In the Biological World, a Foothold for Physical AI

"Seven Years at Enhe: How an Industrial Biotech Company Grew AI Capabilities from Scratch" By Zhiyan Chen

GPT can explain antibiotic resistance mechanisms with flawless logic. But ask it to actually make a better antibiotic, and it has no hands, no lab, can't even hold a petri dish.

The world's top AI companies are already valued in the trillions of dollars, yet nearly all that value exists inside screens. Manufacturing, agriculture, chemicals, materials — the industries that keep the physical economy running — AI still stands outside their doors. It's not that the models aren't smart enough; it's that the physical world hasn't prepared a place for AI to land.

In biology labs, this gap shows up in mundane ways: PhDs still can't escape coming back at midnight to shake flasks. AI can read all 40 million biology papers ever published, reason about protein folding, design metabolic pathways — but it can't walk into a lab, pick up a pipette, and turn a plan into reality. Between cognitive ability and execution ability lies an entire physical world.

A company called Enhe Biotech spent seven years trying to chart a path across it.

Part 01

Build the Digital-Physical Entity First, Then Talk About Intelligence

The first thing Enhe did looked completely unrelated to AI — building a biofoundry.

From the outside, Enhe is a biomanufacturing company centered on industrial strain development. Its founder and CEO, Cheryl Cui, an MIT PhD in medical engineering, wanted to do something deceptively simple: make biological experimentation no longer dependent on individual experience, but into scalable engineering.

In the world of biomanufacturing, the keyword of recent years has been "throughput" — the scale and speed of tasks processed per unit time. If traditional lab R&D is like a handicraft workshop, high-throughput is a modern fully automated assembly line.

Starting in 2019, Cheryl chose the hardest path: instead of building integrated, purpose-built pipelines for specific strains, she built a modular, general-purpose platform. The system had to be flexible enough to catch demands from different industries and different clients.

Those years, a persistent tension ran through Enhe. Programmers leaned toward the internet rhythm of "run a round first, iterate if it doesn't work"; biologists were accustomed to the rigorous discipline of "read extensively, leave nothing to chance." The most typical headache was using automated equipment to synthesize DNA — initially, the success rate was below 10%. It was an excruciatingly dull磨合 period: programmers didn't understand biological complexity, scientists were impatient with code's rigid logic. This tension felt like internal friction at some moments, and at others like some strange kind of fuel. They ground away at it for two years, pushing DNA synthesis success rates above 90%.

Over three to four years, Enhe built out the complete biofoundry, replacing hands with machines, turning experiments into assembly lines: in plasmid construction, a skilled researcher could manually build several dozen plasmids in two weeks; Enhe's automated system pushed that number above 1,000. Sample detection compressed from dozens of minutes to 6-8 seconds per sample. Every fermentation run automatically collected and fed back up to 210,000 data points, with every step highly traceable and reproducible.

These numbers themselves aren't earth-shattering. What truly matters: when Enhe finished building this foundry, it had inadvertently accomplished something far larger — constructing the runtime environment for AI to enter biomanufacturing.

AI Coding Agents exploded not because the models were the smartest, but because the software world had long prepared compilers and execution environments. The foundry Enhe built over four years is, in essence, the "compilation environment" for biomanufacturing.

Part 02

Semiconductor Plants Are Called Foundries. So Are Biofoundries.

With the foundry built, Enhe's next step seemed even riskier at the time: stepping out of the lab and taking over real factories.

In 2020, Alex — now Enhe's Director of AI & Computation — joined this startup of barely a dozen people, a "crossover" from the semiconductor industry. He had previously led global smart manufacturing systems at Micron, witnessing how thousands of process steps in chip fabrication achieved full automation.

During his interview, he handed Cheryl a proposal for a "3-to-5-year architecture." Alex says that document was written mostly on gut feeling — data automation, model closed loops, autonomous driving. Six years later, looking back, he never expected the team to actually walk there together.

