Nvidia's Connectivity Monster: Why Are Carriers Sleeping on the Vera Shift?
The AI arms dealer is moving beyond GPUs to build a CPU infrastructure that dwarfs current datacenter standards. The carriers should be sweating.
For years, the telecom industry has viewed Nvidia as the GPU provider—the vendor of the high-end engines that power the AI revolution. But if you look at the blueprints for the new Vera CPU, it becomes clear that Nvidia isn't just selling engines anymore; they are redesigning the entire garage.
As reported by The Register, Nvidia has launched Vera, a standalone CPU designed to challenge the dominance of Intel and AMD. While the industry has focused on the GPUs, Vera represents a massive leap in connectivity and compute density that should have every carrier on high alert.
Let's look at the numbers. A single Vera CPU features 88 custom Armv9.2 cores (which Nvidia calls Olympus cores) and supports up to 1.5 TB of LPDDR5X memory. But the real monster is the "Vera CPU Superchip." By connecting two Vera CPUs via an NVLink Chip-to-Chip interface, Nvidia is delivering 1.8 TB/s of bidirectional bandwidth. To put that in perspective, the aggregate memory bandwidth for a Superchip is 2.4 TB/s—roughly double what AMD's Turin Epycs offered at their 2024 launch, according to The Register.
Nvidia isn't stopping at the chip level. Their agentic AI reference designs call for packing 128 of these superchips into a single liquid-cooled rack. That is a staggering 256 CPUs, 22,528 cores, and 384 TB of memory in one footprint.
From my perspective as a consumer advocate, this is where the danger lies. We are seeing a fundamental shift in infrastructure requirements. Nvidia is targeting hyperscalers and cloud providers—names like Meta, Oracle, Alibaba, ByteDance, CoreWeave, Lambda, Nebius, and NScale have already signed up—to deploy these chips.
When you move this much data at these speeds, the bottlenecks shift. Vera is designed specifically to quash pipeline and execution bottlenecks, utilizing a custom neural branch predictor to handle branch-heavy workloads like Python scripts. But the physical connectivity required to support 384 TB of memory and thousands of cores in a single rack is a logistical nightmare for traditional datacenter layouts.
Why aren't the carriers sweating? The carriers are the ones who provide the pipes. If the cloud providers—the lairds of the modern digital estate—are deploying hardware that demands this level of throughput and liquid cooling, the legacy infrastructure provided by the carriers will eventually become the choke point.
Nvidia is building a connectivity monster. If the carriers continue to treat AI as just "more traffic" rather than a fundamental shift in how compute is architected, they are going to find themselves providing the narrow straw for a firehose of data.