DP Technology Series C

DP Technology: From the Gordon Bell Prize to National-Level Unicorn Status and the Door of a Hong Kong Listing, the AI-for-Science Leader's Technical Delivery and Commercial Blank

Beijing's AI-for-Science leader built on the Deep Potential machine-learned interatomic potential, the Bohrium science cloud and the SciMaster agent; a further state-backed round of several hundred million renminbi closed in October 2026 and a national-level unicorn rating followed, with a Hong Kong listing in preparation.

Executive Summary

DP Technology, formally Beijing DP Technology Co., Ltd., was founded in 2018 and is one of the most closely watched companies in Chinese AI for Science. It was founded by Zhang Linfeng, chief scientist and holder of a Princeton doctorate in applied mathematics who studied under Academician E Weinan, and Sun Weijie, CEO and an alumnus of Peking University's Yuanpei College, with E Weinan serving as director and chief scientific advisor. Its technical starting point is the Deep Potential molecular simulation method, which as a core contribution won the 2020 ACM Gordon Bell Prize and was named one of China's top ten science and technology advances that year.

On December 24, 2025 the company announced a Series C of more than 800 million renminbi from investors including Fortune Capital, Jingguorui Fund, the Beijing Artificial Intelligence Industry Investment Fund, the Beijing Healthcare Industry Investment Fund, Legend Capital and Oriza Hua. A further state-backed round of several hundred million renminbi was announced on October 12, 2026, from the Beijing Artificial Intelligence Industry Investment Fund and the Zhongguancun Science City company, with proceeds for talent attraction, for deepening full-stack capability from original algorithm innovation through industrial software and industry solutions, and for industrialization and commercialization in biopharmaceuticals, new energy, chemicals and semiconductors. Public reporting puts cumulative funding in the billions of renminbi. Unicorn status is now a recorded certification rather than an inference, following a 2026 national-level unicorn rating and a 2025 national-level potential unicorn rating, and a business data service records 2025 revenue of 168 million renminbi against 206 employees. In February 2026 media reported completion of the corporate form conversion and progress toward a Hong Kong listing that would make it the first Chinese AI for Science listing.

Operating scale is substantial. Bohrium Science Navigation serves more than 1,000 universities and institutions and more than three million scientists worldwide, with nearly a hundred Project 985 and 211 universities onboarded as institutions. It has supported over a thousand research projects, answers about 12 million scientific questions annually and has saved more than two billion minutes of scientists' time. Its scientific intelligence products and solutions serve more than 150 advanced R and D enterprises, including over 70 life science companies and more than 100 pipelines such as Fosun Pharma, Sinopharm, Hansoh Pharma, Huadong Medicine, HEC, Qilu Pharmaceutical, Unilever and Yunnan Baiyao, with materials science customers including the Suzhou National Laboratory, CNPC, Sinosteel, CATL, BYD and GAC Aion.

The product matrix pivots on the Bohrium research station in a cycle of read, compute, experiment and intelligence: Bohrium Science Navigation for literature and data, Bohrium Lebeg computing for compute, Hermite for drug design and Piloteye for battery design as microscale R and D software, Bohrium Cyber Lab for automated experiments and SciMaster as the scientific agent. On models there are the DPA atomic model, Uni-Mol molecular model and Uni-Fold protein model, with the Innovator scientific foundation model added in 2025. In August 2026 the company added the public beta of the Bohrium Science Space, a desktop research work environment, and launched the Bohrium 107 Plan calling for scientific questions worldwide.

In physical AI terms DP Technology occupies a distinctive position: microscale physical AI, using machine learning to solve electronic, atomic and molecular interactions in place of or to accelerate first-principles calculation and molecular dynamics. This differs from macroscopic CFD and structural simulation in that the physical core is quantum and statistical mechanics and the output is material properties, drug binding affinity and ionic conductivity. The value is compressing calculations that once took weeks on supercomputers into hours on GPUs.

