Schrödinger NASDAQ SDGR Public

Schrödinger Deep Dive: Physics-Based FEP+ Simulation and Machine Learning Power Molecular Discovery, with the 2026 Bunsen Agentic Co-Scientist

Schrödinger (NASDAQ: SDGR) powers molecular discovery with physics-based FEP+ simulation and ML; 2026 brings Bunsen, an agentic co-scientist, with NVIDIA and Google Cloud compute.

Executive Summary

Schrödinger (Nasdaq: SDGR) is a computational chemistry company founded in New York in 1990 by Richard A. Friesner and William A. Goddard III and listed on February 10, 2020 through an offering of 11,882,352 shares at 17 dollars. It runs two businesses: software, spanning the Maestro modeling interface, the LiveDesign collaboration platform, FEP+ free energy calculations, Glide docking, Desmond molecular dynamics and Jaguar quantum chemistry; and drug discovery, where it uses its own platform to advance partnered and proprietary programs.

The 2025 financials need to be read line by line. Total revenue was 255.9 million dollars, up 23.3 percent; software revenue was 199.5 million dollars, up 10.6 percent; drug discovery revenue was 56.4 million dollars against 27.2 million; software gross margin was 74 percent; operating expenses were 309.5 million dollars, down 9.3 percent; other income was 64.6 million dollars; net loss was 103.3 million dollars against 187.1 million; and cash, equivalents and marketable securities were 402.3 million dollars at year end. Software annual contract value was 198 million dollars and adjusted EBITDA was negative 114.9 million dollars. The company targets positive adjusted EBITDA by the end of 2028.

The second quarter of 2026 revealed an important structural shift. Annual contract value was 29.6 million dollars, up 27 percent, and 208 million dollars on a trailing four-quarter basis, while software revenue was 32.5 million dollars, down 10 percent. The divergence comes from an accelerated transition from up-front license recognition to hosted revenue recognized over the subscription term, with hosted already 47 percent of software revenue and 30 percent on a trailing basis. The transition depresses short-term revenue and raises deferred revenue while making revenue more predictable. It also pulled software gross margin from 74 percent in 2025 to 71 percent, since hosted delivery carries cloud infrastructure cost.

Drug discovery revenue in the same quarter was 23.0 million dollars against 13.9 million a year earlier, driven mainly by a 10 million dollar collaboration milestone triggered by Eli Lilly's acquisition of Ajax Therapeutics. Contribution revenue was 3.4 million against 4.8 million, total revenue was 58.9 million dollars up 8 percent, operating expenses were 74.0 million down 6 percent, other income was 48.9 million dollars mainly from the Ajax equity gain, and net income was 6.0 million dollars against a 43.2 million dollar loss a year earlier. Cash and marketable securities were 418.8 million dollars.

In physical AI terms the position needs precision. The company does not train general foundation models and does not follow a purely data-driven generate-a-molecule route. Its core asset is physics-based computation of molecular properties through quantum chemistry, molecular mechanics and free energy perturbation, which produces ground-truth data used to train and calibrate machine learning models. CEO Ramy Farid framed it in the second quarter 2026 release as biopharma facing mounting pressure to accelerate discovery and an unprecedented need for ground-truth data, with the company combining accurate physics-based simulation and AI and launching the agentic co-scientist Bunsen to build the computational infrastructure for drug discovery.

Bunsen is the most important product move of 2026. It is an agentic co-scientist that, unlike general-purpose agents, is optimized to orchestrate Schrödinger's validated physics platform, plan and execute complex molecular discovery workflows and interpret results. The company simultaneously announced a strategic software agreement with Bristol Myers Squibb to deploy Bunsen, a global drug discovery and development collaboration with Simcere Pharmaceutical Group, and support from NVIDIA and Google Cloud providing a full-stack AI platform, infrastructure and early customer access to the NVIDIA BioNeMo Agent Toolkit.

One framing point matters. This is a company with an extremely hard physical core, no profitability yet, and an AI narrative that has just begun. Its physics has been validated over three decades and FEP+ accuracy is supported by peer-reviewed work, while Bunsen is in early access with no quantified customer results, and the drug discovery line swings on milestones and equity gains against a 103.3 million dollar 2025 net loss. Assessment must track three lines at once: software annual contract value growth, margin erosion from the hosted transition, and milestone realization in the pipeline.

