With the surge in demand for large-scale AI deployment, the relatively unknown startup, Positron, is stepping up as a formidable competitor to industry giant Nvidia. It aims to do so by providing specialized, energy-efficient, and memory-optimized inference chips designed to alleviate the industry's escalating issues related to cost, power consumption, and availability constraints.

“Our standout feature is our efficiency in running advanced AI models—achieving 2x to 5x better performance per watt and dollar compared to Nvidia,” stated Thomas Sohmers, the co-founder and CTO of Positron, during a recent video call with VentureBeat.

This is undoubtedly promising news for major AI model providers, but Positron's leadership argues that it also benefits a wider range of enterprises, including those integrating AI models into their internal processes rather than offering them as customer services.

“Our chips are deployable in hundreds of existing data centers since they don't necessitate liquid cooling or high power densities,” explained Mitesh Agrawal, Positron’s CEO and former COO of AI cloud inference provider Lambda, in the same VentureBeat interview.

AI Scaling Hits Its Limits

Power limitations, rising token expenses, and inference delays are reshaping enterprise AI. Participate in our exclusive salon to learn how leading teams are:

• Transforming energy into a strategic advantage

• Designing efficient inference for substantial throughput improvements

• Achieving competitive ROI with sustainable AI systems

Reserve your spot to stay ahead: https://bit.ly/4mwGngO

Venture capitalists and initial users are on board with this vision.

Positron recently announced a successful $51.6 million Series A funding round led by Valor Equity Partners, Atreides Management, and DFJ Growth, with contributions from Flume Ventures, Resilience Reserve, 1517 Fund, and Unless.

Positron’s early clientele includes prominent enterprises and companies in sectors heavily reliant on inference. Confirmed deployments feature major security and cloud content networking provider Cloudflare, utilizing Positron’s Atlas hardware in its globally distributed, power-limited data centers, and Parasail, through its AI-native data infrastructure platform SnapServe.

Beyond these, Positron reports adoption across several vital sectors where efficient inference is crucial, such as networking, gaming, content moderation, content delivery networks (CDNs), and Token-as-a-Service providers.

These early adopters are reportedly attracted by Atlas’s capability to provide high throughput and reduced power consumption without the need for specialized cooling or revamped infrastructure, making it an appealing option for AI tasks within enterprise environments.

Entering a challenging market that is decreasing AI model size and increasing efficiency

However, Positron is venturing into a challenging market. The Information recently reported that rival AI inference chip startup Groq — where Sohmers previously held the role of Director of Technology Strategy — has scaled down its 2025 revenue forecast from over $2 billion to $500 million, emphasizing the volatility of the AI hardware sector.

Even well-funded companies face challenges as they vie for data center capacity and enterprise attention against established GPU providers like Nvidia, alongside the emergence of more efficient, smaller large language models (LLMs) and specialized small language models (SLMs) that can operate on devices as compact and low-powered as smartphones.

Yet, Positron's leadership remains optimistic, embracing the trend and dismissing potential impacts on its growth path.

“There’s always been this duality—lightweight applications on local devices and heavyweight processing in centralized infrastructure,” said Agrawal. “We believe both will continue to expand.”

Sohmers concurred, adding: “We envision a future where every individual might possess a capable model on their phone, but those will still depend on large models in data centers to generate deeper insights.”

Atlas is an inference-first AI chip

While Nvidia GPUs played a pivotal role in the deep learning revolution by accelerating model training, Positron argues that inference — the phase where models generate output in production — now represents the actual bottleneck.

Its founders describe it as the most under-optimized segment of the “AI stack,” particularly for generative AI workloads that rely on fast, efficient model serving.

Positron’s answer is Atlas, its first-generation inference accelerator tailored specifically for handling large transformer models.

In contrast to general-purpose GPUs, Atlas is fine-tuned for the specific memory and throughput requirements of modern inference tasks.

The company asserts that Atlas offers 3.5x better performance per dollar and up to 66% lower power usage than Nvidia’s H100, while also achieving 93% memory bandwidth utilization—significantly higher than the typical 10–30% range observed in GPUs.