One framing point matters. This is a company with top-tier technical reputation, unproven commercialization and disputed governance. Technically the Gordon Bell Prize and the DPA and Uni-Mol families are real assets, and the reach of three million scientists and 150 enterprises is real, yet paid conversion, average revenue per account, retention and the split between software and services are undisclosed, with the exception of a registry-caliber 2025 revenue figure of 168 million renminbi, and a 2026 report of a shareholding nominee dispute involving the CEO adds uncertainty to the Hong Kong listing. The key questions are who pays, how much and whether they renew.

Company Overview

From a Gordon Bell Prize to a Company

The technical lineage is academic. In 2018 E Weinan organized the first international AI for Science conference at Peking University, and that year Zhang Linfeng and Sun Weijie founded the company. In 2020 the Deep Potential work, with Zhang as a core contributor, won the Gordon Bell Prize, the highest honor in high-performance computing, for raising molecular dynamics accuracy to first-principles level while cutting cost by orders of magnitude.

The commercial meaning is that molecular simulation had long been trapped between accuracy and cost: density functional theory is accurate but handles only hundreds of atoms, while classical force fields are fast but inaccurate. Deep Potential fits the potential energy surface with a neural network, preserving near-DFT accuracy while scaling to millions of atoms and nanosecond timescales, opening industrial calculation for drugs, materials and batteries.

Team and Governance

Zhang leads technology, Sun leads strategy and commercialization, and E Weinan serves as director and chief scientific advisor. The division is common in AI for Science companies, with a scientist setting the technical route, a commercially minded partner handling deployment and an academician providing endorsement and network.

Scale and qualification data arrive from two directions. A business data service gives 2025 revenue of 168 million renminbi, 206 employees, registered capital of 34.60406 million renminbi and an incorporation date of November 29, 2018, which places revenue per employee at roughly 815,000 renminbi. Recognition records include a 2026 national-level unicorn rating dated July 17, 2026, a 2025 national-level potential unicorn rating dated December 13, 2025, a Beijing enterprise technology center recognition dated September 24, 2026, and first place on the Beijing private enterprise future industry pioneer ranking published by the Beijing Federation of Industry and Commerce on September 17, 2026.

Governance requires recording a dispute. In 2026 media reported a shareholding nominee arrangement involving CEO Sun Weijie, with parties, amounts and litigation status not fully disclosed. Nominee shareholding is a focus in listing reviews for ownership clarity and potential disputes, and unresolved issues could affect the Hong Kong process. This report flags it as a risk without substituting for legal judgment.

Product Matrix and the Four-Part Cycle

The capability system has four parts. Read is Bohrium Science Navigation for literature search, data extraction and frontier tracking. Compute is Bohrium Lebeg and domain software such as Hermite and Piloteye. Experiment is Bohrium Cyber Lab connecting computation with automated experiments. Intelligence is SciMaster orchestrating all of it with large models.

The closed loop is designed to cover the research workflow from finding a direction through computing properties and screening, running experiments and automating the process. Its commercial value is stickiness, since a customer who reads also computes and experiments, but it also means every link must be usable or the loop becomes a patchwork.

Technology

How Deep Potential Works

Understanding the technology requires understanding the basic difficulty of molecular simulation, which is computing interatomic forces that originate in electronic structure. Solving electronic structure accurately means solving the Schrodinger equation through DFT, with cost growing as the cube or worse of atom count and reaching limits at hundreds of atoms. Engineering therefore uses force fields that parameterize interactions analytically, fast but dependent on parameter quality and unable to describe bond breaking and formation.

Deep Potential learns the mapping from atomic configuration to potential energy with a neural network trained on DFT data, subject to two physical constraints: invariance to translation, rotation and permutation, so the same structure in a different place or order has the same energy, and locality, so each atom's energy depends only on its neighborhood. These constraints allow training on small systems and extrapolation to large ones.

Success depends on data efficiency and generalization. A general potential requires training configurations covering element combinations, crystal structures, defects, surfaces, liquids and high temperature and pressure, and generating that data consumes substantial DFT compute. The DPA models iterate toward generality and transferability so that one model covers many elements and structures, reducing retraining for each new system.