Company Overview

One Thread Across Thirty-Five Years

Schrödinger's history can be summarized in one sentence: making molecular-level physical computation engineering-usable. Founded in 1990, it sold its first software to Brookhaven National Laboratory in 1992. Several milestones defined the technical path: Glide for high-throughput virtual screening in 2001, Prime for protein structure prediction in 2003, and WaterMap in 2009, which the company describes as the first rigorous method for computing protein-bound water locations and energies.

Also in 2009 the company co-founded Nimbus Therapeutics with Atlas Ventures, starting a model of incubating companies around its own platform. That model has since been repeated, including Structure Therapeutics, in which it holds equity and whose oral GLP-1 receptor agonist aleniglipron reported positive phase 2b topline data in December 2025, Ajax Therapeutics, acquired by Eli Lilly in 2026 and triggering milestones and equity gains, and Tectora Therapeutics announced in September 2026 in immunology and inflammation.

The economics of that model deserve separate treatment. The company does not only sell software; it trades platform access for equity and milestones in incubated companies. As of 2025 it disclosed more than 650 million dollars of cash generated from drug discovery since 2020 and eligibility for nearly five billion dollars in potential future milestones. These are option value rather than current revenue, and they explain how a company losing about 100 million dollars a year maintains more than 400 million dollars of cash.

Team and Scale

The company is headquartered in New York with Ramy Farid as CEO and Michael Lynton as chairman. Headcount is around 1,000, with third-party sources citing either 1,015 employees or about 900 across 12 offices, a difference noted in the discrepancies. Research teams are distributed across New York, Portland in Oregon, San Francisco, London, Munich, Bangalore and Hyderabad.

The composition of the workforce is central to understanding the company. It employs software engineers, computational chemists mostly with doctorates, and a drug discovery organization of biologists, pharmacologists and clinical staff. That combination is rare in software companies and rare in biotechnology companies, and it is the organizational basis for the platform-plus-pipeline model as well as the reason operating expenses reached 309.5 million dollars in 2025.

How the Two Businesses Interlock

Software and drug discovery are not simply two legs. Software supplies revenue and customer feedback that iterate the platform in real projects, while drug discovery uses the platform to advance molecules whose clinical data in turn validates platform credibility. If molecules designed on the platform perform well in the clinic, the software argument strengthens.

Financially this produces two revenue shapes. Software is predictable subscription revenue at 198 million dollars of annual contract value, while drug discovery is volatile milestones and equity gains, at 56.4 million dollars in 2025 and 23.0 million in a single quarter of 2026. Investors typically value the company as software annual contract value growth plus the net present value of pipeline options rather than consolidated net income.

Technology

Three Layers of Physical Method

The technology stack divides by accuracy and cost into three layers with different methods and uses.

The bottom layer is quantum chemistry. Jaguar performs density functional theory and ab initio calculations to obtain electronic structure, reaction barriers, spectroscopic properties and reference data for force field fitting. Accuracy is highest, within a few kilocalories per mole in relative terms, and cost is highest, limited to tens or hundreds of atoms.

The middle layer is molecular mechanics and dynamics. Proprietary OPLS-family force fields describe bonded and non-bonded interactions, Desmond runs large-scale molecular dynamics, and together they handle solvation, conformational change and protein flexibility. Approximate potentials replace electronic structure calculations at orders of magnitude lower cost, enabling systems of 100,000 to millions of atoms over microsecond timescales.

The top layer is free energy calculation. FEP+ is the flagship method, computing the difference in binding free energy between two molecules and a target through thermodynamic integration or free energy perturbation, thereby predicting relative binding affinity. Its engineering value converts the question of which molecule to synthesize into a computable ranking problem: if FEP+ ranks 100 candidates correctly, only the top 10 need synthesis.

The relationship among the layers is straightforward: quantum chemistry parameterizes and validates force fields, force fields and dynamics provide sampling for free energy calculations, and free energy calculations inform medicinal chemists. What three decades built is not a single algorithm but error control at every link, covering force field parameterization, sampling convergence, protonation state assignment and bound water treatment. Those engineering details determine whether predictions can be trusted and are the hardest part for competitors to replicate.

Accuracy and Limits of FEP+

FEP+ accuracy has published support. In 2026 the company published work in the Journal of Chemical Information and Modeling validating the FEP+ framework for binding prediction. The broader industry consensus is that FEP+ achieves roughly 1 kilocalorie per mole of mean error on relative binding free energies within congeneric series, which is enough to guide lead optimization because a tenfold potency step in medicinal chemistry corresponds to about 1.4 kilocalories per mole.

