The next generation of AI systems will require purpose-built facilities
In Singapore this October, SuperX AI unveiled a rack-scale computing system that does not merely improve upon existing data center technology — it renders much of it obsolete. The GB300 NVL72 delivers 1.8 exaFLOPS of AI performance in a single liquid-cooled rack, a concentration of power so extreme that it demands entirely new approaches to electricity, cooling, and physical infrastructure. It is a moment that marks the closing of one era in computing and the uncertain opening of another — one in which the ambitions of artificial intelligence have finally outgrown the buildings designed to contain them.
- The largest AI models now contain a trillion parameters, and the data centers built to train them were never designed for this scale — the old infrastructure is quietly becoming a bottleneck.
- SuperX's GB300 NVL72 forces a reckoning: its power demands require 800-volt DC distribution systems that most facilities simply do not have, making adoption a question of total reinvention, not upgrade.
- Liquid cooling, a Grace-Blackwell GPU pairing, and 2,304GB of unified memory are engineered together as a single organism — the system only works because every component was designed to work with every other.
- SuperX is packaging this not as a server to be purchased but as a Prefabricated Modular AI Factory — a complete, integrated solution that challenges the fragmented, multi-vendor way data centers have always been assembled.
- Hyperscale operators, sovereign governments, research institutions, and industrial manufacturers are all in the crosshairs, signaling that purpose-built AI infrastructure is no longer a niche ambition but an emerging industry standard.
SuperX AI announced this week a machine that forces a genuine reckoning with how data centers must be built. Unveiled in Singapore on October 16th, the GB300 NVL72 System delivers 1.8 exaFLOPS of AI performance from a single rack — a figure that only becomes meaningful when you consider that existing infrastructure was never engineered to support it.
The system sits at the collision of two long-running forces: the explosive growth of AI models, which now reach a trillion parameters, and the hard physical limits of conventional data centers. Air cooling, standard electrical distribution, and traditional server layouts have all become constraints. SuperX's answer is to abandon those assumptions entirely.
At the heart of the platform are 72 NVIDIA Blackwell Ultra GPUs paired with 36 Grace CPUs in a deliberate 2-to-1 ratio. The Grace processors manage memory-intensive tasks with efficiency; the Blackwell chips supply raw mathematical power. Between them sits 2,304 gigabytes of high-bandwidth memory and a 900GB-per-second chip-to-chip connection — enough to hold trillion-parameter models in place and keep the GPUs continuously fed without bottlenecks.
Keeping this density functional requires liquid cooling, which carries thermal energy away far more efficiently than fans. It also requires 800-volt DC power distribution — once considered an efficiency luxury, now, SuperX argues, a structural necessity. Higher voltage means less current, less heat lost in the wires, and a more compact, reliable power supply.
Rather than selling a server, SuperX is offering what it calls a Prefabricated Modular AI Factory: compute, cooling, and power infrastructure designed as a unified whole. This breaks from the traditional data center model of assembling hardware from separate vendors and hoping for compatibility.
The company is targeting hyperscale operators, national governments building sovereign AI capacity, research institutions running long climate and physics simulations, and manufacturers developing industrial digital twins. Across all of them, the underlying message is the same: the era of incremental improvements to legacy infrastructure is closing, and the next generation of AI will require facilities built from the ground up around its demands.
SuperX AI announced this week the arrival of a machine that represents a genuine shift in how data centers will need to be built. The GB300 NVL72 System, unveiled in Singapore on October 16th, is a single rack of computing hardware that delivers 1.8 exaFLOPS of artificial intelligence performance—a number so large it requires context to mean anything. To put it plainly: this is the kind of concentrated computational power that existing data center infrastructure was never designed to handle.
The system sits at the intersection of two technological currents that have been running in parallel for years. One is the relentless growth of AI models themselves. The largest language models now contain a trillion parameters—a trillion individual adjustable weights that need to be trained, stored, and computed across. The other current is the physical limits of traditional data centers. Air cooling, conventional electrical distribution, the spatial footprint required to house servers—all of these have become constraints. SuperX's platform is built on the premise that you cannot solve a new problem with old infrastructure.