Uni-Mol and Molecular Representation

Uni-Mol is a three-dimensional molecular representation model using a Transformer over atomic coordinates and types to learn structure-to-property mappings, applied to property prediction such as solubility, toxicity and binding affinity, and to conformer generation.

Its relation to Deep Potential is one of scale: Deep Potential computes energies and forces as physical quantities while Uni-Mol computes molecular properties as chemical quantities. Combined they form a chain from energy to property to molecule screening. Competition is sharper for Uni-Mol, since molecular property prediction is the most crowded AI drug discovery track with physics-based approaches from Schrodinger, generative approaches from Isomorphic and many startups.

SciMaster and Agent Positioning

SciMaster is a scientific agent orchestrating research workflows by interpreting goals, retrieving literature, invoking computational tools, analyzing results and proposing next steps. It follows the same paradigm as industrial engineering agents, with a language model orchestrating and professional tools computing, but in research rather than production.

The difficulty is reliability. Research demands reproducibility, verifiability and traceability while language model outputs are stochastic, and conclusions not traceable to a specific calculation or reference will not be adopted. The company's response ties the agent to Bohrium compute so that every conclusion rests on a reproducible calculation, which is a sound design.

Cyber Lab and the Compute-Experiment Loop

Bohrium Cyber Lab runs automated experiments with robots performing synthesis, characterization and testing and feeding results back to models. Its value addresses the biggest bottleneck in AI for Science, which is data, since computation can generate abundant data but final validation must come from experiments, which in turn correct the models.

The engineering difficulty is high, covering hardware integration for liquid handling and analytical instruments, programmatic experiment protocols, exception handling and data management. International leaders include Emerald Cloud Lab and Strateos, with domestic work at XtalPi and university platforms. DP Technology differentiates by integrating computation and experiment rather than selling lab automation alone.

Technical Shortfalls

First, accuracy boundaries of general potentials. DPA models are accurate within training coverage and may fail at extreme conditions such as high temperature, pressure or strong fields, or with new element combinations, precisely where industrial customers often work.

Second, data dependence. Accuracy ceilings are set by the DFT level of training data, and high-level functionals are expensive, limiting data scale.

Third, scale of the experiment loop. Cyber Lab is currently points rather than a network, and how many customers and pipelines it supports is undisclosed.

Fourth, missing commercial metrics. Paid customer counts, revenue per account, retention and the software versus compute split are not public. A business data service now records 2025 revenue of 168 million renminbi and 206 employees from registry-filed annual report data, which the company itself has not published in that form, so it is treated as a registry caliber rather than a company caliber.

Product and Commercial Model

Product Layers

Four layers exist. The platform is Bohrium for compute scheduling, software integration and data management. Tools are vertical software including Hermite for drug design and Piloteye for battery design. Models include DPA, Uni-Mol, Uni-Fold and Innovator. Services include SciMaster, Cyber Lab and enterprise R and D services.

The layering supports a funnel from free to paid: Science Navigation free or cheap to attract three million scientists, compute and software paid on demand to convert enterprises and universities, and R and D services billed by project at high value.

Bohrium Science Space and the Bohrium 107 Plan

In August 2026 the company put the Bohrium Science Space into public beta, a desktop research work environment for researchers and AI scientists that integrates the whole research loop into one workbench. It ships subject agents for biology, drug discovery and materials, including BioMaster, PharmMaster and MatMaster, and connects more than 200 million literature and patent records with more than 50,000 scientific computing tools. The disclosed platform figures include more than 150 categories and 1,800 instrument models reachable through Uni-Lab-OS, a parsing accuracy above 98 percent on multimodal scientific information through Uni-Parser, and a 98.8 percent creation success rate over roughly 1.85 million started tasks in the SciX agent framework. Public beta dates are given as August 21, 2026 by technology media and August 30, 2026 by an encyclopedia entry, and both readings are recorded.

Alongside the beta the company launched the Bohrium 107 Plan, a global call for scientific questions with research value, offering research support through the Bohrium Science Space for the selected problems. The stated positioning is that repetitive work goes to the artificial intelligence and judgement and invention stay with the scientist, which is the clearest public statement yet of the research-workflow agent direction.

