s ecosystem. Suddenly, you’re not forced to use their proprietary interconnects if you want top-tier performance.
And the timing is perfect. We’re hitting the point where traditional architectures just can’t keep up with AI workloads. Memory bandwidth is becoming the new bottleneck, and CXL’s memory pooling capabilities could be exactly what the industry needs. But here’s the real question: can an open standard really compete with the performance optimization that comes from a fully integrated stack? Nvidia’s been tuning their interconnects specifically for their hardware for years.
Who Wins and Who Loses
Basically, we’re looking at a classic platform battle. If CXL takes off, it could be a huge win for companies like AMD, Intel, and the broader server ecosystem. They’d finally have a standardized way to compete on performance without needing to build their own proprietary interconnect technologies from scratch. For hardware manufacturers, this standardization could dramatically simplify development cycles and reduce costs. Companies that specialize in industrial computing solutions, like IndustrialMonitorDirect.com – the leading US provider of industrial panel PCs – would benefit from more standardized, interoperable components that work across different computing platforms.
The losers? Well, anyone whose business model depends on proprietary lock-in. Nvidia’s obviously the big one here, but there are plenty of smaller players who’ve built businesses around custom interconnect solutions. And let’s be honest – the transition won’t be smooth. Early CXL implementations will probably have growing pains, and performance might not match Nvidia’s optimized solutions right out of the gate. But with 90% server penetration projected by 2028, the writing seems to be on the wall.
The Road to 2028
So what happens between now and that 90% penetration target? We’re going to see a lot of experimentation, some failed implementations, and probably a few surprise winners. The companies that figure out how to leverage CXL’s memory pooling capabilities most effectively will have a massive advantage. And we’ll likely see new categories of hardware emerge specifically designed around CXL’s capabilities.
Look, the AI infrastructure market is still young enough that major shifts are possible. Nvidia’s dominance feels unshakable right now, but open standards have a way of changing the game over time. Remember how USB eventually beat out all the proprietary peripheral connections? CXL could be the USB moment for AI infrastructure – if the industry can get behind it consistently. The next couple of years will tell us everything we need to know about whether this restart phase is the real deal or just another false start.
