Building AI factories requires more than powerful hardware—it demands deep system integration expertise.
At a moment when artificial intelligence is crossing the threshold from experiment to industry, PEGATRON arrived at NVIDIA's GTC 2026 conference in San Jose with a family of machines designed not merely to impress, but to endure. The company is wagering that the next chapter of AI infrastructure will be written not by those who build the fastest chips, but by those who can cool them, connect them, and deliver them reliably at scale. In offering systems that cut inference costs tenfold and multiply efficiency fivefold over the previous generation, PEGATRON is quietly making the case that the unglamorous work of manufacturing discipline may matter as much as raw computational power.
- The AI industry has reached an inflection point where the question is no longer whether to build intelligent systems, but how quickly and cheaply they can be deployed at industrial scale.
- PEGATRON's flagship RA4803-72N3 supercomputer delivers 3.6 exaFLOPS of inference performance while slashing costs by up to 10 times compared to the previous Blackwell generation — a number that rewrites the economics of running a data center.
- A modular family of systems — from compact 2U enterprise servers to full 72-GPU rack configurations — means organizations can enter at any level of ambition and scale upward without rebuilding from scratch.
- Liquid cooling is no longer optional: as compute density rises, it becomes the engineering prerequisite that separates a viable product from an expensive heat problem.
- PEGATRON is positioning itself as the critical bridge between NVIDIA's GPU innovation and the practical demands of production deployment, betting that manufacturing expertise is the competitive moat that raw hardware alone cannot provide.
At NVIDIA's GTC 2026 conference in San Jose, PEGATRON unveiled a lineup of AI infrastructure platforms built for the moment when artificial intelligence stops being experimental and starts being industrial. The company, with decades of server design and manufacturing behind it, is betting that the next wave of AI infrastructure will belong to whoever can pair raw computing power with the disciplined work of keeping it cool, connected, and reliable.
The centerpiece is the RA4803-72N3, a liquid-cooled supercomputer housing 72 NVIDIA Rubin GPUs in a single rack, delivering 3.6 exaFLOPS of inference performance and 20.7 terabytes of high-bandwidth memory. Compared to the previous Blackwell generation, it trains certain AI models using one-fourth as many GPUs, cuts inference costs by up to 10 times, and achieves five times higher throughput per watt — translating directly into smaller footprints, lower electricity bills, and faster deployment for data center operators.
Beyond the flagship, PEGATRON introduced a modular family of systems. The AS210-2T1-8H3 is a 2U server built for enterprises seeking AI acceleration without supercomputer-scale investment, and these units can be stacked into rack configurations supporting 64 or 72 GPUs. For workloads demanding flexibility over maximum density — generative AI, simulation, visualization — servers built around NVIDIA's RTX PRO Blackwell Server Edition GPUs round out the portfolio, pairing high-memory cards with AMD EPYC or Intel Xeon processors.
Liquid cooling runs through every system as a practical necessity rather than a premium feature: at this level of compute density, air cooling simply cannot keep pace. PEGATRON's CTO Dr. James Shue framed the broader challenge plainly — building AI factories demands system integration expertise and manufacturing discipline that only comes from years of doing it at scale. The company is presenting itself not as a chip designer, but as the bridge between NVIDIA's GPU innovation and the hard reality of deploying AI infrastructure in production environments where reliability and cost matter as much as performance.
At NVIDIA's GTC 2026 conference in San Jose, PEGATRON walked through the doors with a lineup of machines built for the moment when artificial intelligence stops being experimental and starts being industrial. The company, which has spent decades learning how to design and manufacture servers at scale, is now betting that the next wave of AI infrastructure will belong to whoever can integrate raw computing power with the unglamorous work of keeping it cool, connected, and reliable.
The centerpiece is the RA4803-72N3, a liquid-cooled supercomputer that packs 72 NVIDIA Rubin GPUs into a single rack. The machine delivers 3.6 exaFLOPS of inference performance—the speed at which it can process trained AI models—and carries 20.7 terabytes of high-bandwidth memory. What matters most to the people who will actually buy this thing is not the raw number but the efficiency math underneath it. Compared to NVIDIA's previous Blackwell generation, this system trains certain types of AI models using one-fourth as many GPUs. It cuts inference costs by up to 10 times. It achieves five times higher throughput per watt of power consumed. For a data center operator, that translates to smaller physical footprint, lower electricity bills, and faster time to deploy new AI services.