User Scale and Paid Conversion

Disclosed scale is large at more than 1,000 universities and institutions, three million scientists, nearly a hundred Project 985 and 211 universities onboarded, 12 million questions answered annually and two billion minutes saved. These show breadth of use rather than depth of payment.

Research has a distinctive pattern: many individual users, weak paying capacity and long decision chains. University software purchasing runs through grants or centralized procurement with long cycles and limited amounts, while enterprise customers pay more but demand quantifiable returns. A university-heavy user base means large scale and difficult conversion.

Enterprise Customers and Efficiency Data

Enterprise efficiency figures include a hundredfold gain in literature review, a 76 percent reduction in wet lab demand and cost from AI computation, more than tripling of instrument utilization and throughput from smart labs, and about a 70 percent reduction in repetitive work for scientists.

If valid, these figures carry high commercial value, since R and D return on investment is what pharmaceutical and materials companies care most about. Statistical baselines, sample sizes and customer verification are undisclosed, making them company figures.

One named pharmaceutical collaboration was signed on September 16, 2026 with AnkeBio, covering the design and optimisation of antibody drug conjugate toxin payload molecules, where the company contributes molecular simulation, free energy calculation, quantitative structure-activity relationship analysis and ADMET prediction for candidate generation, virtual screening, key physicochemical property prediction and transporter-mediated resistance risk assessment, inside a compute-design, experiment-validate, data-reflow loop. AnkeBio had previously worked with the Precision Intelligent Chemistry key laboratory of the University of Science and Technology of China. This is a technical cooperation agreement without a disclosed contract value, so it is recorded as an ecosystem and pipeline signal rather than as a revenue disclosure.

Revenue Mix and Business Model

The model is likely a mix of subscriptions by seat or usage, compute consumption by GPU hour and project services. Such mixes are common in research cloud, but margin structures differ: compute is low margin near cost-plus, software is high margin and services are intermediate.

If revenue is compute-dominant the company resembles a research cloud provider rather than an AI software company, with different valuation logic. This is undisclosed but critical.

Competitive Landscape

International Comparison

International benchmarks in molecular simulation and computational chemistry include Schrodinger with FEP+ free energy, quantum chemistry and molecular dynamics at 256 million dollars of 2025 revenue, Dassault BIOVIA, and AI drug discovery companies such as Isomorphic Labs, Relay Therapeutics and Recursion.

The difference from Schrodinger is technical route. Schrodinger centers on physics with quantum chemistry, force fields and free energy perturbation and three decades of error-control engineering, giving interpretability and clear boundaries. DP Technology centers on machine learning potentials, fast and scalable but with extrapolation dependent on training coverage.

Domestic Comparison

Domestic competitors include XtalPi with quantum physics, AI and lab automation already listed in Hong Kong, BioMap in biological computation, and university teams. XtalPi is closest in positioning, also building compute plus experiment loops for pharmaceutical customers.

The difference is emphasis. XtalPi is strong in crystal form prediction and lab automation while DP Technology is strong in the physical core of molecular simulation and in research platforms. XtalPi is listed with public financials while DP Technology is in the listing process.

Research Cloud Competition

Bohrium competes with research computing from Alibaba Cloud, Huawei Cloud, AWS and Azure. Cloud vendors have compute resources and pricing advantages but do not understand research workflows, software needs or data formats. DP Technology differentiates on workflow plus domain software, which is its moat.

Relationship With Open Source

The underlying DeePMD-kit is open source, a major reason for its academic influence. Open source builds standards and ecosystems quickly, but commercialization then depends on managed services, enterprise editions and domain software, since customers can self-deploy and pay only for compute, support and templates.

Financing and Capital Structure

Funding History

Cumulative funding is in the billions of renminbi. The Series C of more than 800 million was announced December 24, 2025 from Fortune Capital, Jingguorui Fund, two Beijing industry funds, Legend Capital and Oriza Hua. A further state-backed round was announced in October 2026 and is described in the next subsection. The structure combines local industry funds, market venture capital and industrial capital, with the Beijing funds providing policy and resources, Fortune and Legend providing market endorsement and Oriza Hua providing semiconductor and hard technology perspective.

