Calice and Virtual Field Trials

A few months ago, I shared a list of questions on Linkedin asking start-ups to fill out the questionnaire so that I could learn more about new companies operating in the industry. Calice filled out the questionnaire and stood out as a really interesting company to learn more about. I recently connected with Chief Commercial Officer, Alejo Valverde, to dig deeper in to organization.

Index:

Calice, at it’s most basic, runs virtual field trials.

Its system takes genotype, environment, and management variables as inputs and simulates their interactions across thousands of combinations, producing performance predictions for products and varieties in environments where they have never been physically tested.

Agribusinesses use the output to screen variety candidates earlier, shrink and redesign trial networks, and decide where a product should be positioned commercially.

According to a recent AgFeed profile, the company initially focused on gene discovery and editing before pivoting to computational modeling for trialling.

Within the industry, we validate and learn about products by doing trials.

A new variety, trait, or biological input gets planted across hundreds of locations spanning different climates, soils, and management systems, and the company waits a full season for results, then repeats the process to obtain adequate data.

Multi-environment trial programs run into the millions of dollars per product, and even at that cost they sample only a small fraction of the genotype-by-environment-by-management combinations a product will encounter commercially.

Not testing enough runs into the obvious risk of a lack of understanding, such as finding a product that performs well in a dry spring, but under performs in a wet one or a variety that does well in one soil type, but underperforms in another.

Companies frequently discover this nuance after launch, when the issues have already shown up in the field, and when dealer and farmer confidence has already been impacted.

There is also a knowledge retention problem on top of the cost problem. The industry runs hundreds of millions of dollars of trials annually, and most of what those trials reveal about how products interact with environments never gets systematically captured or reused. Each program largely starts from scratch. Calice's insight was that this knowledge is recoverable and usable for the future.

Calice describes its system as an environmental engine built in layers.

The enviromics layer constructs what the company calls environmental fingerprints: temporal models of environmental conditions aligned to crop development stages, capturing the sequence of events a crop experiences rather than seasonal averages. They describe the principle “as if we were looking at the environment through the eyes of the plant. A soybean plant doesn't react to the same climatic effects as a corn plant. Corn is more sensitive during flowering, soybeans during grain filling. It's not the average environment that matters, but the sequence of events that happens during each phase of the crop."

The P×E×M (P for ”product,” usually stated as GxExM) simulation layer sits on top of that structured environmental representation and predicts performance across untested combinations. The company's central claim is that its machine learning operates inside a biologically structured model rather than searching for statistical patterns in flat trial tables, and that this is what allows extrapolation to conditions the system has never seen.

Clients access the engine three ways: a web platform, decision-ready reports, or API integration into their own internal systems:

Notably, the system requires no dedicated data generation. It works on trial and environmental data clients already have.

Two things are worth highlighting about this architecture.

First, the underlying premise has solid scientific footing. Enviromics is an established and growing research area, and the broad finding that environment and management jointly explain more yield variance than genotype/product in multi-environment trial data is well supported in the scientific literature, even if Calice's specific 30-60% environment, 20-40% management, 10-25% genetics/product split will vary considerably by crop, trait, product and dataset.

Second, none of the component ideas are new. Process-based crop models like DSSAT have existed for decades, envirotyping is an active academic field, and every major company has internal teams doing some version of environmental classification for product placement.

Calice's differentiation is in execution: the specific modeling architecture, the cross-crop and cross-product-category breadth, and the accumulated calibration from running cycles against client data.

The current development priority is reducing the personnel dependency. Today, engagements require Calice's own scientists to configure and run analyses, which means client growth requires headcount growth. The company is building agentic pipelines intended to run the full reasoning stack autonomously, compressing turnaround from weeks to minutes. This has similarities to what Corteva has described on the genetics side, where AI agents run breeding programs and make decisions across datasets, and it points to the same direction crop protection is likely heading. No crop protection company has used the "agentic" label explicitly yet, but Syngenta's "Data for Synthesis" platform is building the substrate that agentic workflows require.

Last year, I wrote about Competing on Analytics, and how the book of that name outlines that companies that lead in their sectors don’t just use data for reporting, they embed analytics into decision-making at every level.

These “analytical competitors” do three things exceptionally well according to the book:

Netflix is an example that is prominent in the book.

Netflix gathers granular behavioral data, applies advanced algorithms to recommend and acquire content, and integrates those insights into what shows it makes (capital allocation), what shows it recommends (customer experience), pricing (value maximization), and marketing decisions (customer acquisition). Its advantage is inseparable from its analytics capability.

I believe this is true in all competitive markets: analytics are not a nice to have. They are the capability required to outsmart the competition and effectively deliver distinct experiences and outcomes to customers.

Calice sits well inside this logic as a tool to differentiate as they can help with resource allocation, product recommendation, marketing decisions and more.

The most compelling part of Calice is how clients test it.

Several large companies handed over historical trial datasets with outcomes withheld and asked the engine to predict what had already happened.

In one case, Calice analyzed pre-commercial trial data for a corn variety that had already launched and flagged that it was mismatched to specific target environments. The variety had already failed in exactly those markets. Blind retrodiction of this kind is now a standard part of how the company engages new clients, and it is a credible validation method because the client has the answers already.

With that said, retrodiction does have limits as a legitimizer.

Predicting held-out historical outcomes within the distribution of environments a company has already trialed is an easier problem than the other problem Calice is positioning for: predicting performance in untested environments for products that do not yet exist commercially. For example, a novel biological in a new commercial area.

