Nvidia's breakneck pace of releasing new AI hardware every single year just hit a massive reality check. Three months ago at GTC, CEO Jensen Huang stood on stage and proudly showed off the Kyber NVL144 rack architecture. It looked like the ultimate flex—a massive server cabinet packing 144 next-generation chips designed to handle the absolute heaviest AI workloads for 2027.
Now, that plan is in tatters.
According to research firm SemiAnalysis, the Kyber NVL144 has been delayed by more than 12 months, pushing its actual arrival into 2028. This isn't just a minor shipping hiccup. It's a structural bottleneck that alters how cloud providers build data centers and gives serious leverage to Nvidia's rivals. If you think the AI infrastructure boom can just keep doubling performance every 12 months without breaking a sweat, this delay proves otherwise.
The Midplane Mess That Stalled Kyber
To understand what went wrong, you have to look at how these massive AI systems are built. The Kyber architecture flips the script on traditional server design by mounting graphics processing units vertically inside compute trays. This layout drastically increases density, lowers latency, and uses a dense, all-copper NVLink interconnect to make 144 chips act as one giant, cohesive computer.
The component holding up the entire show is the midplane printed circuit board (PCB), which Nvidia calls the orthogonal backplane.
Manufacturing this specific board is an absolute nightmare. It requires a highly complex hybrid material stack composed of M9-grade copper-clad laminate, quartz fabric (Q-fabric), and PTFE (Teflon). The sheer number of layers and the extreme precision required to route signals without degradation mean that high-end PCB suppliers simply cannot yield these boards at scale right now. It's a classic case of cutting-edge physics colliding with the messy reality of factory floors.
The Scrapped Backup Plans and Shrunk Chips
Nvidia knew this was going to be an uphill battle, so they engineered a backup plan called the NVL72x2. The idea was simple on paper: bypass the problematic Kyber midplane by bolting two current-generation Oberon racks together back-to-back, using pure copper NVLink cabling to scale the domain.
It blew up in their faces.
Hyperscalers and cloud service providers (CSPs) flatly rejected the NVL72x2. The back-to-back design was awkward, took up an uncomfortable footprint, and introduced a massive operational burden for data center technicians who actually have to service these liquid-cooled monstrosities. Because of that heavy pushback, Nvidia canceled the NVL72x2 entirely.
Worse yet, the pain isn't isolated to the rack infrastructure. The upcoming Rubin Ultra chip itself has been significantly scaled back. Nvidia scrapped the ambitious four-compute-die version of the Rubin Ultra. Instead, they're moving forward with a smaller two-compute-die version, which basically halves the raw per-chip performance compared to the original roadmap.
To bridge the gap, Nvidia is forcing a pivot. They plan to aggressively push sales of standard Oberon Rubin and Oberon Rubin Ultra racks. But because they currently lack a proven, working solution to scale the interconnect domain beyond those standard configurations, their tech roadmap has a glaring vulnerability.
An Open Door for AMD and Google
For the last few years, Nvidia's dominance felt bulletproof because their software stack (CUDA) and fast hardware release cadence left everyone else gasping for air. This delay changes the math.
With Kyber pushed to 2028 and the larger NVL576 system—which relies on fragile Co-Packaged Optics (CPO) tech—likely restricted to low-volume trickle shipments, the scaling ceiling for Nvidia hardware is temporarily capped.
This creates a massive window of opportunity for two specific players:
- AMD: Their upcoming MI500X architecture suddenly looks a lot more attractive to cloud providers who can't get the ultra-scale clusters they were promised by Team Green.
- Google: The tech giant has been quietly building out its own infrastructure, recently teaming up with Blackstone to establish massive data centers using Google's TPUv8i Broadfly chips instead of relying solely on Nvidia.
If Google or AMD can deliver superior cluster-scale performance while Nvidia untangles its midplane manufacturing knots, the competitive landscape will look wildly different by 2027.
What Infrastructure Teams Need to Do Next
If you're managing hardware procurement, capacity planning, or budget allocation for large-scale machine learning operations, you can't just sit around waiting for 2028.
Shift your procurement forecasts immediately. Do not bank on acquiring Kyber-level density for your 2027 training runs. You need to maximize your current allocations for Oberon Rubin systems and start evaluating the operational costs of horizontal scaling across standard NVL72 setups.
At the same time, open the door to alternative architectures. Run workloads on Google's latest TPUs or pilot AMD's hardware to ensure your software stack isn't completely locked into an ecosystem that's hitting physical manufacturing limits. Diversification isn't just a cost-saving strategy anymore; it's a basic requirement for operational survival.