October 2026 Round and Unicorn Recognition

On October 12, 2026 the company announced a new round of several hundred million renminbi, invested by the Beijing Artificial Intelligence Industry Investment Fund and the Zhongguancun Science City company. The stated use of proceeds is attracting top talent, deepening full-stack capability from original algorithm innovation through industrial software and industry solutions, and pushing the technology toward industrialization and commercialization in biopharmaceuticals, new energy, chemicals and semiconductor scenarios. The round carries no disclosed stage label and no amount beyond the range, and a business data profile still records the latest financing date as December 24, 2025 at stage C, so this report does not label the round as a later stage.

Unicorn status is now a recorded certification rather than an inference. The company was rated a national-level unicorn enterprise in 2026 and had been rated a national-level potential unicorn in 2025, both by Great Wall Strategic Consulting, with a separate Beijing enterprise technology center recognition dated September 24, 2026. The valuation readings behind that label do not agree: an industry ranking record published in January 2026 gives a 2025 valuation of 1.231 billion US dollars, the database profile gives a C-round valuation of 2.6 billion renminbi, and the unicorn framing itself implies a figure above one billion US dollars. This report keeps the readings parallel rather than reconciling them.

Corporate Conversion and Hong Kong Listing

In February 2026 media reported completion of the corporate form conversion and progress toward a Hong Kong listing. Hong Kong values AI for Science companies differently from mainland markets, emphasizing revenue growth and international customers with higher tolerance for losses, whether under Chapter 18C or the main board.

The listing hinges on three things: revenue scale and growth, customer concentration and governance including resolution of the nominee shareholding dispute.

Valuation and Pace

Valuation is not formally disclosed by the company. The unicorn rating is now a recorded certification, but the readings behind it disagree: an industry ranking record gives 1.231 billion US dollars for 2025, a database profile gives a C-round valuation of 2.6 billion renminbi, and the unicorn framing itself implies a figure above one billion US dollars. This report keeps the three readings parallel rather than picking one, and notes that the valuation reflects the AI for Science platform narrative rather than current revenue, which a business data service records as 168 million renminbi in 2025.

DISCREPANCIES

1. Framing of cumulative funding. The Series C announcement cites more than 800 million renminbi while media cite cumulative funding in the billions. The two differ as single round versus cumulative, and billions is approximate. This report separates the two. A further round of several hundred million renminbi announced in October 2026 is recorded in item 9. Source tier: OFFICIAL (Series C announcement) and MEDIA (cumulative framing).

2. Three user scale figures. The company cites more than 1,000 universities and institutions, more than three million scientists and nearly a hundred Project 985 and 211 universities onboarded, without defining the counting method, whether scientists include students and inactive registrations, or what institutional onboarding means in procurement terms. This report retains the figures and flags undefined definitions. Source tier: OFFICIAL (company statements).

3. Missing methodology for efficiency figures. The hundredfold literature gain, 76 percent wet lab cost reduction, tripled instrument throughput and 70 percent reduction in repetitive work are point estimates without baseline, sample size, named customers or verification. This report retains them as company figures. Source tier: company statements, UNVERIFIED.

4. Description of E Weinan's role. Some sources call him a founder while the registered founders are Zhang Linfeng and Sun Weijie, with E Weinan as director and chief scientific advisor. This report uses director, chief scientific advisor and key driver of the founding. Source tier: OFFICIAL (registry) over MEDIA summary.

5. Disclosure on commercial quality. The company discloses scale of 1,000 universities, three million scientists and 150 enterprises. A business data service now records 2025 revenue of 168 million renminbi and 206 employees from registry-filed annual report data, which the company itself has not published in this form, so it is treated as a registry caliber rather than a company caliber. Paid customer counts, revenue per account, retention and the software versus compute split remain without a public source, and media have raised commercialization concerns. This report lists the remaining gaps as key and does not infer commercial success from scale. Source tier: OFFICIAL (scale data) with registry-caliber revenue data added and quality data still absent.