But PEGATRON is not just selling one machine. The company unveiled a family of systems designed to fit different scales of ambition. The AS210-2T1-8H3 is a more modest 2U server—the kind that fits in a standard rack alongside dozens of others—built around NVIDIA's HGX Rubin NVL8 GPUs and dual Intel Xeon 6 processors. It's designed for enterprises that want to accelerate AI workloads without building an entire supercomputer. Stack these servers together, and you get the RA4800-64H3 or RA4800-72H3, which can hold eight or nine of these 2U units per rack, scaling to 64 or 72 GPUs respectively. The architecture is modular by design: you can start small and grow as demand grows.
For workloads that don't need the full horsepower of Rubin but still demand serious acceleration—generative AI, data processing, simulation, visualization—PEGATRON offers servers built around NVIDIA's RTX PRO Blackwell Server Edition GPUs. The MS303-2A1G-P60 holds up to four of these cards, each with 96 gigabytes of memory, paired with dual AMD EPYC processors. The AS205-2T1 takes a different approach, fitting four RTX PRO 4500 cards with 32 gigabytes each into a single-slot form factor alongside Intel Xeon 6 processors. These are the machines for organizations that need flexibility and efficiency over maximum density.
All of these systems share a common thread: liquid cooling. It's not a luxury feature. When you're running this much compute in this tight a space, air cooling hits a wall. Liquid carries heat away more efficiently, which means you can pack more performance into the same physical footprint and keep the power bill from becoming astronomical. PEGATRON has built its reputation on understanding how to manufacture these systems reliably and at scale—not as one-off engineering projects but as products that can be deployed across dozens or hundreds of data centers.
Dr. James Shue, PEGATRON's senior vice president and chief technology officer, framed the challenge plainly: building AI factories requires more than just powerful hardware. It requires the kind of system integration expertise and manufacturing discipline that only comes from years of doing it. The company is positioning itself as the bridge between NVIDIA's GPU innovation and the practical reality of deploying AI infrastructure in production. You can have the fastest chips in the world, but if you can't cool them, connect them, and manufacture them reliably, they're just expensive paperweights.
The announcement arrives at a moment when the AI industry is shifting from proof-of-concept to scale. Companies are no longer asking whether they can build AI systems; they're asking how fast they can build them, how much they'll cost to run, and whether they can do it without completely redesigning their data centers. PEGATRON's answer is that the machines are ready. The question now is whether the market will move as fast as the hardware can.
Notable Quotes
Building AI factories requires more than powerful hardware—it demands deep system integration expertise and execution at scale. With the launch of our platforms powered by NVIDIA Vera Rubin NVL72 and NVIDIA HGX Rubin NVL8, combined with proven deployment capabilities, Pegatron enables customers to move from infrastructure build-out to AI production faster, more efficiently, and with greater operational confidence.— Dr. James Shue, SVP & CTO of PEGATRON
The Hearth Conversation Another angle on the story
Why does it matter that PEGATRON is showing these systems now, at this particular moment?
Because we're at an inflection point. For the last few years, AI has been about raw performance—who has the fastest chips. Now it's about efficiency and scale. If you're a data center operator, you don't just care about FLOPS. You care about FLOPS per watt, FLOPS per dollar, and whether you can actually fit the thing in your building without rewiring the entire power infrastructure.
The 10X cost reduction for inference—is that real, or is that marketing math?
It's real, but it's specific. They're comparing to Blackwell on a particular type of workload—inference on mixture-of-experts models. That's a real use case, and the improvement is genuine. But it's not magic. It's the result of better memory bandwidth, more efficient GPU design, and smarter system integration.
What's the actual competitive advantage for PEGATRON here? NVIDIA makes the chips. Why not just buy directly from them?
Because NVIDIA doesn't manufacture servers. They design GPUs. PEGATRON takes those GPUs and builds the entire system around them—the cooling, the networking, the power delivery, the chassis. They've been doing this for decades. They know how to make it reliable, how to manufacture it at scale, how to support it in the field.
Liquid cooling sounds expensive and complicated. Why would a typical enterprise choose that over air cooling?
At this density, you don't have a choice. Air cooling maxes out around a certain power density. Once you exceed that, the fans can't move enough air fast enough. Liquid cooling is more complex upfront, but it's actually more cost-effective at scale because it lets you pack more compute into the same space and keeps your power bills lower.
Who actually buys these machines?
Cloud service providers building massive AI data centers, and large enterprises that need serious AI infrastructure. Not startups. Not small companies. This is for organizations that are already thinking in terms of racks and power budgets and deployment timelines measured in months.