Today, with top-tier graphics cards consuming hundreds of watts and AI server racks requiring kilowatts per unit of equipment, the power supply has transformed from a secondary component into the foundation of the entire system. This is why FSP Group, which has become one of the world leaders in this field over three decades, came to Computex 2026 in Taipei not just with an updated product range, but with a clearly defined shift in vector. Its booth was dedicated to the idea of “Powering AI Together” — and behind this slogan was not an abstract declaration, but a systemic portfolio of solutions covering everything from server racks to cases for home workstations.
Server Solutions for AI
The central place in the exhibition was occupied by the CANNON PRO 3300W power supply. It was created not for gaming personal computers, but for local AI stations — systems equipped with several flagship graphics accelerators. We are talking about machines capable of simultaneously powering up to six RTX 5090 or RTX 5080 level video cards. Such solutions are in demand not in data centers, but in research laboratories, universities, and small companies that, for reasons of confidentiality or cost control, prefer to deploy their own infrastructure rather than rent cloud capacities. For such scenarios, FSP also offers 4U cases that can be used in both tower and rack configurations.
The point is not that FSP engineers simply decided to make a very powerful unit. The main thing is how they achieved this. This involves not just a simple increase in components, but a rethinking of the physics of the process, driven by the AI infrastructure's requirements for uninterrupted operation and power density. Therefore, CANNON PRO is not just a power supply, but a well-thought-out power platform capable of operating at its limits 24 hours a day, 7 days a week, which is a critically important requirement for systems training neural networks.
Using GaN and Totem-Pole PFC
The key role in achieving such performance is played by the component base and topology. Instead of conventional silicon diodes, gallium nitride (GaN) based semiconductors are used here. Thanks to these components, an efficiency of up to 99.4% was achieved even at full load of 3300W, while fitting all the high-performance components into a case measuring 150×200×86 mm.
However, the main feature of CANNON PRO is related to the use of Totem-Pole PFC (bridgeless active power factor correction) technology with continuous conduction mode (CCM). This solution, previously characteristic of industrial equipment, provides high efficiency in continuous mode. Traditional PFC topologies with a diode bridge have significant conduction losses. Bridgeless Totem-Pole PFC minimizes these losses, as at any given time, current passes through only one high-frequency and one low-frequency transistor.
However, this topology presents a serious technical challenge: in it, MOSFET transistors operate in synchronous rectifier mode, and their built-in diodes are subjected to extreme loads. With classic silicon MOSFETs with a super-junction (SJ), the reverse recovery time of the diode is long, which, during hard switching, leads to enormous losses and can even destroy the transistor. This is where the advantage of GaN transistors is revealed — they practically have no reverse recovery. This allows for the implementation of a Totem-Pole PFC circuit without the risk of failure and with minimal losses, operating in hard-switching mode at high frequencies, which, in turn, increases the power density of the unit.
In addition, instead of one large transformer, which in traditional units is the main source of heat, CANNON PRO uses three separate transformers. This allows for a more even distribution of heat dissipation throughout the case volume and significantly improves cooling efficiency, which is critical for systems with multiple GPUs that generate an enormous amount of heat. The unit's printed circuit board is covered with a special protective varnish — conformal coating, which protects components from moisture, dust, and contaminants. Previously, this technology was the domain of military and industrial orders; now it is becoming a standard for new FSP consumer products.
The unit is equipped with six native 12V-2x6 connectors (ATX 3.1 and PCIe 5.1 standard), allowing modern cards to be connected directly, without adapters. Four Japanese capacitors at 450V with an operating temperature range up to +105°C are used, ensuring parameter stability under high loads. The entire resulting design is aimed at providing high reliability (MTBF) and stability, since for AI systems, a sudden power failure means lost computation time, which is absolutely unacceptable. Sales in the CIS region are scheduled to start at the end of 2026.
TWINS PRO and the New Generation of CRPS
The TWINS PRO series was an equally successful novelty. These are the first power supply units with hardware redundancy, made in a standard ATX case (PS2 ATX form factor). Previously, such solutions were the domain of server racks, but now two independent power modules are combined in one case that can be installed in a regular workstation.
The principle of operation is simple: if one of the modules fails, the second one almost instantly takes over the entire load, ensuring uninterrupted power (1+1 architecture). The TWINS PRO 1400W model is 80 PLUS Platinum certified, supports hot-swapping of modules, and monitoring via PMBus and USB with control through proprietary FSP Guardian software. The main purpose of such systems is SMBs, IoT infrastructure, edge computing, and 5G equipment, where system downtime is unacceptable.
CRPS Solutions for Data Centers and High-Density Systems
For the professional segment, CRPS (Common Redundant Power Supply) solutions were presented at the exhibition, complying with OCP/DC-MHS (Open Compute Project / Data Center Management Hardware Specification) specifications. Here, FSP demonstrated a full range — from modules to entire power “shelves,” forming the basis of data center-level AI infrastructure. These systems support N+1 redundancy and hot-swapping, ensuring uninterrupted operation of critical AI clusters.
- The TWINS CRPS 3200W module with two redundant units and over 94% efficiency (80 PLUS Titanium certificate) — for dense AI servers where every watt and thermal footprint are critical.
- Modules with an output voltage of 54V and up to 3600W with N+1 redundancy support — this solution reduces energy transmission losses within the rack, as it allows for thinner cables to transmit more power, which is especially relevant for racks with high density GPU accelerators.
- And finally, the 33-kilowatt Power Shelf — a rack-level solution that consolidates power for the entire rack. It contains six 5500W modules and achieves a peak efficiency of 97.5%, confirmed by the 80 PLUS Ruby standard. This reduces the number of single points of failure and simplifies capacity planning for large clusters.
These products confirm that FSP today has a portfolio of solutions for data centers of almost any scale — from isolated servers to large clusters.