XtalPi’s July 29 launch of XtalPi Science is a bid to turn AI for Science from a collection of model demos and laboratory point tools into a metered, end-to-end research platform—one that can plan work, call specialized models, operate automated experiments and learn from the results. As reported by 36Kr, the company is pairing the platform with “Science Token,” a usage model intended to price access to models, data, workflow agents and robotic laboratory capacity as a unified service.
The timing matters because AI’s next commercial proving ground may not be another coding assistant. Claude Code and similar tools benefited from an unusually favorable environment: code can be compiled, tested and deployed, producing relatively fast and objective feedback. Scientific research has a tougher verifier—the physical experiment—but the prize is much larger. The European Commission’s 2025 Industrial R&D Investment Scoreboard puts 2024 spending by the world’s 2,000 largest corporate R&D investors at €1.446 trillion.
For Windows professionals, developers and enterprise architects, the relevance is not limited to drug discovery. XtalPi Science represents a broader shift toward physical AI workflows: systems in which an AI agent does not merely summarize papers or generate a proposed molecular structure, but orchestrates software, data services, instruments and evidence trails around a measurable research task.
The central idea behind the platform is the DMTA loop: Design, Make, Test and Analyze. In software, compilation and automated testing can identify many failures before humans review the code. In science, a model’s proposed molecule, synthesis path or material composition must eventually survive a real-world test.
That distinction is why AI for Science has moved more slowly than AI coding despite years of impressive demonstrations in protein prediction, molecular generation and simulation. A plausible answer from a generative model is not a validated answer. The missing piece has been a closed feedback system that records what was attempted, what failed, what anomalous conditions occurred, and what the next experiment should be.
According to 36Kr, XtalPi Science is designed around that feedback chain. Its Genius Agents are intended to decompose research goals, invoke scientific models and software, coordinate experimental resources, consolidate results and retain project knowledge for subsequent work. The company says its platform spans drug molecules, proteins, chemical reactions, materials formulas, solar cells and industrial experiments.
That is an ambitious claim, and it is more useful to view it as an operating-model proposition than as evidence that one agent can master every scientific discipline. The common denominator across those fields is not the underlying science; it is the workflow: define an objective, search available knowledge, generate candidates, execute a test, interpret the outcome and repeat.
This is also where the comparison to coding is strongest. The future competitor to Claude Code may not be a chatbot with a science-themed interface. It may be an agentic system that owns the entire evidence loop, including the expensive handoff from digital recommendation to physical validation.
The distinction is important. Specialized models can increasingly be bought, licensed or called through an API. Compute can be rented. A lab can adopt another vendor’s data-management package. What is difficult to replicate is the connective tissue: validated procedures, instrument integrations, failure records, permissions, data lineage, lab scheduling and the accumulated operational knowledge of how an organization actually runs research.
36Kr describes XtalPi Science as an attempt to package those pieces into reusable infrastructure. The platform reportedly connects large language models, vertical scientific models, scientific software and large-scale robotic laboratories rather than treating each as a separate product. That approach could help researchers avoid the familiar enterprise problem of moving between incompatible databases, spreadsheets, instrument consoles, local scripts and bespoke consultancy engagements.
NVIDIA is advancing a related, though differently positioned, strategy. The company announced its BioNeMo Agent Toolkit on June 23, describing a collection of agent-ready skills across biology, chemistry, genomics and drug discovery. NVIDIA says laboratory automation and instrumentation companies including Thermo Fisher, Tecan, Automata and HighRes are connecting to those workflows. Anthropic’s Claude Science, launched in late June, is also integrating NVIDIA’s BioNeMo capabilities, according to NVIDIA and reporting by STAT and TechCrunch.
The industry direction is clear: AI vendors are trying to make research systems more capable by giving agents structured tools rather than asking a general-purpose language model to improvise its way through science. But there is a meaningful difference between providing a toolkit and operating a tightly integrated experimental loop. XtalPi’s commercial thesis depends on proving that integration is a durable advantage.
Public research literature is overwhelmingly optimized around success: a promising compound, a viable process, a result worth publishing. In industrial research, however, the most useful internal records can be negative results—routes that did not scale, conditions that produced poor yield, material formulations that degraded, and instrument readings that exposed an assumption as wrong.
Those records are difficult to gather, standardize and share. They may also be commercially sensitive. Yet they are essential for reducing the tendency of AI systems to make polished but impractical scientific suggestions.
XtalPi told 36Kr that its autonomous laboratory generates more than 50,000 reaction-yield records and 300,000 process records a month, with more than 500,000 experimental records accumulated. The company also cited internal evaluations of its SureRXN system, including a 4.6% chemical-hallucination rate and claimed first-route synthesis accuracy of 51.7%. Those figures are company-reported rather than independently audited, so they should be treated as indicators of the company’s direction, not definitive industry benchmarks.