Across its client base, Calice reports an average 60% reduction in trial network size without information loss, 46% more candidates reaching the next stage gate, 2.5x more efficient product development, and decisions arriving 30 to 45 days earlier.

All of these figures originate from the company, from private engagements, with no published methodology and I did not vet these with any of their clients, but I have no specific reason to doubt them, and the client list implies sophisticated buyers found the outputs worth paying for repeatedly, but how the numbers progress as the company grows will be interesting to see.

To Calice's credit, the company emphasizes that the engine does not replace field trials. It compresses the path to them and redesigns what gets planted.

Calice is promising fewer, better-designed trials, which is both more believable and more useful — it is unlikely that we can ever eliminate physical trialling altogether, but we can improve at the the trialling already being done.

Calice sells enterprise direct, exclusively to large organizations: seed companies, crop protection companies, biological manufacturers, and grain origination businesses like malting companies.

Engagements start as paid proofs of concept tied to a decision the client is already making, then expand from one team to adjacent teams toward multi-year enterprise licenses. They shared with me that free pilots have failed, which I have found consistent no matter the industry segment, and that selling the technology rather than the decision stalls deals — reinforcing that it is about emphasizing the problem they are solving for the prospective client first.

The client list is pretty significant: Syngenta, BASF, GDM, Nufarm, TMG, Boortmalt, CoverCress, and CIMMYT, Corteva, Nufarm, and Puna Bio, across soy, corn, wheat, cotton, sugarcane, barley, canola, carinata, potatoes, and sorghum. The Boortmalt and GrainCorp (investor in Calice) relationships are interesting to me, because they extend the engine past input manufacturers into sourcing and origination and can lead into quality based forecast parameters.

The second revenue stream is interesing: Tech Partnerships, where Calice co-develops products with breeding partners and takes royalties on commercialized products instead of fees.

If the engine's predictions are as good as claimed, royalty exposure to the products it helps create is where the asymmetric upside lives, because annual licensing contracts will never fully capture the value of what a correct placement decision is worth on a successful hybrid. It also functions as a signal: a company confident in its predictions should want payment contingent on outcomes, and Calice seemingly does.

On the capital side of things, Calice has roughly $3.5 million raised to date, comprising a $750K pre-seed in 2023 plus about $300K in grants, and a $2.5 million seed in 2025 led by Astanor with participation from Draper Cygnus, Artesian, GrainCorp Ventures, and others. Revenue is undisclosed, described as growing consistently, and the company is not claiming profitability.

For context, that is a very small amount of capital against a stated three-year goal of becoming operationally embedded inside the world's largest agricultural companies.

The competitive set has three segments.

The specialized startups are one portion of competitors.

Computomics does genomics and microbiome analytics, NoMaze does breeding program management and genomic prediction, INTENT and INNOVA combine contract research with market analytics. Many operate on a narrower scope, typically crops only, and most require dedicated data generation or sequencing pipelines that clients fund upfront. Calice's no-new-data-required approach and its breadth across seeds, biologicals, and chemistry are advantages. Heritable Agriculture, the Alphabet X spinout that entered a collaboration with Syngenta Vegetable Seeds in late 2025, is another segment. It approaches the problem genetics-first where Calice approaches it environment-first. The two are not direct substitutes today, but they likely pull from the same budget inside the same buyers.

The second group is internal. Every major company has data science and bioinformatics teams running computational analysis on their own trial data, and Bayer in particular has spent a decade building environmental classification and seed placement capability through Climate Corp and its breeding organization. Calice's counterargument is that internal teams are crop-siloed, method-conservative, and lack the cross-domain architecture. Displacing or complementing internal teams requires organizational trust built over multiple seasons, and Calice itself acknowledges this is a barrier that needs to be overcome. An internal data science lead has career incentives to build internally rather than buy and “Not Invented Here Bias” is a strong one.

The third tier is general-purpose AI capability. Calice's advantage rests on embedding ML inside biologically structured models. If foundation models keep improving at scientific reasoning and simulation, the specialized-architecture of Calice also has to keep improving to justify the cost and usage.

What Calice has, in my view, is a moat-in-progress. The engine calibrates to each client's data, crops, and decision patterns over multiple seasons, so the model running in year three of a relationship is meaningfully better than the first year version, and switching away means abandoning that accumulated calibration. Multi-cycle embedding has to be a significant component of their strategy.

The other interesting question is around data governance: Calice says knowledge accumulates in the engine with every engagement, and its clients are direct competitors of one another. How much learning transfers across accounts, and how comfortable Syngenta is for example with an engine that also runs Corteva's data, is something that I am sure will constantly arise.

There are several groups that I think would be interested in digging more into Calice.

I will also be curious how the larger crop input manufacturers will leverage system like Calice. They are stated targets and current clients (eg: Corteva), but they are also the groups most able to build internally and most likely to view an external engine as unneeded risk.

Calice's own exit logic acknowledges this as well, as the natural acquirers are the clients, because the capability is worth more as an owned asset rather than licensed if it is truly differentiated. That is a plausible acquisition thesis and a reason a larger entities might hesitate to let a competitor-shared platform get embedded in their organization.

Calice is working on a problem with a legitimate value when solved. Trial economics are tough, most of what trials teach never gets reused, and the science behind environment-first modeling is well established.

What the company still has to prove is that the engine extrapolates as well as it retrodicts, because predicting held-out history inside familiar environments is a different task than placing a product that does not exist yet in a geography no one has trialed.

The strategy hinges on multi-cycle embedding, which is the one variable Calice does not control, and it runs against buyers who are simultaneously its best customers, its most capable internal competitors, and its most likely acquirers.