The limits are equally clear. First, it computes relative rather than absolute binding free energy, and absolute predictions carry much larger error and need experimental anchoring. Second, it depends on high-quality protein structures, and error grows when the target structure is predicted rather than experimentally determined. Third, it cannot handle large scaffold hops, because the perturbation path breaks when binding modes differ. Fourth, covalent inhibitors, allosteric modulators and metalloenzymes require special treatment, limiting generality.

These limits are not secret and the company documents them in training. Understanding them is itself evidence of professionalism: FEP+ is a tool that is trustworthy under specific conditions, and used correctly it saves substantial synthesis and assay cost.

Why Physics Matters More in the AI Era

This is the most important part of the Schrödinger narrative. The dominant drug AI route is data-driven, training models on known compound-activity data to predict properties of new molecules. Its bottleneck is data quality and distribution, because public activity data suffers publication bias toward positive results, assay methods differ across laboratories, and chemical space coverage is sparse.

Physics has the advantage of not depending on historical data. Given a target structure and a molecule, quantum chemistry and free energy calculations compute binding directly without having seen a similar molecule. That allows extrapolation into entirely new chemical space and lets physics generate high-quality labeled data for machine learning.

The company frames this as generating ground-truth data: compute a set of high-confidence data points by physics, train AI models on them, and let those models cover the vast space physics cannot afford. This loop of physics generating data and AI extending coverage is the fundamental difference from purely AI-driven drug discovery companies and the most persuasive positioning under a physical AI framework.

The cost structure of this loop matters. Physical computation cost grows linearly with the number of molecules while AI inference is nearly free, so loop efficiency depends on how many physics data points are needed to train a usable model. If thousands suffice the loop works, and if millions are required the physics cost would break the model. Public information does not disclose that ratio, which is the key gap in assessment.

Bunsen and What the Agent Actually Does

Bunsen is the agentic co-scientist launched in 2026 and in early access. Per the official description it differs from general agents by being optimized to orchestrate Schrödinger's validated physics platform, planning and executing complex molecular discovery workflows and interpreting results; it helps computational chemists run concurrent research sessions across multiple programs; and it lets drug hunters without computational chemistry experience use advanced methods.

The positioning is technically sound. Drug discovery workflows are long, covering target structure preparation, binding site identification, virtual screening, free energy calculation, selectivity assessment, developability prediction and synthetic route planning. Each step requires different tools and parameter choices, and the difference between experts and non-experts lies mainly in knowing which tool to use, how to set parameters and how to interpret results. An agent encodes that knowledge into executable processes and lowers the barrier.

The value has two faces. For experts it is parallelization, since an agent handles repetitive execution and summarization while a chemist watches several programs. For non-experts such as medicinal chemists and biologists it is accessibility, since they can ask in natural language for selective scaffolds that cross the blood-brain barrier without learning Maestro operations.

The boundaries are equally clear. First, the agent does not validate physics results and only orchestrates tools, so wrong inputs or parameters produce wrong outputs. Second, core judgments such as target selection, acceptable selectivity risk and PK/PD trade-offs remain human decisions with the agent supplying evidence. Third, early access has no published customer quantification, so actual time savings cannot be judged.

Partnership With NVIDIA and Google Cloud

Bunsen infrastructure comes from NVIDIA and Google Cloud through a full-stack AI platform, compute and early customer access to the NVIDIA BioNeMo Agent Toolkit. Technically this means agent throughput far exceeds a single machine and requires GPU clusters for both physical computation and model inference, while BioNeMo supplies biological foundation models and agent tooling so Schrödinger need not build everything.

Commercially it is channel. NVIDIA and Google Cloud have strong enterprise reach, and Bunsen as a physics-plus-AI showcase can enter their customer recommendations. For a software company with a couple of hundred million dollars of revenue, that is material help.

Technical Shortfalls

First, the cost ceiling of physical methods. A single FEP+ calculation still takes hours to days on GPUs, and an optimization cycle over a hundred molecules requires substantial compute, limiting adoption among smaller biotechs without clusters where cloud costs are high.

Second, target coverage. Physics requires a known three-dimensional structure and cannot address targets lacking structural information, including many membrane proteins and intrinsically disordered proteins, which are a meaningful share of current discovery.

Third, the depth of AI-physics integration is unclear. The company says it combines simulation and AI but does not specify where AI sits: as a surrogate replacing part of physical computation, as an improvement to sampling efficiency, or as post-processing for property prediction. Technical substance and effect differ greatly across those uses.

Fourth, Bunsen's verifiability. Agent effectiveness requires clinical or project-level evidence that early access cannot provide, and such products risk impressive demos that disappoint in production.