"Semiconductor plants are called foundries. The core of biomanufacturing is also called biofoundry." Alex discovered the two are highly isomorphic in logic: both combine components to produce products with specific functions. Only biological systems are harder than semiconductors — because they're alive.

Cheryl decided to enter the industrial validation phase. In her view, if technology stayed in the lab, it held no real value. Enhe made a bold decision: revitalized a fermentation enterprise in Shandong; built a large-scale production base in Ordos. This gave Enhe end-to-end capability from lab to factory — the data flywheel spinning from microliter well plates all the way to hundred-ton fermentation tanks.

This step's significance deserves particular emphasis in the context of industrial AI. Most AI-for-Science companies' data stops at the lab phase; between lab and factory lies a massive black box. How a strain performs in a 0.5-liter fermentation tank and how it behaves at 30-ton production scale can be completely different. Breaking through this black box doesn't rely on algorithms — it requires actually building factories, running production lines, making deliveries.

With end-to-end delivery capability, Enhe partners with clients through joint R&D and material supply. To date, Enhe has earned recognition from top-tier strategic partners both internationally and domestically, including NHU, Yili, Proya, Pechoin, BASF, and Syensqo, with Estée Lauder also deeply engaged in its project collaboration pipeline. This cooperation isn't simple buying and selling, but co-research and co-creation from source demand in food, nutrition, and personal care.

As of now, Enhe has cumulatively delivered 23 projects, holds 15 commercial product pipelines, with over 44 total projects, establishing a global network spanning Hangzhou, Yucheng, and Ordos in China, as well as the United States and the Netherlands overseas.

"AI without real scenarios is just a tool," Alex told An Yong Waves. It was only when the digital foundry, physical factories, and cross-boundary team were all in place that Enhe's AI capabilities finally had soil to grow in.

Part 03

Where Does Data Come From?

For Chinese companies today, nearly every CEO faces the same question: "How do we grow AI capabilities?" Most answers point to the same starting point: data.

But where does data come from?

Internet AI companies' approach is to burn money — hire labeling teams, buy datasets, generate synthetic data. But for industrial companies rooted in the physical world, another path exists.

"Data is a moat, but the more critical thing is the virtuous cycle of 'accumulating data without burning money.'" This is Cheryl's deepest insight from these years. Enhe's logic: collect data imperceptibly while creating commercial value (delivering projects).

Cheryl draws an analogy to Tesla — when Tesla drives on the road, it's collecting autonomous driving data; when Enhe's biofoundry develops strains for clients, it's accumulating the most authentic closed-loop experimental data. Enhe now possesses over ten million end-to-end closed-loop data entries, all structured "AI Ready" assets.

These data are valuable not just for their quantity, but for their structure — for the same strain, Enhe has its lab data on microplates, its performance in 2-liter pilot fermenters, its changes at 1,000-liter scale-up, its stability at 30-ton production scale. This end-to-end continuity from lab to factory is what most lab-only companies lack. General-purpose large models trained on public literature can read the final published results; what they can't read are the failed cases in process, parameter sensitivities, hidden patterns in scale-up — precisely the most critical knowledge in biomanufacturing.

Data is the foundation. What truly shapes AI capability is organizational adaptation. At Enhe, you'll see extreme cross-boundary scenarios: computation teams reassigned to digitize finance, programmers required to complete MIT's synthetic biology open course upon joining, biologists learning to read code logic.

"We want everyone to recognize each other's value," Cheryl says. "If you only want people like yourself, this won't work."

With the foundry running for four years and data accumulating to the tens of millions, Enhe's AI capabilities finally reached the stage where they could be defined and named.

In March 2026, Enhe officially launched SAION AI — the world's first Physical AI platform for biomanufacturing. What it does can be understood through the VLA architecture (Vision-Language-Action) from autonomous driving: AI first comprehends the biological world, then translates that understanding into standardized instructions, and finally executes in closed loops in the physical lab.