6. Completeness of the nominee shareholding dispute. A 2026 media report describes a nominee shareholding dispute involving CEO Sun Weijie, without full disclosure of parties, amounts, litigation status or company response. This report lists it as a governance risk without factual determination. Source tier: MEDIA (single source), to be confirmed.

7. The first Chinese AI for Science listing claim. Media describe a successful Hong Kong listing as making it the first Chinese AI for Science listing, which depends on how the category is defined given that XtalPi is already listed. This report does not use the label. Source tier: MEDIA, UNVERIFIED.

8. Relationship between open source and commercialization. DeePMD-kit is open source while commercial products are Bohrium, Hermite and Piloteye, and some interpretations equate open source influence with commercial market share. This report separates ecosystem influence from commercial revenue. Source tier: OFFICIAL (open source fact) distinguished from inference.

9. Framing of the October 2026 round. The company and an industry alliance bulletin describe a new round of several hundred million renminbi completed on October 12, 2026, from the Beijing Artificial Intelligence Industry Investment Fund and the Zhongguancun Science City company, without a stage label or an exact amount. A business data profile instead still records the latest financing as December 24, 2025 at stage C with a valuation of 2.6 billion renminbi. Adjudicated: both readings are stated, the announcement is used for the fact of a new state-backed round and the database record is used for the undisclosed stage and amount. This report does not label the round as a later stage.

10. Valuation readings. The unicorn framing implies a private valuation above one billion US dollars, an industry ranking record published in January 2026 gives 1.231 billion US dollars for 2025, and a database profile gives a C-round valuation of 2.6 billion renminbi, which is well below four hundred million US dollars. Adjudicated: all three are stated with their source and none is presented as the company figure. The valuation is not formally disclosed by the company, and the spread between the readings is itself a finding.

11. Date of the Bohrium Science Space public beta. Technology media report August 21, 2026, while an encyclopedia entry gives August 30, 2026. Adjudicated: the month is used and the two days are recorded in parallel. No dated company announcement was located to settle it, and this report does not convert either date into a firm fact.

12. Unicorn status as fact rather than inference. The earlier version of this report derived the unicorn description from the funding scale and treated it as a media framing. Registry-tier records now show a 2026 national-level unicorn rating dated July 17, 2026 and a 2025 national-level potential unicorn rating dated December 13, 2025, both issued by Great Wall Strategic Consulting, alongside third-party list entries from IT Orange and Hurun Research Institute. Adjudicated: the status is stated as a recorded certification with its rating body and date, while the underlying valuation remains in dispute as recorded in item 10.

13. Founding year, 2018 against 2019. The registry incorporation date of November 29, 2018 and the company's own account both give 2018, while some media covering the October 2026 financing describe the company as founded in 2019. Adjudicated: 2018, on the strength of the registry record and the company's own statement, with the media reading noted as a caliber difference that does not change the finding.

Risks and Diligence Gaps

First, unclear commercial quality. Three million scientists and 1,000 universities reflect breadth of use rather than depth of payment, research customers have weak paying capacity and long decision chains, and compute-heavy revenue would limit margins and scale effects.

Second, governance risk. An unresolved nominee shareholding dispute could affect Hong Kong listing review, where ownership clarity is a hard criterion.

Third, technical boundaries. Accuracy of general potentials at extreme conditions and with new element combinations determines entry into industrial design and certification, and public validation data are limited.

Fourth, intensifying competition. Internationally Schrodinger's physics and Isomorphic's generative approaches compete, domestically XtalPi is already listed, and cloud vendors compete on research compute.

Fifth, capital consumption in the experiment loop. Cyber Lab is asset heavy in hardware, facilities and staff, and slower scaling would pressure cash flow.

Sixth, the listing window. Hong Kong valuations for AI companies are sentiment sensitive, and a closed window would affect financing pace.

Seventh, data compliance. Research data involves intellectual property and cross-border flow, especially with international customers, making compliance a hidden cost.

