We are entering a new era of tech innovation driven by intelligent hardware platforms – innovation catalyzed by several converging factors in the tech ecosystem that are enabling both breakthrough functionality and lower cost for these new companies.

This might sound like a contrarian view. For the last two decades, VC conventional wisdom has favored software-only business models because of their higher gross margins, rapidly scalable distribution, and customer lock-in. Notwithstanding the physicality of the “silicon” in Silicon Valley and notable exceptions such as Tesla, software-only businesses have been the dominant recipients of the last two decades of venture capital funding.

But this tide is beginning to shift with AI.

While machine learning has been, definitionally, at the heart of robotics, autonomous vehicles, and IoT systems from the start, the dramatic success of language-based AI is inspiring and enabling founders working in the physical world to challenge that dominance. If LLMs can turn language into useful reasoning — summarization, translation, and now, increasingly, actual problem-solving — then physical AI systems should be able to make increasingly sophisticated use of sensor data to predict and engage with the real world. We believe the potential applications across every sector of the economy are enormous.

We believe the next “inning” in this game is hardware-enabled software: an inversion of the “software-enabled hardware” category that has taken a backseat to software-only business models. Hardware-enabled software is a category of vertical software that anchors intelligence in first-party data gathered by hardware deployed by the company. Examples include:

• Coram AI*, a company whose series A we led in January last year, provides a physical security platform including AI-enabled cameras; a video management system (VMS) with built-in key detections (such as for slip-and-fall incidents, or a weapon being carried in a workplace or school); biometrics-based access control; visitor intercom; and natural language-based incident investigation of video.

• GenLogs*, whose series B we led in December last year, uses a proprietary network of specialized highway and port cameras, along with satellite data, to identify individual trucks, derive shipping lanes, and provide third-party logistics suppliers, shippers, insurers and other customers of with a ground-truth visualization of freight movement. Use cases include fraud prevention, freight-carrier onboarding, insurance underwriting, DoT compliance, and law enforcement.

• Kargo deploys purpose-built camera towers to capture pallet-level data at warehouse dock doors and reconcile inbound and outbound shipments with advanced shipping notices (ASNs) and warehouse management systems (WMSs).

• Sage*, whose Series B we participated in in 2024, uses sensors and cameras in senior living facilities to unify nurse call, fall management, and clinical data into a single view combining real-time accident monitoring with longer-term health and service analysis.

Each of these companies uses specialized cameras and other sensors to capture data as the on-ramp to initiating or managing important workflows. In some cases, such as with Kargo’s technology, there is some displacement of manual labor. But the greater value generated by these companies typically comes from real-time monitoring, data assimilation, and workflow orchestration far beyond what humans are doing today in those contexts. Critical to each platform is the vertically-specific combination of physical world input with AI detection, AI analysis, and conventional software features such as data visualization and workflow.

What makes this different from the IoT wave?

Earlier-generation CV and IoT systems had similar aspirations with their instrumentation of the physical world, but the outputs still required human interpretation. In contrast, the hardware-enabled software systems emerging today leverage vision transformers and multimodal models to convert video and sensor data into embeddings (structured representations) that capture context, sequence, intent, and causality. This understanding, in turn, enables software to begin to “reason” about what is happening within the physical world and take action from there.

While still early, this evolution is analogous to the breakthrough from decision trees based on procedural rules to open-ended AI reasoning based on natural language and reinforcement learning. And just as the increasing capability of general purpose LLM models is enabling a new wave of AI-native application companies, we foresee that the general-purpose, physical-world models will propel the same for hardware-enabled software startups.

The economics of hardware have also quietly gotten better.

Manufacturing costs for cameras and sensors have fallen sharply, and edge-compute connectivity is cheaper than ever. These dynamics make “always-on” data collection now economically feasible across large, physical environments. In parallel, hardware financing structures have matured, allowing startups to pay for deployments over time and to match them to multi-year software contracts. In other words, the bottleneck to building and deploying hardware systems is no longer the cost of the hardware itself: Today, hardware-enabled software can be embedded into physical contexts with far less capital than ever before.

And the moats are just as strong, if not stronger, than software.

Once they overcome operational and deployment switching costs, hardware implementations can create stickiness just as durable as the most-adopted software platforms by fitting themselves into the operational fabric of their customers. From there, value compounds through automation and data aggregation. A universe of data awaits capture and interpretation: While the public internet only offers one to two petabytes of high quality text for training LLMs, the physical world is essentially an infinite training ground.

The best companies in this category are now showing characteristics long associated only with SaaS: long-term recurring contracts, mission criticality, and strong retention and expansion. Already, venture investors have seen multi-billion dollar valuations and outcomes in the category from the likes of Samsara, Motive, Flock Safety and others — and we believe the next five to 10 years will yield many more.

General purpose robotics is exciting. But vertical, purpose-built, hardware-enabled software is the pragmatic path for AI to enter the physical world today. We’re actively looking to partner with deeply technical founders who understand the physical domains they are building for and have a pragmatic thesis about bringing modern AI capabilities to bear within them.