MSI’s $5,090 RTX 5090 Lightning Z cracks from thermal shock during 2,500W BIOS test

MSI’s GeForce RTX 5090 Lightning Z isn’t built for everyday gaming rigs. With a $5,090 price tag, the card is aimed either at hardcore collectors or — more realistically — extreme overclockers chasing benchmark records. In the hands of YouTuber Alva Jonathan, it did exactly that, until an aggressive BIOS and a sudden thermal imbalance cracked the GPU core, killing one of the samples mid-experiment.

Jonathan was involved with the RTX 5090 Lightning Z from an early stage. MSI Taiwan contacted him in August 2025 to consult on the card’s development, first sending two early PCB samples without coolers, followed by three retail units. For initial testing, he built a custom cooling solution before eventually moving to liquid nitrogen (LN2).

The card stands apart from standard RTX 5090 models in both price and design. On paper, it is clearly engineered for competitive overclocking. Key features include:

• Dual 12V-2x6 power connectors;
• Up to 1,000W total power limit (600W per connector);
• 40-phase VRM for high current delivery;
• A dedicated 2,500W XOC BIOS for competitive overclocking;
• An integrated 8-inch display replacing the traditional backplate for telemetry readouts.

Jonathan began with what could almost be described as a modest overclock for hardware of this class: 3.25 GHz at 1.05V. Even at that level, the card was already drawing more than 700W.

With an 800W power limit in 3DMark Port Royal, the Lightning Z achieved 43,112 points. For context, MSI’s RTX 5090 Suprim Liquid previously scored in the 40,000–41,000 range, while a standard RTX 5090 typically lands around 36,000–37,000 points. During that run, the GPU peaked at 772W, with power distributed evenly across both 16-pin connectors.

To push further, Jonathan teamed up with ARX (arxidmedia) and transitioned to LN2 cooling. Even under liquid nitrogen, keeping the GPU within a stable operating window proved difficult. The heatsink in contact with the core measured -40°C, yet the GPU itself remained in positive territory, climbing as high as 9°C under load. At 1.12V and 3.42 GHz, power consumption exceeded 1,000W.

Under LN2, the safe operating range appears narrow — roughly between 0°C and 15°C. In one run, temperatures reached 21°C, and the system immediately crashed.

Some benchmarks were more forgiving. In GPUPI, Jonathan briefly hit 3.6 GHz at around 0°C. The team ultimately settled at 3.5 GHz as a stable operating point and broke the HWBot world record for GPU compute performance in Geekbench 5, posting a score of 683,433. That result still stands.

The turning point came when the team switched to the 2,500W XOC BIOS. They were working with an earlier revision that reportedly applied too much voltage too quickly.

At just 1.2V — manageable under LN2 but risky at ambient temperatures around 25°C — one of the GPUs failed catastrophically. The core visibly cracked, likely due to thermal shock. One part of the die remained extremely cold while another developed a hotspot, creating a temperature imbalance the silicon could not withstand.

Roughly $5,000 was lost in a single moment. The rest of the board components remain intact, meaning the card could theoretically be revived with a donor core.

Despite securing the Geekbench 5 world record, Jonathan did not manage to break other targets, including 3DMark Solar Bay Extreme. Jonathan still has four additional samples for continued testing and plans to improve cooler mounting. He also indicated that he may revert to a more stable retail BIOS for future attempts.