Composite Assessment

Technology scores 8.5. Deep Potential and the Gordon Bell Prize are international-grade technical assets, DPA, Uni-Mol and Uni-Fold form a complete microscale model system, and the four-part cycle covers the research workflow. Deductions for accuracy at extremes and unverified experiment loop scale.

Business scores 6.0. Three million scientists, 1,000 universities and 150 enterprises represent real market reach, and the efficiency figures would be valuable if valid. Deductions for undisclosed paid conversion and revenue mix, weak paying capacity among research customers and margin limits if the model is compute-dominant.

Finance scores 7.5. Cumulative funding in the billions, an 800 million Series C, a further state-backed round of several hundred million renminbi in October 2026, national-level unicorn recognition and an investor base of local industry funds plus market venture capital provide ample capital. Deductions for undisclosed profitability, an asset-heavy experiment loop, a listing process affected by governance, and the wide spread between the valuation readings behind the unicorn label.

Product scores 7.5. Bohrium, Hermite, Piloteye, SciMaster and Cyber Lab form a complete matrix from tools to closed loop, and the open-source DeePMD-kit established standard influence. Deductions for unverified depth of product synergy and insufficient disclosure of enterprise features such as permissions, audit and compliance.

Market scores 8.0. AI for Science draws both policy and capital attention, drug discovery and new materials demand are rigid, and the company has strong academic brand. Deductions for early commercial validation, deep international competition and limited growth in research budgets.

Overall 7.5. This is the hardest technology and least verified commercialization in Chinese AI for Science: Deep Potential resolved the accuracy-cost dilemma in molecular simulation, Bohrium productized the capability and Cyber Lab added the experiment loop. Its weaknesses are opaque commercial quality and governance risk. The judgment on physical AI is that its physical core in quantum and statistical mechanics plus machine learning potentials is real and internationally leading and that microscale physical AI is a second main line beyond macroscopic simulation, while moving from academic influence to industrial revenue requires crossing paid conversion, validation systems and governance, which the Hong Kong listing will test in one pass.

