Pearl Network Allegedly Consumes 112 MW on 320,000 GPUs for Random Matrix Math

*A preprint study asserts that Pearl’s claimed AI computation network delivers no verified useful output while driving up GPU rental prices.*

A preprint claims Pearl operates the equivalent of 320,000 RTX 3090-class GPUs and draws 112 megawatts for work that amounts to random matrix math. The same report links the network’s activity to a 38 percent rise in GPU rental costs.

The study examined Pearl’s operations and found no evidence of training runs, inference workloads, or other productive AI tasks. Instead, the GPUs appear occupied with arbitrary linear-algebra exercises that match patterns seen in certain cryptocurrency mining schemes rebranded as AI services.

Power draw at this scale equals the output of a small fossil-fuel plant running continuously. The authors note that the claimed AI results cannot be reproduced or audited, leaving the network’s actual contribution at zero under independent review.

GPU rental markets tightened as the reported demand from Pearl and similar operations absorbed available capacity. Prices rose 38 percent in the period covered by the data, affecting researchers and smaller teams that rely on short-term cloud instances.

No public response from Pearl appears in the preprint or accompanying coverage. The authors present their findings as an observational analysis of network traffic and power signatures rather than a direct audit of internal systems.

The episode shows how easily large GPU fleets can be assembled under an AI label while performing tasks that consume electricity and hardware without advancing models or applications. Operators who rent capacity now face higher costs and tighter availability even when their own workloads have clear scientific or commercial value. Sustained misallocation at this level also adds measurable load to power grids already under pressure from data-center growth.

---

Sources:

• Tom's Hardware

{
  "excerpt": "A preprint claims Pearl’s network runs 320,000 RTX 3090 GPUs at 112 MW on random matrix math with no useful AI output, pushing rental prices up 38%.",
  "suggestedSection": "ai",
  "suggestedTags": ["ai-mining", "gpu-power"],
  "imagePrompt": "Rows of server racks stretch into darkness under a low ceiling, their indicator lights forming faint grids against matte metal panels. A single thick power cable snakes across the foreground floor, lit by a cool overhead source. muted color palette, cinematic lighting, 16:9"
}