The most critical layer among the three, and also the scarcest in the industry, is the middle orchestration layer. For this, Enhe self-developed BPL (Biology Protocol Language), a standardized protocol language for biological experiments, functioning similarly to EDA in the semiconductor industry: translating scientists' experimental intentions into instructions machines can execute with precision. In internal testing, BPL-compiled experimental protocols achieved 99.4% structural consistency, versus only 43% for natural language descriptions of similar protocols. With this layer, the upstream cognitive model (connected to 300+ scientific research tools and 40 million papers, surpassing GPT and Claude on multiple benchmarks) and downstream automated foundry (24/7 operation, processing tens of thousands of operations daily) can truly form a closed loop — experimental results feeding back in real time to drive the next iteration.

With SAION AI's intervention, traditional "trial-and-error" biological R&D is replaced by "intelligent orchestration." Efficiency changes can be quantified: annual lab throughput grew from hundreds or thousands of strains to over a million; project R&D cycles compressed from 6-8 years to 1-3 years.

Take one concrete example: zeaxanthin. This molecule widely used for eye health was previously highly dependent on natural plant extraction, vulnerable to climate and supply chain fluctuations. After receiving market demand, SAION AI not only used large model reasoning to find optimal enzyme mutation sites, boosting conversion efficiency by over 60%, but also shortened each generation of strain iteration cycles to within one quarter. The product has completed US Self-GRAS evaluation, achieving full industry chain penetration.

"The core of Physical AI lies in effective closed-loop data collection," Cheryl says. It is no longer merely prediction on screens, but can truly enter the physical world, optimize production processes, and redefine efficiency boundaries. When cognition, orchestration, and execution are all connected, AI is no longer just a tool, but a system capable of self-evolution.

Part 04

Merely "Day 0"

Despite zeaxanthin's success validating the path's feasibility, for Enhe this remains merely "Day 0." SAION AI is currently at L2 stage — AI executes tasks, humans make final decisions. The next target is L3: AI leads experimental design, humans only set direction. The farther vision of L4 — end-to-end autonomous biological research — still requires continuous breakthroughs across multiple dimensions.

Currently, AI mostly handles task orchestration within known frameworks, while true "AI scientists" need to evolve to identifying and proposing correct questions in unknown territories. Physical world perception and interaction remain difficult — giving AI hardware limbs that can handle physical deviations, possess tactile sensing, and are sufficiently flexible is still a hardcore challenge. How to extend AI's decision-making reach from labs to every valve and parameter in large-scale industrial production, achieving real-time intelligent scheduling across geographically distributed factories, is also a focus of next-stage breakthroughs.

But if we pull back to a more macroscopic view, what Enhe's practice points to is a far larger question.

Today's world sits in a delicate instability. Energy restructuring, geopolitical fracturing, repeated supply chain ruptures. All point to the same underlying reality: human civilization remains highly dependent on petroleum. Plastics, fibers, chemicals, fuels — the materials constituting modern life's infrastructure, almost without exception.

Biomanufacturing offers another possibility. Microorganisms can use sugar, starch, even industrial byproducts as feedstock to produce functionally equivalent alternatives. But for a long time, this path was blocked by one practical problem: too slow, too expensive, too uncertain. From concept to commercialization, a strain routinely took 6-8 years and tens of millions of dollars.

What Physical AI is doing is cutting the most critical time cost on this path. The global bioeconomy track holds potential scale exceeding $6 trillion, but the entry ticket to this market has never been smarter algorithms — it's the physical infrastructure capable of carrying algorithms: standardized experimental environments, end-to-end data closed loops, complete execution capability from lab to factory.

Enhe's seven-year journey offers one template: industrial company AI capabilities grow out of practice in the physical world. Build the environment first, let intelligence grow after — this sequence may be the correct way to open industrial AI.

"I hope biology becomes an engineering discipline that can scale." Cheryl has said this for many years. But what she wants to say now goes beyond scaling. When AI truly intervenes in every link of biomanufacturing, biology could become an intelligently generalizable platform — not just making a better strain, but building a general capability to understand and engineer life.

This isn't finished. But it's no longer just an idea.

Layout: Zhixin Han | Image source: Company provided

Recommended Reading

The flow of money, the rise and fall of people