From Atlas to Titan, supporting multi-trillion parameter models

Introduced just 15 months after its founding — with only $12.5 million in seed funding — Atlas is already being shipped and utilized in production.

The system supports up to 0.5 trillion-parameter models in a single 2kW server and is compatible with Hugging Face transformer models through an OpenAI API-compatible endpoint.

Positron is now gearing up to launch its next-generation platform, Titan, in 2026.

Built on custom-designed “Asimov” silicon, Titan will feature up to two terabytes of high-speed memory per accelerator and support models up to 16 trillion parameters.

Today’s leading models are in the hundred billions and single-digit trillions of parameters, but newer models like OpenAI’s GPT-5 are believed to be in the multi-trillions, and larger models are currently thought to be necessary to achieve artificial general intelligence (AGI), AI that surpasses humans in most economically valuable tasks, and superintelligence, AI that exceeds human understanding and control.

Crucially, Titan is engineered to operate with standard air cooling in conventional data center environments, sidestepping the high-density, liquid-cooled setups that next-gen GPUs increasingly demand.

Engineering for efficiency and compatibility

From the beginning, Positron designed its system to be a seamless replacement, allowing customers to use existing model binaries without code alterations.

“If a customer had to alter their behavior or actions in any way, that was a hurdle,” said Sohmers.

Sohmers elaborated that instead of building a complex compiler stack or restructuring software ecosystems, Positron concentrated narrowly on inference, developing hardware that directly utilizes Nvidia-trained models.

“CUDA mode isn’t something to oppose,” said Agrawal. “It’s an ecosystem to engage with.”

This practical approach enabled the company to swiftly launch its first product, validate performance with real enterprise users, and secure substantial follow-on investment. Additionally, its focus on air cooling versus liquid cooling makes its Atlas chips the sole option for some deployments.

“We’re committed to purely air-cooled deployments… all these Nvidia Hopper- and Blackwell-based solutions moving forward necessitate liquid cooling… The only locations for those racks are data centers currently being built in remote areas,” said Sohmers.

Overall, Positron’s ability to execute quickly and cost-effectively has helped set it apart in a crowded AI hardware market.

Memory is what you need

Sohmers and Agrawal highlight a fundamental shift in AI workloads: from compute-bound convolutional neural networks to memory-bound transformer architectures.

While older models demanded high FLOPs (floating-point operations), modern transformers require extensive memory capacity and bandwidth to operate efficiently.

While Nvidia and others continue to focus on compute scaling, Positron is betting on a memory-first design.

Sohmers noted that with transformer inference, the ratio of compute to memory operations approaches 1:1, meaning that enhancing memory utilization has a direct and significant impact on performance and power efficiency.

With Atlas already surpassing contemporary GPUs on key efficiency metrics, Titan aims to push this further by offering the highest memory capacity per chip in the industry.

At its launch, Titan is expected to provide a substantial increase over typical GPU memory configurations — without requiring specialized cooling or exclusive networking setups.

U.S.-built chips

Positron’s production pipeline is proudly domestic. The company’s first-generation chips were manufactured in the U.S. using Intel facilities, with final server assembly and integration also conducted domestically.

For the Asimov chip, production will transition to TSMC, though the team aims to maintain as much of the rest of the production chain in the U.S. as feasible, contingent on foundry capacity.

Geopolitical resilience and supply chain stability are increasingly becoming crucial purchasing factors for many customers — another reason Positron believes its U.S.-made hardware presents a compelling alternative.

What’s next?

Agrawal noted that Positron’s silicon is designed not just for broad compatibility but for maximum utility across enterprises, cloud services, and research labs alike.

While the company has not yet named any frontier model providers as customers, he confirmed that outreach and discussions are ongoing.

Agrawal emphasized that selling physical infrastructure based on economics and performance—not tying it to proprietary APIs or business models—contributes to Positron’s credibility in a skeptical market.

“If you can’t convince a customer to deploy your hardware based on its economic value, you’re not going to be profitable,” he said.