Product and Commercial Model

Software Portfolio

On the life science side the core includes Maestro for modeling, LiveDesign as a web-based collaboration platform connecting computational chemists, medicinal chemists and biologists, FEP+ for free energy, Glide for docking, Desmond for dynamics, Jaguar for quantum chemistry, Prime for structure prediction, WaterMap for bound water analysis, PyMOL for visualization and RetroSynth for AI-driven synthetic route planning launched in January 2026.

Materials science includes the Materials Science Suite covering organic electronics, polymers, consumer packaged goods, catalysis, thin films, energy storage, metals, alloys and ceramics. Customers include chemical, energy and aerospace companies such as Sabic and Boeing.

LiveDesign deserves separate mention commercially. It integrates dispersed computational tools and experimental data into one collaborative environment so that computational chemists, medicinal chemists and biologists see the same data. A collaboration layer of this kind has far higher stickiness than an individual computational tool, because once team workflows run on LiveDesign switching costs become high. The January 2026 collaboration with Lilly TuneLab makes LiveDesign the priority interface for biotechs accessing TuneLab workflows, combining Lilly federated learning models with Schrödinger physics.

Business Model and Key Metrics

Software is measured by annual contract value rather than current revenue because license recognition timing distorts revenue. Software annual contract value was 198 million dollars in 2025, 29.6 million in the second quarter of 2026 up 27 percent, and 208 million on a trailing four-quarter basis, or 22.6 million and 196 million excluding contribution annual contract value. Full year 2026 guidance is 218 to 228 million dollars, growth of 10 to 15 percent.

That guidance needs comparison with realized figures. Trailing four-quarter annual contract value already reached 208 million dollars, so a midpoint of 223 million implies roughly 7 percent of incremental growth. In other words the company expects 2026 growth to decelerate from 27 percent in the second quarter to 10 to 15 percent for the year. Management typically attributes this to renewal timing and large deal phasing, but investors should watch whether it signals softer demand.

Drug discovery revenue has a completely different shape. The 56.4 million dollars in 2025 came mainly from collaboration milestones and equity gains, and 10 million of the 23.0 million in the second quarter of 2026 came from the Ajax milestone. The company raised 2026 drug discovery revenue guidance from 55 to 65 million dollars to 65 to 75 million precisely because of that one-time milestone, and such revenue cannot be extrapolated linearly.

Financial Impact of the Hosted Transition

The company is accelerating the shift to hosted software licensing, with three financial consequences.

The first is revenue recognition. Hosted revenue is recognized over the subscription term while traditional licenses are recognized at signing, so the transition period shows annual contract value growth alongside revenue decline, exactly as in the second quarter of 2026 with annual contract value up 27 percent and software revenue down 10 percent.

The second is gross margin. Hosted delivery carries cloud infrastructure cost, and software gross margin fell from 74 percent in 2025 to 71 percent in the second quarter of 2026, after 81 percent in the fourth quarter of 2025 when one-time high-margin license revenue was recognized. Over time hosted margin should stabilize between 70 and 75 percent, below the 80 percent or more of pure licensing.

The third is cash flow. Hosted contracts usually bill annually in advance, increasing deferred revenue and making operating cash flow better than reported profit, which materially helps a loss-making company.

Pipeline and Partners

As of the 2025 results the drug discovery portfolio included more than 25 first-in-class, best-in-class or first-in-modality programs across development stages, more than 10 in clinical stages and 16 eligible for sales royalties, with 20 partners since 2018 and 16 ongoing royalty-eligible programs.

Proprietary clinical candidates include SGR-1505, a MALT1 inhibitor for relapsed or refractory B cell malignancies, and SGR-3515, a Wee1 and Myt1 inhibitor for solid tumors, both in phase 1, with the company completing clinical packages and exploring strategic partnerships.

Incubated companies are the hidden asset of this line: Nimbus, co-founded with Atlas in 2009, which sold an ACC inhibitor to Gilead in 2016 in a deal worth up to 1.2 billion dollars; Structure Therapeutics, in which the company holds equity and whose oral GLP-1 receptor agonist aleniglipron reported positive phase 2b topline data in December 2025; Ajax, acquired by Eli Lilly in 2026; and Tectora, announced in September 2026. The value of these stakes does not appear in software revenue but flows through other income, which was 64.6 million dollars in 2025 and 48.9 million in the second quarter of 2026, mainly from fair value changes and disposals of these holdings.