At its core, the GB300 NVL72 combines 72 of NVIDIA's Blackwell Ultra GPUs with 36 Grace CPUs in a carefully balanced 2-to-1 ratio. The Grace processors handle memory-intensive work and general computation with exceptional efficiency; the Blackwell chips provide raw processing power for the mathematical operations that define modern AI. Between them sits 2,304 gigabytes of high-bandwidth memory—enough to hold the weights and intermediate calculations of the largest models without forcing the system to constantly shuffle data in and out of slower storage. A 900 gigabyte-per-second connection between the Grace and Blackwell components ensures that this data moves at speeds that keep the GPUs fed and working.
But raw performance numbers tell only half the story. The real innovation lies in what it takes to keep this much hardware functioning in a single physical space. Traditional air cooling cannot dissipate the heat generated by 72 high-end GPUs running continuously. SuperX's solution is liquid cooling—pumping coolant directly through the hardware to carry away thermal energy far more efficiently than fans ever could. This allows the company to pack vastly more computing power into a standard rack footprint, the kind of equipment that already fits through data center doors and onto standard floor space.
Equally critical is the power delivery system. A conventional data center distributes electricity at relatively low voltages—typically 480 volts—and then steps it down further at each server. This works fine for modest power draws. But a single GB300 NVL72 rack consumes enormous amounts of electricity. SuperX has designed the system around 800-volt direct current power distribution, a technology that was once considered a luxury efficiency improvement but is now, the company argues, a fundamental requirement. Delivering power at higher voltage means less current flowing through the same wires, which means less energy lost as heat in the distribution system itself. It also means the power supply can be more compact and more reliable.
The company frames the GB300 NVL72 not as a standalone product but as the centerpiece of what it calls a Prefabricated Modular AI Factory—essentially a complete, integrated solution that includes the cooling system, the power infrastructure, and the compute hardware all designed to work together from the start. This is a significant departure from how data centers have traditionally been built, where you buy servers from one vendor, cooling from another, and power distribution from a third, then hope they all work together.
SuperX is targeting three broad categories of customers. Hyperscale operators and national governments building sovereign AI infrastructure need the kind of compute density this system provides to train the largest models at reasonable cost. Research institutions tackling problems in physics, materials science, and climate modeling need both the raw power and the efficiency—these simulations can run for weeks, and energy costs become a major factor. And manufacturers building digital twins—virtual replicas of physical systems used for testing and optimization—need the combination of CPU and GPU performance that the Grace-Blackwell pairing provides.
The announcement signals something larger about the trajectory of AI infrastructure. The era of incremental improvements to existing data center designs appears to be ending. The next generation of AI systems will require purpose-built facilities, designed from the ground up around the specific demands of training and running trillion-parameter models. SuperX is betting that companies willing to rethink their entire infrastructure approach will find themselves with a significant competitive advantage.
Notable Quotes
Legacy infrastructure cannot adequately support these next-generation workloads— SuperX AI Technology Limited
Advanced power solutions, particularly 800VDC, have shifted from an efficiency advantage to a fundamental necessity— SuperX AI Technology Limited
The Hearth Conversation Another angle on the story
Why does a single rack of computers require an entirely new approach to power distribution?
Because the concentration of compute is so extreme. You're fitting the equivalent of thousands of traditional servers into a space that used to hold maybe a dozen. The power density—watts per square foot—becomes almost incomprehensible with air cooling. Direct current at 800 volts isn't just more efficient; it's the only way to reliably deliver that much electricity without the distribution system itself becoming a bottleneck or a fire hazard.
The specs mention 2,304 gigabytes of memory. Is that unusually large?
For context, most high-end consumer computers have 32 or 64 gigabytes. This system has roughly 36,000 times that amount. It's designed so that a trillion-parameter model can sit entirely in memory during training, along with all the intermediate calculations. Without that, you're constantly moving data between fast memory and slower storage, which kills performance.
Who actually needs this? What company or institution looks at this and thinks, yes, we must have it?
A cloud provider training the next generation of large language models. A government building national AI capability. A research lab running climate simulations that take weeks. These aren't academic exercises—they're the organizations that will define what AI can do in the next five years. If you're not in that category, you probably don't need it.
The company calls this a modular AI factory. What does that mean in practice?
It means you're not assembling components from different vendors and hoping they work together. You're buying a complete, integrated system where the cooling, power, and compute are all designed as one unit. You can theoretically link multiple racks together, but each one is self-contained and optimized.
What happens to older data centers when this becomes standard?
They become less competitive for the most demanding workloads. If you're training the largest models, you need this density and efficiency. Older facilities can still run inference—serving already-trained models to users—but the cutting edge of AI development will migrate to infrastructure built for it.