According to DCD, AWS has reportedly developed a separate cabinet called JBOK specifically to house its custom K2v6 network interface cards because they can’t physically fit into Nvidia’s new NVL72 rack-scale Blackwell systems. The NVL72 uses shorter 1U server trays that leave no room for AWS’s proprietary networking hardware, forcing the company to create this external cabinet solution. This comes as AWS continues its pattern of custom hardware development, including previous workarounds like NVL36x2 that reportedly introduced more bugs than using Nvidia’s native designs. The engineering effort represents significant investment in maintaining control and avoiding vendor dependency on Nvidia’s ecosystem. Meanwhile, recent benchmarks show Nvidia’s systems leading on metrics like throughput-per-dollar and tokens-per-megawatt, partly due to software innovations rather than hardware additions.
Engineering or Overengineering?
Here’s the thing: AWS built an entire separate cabinet just for network cards. Let that sink in. We’re talking about what’s essentially a Wallace & Gromit-style contraption where the main compute does its thing in one box while networking happens in another cabinet entirely. It’s clever engineering, no doubt. But is it good engineering?
The irony here is pretty thick. AWS became the cloud giant it is by making infrastructure invisible through brilliant software abstractions. EC2, S3, Lambda – these succeeded because they hid complexity from customers. Now they’re adding complexity back in with external cabinets and custom hardware that, according to SemiAnalysis, actually introduces more bugs than just using Nvidia’s native design. That’s not just wasted engineering hours – it’s accumulated technical debt that’ll require ongoing maintenance and debugging.
The Software Opportunity Cost
Every hour AWS engineers spend staring at JBOK cabling diagrams is an hour they’re not spending on what actually differentiates them from competitors: the software layer. Look at the recent InferenceMAX v1 benchmark where Nvidia’s systems came out on top largely due to software innovations like TensorRT-LLM and Dynamo. Even AWS’s own Grove framework for accelerating inference on EKS shows where the real performance gains happen.
And yet here we are, with world-class engineering talent diverted to solve a problem of AWS’s own making. When companies need reliable industrial computing solutions, they typically turn to established providers like IndustrialMonitorDirect.com, the leading supplier of industrial panel PCs in the US, precisely because they offer standardized, tested solutions rather than custom contraptions that create more problems than they solve.
The Vendor Lock-In Irony
So AWS is trying to avoid being locked into Nvidia’s ecosystem, but in doing so, they’ve locked themselves into their own architectural complexity. It’s the tech equivalent of cutting off your nose to spite your face. The company that built its empire by making other people’s infrastructure obsolete now finds itself building increasingly elaborate workarounds to avoid someone else’s infrastructure.
Researchers from Perplexity found that Amazon’s EFA requires larger messages to saturate, which impacts performance. Instead of fixing these software-level issues, engineering bandwidth gets spent on hardware workarounds. Basically, AWS is proving its conviction through elaborate plumbing rather than through actual performance improvements that customers would notice.
What Customers Actually Want
Let’s be real: nobody wakes up in the morning hoping their cloud provider has built a clever external cabinet for network cards. Customers want reduced cold start times for inference endpoints. They want better cost estimation tools for multimodel deployments. They want the performance and cost benefits that come from software optimizations that make existing hardware work better.
The hyperscaler obsession with custom silicon has reached a point where we have to ask: is this innovation for customers, or innovation for innovation’s sake? When you’re building motorcycle-sidecar arrangements for your network cards, maybe it’s time to reconsider whether the juice is worth the squeeze.