This also explains an easily misread phenomenon. Net income of 6.0 million dollars in the second quarter of 2026 looks like a return to profit, but 48.9 million dollars of it came from other income. Excluding that, the operating business remains loss-making, and assessment must separate operating improvement from investment gains.

Competitive Landscape

Direct Rivals in Computational Chemistry Software

In molecular modeling and computational chemistry software the main rivals are Dassault Systèmes BIOVIA with Discovery Studio and Pipeline Pilot, Chemical Computing Group with MOE, Cadence OpenEye with shape-based screening, Cresset with electrostatic methods, and open-source tools including RDKit, Open Babel, GROMACS and AMBER.

Schrödinger differs in two ways. First, method completeness: quantum chemistry through force fields through free energy are all developed in house with controlled error, whereas rivals are typically strong in one link and must integrate third-party tools. Second, the closed loop of physics, data and experiment: its drug discovery business provides real project validation that pure software vendors lack.

The weaknesses are price and openness. Schrödinger licenses are high-end for the industry and hard for smaller biotechs to afford, and its file formats and workflows are relatively closed with less interoperability with open-source ecosystems. Both weaknesses amplify in budget-constrained environments.

Relationship With AI Drug Discovery Companies

Purely AI-driven drug discovery companies such as Recursion, Isomorphic Labs, Insilico Medicine and Relay Therapeutics follow data and algorithm routes, training models on large experimental datasets or designing molecules with generative models. Their narrative is that AI can find molecules faster and cheaper than physical simulation.

Schrödinger's counter is that AI models need high-quality labeled data and physics is the means to generate it, and that purely data-driven models fail outside the training distribution while physics does not. The counter is technically reasonable but not zero-sum commercially, since many AI companies are themselves Schrödinger customers using FEP+ for validation and others build in-house physics.

The realistic relationship is layered coexistence: physics provides accuracy and interpretability, AI provides speed and coverage, and leading organizations use both. Schrödinger's opportunity is to become the de facto standard at the physics layer rather than to replace the AI layer.

Relationship With Chinese Vendors

China's computational chemistry market includes local vendors and in-house tools at contract research organizations, but Schrödinger remains the mainstream high-end choice. Geopolitical factors such as export controls and data security affect scientific software less than semiconductor equipment but are not risk-free. Teams in Bangalore and Hyderabad indicate a strategy of competing with a combination of services and software in cost-sensitive markets.

Financing and Capital Structure

Listing and Share Count

The company listed on February 10, 2020 with 11,882,352 shares at 17 dollars, raising about 202 million dollars. Subsequent financings and equity compensation expanded the share count. As of the second quarter of 2026, cash, equivalents, restricted cash and marketable securities totaled 418.8 million dollars, against 402.3 million at the end of 2025 and 367.5 million at the end of 2024.

Cash stability comes from two sources: subscription cash flow from software at 198 million dollars of annual contract value and 71 to 74 percent gross margin, and drug discovery milestones and equity disposals exceeding 650 million dollars since 2020. That explains why a company losing about 100 million dollars a year has not needed emergency financing.

Loss Structure and the Path to Profitability

The 2025 net loss was 103.3 million dollars against 187.1 million in 2024, and adjusted EBITDA was negative 114.9 million against negative 152.5 million. The narrowing came from a 9.3 percent reduction in operating expenses from 341.2 to 309.5 million dollars and from drug discovery revenue growth.

The company targets positive adjusted EBITDA by the end of 2028. The path depends on three variables: annual contract value growing at a double-digit rate as guided at 10 to 15 percent for 2026, gross margin stabilizing above 70 percent after the hosted transition, and continued expense control. Of the three, annual contract value growth matters most, since at 220 million dollars of annual contract value and 75 percent gross margin, gross profit of about 165 million dollars must cover roughly 300 million dollars of operating expense, requiring drug discovery contribution or further expense reduction.

Valuation Logic

Such companies are typically valued as enterprise value over annual contract value for the software business plus the net present value of pipeline options. Software at 220 million dollars of annual contract value and ten times enterprise value to annual contract value implies about 2.2 billion dollars, with pipeline options, 16 royalty-eligible programs, nearly five billion dollars of potential milestones and incubated equity, added separately. The framework is highly sensitive: growth falling from 15 to 10 percent could compress the multiple from ten to seven times, and any pipeline failure would force a re-rating of option value.

DISCREPANCIES

1. Two versions of headcount. One third-party encyclopedia records 1,015 employees, while another source describes about 900 across 12 offices. The company does not give a single figure in press releases, though the annual report will. This report uses a range around 1,000 and labels both sources. Source tier: MEDIA third-party sources with the OFFICIAL annual report figure to be confirmed.

