The artificial intelligence revolution hinges on hardware, and for years, Nvidia has reigned supreme with its versatile GPUs. But as AI workloads, particularly large language models (LLMs), demand ever-greater efficiency, a new contender is emerging.

Positron, a Reno-based startup founded in 2023, is led by Thomas Sohmers and Mitesh Agrawal, a duo leveraging purpose-built chips to take on Nvidia’s multipurpose dominance. With energy costs soaring, market concentration under scrutiny, and tariffs reshaping supply chains, their vision could redefine AI infrastructure.

The Industry Landscape: A Shifting Paradigm

Nvidia’s GPUs, like the H100, are engineering marvels — multipurpose tools designed for gaming, scientific computing, and now AI. Their adaptability has made them the backbone of the AI boom, powering everything from ChatGPT to autonomous vehicles, with a commanding 90% market share.

But that generalist nature comes with trade-offs: high power consumption and costs that strain data centers as inference workloads for LLMs surge. Unlike training, where raw compute reigns, inference prioritizes speed and efficiency — areas where Nvidia’s broad-focus chips weren’t originally optimized.

The International Energy Agency projects that data centers will double their electricity use by 2030. That’s a daunting challenge, especially as tech giants like Meta spend up to $65 billion on AI infrastructure in 2024 alone.

Meanwhile, tariffs on foreign chipmakers and U.S. policy efforts like the CHIPS Act are pushing companies to rethink their reliance on overseas supply chains. This opens the door for domestic innovators like Positron, whose chips are tailored specifically for AI inference, a focus they believe gives them an edge over Nvidia’s jack-of-all-trades approach.

The Duo: Sohmers and Agrawal

Thomas Sohmers, Positron’s founder and CTO, is a Thiel Fellow with a history of defying convention. His work at AI hardware startups like Groq and Lambda Labs sharpened his obsession with efficiency.

“Instead of slowing down AI progress, I think we should accelerate it — but by doing it using more energy-efficient means, by making it cheaper,” says Sohmers.

Mitesh Agrawal, who became CEO in early 2025, brings deep operational chops from Lambda Labs, where he scaled revenue from $500,000 to nearly $500 million and raised over $1 billion. Agrawal’s firsthand experience with Nvidia’s ecosystem, Lambda, which relied heavily on its GPUs, uniquely equips him to challenge it.

“The curve of technology for inference is just going up,” says Agrawal.

Together, their technical and scaling expertise forms a potent partnership — backed by $23.5 million raised in February 2025 from Flume Ventures, Valor Equity Partners, and Atriedes Management.

The Product: Purpose-Built for AI

Positron’s flagship product, the Atlas system, is now shipping to U.S. customers, targeting transformer model inference — the engine behind LLMs. Unlike Nvidia’s GPUs, which juggle multiple tasks, the Atlas is engineered solely for AI inference.

Built on Intel’s Altera line of FPGAs, Atlas claims:

  • 70% faster performance than Nvidia’s Hopper-based systems
  • 3.5x better performance per dollar and watt
  • Over 93% memory bandwidth utilization, compared to 10–30% on GPUs

Available as a 4U appliance with four FPGA cards or standalone PCIe cards, Atlas integrates with platforms like Hugging Face and OpenAI APIs — sidestepping Nvidia’s CUDA lock-in.

“We believe this appliance model — just tokens in, tokens out, a black box — is the easiest way for customers to purchase our hardware,” says Sohmers.

By focusing exclusively on inference, Positron sidesteps the inefficiencies of general-purpose chips, delivering what it claims is a 66% reduction in power use — a major benefit for energy-strapped data centers.

The FPGA-first strategy is intentional, allowing rapid iteration based on customer feedback.

“We don’t want to spend the massive amount of time and money on building an ASIC until we are absolutely sure,” says Sohmers.

Hundreds of PCIe cards are already deployed, and Positron continues to scale up production. Looking to 2026, Positron plans a custom ASIC with the highest memory capacity available of any accelerator in the world, allowing it to outperform NVIDIA’s announced Ruben platform on performance per dollar.

Riding Macro Waves

Positron’s timing is strategic. Nvidia’s dominance is under scrutiny as customers balk at high prices and search for inference-specific solutions.

“The industry is finally waking up to the dangers of Nvidia controlling 90% of the market,” says Sohmers.

With data centers reaching energy capacity, Positron’s reduced power consumption could extend the life of existing infrastructure and cut operational costs by up to 50%, the company estimates.

Positron’s U.S.-based supply chain — FPGAs made in Arizona, with ASICs planned via Intel Foundry Services — also dodges tariffs and aligns with “Made in America” incentives. In a climate rocked by export controls, that could prove decisive, especially as industries like blockchain mining and decentralized compute face similar geopolitical pressures to localize infrastructure.

A Reno Renaissance

Sohmers and Agrawal are also betting on Northern Nevada as a tech hub.

“It’s very vital that we have this technology in as many places as possible. And because we are from here, we really wanted to bring that ecosystem here,” says Agrawal.

With plans to grow from 21 to 35–40 employees in 2025 — and double thereafter — Positron is drawing top talent to Reno. The startup has earned support from Governor Joe Lombardo and University of Nevada, Reno President Brian Sandoval.

This local focus could decentralize AI innovation, much like how blockchain infrastructure pushed compute closer to the edge, away from centralized cloud dominance.

The Stakes Ahead

Positron’s Test Flight program, offering free trials of Atlas servers, lowers the barrier to entry. Early shipments are already building momentum.

Still, Nvidia remains a Goliath, and scaling hardware is no easy feat. But the trends are on Positron’s side: a market hungry for inference efficiency, energy ceilings limiting growth, and a national push for domestic chipmakers.

By designing for AI’s specific needs — not adapting to them — Positron is carving out a compelling new path.

Their next moves will determine whether they redefine the future of AI infrastructure. But the purpose-built revolution has already begun.