Analysis
Anthropic launched a research preview of the Model Hardware Standard on Thursday, a protocol for letting AI agents directly operate physical lab and manufacturing equipment -- microscopes, robotic arms, laser calibration tools -- through a single, standardized interface, CNBC reported. Fortune's coverage frames it as Anthropic's first real move into physical AI, a category it had mostly ceded to robotics-focused labs until now.
The pitch is deliberately mundane: Anthropic compares MHS to USB-C, a connector standard that removes the need for custom software builds every time a lab or factory wants an AI agent to run a new instrument. Today, wiring an AI model into a piece of scientific or manufacturing hardware typically requires weeks or months of bespoke engineering by specialists who understand both the model's tool-calling behavior and the device's control interface. Anthropic claims MHS collapses that to hours.
The standard is model-agnostic by design -- it isn't locked to Claude, which sets it apart from the more common pattern of AI labs building hardware integrations as a moat for their own models. That's a notable strategic choice: Anthropic is betting that owning the protocol layer that connects agents to physical equipment is more valuable long-term than exclusivity to its own model family, a wager comparable to how Anthropic's own Model Context Protocol became a de facto industry standard for software tool-calling rather than a Claude-only feature.
The competitive field in physical AI includes Physical Intelligence, Figure AI, and Google DeepMind's robotics group, all of which have focused on humanoid or embodied robotics rather than lab-instrument integration specifically. Anthropic's angle -- scientific and manufacturing equipment rather than humanoid robots -- targets a narrower but more immediately monetizable use case: pharma companies running automated drug-discovery experiments, semiconductor labs doing laser calibration on quantum hardware, and manufacturers wiring AI into existing production lines rather than replacing workers with robots outright.
MHS ships initially to a small group of partner organizations in science, robotics and manufacturing as a research preview, with Anthropic saying it plans to eventually open-source the full standard so any device maker can adopt it without Anthropic's involvement. That's the same playbook MCP followed -- narrow release, ecosystem adoption, then open governance -- and if MHS follows a similar adoption curve, it positions Anthropic as infrastructure across an entirely new category of AI deployment beyond chat and coding assistants.
The unresolved question is safety and liability. Handing an AI agent direct control over lab equipment and manufacturing hardware raises stakes that chat-based tool use doesn't: a miscalibrated laser or a mishandled robotic arm has physical consequences a bad code suggestion doesn't. Anthropic hasn't detailed what safeguards or human-approval gates sit inside MHS by default, and that gap will matter more the further the standard spreads beyond controlled research-preview partners into general manufacturing use.
For VCs, MHS matters less as a Claude feature and more as a potential land grab on the interface layer between AI models and the physical world -- the kind of infrastructure position that, if it sticks, outlives any single model generation.