2. Annual contract value up 27 percent alongside software revenue down 10 percent. In the second quarter of 2026 annual contract value was 29.6 million dollars up 27 percent while software revenue was 32.5 million down 10 percent. The directions differ because the accelerated hosted transition changes recognition timing. Some coverage cites only annual contract value growth or only revenue decline and draws opposite conclusions. This report presents both and explains the cause. Source tier: OFFICIAL (financial release).

3. Nature of the 6.0 million dollar second quarter net income. Net income was 6.0 million dollars against a 43.2 million dollar loss a year earlier, but other income of 48.9 million dollars came mainly from the Ajax equity gain, so the operating business remains loss-making. Some coverage headlined a return to profit. This report separates operating improvement from investment gains. Source tier: OFFICIAL.

4. Attribution of founding. Some sources describe the company as founded by Richard A. Friesner alone, while official and multiple sources agree on Friesner and William A. Goddard III as co-founders. This report uses the co-founder description. Source tier: OFFICIAL over one-sided MEDIA.

5. Relationship between 2026 annual contract value guidance and realized value. Guidance is 218 to 228 million dollars, growth of 10 to 15 percent, while trailing four-quarter annual contract value already reached 208 million and the second quarter grew 27 percent. Implied guidance growth is below the single-quarter rate, which the company attributes to renewal timing and large deal phasing without detailed decomposition. This report presents both without one-sided inference. Source tier: OFFICIAL.

6. Reason for raising drug discovery revenue guidance. Guidance rose from 55 to 65 million dollars to 65 to 75 million because of a 10 million dollar one-time milestone from the Ajax acquisition. The raise therefore does not represent acceleration across the pipeline. This report labels the milestone as one-time. Source tier: OFFICIAL.

7. Status of the end-2028 adjusted EBITDA target. This is a stated financial objective rather than guidance and comes without quantified assumptions on the path. This report labels it a target rather than a forecast. Source tier: OFFICIAL (company objective statement).

8. Three software gross margin figures. Full year 2025 at 74 percent, fourth quarter 2025 at 81 percent and second quarter 2026 at 71 percent correspond to different periods and revenue mixes, with the fourth quarter figure lifted by one-time high-margin license revenue and the second quarter figure reflecting the hosted transition and cloud costs. This report labels the period whenever a figure is cited and does not summarize with a single number. Source tier: OFFICIAL.

Risks and Diligence Gaps

First, long-term margin erosion from the hosted transition. The move from 74 to 71 percent is only the beginning, and as hosted share rises further the margin may fall again. Cloud infrastructure cost is a rigid variable cost, and scale does not necessarily improve margin. If margin settles at 68 to 70 percent, the 2028 EBITDA target becomes harder.

Second, sustainability of annual contract value growth. The gap between 27 percent in the second quarter and 10 to 15 percent guidance for the year is significant. If growth falls below 10 percent in the second half, the narrative of AI-driven demand acceleration weakens. Renewal and large deal data in the third and fourth quarters need watching.

Third, the Bunsen validation gap. Early access provides no quantified customer results, so whether it improves research efficiency cannot be judged. Agent products typically risk strong demos and weak production performance, especially in complex enterprise workflows.

Fourth, non-recurring drug discovery revenue. The 56.4 million dollars in 2025 and 23.0 million in the second quarter of 2026 depended heavily on individual milestone events. Such revenue cannot be forecast linearly or used to support valuation, and proprietary programs SGR-1505 and SGR-3515 remain in phase 1 with at least five years to commercialization.

Fifth, the cost constraint of physical computation. A single FEP+ calculation takes hours to days of GPU time and customers have limited compute budgets. If AI surrogates cannot materially reduce the need for physical computation, the loop of physics generating data and AI extending coverage becomes expensive, which is the key question for scaling the model.

Sixth, dependence on target structures. Physics requires a known three-dimensional structure and cannot address targets without one. Structure prediction coverage is rising with AlphaFold-class methods, but whether predicted structures are accurate enough for FEP+ level work remains debated.

Seventh, competition and price pressure. Open-source tools including RDKit, GROMACS and OpenMM are already capable for basic tasks, and AI companies increasingly build in-house physics. Schrödinger's premium pricing needs continuous proof through accuracy and workflow completeness.