References

  1. DP Technology Series C announcement, December 24, 2025: more than 800 million renminbi from Fortune Capital, Jingguorui Fund, the Beijing Artificial Intelligence Industry Investment Fund, the Beijing Healthcare Industry Investment Fund, Legend Capital and Oriza Hua; proceeds for talent, iterating the scientific discovery engine, consolidating full-stack capability and accelerating market expansion
  2. Product and scale data from the same announcement: a four-part capability system of read, compute, experiment and intelligence pivoting on the Bohrium research station; products including Bohrium Science Navigation, Bohrium Lebeg computing, Hermite, Piloteye, Bohrium Cyber Lab and SciMaster; Science Navigation serving more than 1,000 universities and institutions and more than three million scientists with nearly a hundred Project 985 and 211 universities onboarded; more than a thousand research projects supported, 12 million scientific questions answered annually and over two billion minutes saved
  3. Enterprise customers and efficiency data from the same announcement: more than 150 advanced R and D enterprises including over 70 life science companies and more than 100 pipelines such as Fosun Pharma, Sinopharm, Hansoh Pharma, Huadong Medicine, HEC, Qilu Pharmaceutical, Unilever and Yunnan Baiyao; materials customers including the Suzhou National Laboratory, CNPC, Sinosteel, CATL, BYD and GAC Aion; over 50 high-value scientific assets created; a hundredfold gain in literature review, 76 percent lower wet lab demand and cost, tripled instrument utilization and throughput and about 70 percent less repetitive work
  4. Company history and team from public sources and media: founded 2018; Zhang Linfeng as founder and chief scientist with a Princeton applied mathematics doctorate under E Weinan and Sun Weijie as founder and CEO from Peking University's Yuanpei College; E Weinan as director and chief scientific advisor who drove the founders to return and start the company in 2018; the Deep Potential work winning the 2020 ACM Gordon Bell Prize and being named a top ten Chinese science and technology advance
  5. Model system from public sources: DPA atomic model, Uni-Mol molecular model and Uni-Fold protein model forming the Shenshi Yuzhi scientific model system; the Innovator scientific foundation model and SciMaster scientific agent added in 2025; work in 2026 combining intelligent research with experiment automation
  6. Media report, 2026: cumulative funding in the billions of renminbi; corporate conversion and Hong Kong listing news in 2026; a reported commercialization challenge noting the absence of public data on paid customer share, software versus project revenue and retention; and a reported nominee shareholding dispute involving CEO Sun Weijie without full detail
  7. Listing progress from media, February 2026: corporate form conversion completed and a Hong Kong listing being pursued, described by media as potentially the first Chinese AI for Science listing
  8. Technical background from public sources: the accuracy-cost dilemma in molecular simulation, with DFT accurate but limited to hundreds of atoms and classical force fields fast but inaccurate; Deep Potential fitting the potential energy surface with neural networks trained on DFT data subject to invariance and locality constraints, enabling training on small systems and extrapolation to large ones; the DPA line iterating toward generality and transferability
  9. Technical background from public sources: Uni-Mol using a Transformer over atomic coordinates and types to learn structure-to-property mappings for property prediction and conformer generation; its relation to Deep Potential being one of scale between energy and force versus molecular property
  10. Competitive landscape from public sources: international benchmarks including Schrodinger at 256 million dollars of 2025 revenue centered on FEP+ free energy, quantum chemistry and molecular dynamics, Dassault BIOVIA, Isomorphic Labs, Relay Therapeutics and Recursion; domestic peers including XtalPi listed in Hong Kong with quantum physics, AI and lab automation, and BioMap; research cloud competition from Alibaba Cloud, Huawei Cloud, AWS and Azure
  11. Industry background from public sources: international leaders in automated laboratories include Emerald Cloud Lab and Strateos, with engineering difficulty in hardware integration, programmatic protocols, exception handling and data management; the compute plus experiment loop addresses the data bottleneck in AI for Science
  12. Hong Kong listing background from public sources: Hong Kong valuations for AI software companies emphasize revenue growth and international customers with higher tolerance for losses; listing review focuses on ownership clarity, and nominee shareholding is a key verification item
  1. Zhongguancun Financial Technology Industry Development Alliance weekly bulletin, 2026-10-14 issue, together with the Niutoushe report of the same date, for the October 12, 2026 announcement of a new round of several hundred million renminbi from the Beijing Artificial Intelligence Industry Investment Fund and the Zhongguancun Science City company and for the stated use of proceeds
  2. Qichacha business profile for Beijing DP Technology Co., Ltd., updated October 2026, for the 2026 national-level unicorn and 2025 national-level potential unicorn ratings, the January 2026 industry ranking valuation of 1.231 billion US dollars, the registry caliber figures of 168 million renminbi revenue and 206 employees for 2025, registered capital of 34.60406 million renminbi, the incorporation date of November 29, 2018, the C-round valuation of 2.6 billion renminbi, the latest recorded financing date of December 24, 2025, the Beijing enterprise technology center recognition and the public beta of the Bohrium Science Space with the Bohrium 107 Plan and the AnkeBio cooperation
  3. Baidu Baike entries for the Bohrium Science Space and for SciMaster, together with technology media reporting of the August 2026 public beta, for the desktop work environment, the BioMaster, PharmMaster and MatMaster agents, more than 200 million literature and patent records, more than 50,000 scientific computing tools, more than 150 categories and 1,800 instrument models, the parsing accuracy above 98 percent, the 98.8 percent task creation success rate, and the August 21 against August 30 date difference
  4. AnkeBio official announcement and the Shanghai Securities News report of September 16, 2026, for the strategic cooperation agreement on antibody drug conjugate payload molecule design and optimisation and for the technical scope of the company's contribution
This report is compiled from public sources for informational purposes only and is not investment advice. Data is current as of 2 November 2026. The company is private and funding and user data come from company announcements, registry records and media; efficiency figures are company-reported without third-party verification; the nominee shareholding dispute is media-reported with details unconfirmed and does not constitute a legal judgment.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.