Still, the theory is sound. If a platform can capture a complete chain—from goal and model prediction through route selection, experiment execution, anomalies and final results—it can produce a data flywheel that a standalone model vendor cannot easily duplicate. The platform gets better not simply because it trains on more documents, but because it observes more real decisions and their consequences.
For enterprise IT teams, that introduces familiar governance questions. Research data needs provenance, access control, retention policies, audit logging and a clear separation between customer-owned project data and provider-improving telemetry. A scientific agent that touches proprietary molecular structures, unreleased material formulas or lab-control systems cannot be governed like a public chatbot.
This is effectively a cloud-computing model for experimental research. Instead of separately procuring model licenses, simulation tools, consulting work, lab time and equipment access, a customer would purchase an outcome-oriented unit of capacity. A research task might consume tokens across multiple resources: a foundation model for planning, a chemistry model for prediction, a database query, a simulation run, a robotic synthesis operation and an analytical test.
That abstraction could simplify procurement, but it also creates the classic metering challenge. Cloud consumption is relatively easy to describe in compute hours, storage and network traffic. Science is less neat. A failed experiment may still be valuable; a single sample can consume different levels of scarce instrument time; and the cost of a result can depend on regulatory, safety and quality requirements rather than raw compute.
The practical test will be whether Science Token makes research more predictable rather than merely more opaque. Customers will need clear visibility into what a task invoked, why it selected a resource, what failed, what data was retained and what they are paying for. Without that transparency, a token system risks becoming a premium wrapper around custom R&D services.
XtalPi reportedly offered 100 million Science Tokens for trials to alliance members and selected university and research partners, while nearly 200 organizations had applied to test the platform as of July 29. Those early figures show interest, but not product-market fit. The harder evidence will be recurring paid usage, external laboratory throughput, repeatable outcomes across customers and the degree to which third-party models and equipment providers actually build on the platform.
On a research workstation, that will mean agents with access to scientific applications, local or private-cloud data, instrument interfaces and high-performance compute. In a regulated enterprise, it will mean identity integration, delegated permissions, approval gates and immutable activity records. The trusted system will not be the one that writes the most persuasive paragraph about a molecule; it will be the one that can show how a recommendation was produced and how it was physically tested.
XtalPi has now put that proposition into the market. Its next milestone is not another platform announcement, but proof that Science Token can turn robotic-lab capacity and scientific agents into a repeatable service customers are willing to buy.
For Windows professionals, developers and enterprise architects, the relevance is not limited to drug discovery. XtalPi Science represents a broader shift toward physical AI workflows: systems in which an AI agent does not merely summarize papers or generate a proposed molecular structure, but orchestrates software, data services, instruments and evidence trails around a measurable research task.
The “Scientific Compiler” Is the Laboratory
The central idea behind the platform is the DMTA loop: Design, Make, Test and Analyze. In software, compilation and automated testing can identify many failures before humans review the code. In science, a model’s proposed molecule, synthesis path or material composition must eventually survive a real-world test.That distinction is why AI for Science has moved more slowly than AI coding despite years of impressive demonstrations in protein prediction, molecular generation and simulation. A plausible answer from a generative model is not a validated answer. The missing piece has been a closed feedback system that records what was attempted, what failed, what anomalous conditions occurred, and what the next experiment should be.
According to 36Kr, XtalPi Science is designed around that feedback chain. Its Genius Agents are intended to decompose research goals, invoke scientific models and software, coordinate experimental resources, consolidate results and retain project knowledge for subsequent work. The company says its platform spans drug molecules, proteins, chemical reactions, materials formulas, solar cells and industrial experiments.
That is an ambitious claim, and it is more useful to view it as an operating-model proposition than as evidence that one agent can master every scientific discipline. The common denominator across those fields is not the underlying science; it is the workflow: define an objective, search available knowledge, generate candidates, execute a test, interpret the outcome and repeat.
This is also where the comparison to coding is strongest. The future competitor to Claude Code may not be a chatbot with a science-themed interface. It may be an agentic system that owns the entire evidence loop, including the expensive handoff from digital recommendation to physical validation.
XtalPi Is Selling Orchestration, Not Just a Better Model
Shanghai Securities News reported before the event that XtalPi would unveil XtalPi Science and its Genius Agents on July 29, alongside an open AI-for-Science ecosystem initiative. XtalPi’s own technology materials describe Genius Agents as AI-agent and automation technology aimed at connecting experimental workflows across research scenarios.The distinction is important. Specialized models can increasingly be bought, licensed or called through an API. Compute can be rented. A lab can adopt another vendor’s data-management package. What is difficult to replicate is the connective tissue: validated procedures, instrument integrations, failure records, permissions, data lineage, lab scheduling and the accumulated operational knowledge of how an organization actually runs research.
36Kr describes XtalPi Science as an attempt to package those pieces into reusable infrastructure. The platform reportedly connects large language models, vertical scientific models, scientific software and large-scale robotic laboratories rather than treating each as a separate product. That approach could help researchers avoid the familiar enterprise problem of moving between incompatible databases, spreadsheets, instrument consoles, local scripts and bespoke consultancy engagements.