Composite Assessment

Technology scores 9.0. The in-house chain from quantum chemistry through force fields to free energy is top tier in computational chemistry, FEP+ accuracy is supported by peer-reviewed literature at roughly 1 kilocalorie per mole, sufficient to guide lead optimization, and three decades of error control engineering across force field parameterization, sampling convergence, protonation and bound water treatment constitute a real moat. Deductions for the cost ceiling of physical computation and dependence on target structures.

Business scores 7.5. Software annual contract value of 198 million dollars grows at double digits with high retention, LiveDesign's collaboration layer is sticky, and drug discovery provides option value through milestones and equity. Deductions for the hosted transition depressing near-term revenue and margin, annual contract value guidance below the single-quarter rate, and non-recurring drug discovery revenue.

Finance scores 6.5. Cash of 418.8 million dollars, a 9.3 percent reduction in operating expenses and a loss narrowing from 187.1 to 103.3 million dollars give ample runway. Deductions for continued losses, adjusted EBITDA of negative 114.9 million dollars, and a 2028 target dependent on several assumptions holding at once.

Product scores 8.0. Maestro, LiveDesign, FEP+, Glide, Desmond and Jaguar form a complete toolchain, RetroSynth fills synthetic route planning, Bunsen's direction of encoding expert knowledge into executable processes is correct, and the NVIDIA and Google Cloud partnership solves compute and channel needs. Deductions for missing quantified Bunsen validation, high pricing and limited interoperability with open-source ecosystems.

Market scores 8.0. The top global pharmaceutical companies are essentially all customers, materials science covers chemicals, energy and aerospace, the AI drug discovery boom reinforces the value of physics-based ground truth, and partnerships with BMS, Simcere and Lilly TuneLab deepen platform embedding. Deductions for tighter biotech funding affecting smaller customer budgets and substitution pressure from open source and AI companies.

Overall 7.9. This is the company with the hardest physical core in this series: thirty-five years turned molecular-level physical computation into something engineering-usable, and it found a positioning in the AI era of physics generating truth and AI extending coverage. The weakness is financial, with continued losses, margin eroded by the hosted transition and drug discovery revenue swinging on milestones. The judgment on physical AI is that the physical layer is real and hard to replace and that Bunsen's direction of agents orchestrating validated physical tools is correct, while whether the shift from selling tools to selling an AI platform succeeds depends on Bunsen producing quantifiable customer benefit and on physical computation cost falling with scale.

