Servermall UAB

At GTC 2026, Pegatron showcased an updated lineup of NVIDIA servers for AI — from the massive NVL72 racks to compact 2U GPU servers.

We’ve put together the most important highlights for you:

1. NVL72 — data center-level density and scalability.

Previously, training large models required dozens of separate servers, creating network and storage bottlenecks. The NVL72 combines 72 Rubin GPUs and 36 Vera CPUs in a single rack with NVLink and liquid cooling. The result is hyperscale performance in a single unit, faster model training, and the ability to scale clusters without added latency between GPUs.

2. DPU and CPU offload.

The RA4803-72N3 and other nodes use BlueField-4 DPUs to handle networking, storage, and infrastructure tasks. This frees the CPUs to focus almost entirely on AI workloads, boosting model performance, reducing latency, and making the infrastructure more predictable under heavy loads.

3. MGX and HGX modularity.

Servers used to be fixed configurations — the number of GPUs dictated the setup. MGX allows combining CPUs and GPUs to match specific workloads. For end users, this means computing resources can be tailored precisely to their needs, accelerating AI deployment and making better use of budget.

4. High-density storage and data acceleration.

MS303 and SS201 servers leverage E3.S SSDs and GPU acceleration to handle large models. This lets users store and process data in a single node, reducing latency between computation and storage, speeding up training and inference, and simplifying overall operations.

Bottom line.
Early AI infrastructure used to rely on scattered servers and storage, which limited density and speed. But this new solutions provide integrated, high-density, and scalable hardware, enabling clients to deploy full AI clusters faster, accelerate model training and inference, and reduce infrastructure costs without compromising performance.

The new batch of servers is finishing testing and waiting for final disk installation before shipment.

Germany, we’re on our way! 🇩🇪🚚

Latency Is the Profit Killer in Fashion IT
By Dmitry Onosh | Chief Business Development Officer

When I work with fashion retailers whose turnover is measured in billions, I am often struck by the same contradiction.

On paper, their IT estates appear beyond reproach, well funded, carefully provisioned, and technically sound. And yet, when demand surges and every moment should convert into revenue, performance slips not because systems fail, but because they hesitate.

Latency, almost imperceptible to dashboards and reports, intervenes between intent and ex*****on, and in that brief pause, value is steadily lost.

Take a recent case: a multi-country fashion chain consolidated its core e-commerce, OMS, and inventory systems to cut costs and simplify operations. On the infrastructure side, resource utilization was excellent.

In practice, regional teams faced 150–300 ms delays on key operations. Pages loaded slower, checkout retries increased, and inventory updates lagged behind actual transactions. The result: bottlenecks weren’t in servers or networks, but in the way critical business processes were coupled to central systems.

The impact was tangible: a projected €25–30 million online revenue stream fell short by €2–3 million annually — not because of demand or pricing errors, but because the systems reacted slower than customers shopped and stores moved inventory.

Margin leaks were compounded by overstocking and extra operational work, quietly inflating costs by 10–15% in working capital.

From my experience, addressing this requires three actions:

1. Measure latency as a business metric, not just IT. Quantify the revenue impact of delays in checkout, stock updates, and POS writes. When 100 ms costs conversion, it belongs in P&L discussions, not only network dashboards.

2. Localize critical transactions. Run POS, stock writes, and order confirmation at the edge, close to the store or data source, while centralized analytics and forecasting remain in the cloud. This separation preserves speed where it matters most without duplicating entire platforms.

3. Align infrastructure with business cycles, not geography. Peaks in fashion are commercial events — collection drops, flash sales, viral trends — not just regional traffic spikes. Designing ownership and resource allocation around these events ensures agility without overprovisioning.

Implementing these steps allows retailers to recover 20–25% of lost digital margin within 18–24 months, stabilize operations during peak periods, and maintain performance without expanding overall infrastructure spend.