As demand for AI accelerates, power consumption has become a growing concern. Last year alone, Nvidia’s AI processors consumed more power than 1.3 million American homes. Industry projections suggest generative AI could account for over 1.5% of global electricity use by the end of the decade. Speslab, a finalist at the 2025 Startup World Cup Cyprus Finals, believes the industry has been thinking about this problem all wrong.
Their solution sounds almost too good to be true: a processor that draws just 20 watts (the same as an energy-saving light bulb) while supporting real-time AI inference for applications like computer vision and threat detection. Unlike traditional GPU-based solutions, Speslab’s architecture requires no expensive cooling infrastructure and runs efficiently in high-heat environments. “It can even work on a pole in the sun,” remarks founder Oleg Grishanin.
Follow THE FUTURE on LinkedIn, Facebook, Instagram, X and Telegram
The Electricity Bill
Speslab’s early work focused on video analysis for threat detection. But while the software showed promise, the hardware didn’t hold up.
“Everyone uses video cameras only to watch how they were robbed or killed,” says founder Oleg Grishanin. “That always seemed like a strange use case to us. From the beginning, we looked for ways to recognize threats before they happen.”
What they quickly found was that the processors needed to run these kinds of models weren’t just expensive, they were wildly inefficient. “When you need 2,000 graphics cards and you’re spending $2 million a year just on electricity, something’s not right,” Grishanin explains. “And that’s before cooling.”
Rather than optimize around existing hardware, they started from scratch. The result is what they call their “SHORT DATA” platform, a fundamentally different approach to how machines process visual information.
The Human Vision Model
SHORT DATA borrows from human perception: we don’t store everything we see, we extract meaning. “Like humans, computers get most of their data through vision,” Grishanin explains. “But instead of storing everything, we teach our systems to extract only what’s important and on the spot.”
It’s a major departure from the standard model of cloud-based AI, where vast amounts of data are collected, stored, and processed off-site. Speslab’s processors are designed to do the job locally and efficiently, using just 20 watts of power, roughly the same as a household light bulb.
For comparison, some AI accelerators consume more than 1,800 watts, with projections reaching 15,000 watts for future models, according to research from the Korea Advanced Institute of Science and Technology (KAIST). By operating at a tiny fraction of that energy, Speslab’s hardware can run on-site, without dedicated cooling systems or constant cloud access.
From Product to Platform
Speslab isn’t just building devices. It’s creating the tools for others to build with. What began as a focused effort to improve video-based threat detection has grown into a wider vision: a flexible hardware-software stack that others can adapt to their own AI applications.
“We’re not interested in selling pre-made solutions,” Grishanin says. “Our goal is to let others create their own neural systems without needing huge data centers or deep technical teams.”
According to the team, their platform is already “75% complete,” and they’re working toward making it developer-friendly.
Still, the company knows it faces challenges that go beyond engineering.
2025 Startup World Cup Cyprus Finals
The Speslab team recently relocated to Cyprus, drawn in by its growing tech hub and the quality of life. “Everyone who could move, moved,” says founder Oleg Grishanin.
Cyprus has offered proximity to key markets and a startup-friendly environment, but like many deep-tech teams, Speslab has found that visibility is often harder than invention.
“We desperately need marketing,” Grishanin says. “No one sees or hears us. We’re inventors by nature, we don’t know how to sell.”
Their appearance at the Startup World Cup Cyprus finals was part of an effort to break through that invisibility. The competition attracted over 100 applicants this year, with Speslab earning a spot among the top ten. For a team that describes typical reactions to their claims as “bewilderment that this is possible,” making the finals offered important validation.
“Few people believed that an airplane could fly,” their founder observes, referring to early market skepticism of their solution. It’s a sentiment that resonates in hardware development, where revolutionary claims often face intense scrutiny before acceptance.
Looking Ahead
Building chips is one of the most difficult verticals in tech, and potentially one of the most promising. While companies across industries adopt AI tools for everything from logistics to fashion, the infrastructure underneath remains inefficient and expensive.
Speslab’s approach represents a growing recognition that AI’s energy crisis isn’t just a cost problem but also an access problem. If only funded organizations can afford to run sophisticated AI systems, the technology’s benefits remain limited to the few.
“Our goal isn’t just to offer an alternative,” Grishanin explains. “It’s to make AI truly accessible, by rethinking everything from the chip up.”
The company’s long-term vision is characteristically bold: “We want to replace all data centers in the world with our equipment.” Rather than providing pre-built AI solutions, they want to create tools that let others build their own intelligent systems.
Whether their 20-watt processor can deliver on these promises remains to be seen. Independent testing and real-world deployment will ultimately determine if Speslab’s approach can challenge the GPU-dominated status quo.
But then again, sometimes the biggest breakthroughs come not from doing things better. They come from asking whether we’re doing the right things at all.