References

  1. Schrödinger second quarter 2026 results release, August 5, 2026: annual contract value 29.6 million dollars up 27 percent and 208 million on a trailing four-quarter basis; 22.6 million and 196 million excluding contribution annual contract value; software revenue 32.5 million dollars down 10 percent with hosted at 47 percent of software revenue and 30 percent on a trailing basis; drug discovery revenue 23.0 million versus 13.9 million including a 10 million dollar milestone from the Ajax acquisition by Eli Lilly; contribution revenue 3.4 million versus 4.8 million; total revenue 58.9 million dollars up 8 percent; software gross margin 71 percent; operating expenses 74.0 million down 6 percent; other income 48.9 million; net income 6.0 million versus a 43.2 million loss; cash and marketable securities 418.8 million
  2. Schrödinger 2026 guidance in the same release: full year annual contract value 218 to 228 million dollars, growth of 10 to 15 percent; drug discovery revenue 65 to 75 million dollars against prior expectations of 55 to 65 million due to the 10 million milestone; operating expenses below 2025; third quarter annual contract value 41 to 45 million excluding contribution annual contract value against 38.3 million a year earlier including 2.2 million of contribution
  3. Schrödinger second quarter 2026 product and partnership announcements in the same release: early access launch of Bunsen as an agentic AI co-scientist optimized to execute the validated physics-based platform, plan and run complex molecular discovery workflows and interpret results; NVIDIA and Google Cloud providing a full-stack AI platform, infrastructure and early customer access to the NVIDIA BioNeMo Agent Toolkit; a strategic software agreement with Bristol Myers Squibb to deploy Bunsen; a global drug discovery and development collaboration with Simcere Pharmaceutical Group where Schrödinger leads design and optimization in the joint research phase and Simcere leads preclinical and clinical development, with milestone payments and tiered royalties; publication in the Journal of Chemical Information and Modeling validating FEP+ for binding prediction
  4. Schrödinger fourth quarter and full year 2025 results release, February 25, 2026: total revenue 255.9 million dollars up 23.3 percent, software revenue 199.5 million up 10.6 percent, drug discovery revenue 56.4 million versus 27.2 million, software gross margin 74 percent, operating expenses 309.5 million down 9.3 percent, other income 64.6 million, net loss 103.3 million versus 187.1 million, year end cash and marketable securities 402.3 million versus 367.5 million, full year adjusted EBITDA negative 114.9 million versus negative 152.5 million, software annual contract value 198 million; targets positive adjusted EBITDA by the end of 2028, an accelerated transition to ratable hosted software revenue, and commercial launch of a predictive toxicology solution in 2026
  5. Schrödinger fourth quarter 2025 data in the same release: total revenue 87.2 million down 1.2 percent, software revenue 69.3 million down 13 percent due to accelerated recognition of up-front revenue from multi-year agreements signed in 2024, drug discovery revenue 18.0 million versus 8.7 million, software gross margin 81 percent, operating expenses 74.5 million down 12.2 percent, other income 50.1 million, net income 32.5 million versus a 40.2 million loss, fourth quarter adjusted EBITDA negative 5.2 million
  6. Company background from third-party sources: founded 1990 in New York by Richard A. Friesner and William A. Goddard III; first software sale to Brookhaven National Laboratory in 1992; Glide launched 2001, Prime 2003 and WaterMap 2009; Nimbus Therapeutics co-founded with Atlas Ventures in 2009; initial public offering February 10, 2020 of 11,882,352 shares at 17 dollars; headcount around 1,000, with one source citing 1,015 and another about 900 across 12 offices; Ramy Farid as CEO and Michael Lynton as chairman
  7. Product and business structure from third-party sources: Maestro as the molecular modeling interface, LiveDesign as a web-based collaborative informatics platform, FEP+ for free energy, Glide for docking, Desmond for molecular dynamics, Jaguar for quantum chemistry, Prime for structure prediction, WaterMap for bound water analysis, PyMOL for visualization, and RetroSynth launched in January 2026 for AI-driven synthetic route planning; the materials science suite covering organic electronics, polymers, consumer packaged goods, catalysis, thin films, energy storage, metals and ceramics; customers including Bayer, Takeda, Boeing and Sabic
  8. Pipeline and incubation from third-party sources based on 2025 disclosure: more than 25 first-in-class, best-in-class or first-in-modality programs with more than 10 in clinical stages and 16 royalty-eligible; 20 partners since 2018; more than 650 million dollars of cash generated from drug discovery since 2020 with potential future milestones approaching five billion dollars; proprietary candidates SGR-1505 as a MALT1 inhibitor and SGR-3515 as a Wee1 and Myt1 inhibitor both in phase 1; incubated companies including Nimbus, which sold an ACC inhibitor to Gilead in 2016 in a deal worth up to 1.2 billion dollars, Structure Therapeutics with positive phase 2b topline data for aleniglipron in December 2025, Ajax acquired by Eli Lilly in 2026, and Tectora announced in September 2026
  9. Schrödinger website, accessed 2026: positioning as a physics-powered computational platform serving life science and materials science; the therapeutics group selecting high-value targets with human genetics or clinical validation to advance proprietary and collaborative programs; the September 9, 2026 announcement of a licensing and collaboration agreement with Tectora Therapeutics, a company it co-founded, in immunology and inflammation
  10. January 2026 products and partnerships from third-party sources: RetroSynth introduced for identifying efficient synthetic routes for novel molecules; collaboration with Lilly TuneLab making LiveDesign the priority interface for biotechs accessing TuneLab workflows and combining Lilly federated learning models with Schrödinger physics; a strategic agreement with Manas AI granting access to the computational platform
  11. Technology and competitor information from public sources: Jaguar performing density functional theory and ab initio calculations; OPLS-family force fields describing intermolecular interactions; Desmond for large-scale molecular dynamics; FEP+ using thermodynamic integration or free energy perturbation for relative binding free energy with mean error around 1 kilocalorie per mole on congeneric series; competitors including Dassault BIOVIA, Chemical Computing Group MOE, Cadence OpenEye and Cresset, plus open-source tools RDKit, GROMACS, AMBER and OpenMM
  12. Industry background from public sources: the data bottleneck in AI drug discovery arises from publication bias, inconsistent assay methods and sparse chemical space coverage; physics does not depend on historical data and can extrapolate into new chemical space but its cost grows linearly with molecule count; FEP+ requires a three-dimensional target structure and needs special handling for scaffold hops, covalent inhibitors, allosteric modulators and metalloenzymes
This report is compiled from public sources for informational purposes only and is not investment advice. Financial figures follow company announcements and SEC filings; pipeline milestones and equity gains are uncertain and valuation judgments are analytical.

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.