After a year in stealth mode, Hyphastructure has emerged with an ambitious mission: to bridge the gap between artificial intelligence and the physical world. The company, backed by notable investors from the power infrastructure space, is applying a 15-year playbook of deploying distributed and decentralized power systems to the compute industry. Their vision? Harmonize human and machine cognition in real-world environments where autonomous drones, robots, and smart city systems operate�not just on screens, but in the physical spaces where latency can mean the difference between success and catastrophic failure.
The timing couldn’t be more critical. While AI language models have scaled rapidly through well-established training laws, physical AI has lagged behind due to limited training data and deployment friction. Hyphastructure’s newly launched platform addresses this fundamental challenge with its Hyphagrid software�a latency router engineered from the ground up to deliver zero-variance, sub-10ms compute for geospatially distributed AI workloads. As NVIDIA CEO Jensen Huang declared physical AI “the next big thing” at CES 2025, and the market races toward $124 billion by 2030, Hyphastructure’s CEO Michael Huerta sat down with VMblog to discuss how his company plans to power the next generation of autonomous vehicles, smart cities, and edge-deployed AI systems.
VMblog: Hyphastructure launched a couple of weeks ago – can you share some details about the company?
Michael Huerta: The company had been in stealth for about a year, and we are very excited for the launch.
About 12 months ago we identified pockets of customer demand for AI edge inferencing. Basically, these were workloads that are latency sensitive and that have geospatial requirements. The founding team and I have been deploying distributed and decentralized power infrastructure for about 15 years, and we are applying that playbook to compute. The company is backed by notable investors from the power infrastructure space.
We will also be announcing recent hires, some of which come from the largest companies in the AI and computing business. Hyphastructure’s vision is to harmonize human and machine cognition in the real, physical world.
VMblog: Hyphastructure just launched the world’s first distributed Edge Cloud Network purpose-built for physical AI. Please give us some insight into this and what it offers.
Huerta: We believe physical AI is just not getting the same traction as AI language models. And that is because the scaling laws that have powered training for AI language models have not yet been applied for physical AI. Specifically, multi-modal and visual-language-action models. The reason for that is a lack of training data and friction in use cases. It’s a lot harder and more expensive to deploy AI in the real world. And the flywheel of training, inference, and adoption is a lot slower.
Robots and autonomous drones do not exist as words on a screen. The models powering these physical devices require high density inferencing compute with zero-variance latency in order to handle all the entropy of the real world. And typically, that real world use case has mobile, geospatial requirements. It turns out that on-board compute and centralized data centers are not that good for that. Our network has been engineered from the ground up to serve these kinds of workloads to help the physical AI community achieve the same scaling laws that have powered the LLMs. That is, getting the compute off the physical machine and under zero-variance, low latency.
VMblog: What are some of the key customer benefits?
Huerta: Hyphastructure’s software, Hyphagrid, is basically a latency router for real world inferencing. What we’re doing is powering the geospatial and zero-variance latency demands of physical AI. Our customer picks a geospatial and latency subscription and Hyphagrid automatically orchestrates an SLA to meet the subscription based on customer latency budget, geospatial requirements, inferencing and other requirements. Hyphagrid leverages its own AI infrastructure on Hyphastructure’s balance sheet and third party-owned infrastructure at select colocation facilities to fulfill the subscription SLA for the customer.
Specifically, what this means is that customers get:
- A cloud console that orchestrates deployment of AI inference workloads according to the lowest latency possible, along other parameters, (i.e., size of the infrastructure available, energy available etc.)
- Software defined networking that allows seamless movement of AI workloads from site to site (for mobile AI Inference use cases)
- Storage integration at the edge to cut even more the latency (from external S3)
- Tight integration with Intel’s Xeon and Gaudi AI Accelerators (and possibly others in the future)
VMblog: Can you share some specific use cases for this type of technology?
Huerta: The physical AI market is projected to grow from $12 billion to $124 billion by 2030, with NVIDIA CEO Jensen Huang declaring physical AI “the next big thing for AI” at CES 2025. Many of these applications will be latency-sensitive, and it’s not practical for all of this compute to be onboard the devices. Similarly, not every location can build its own micro data center. We anticipate a massive edge network will be required to support these workloads.
Hyphastructure supports several breakthrough use cases across diverse industries. For smart cities, it delivers the first edge-based architecture capable of coordinating traffic, crime, and emergency services across urban areas. In smart retail, it enables real-time shelf monitoring and personalized offers at the edge without relying on on-premises resources. For autonomous systems and robotics, it is the first decentralized platform to support collision-avoidance inference through V2V networks, a capability critical for next-generation vehicles, robotics, and drones. Additionally, in gaming and interactive media, Hyphastructure powers the first sub-10ms edge compute service, powering immersive AR/VR experiences that cannot tolerate the latency of centralized cloud systems.
VMblog: Why Intel Gaudi instead of AMD or NVIDIA?
Huerta: At Hyphastructure, we are focusing on scalable, cost-efficient AI infrastructure that can be deployed quickly and flexibly. Intel Gaudi gives us the performance we need for large-scale inference and training workloads, but at a fraction of the cost and power draw of comparable NVIDIA or AMD. Just as important, it’s built on an open architecture that avoids vendor lock-in and allows us to innovate freely across our stack. Gaudi lets us deliver high throughput, predictable performance, and better economics for our customers, all without compromising capability or flexibility.
VMblog: What is next for Hyphastructure?
Huerta: Right now, self-driving relies on each vehicle being its own data center. This is because camera sensors on autonomous vehicles must detect and recognize their surrounding environment within 3ms with 99.9999% reliability to meet safety standards. We believe Hyphastructure can complement these workloads as well as facilitate V2V communication to improve outcomes and unit economics for all vehicles. We are also busy now installing the first POC system where potential clients could do their tests of the platform. In addition, we are looking to deploy 26 MW during 2026 in many distributed data centers.
##