NVIDIA is advancing a related, though differently positioned, strategy. The company announced its BioNeMo Agent Toolkit on June 23, describing a collection of agent-ready skills across biology, chemistry, genomics and drug discovery. NVIDIA says laboratory automation and instrumentation companies including Thermo Fisher, Tecan, Automata and HighRes are connecting to those workflows. Anthropic’s Claude Science, launched in late June, is also integrating NVIDIA’s BioNeMo capabilities, according to NVIDIA and reporting by STAT and TechCrunch.
The industry direction is clear: AI vendors are trying to make research systems more capable by giving agents structured tools rather than asking a general-purpose language model to improvise its way through science. But there is a meaningful difference between providing a toolkit and operating a tightly integrated experimental loop. XtalPi’s commercial thesis depends on proving that integration is a durable advantage.
Failure Data May Be More Valuable Than Another Benchmark
One of the more consequential claims in 36Kr’s reporting is not about an AI model’s score. It is about the data that a robotic lab can generate when experiments fail.Public research literature is overwhelmingly optimized around success: a promising compound, a viable process, a result worth publishing. In industrial research, however, the most useful internal records can be negative results—routes that did not scale, conditions that produced poor yield, material formulations that degraded, and instrument readings that exposed an assumption as wrong.
Those records are difficult to gather, standardize and share. They may also be commercially sensitive. Yet they are essential for reducing the tendency of AI systems to make polished but impractical scientific suggestions.
XtalPi told 36Kr that its autonomous laboratory generates more than 50,000 reaction-yield records and 300,000 process records a month, with more than 500,000 experimental records accumulated. The company also cited internal evaluations of its SureRXN system, including a 4.6% chemical-hallucination rate and claimed first-route synthesis accuracy of 51.7%. Those figures are company-reported rather than independently audited, so they should be treated as indicators of the company’s direction, not definitive industry benchmarks.
Still, the theory is sound. If a platform can capture a complete chain—from goal and model prediction through route selection, experiment execution, anomalies and final results—it can produce a data flywheel that a standalone model vendor cannot easily duplicate. The platform gets better not simply because it trains on more documents, but because it observes more real decisions and their consequences.
For enterprise IT teams, that introduces familiar governance questions. Research data needs provenance, access control, retention policies, audit logging and a clear separation between customer-owned project data and provider-improving telemetry. A scientific agent that touches proprietary molecular structures, unreleased material formulas or lab-control systems cannot be governed like a public chatbot.
Science Token Is an Attempt to Make Research Capacity Consumable
The most commercially revealing piece of the launch is Science Token. 36Kr characterizes it as a mechanism to schedule and measure access to AI models, proprietary data, professional tools, agents and automated laboratory resources.This is effectively a cloud-computing model for experimental research. Instead of separately procuring model licenses, simulation tools, consulting work, lab time and equipment access, a customer would purchase an outcome-oriented unit of capacity. A research task might consume tokens across multiple resources: a foundation model for planning, a chemistry model for prediction, a database query, a simulation run, a robotic synthesis operation and an analytical test.
That abstraction could simplify procurement, but it also creates the classic metering challenge. Cloud consumption is relatively easy to describe in compute hours, storage and network traffic. Science is less neat. A failed experiment may still be valuable; a single sample can consume different levels of scarce instrument time; and the cost of a result can depend on regulatory, safety and quality requirements rather than raw compute.
The practical test will be whether Science Token makes research more predictable rather than merely more opaque. Customers will need clear visibility into what a task invoked, why it selected a resource, what failed, what data was retained and what they are paying for. Without that transparency, a token system risks becoming a premium wrapper around custom R&D services.
XtalPi reportedly offered 100 million Science Tokens for trials to alliance members and selected university and research partners, while nearly 200 organizations had applied to test the platform as of July 29. Those early figures show interest, but not product-market fit. The harder evidence will be recurring paid usage, external laboratory throughput, repeatable outcomes across customers and the degree to which third-party models and equipment providers actually build on the platform.
The Windows and Enterprise Angle Is Workflow Control
XtalPi Science is not a Windows product announcement, but it belongs in the same enterprise conversation as Copilot, Claude Code, Azure AI Foundry and laboratory information-management systems. The competition is increasingly about who controls the workflow layer where AI actions become production work.On a research workstation, that will mean agents with access to scientific applications, local or private-cloud data, instrument interfaces and high-performance compute. In a regulated enterprise, it will mean identity integration, delegated permissions, approval gates and immutable activity records. The trusted system will not be the one that writes the most persuasive paragraph about a molecule; it will be the one that can show how a recommendation was produced and how it was physically tested.
XtalPi has now put that proposition into the market. Its next milestone is not another platform announcement, but proof that Science Token can turn robotic-lab capacity and scientific agents into a repeatable service customers are willing to buy.
References
- Primary source: 36 Kr
Published: Fri, 31 Jul 2026 23:52:13 GMT
Loading…
eu.36kr.com