TheStage AI is betting that the next wave of AI runs on phones rather than servers

Every AI company discovers the same thing at roughly the same moment. The demo works, users appear, and the infrastructure invoice starts growing faster than the revenue attached to it. TheStage AI, a United States-registered deep tech company built by four university friends holding doctorates in mathematics and neuroscience, sells software aimed squarely at that moment.

“The core problem we solve is that AI is becoming increasingly expensive and complex to deploy at scale,” said Kirill Solodskikh, Co-Founder and Chief Executive Officer of TheStage AI. “Many companies can prototype with open-source models or APIs, but once they move into production, they face rising GPU costs, latency constraints, power-consumption limits, hardware fragmentation, and difficult trade-offs between model quality, speed, and cost.”

The company’s platform optimises trained neural networks so they run faster and cheaper across NVIDIA GPUs, Apple Silicon, Qualcomm NPUs and edge hardware such as NVIDIA Jetson boards. Its core product, ANNA, or Automated Neural Networks Accelerator, grew out of research the founding team presented at CVPR in 2023, work that earned an award nomination and drew on years at Huawei’s Noah’s Ark Lab optimising camera models for the P50 and P60 smartphones. TheStage AI raised $4.5 million in April 2025 from backers including Atlantic Labs, DVC, AAL VC and DFINITY founder Dominic Williams.

Inference costs now decide which AI products survive contact with real users

McKinsey has estimated that GPU infrastructure accounts for as much as 70% of the cost of deploying AI systems, which converts optimisation from an engineering nicety into a question of whether a business model works at volume. The pain concentrates in products where every user interaction triggers a model call: voice agents, tutoring apps, generative media tools and mobile assistants.

Praktika, an English-learning app whose AI avatar tutors require real-time inference, saw GPU costs climb alongside its user growth. TheStage AI reduced the company’s GPU load by three times, saving it tens of thousands of dollars each month. The generative design company Recraft has worked with TheStage AI on diffusion model acceleration, and Brilliant Labs, Neuphonic and Mirelo AI sit among current and target customers.

“At a high level, our mission is to increase the intelligence density of AI models: more useful AI performance per dollar, per watt, and per device,” Solodskikh said.

The same economics are pulling workloads onto devices. The edge AI market was valued at around $30 billion in 2026 and is forecast to reach $118.7 billion by 2033, while on-device inference silicon from Apple, Qualcomm and Intel accounts for a separate $25 billion to $35 billion market. Devices certified for 40 or more TOPS of on-device processing are expected to exceed 100 million units by 2027. TheStage AI is building an SDK for hybrid voice assistants that runs inference locally on smartphones and laptops, targeting NPUs so the GPU and CPU stay free for the application, and says its Apple SDK delivers close to twice the inference speed of Apple’s CoreAI runtime with up to six times lower power draw than 4-bit MLX Metal inference.

The competitive field is crowded, extremely well capitalised, and consolidating fast

The inference market has become one of the most aggressively funded categories in technology. Fireworks AI raised $1.5 billion in July 2026 at a $17.5 billion valuation with annualised revenue past $1 billion. Together AI raised $800 million at $8.3 billion the same month, and Baseten closed a Series F of comparable scale. Cerebras completed the largest technology listing of 2026, NVIDIA licensed Groq’s LPU technology for roughly $20 Billion, and Cloudflare absorbed Replicate. Technavio forecasts the inference-as-a-service market growing at a 22.1% compound annual rate through 2030.

Those companies mostly rent GPUs, add serving software and resell the result as an API, which leaves their margins exposed as AWS, Azure and Google Cloud expand their own managed inference services. TheStage AI has taken a different route by selling the optimisation layer itself rather than the capacity underneath it, which lets customers keep control of where models run. Its revenue comes from a usage-based inference runtime billed on tokens, requests, GPU or NPU time, or device usage, with integration fees for larger enterprise deployments.

“Through customer development, we also learned that model owners often do not want to share customised or fine-tuned models, so supporting on-premise and customer-controlled deployments became important to adoption,” Solodskikh said.

The exposure is obvious enough. A four-year-old company with $4.5 Million behind it is competing for attention against rivals holding billions, and hardware vendors keep absorbing optimisation techniques into their own free toolchains. The counter-argument is that neither NVIDIA nor Apple has much incentive to make models run well on the other’s silicon.