Anthropic, makers of Claude AI, opened a research preview Thursday of what it calls the Model Hardware Standard, a shared specification that lets AI agents discover, understand and operate physical laboratory and manufacturing equipment.
The first group of participants is scientific research labs and advanced manufacturers, and the early results the company published are specific enough to be worth a Puerto Rico manufacturing executive’s attention.
The reason has less to do with the technology itself than with arithmetic. Puerto Rico industrial customers paid an average 23.65 cents per kilowatt-hour in 2024, according to the U.S. Energy Information Administration, roughly triple the U.S. industrial average. One analysis of island manufacturing operations puts energy at 40% to 60% of total operating costs, against 15% to 25% for a mainland plant.
Any technology that reduces energy consumption or equipment downtime is therefore worth two to three times more, proportionally, to a plant in Barceloneta than to the same plant in North Carolina. Based on an exclusive Caribbean Business analysis, that is the frame through which the island’s manufacturers, and Invest Puerto Rico’s reshoring pitch, should read this announcement.
What the standard actually does
The way Anthropic explains it, laboratory and production instruments generally do not talk to each other. Each has its own programming interface, its own data format, its own vendor software. Connecting them requires specialists writing bespoke integration code, work Anthropic says typically takes weeks or months per facility.
MHS introduces a standardized driver using simple commands, read and write, that any device can act on, plus a way for devices to describe themselves in plain language, including physical characteristics that never appear in code, such as the weight of a robotic arm.
The result is a reference file telling an AI agent what a machine can measure, what can be adjusted and what safety limits apply. The standard is model-agnostic and works through the now standard Model Context Protocol, also created by the company for wide industry use, meaning it is not restricted to Anthropic’s own models.
The published results are concrete. Carnegie Mellon researchers built drivers from scratch for a liquid handler, a plate reader, a robotic arm and monitoring cameras spread across three computers with incompatible interfaces, then ran a full dose-response protocol autonomously. Total time from unautomated equipment to a completed result, including one autonomous rerun the system decided on its own: eight hours, against the multiple weeks a vendor-built integration typically requires.
At quantum computing firm QuEra, an agent working overnight rewrote a laser recovery procedure that a four-person team had spent months building, another example cited by Anthropic. The original script worked 58% of the time and took about 150 seconds. The agent’s version, tested blind across 700 trials, recovered correctly 695 times, a 99.3% success rate, in roughly six seconds.
Separately, the agent retuned 12 interdependent control parameters that a specialist had already set, reducing residual error roughly tenfold and holding the laser locked for 19 hours without a single failure, against about 1.6 failures per hour under the expert tune.
The Genentech case
For Puerto Rico, the most directly transferable example is Genentech’s. Researchers there used MHS to automate the BCA protein assay, a standard procedure requiring coordination across a liquid handler, a robotic arm and a microplate reader. Claude optimized liquid-transfer flow rates independently for water and for viscous protein solutions, arriving at parameters the company’s automation experts confirmed were reasonable.
That is drug discovery rather than commercial manufacturing, and the distinction matters. But the underlying problem, orchestrating incompatible instruments in a regulated pharmaceutical environment, is recognizable to anyone running a plant in Juncos, Carolina, Barceloneta or Humacao.
Pharmaceuticals accounted for $42.3 billion of Puerto Rico’s exports in 2025, 66.6% of the island’s total, and the island produces eight of the world’s 15 top-selling drugs.
Turning disadvantage into opportunity
The broader industry research points in a consistent direction. McKinsey and World Economic Forum Lighthouse Network data cited across recent industrial AI analyses put productivity gains from scaled AI deployment at 20% to 30%, downtime reductions at up to 50%, and energy cost reclamation at around 25%.
Siemens has reported a 69% productivity improvement at its Erlangen facility. McKinsey has documented reductions above 20% in inventory and logistics costs from autonomous routing.
Apply the energy figure to Puerto Rico’s cost structure and the result is significant. A 25% cut in energy consumption at a mainland plant where energy is 20% of operating costs reduces total costs by roughly 5%. The same 25% cut at an island plant where energy is 50% of operating costs reduces total costs by roughly 12.5%. The island’s most cited competitive weakness is, in this narrow sense, the line item where AI optimization pays the largest proportional dividend.
That has direct bearing on the island’s reshoring pitch. The persistent counterargument has always been energy and operating expenses. Technology that narrows that gap strengthens the pitch at precisely its weakest point.
The capital is already moving. Amgen committed $650 million to its Juncos biologics facility with roughly 750 jobs attached, then added a further $300 million. Eli Lilly pledged more than $1.2 billion to expand and modernize Lilly del Caribe in Carolina. Those are the expansions that keep the island in the conversation, and they are precisely the kind of highly instrumented, capital-intensive operations where instrument orchestration savings compound.
Beyond pharmaceuticals
Medical devices, Puerto Rico’s second major manufacturing cluster, run on the same regulated, heavily instrumented production model as pharmaceuticals. The aerospace and precision manufacturing operations that have grown around Aguadilla depend on calibration and quality-assurance workflows of the sort Doosan Robotics is testing MHS against.
Food and beverage processing, electronics assembly and contract manufacturers with mixed equipment fleets face the same integration problem at smaller scale.
The vendor list matters here as much as the technology. Amazon Web Services is supporting MHS through its Strands Robots library. Tecan is adding support for its Fluent liquid handling platforms, Universal Robots for its robotics platform, QIAGEN for its nucleic acid purification systems, and Danaher is exploring applications across its instrument portfolio.
Those are companies whose equipment already sits on Puerto Rico factory floors. Support arriving through the vendor rather than through custom integration work is what would make this practical for a mid-sized island manufacturer rather than only for a multinational with a dedicated automation team.
The gap between interest and readiness
To be sure, industry information points to two potential obstacles. First is adoption readiness. A 2026 Redwood Software survey found 98% of manufacturers exploring AI but only 20% fully prepared to deploy it. McKinsey’s latest global survey similarly reports 88% of organizations using AI in at least one function while most have yet to scale it.
Gartner has warned that more than 40% of agentic AI projects will be canceled by 2027, attributing most failures to inadequate foundations rather than to the technology itself. Data infrastructure, not model capability, is the binding constraint. Will Puerto Rico’s plants buck that trend? We will surely follow up this story with interviews to find out.
Second is regulatory. Puerto Rico’s flagship manufacturing sectors operate under FDA current good manufacturing practice requirements, where changes to validated production processes carry documentation and revalidation burdens that a research laboratory does not face.
The eight-hour integration Carnegie Mellon achieved is a research result. Realizing anything close to it inside a GMP-regulated commercial plant is a materially harder problem, and one neither Anthropic nor its partners have yet demonstrated. Anthropic says it will open-source MHS after the research preview and is inviting participants through a waitlist. For island manufacturers, the sensible posture is neither to dismiss this as distant nor to treat it as imminent, but to understand that the standard is being built now, that the equipment vendors already serving Puerto Rico plants are building support into their products, and that the economics of adoption favor Puerto Rico more than most places precisely because operating here costs more. The rest is up to local executives to capitalize on.
