Jonathan Horn, CEO and Co-Founder of Treefera – Interview Series

Jonathan Horn, CEO and Co-Founder of Treefera, is a technology entrepreneur and former investment banking executive with deep expertise in risk management, artificial intelligence, and large-scale data analytics. Prior to founding Treefera in 2022, he held senior leadership roles at both J.P. Morgan and Citigroup, where he focused on risk, data, and complex financial systems. Drawing on his background in financial-grade risk modeling, Horn launched Treefera to address one of the most persistent blind spots in global supply chains: the “first mile,” where raw materials originate. Under his leadership, the company has rapidly grown into a leading provider of AI-powered supply chain intelligence, helping enterprises gain real-time visibility into sourcing, environmental risks, compliance requirements, and operational resilience.

Founded in London in 2022, Treefera is an AI-enabled supply chain intelligence company focused on bringing transparency to the first mile of global commodity supply chains. Its proprietary data fabric combines satellite imagery, spatial and temporal data, AI models, and risk analytics to provide organizations with real-time insights into sourcing, compliance, sustainability, and supply chain risk. The platform helps businesses monitor everything from deforestation exposure and carbon impacts to commodity sourcing risks, enabling more informed decision-making across procurement, finance, and operations. By transforming fragmented environmental and supply chain data into actionable intelligence, Treefera aims to strengthen supply chain resilience in an increasingly volatile and regulated global economy

You founded Treefera after senior risk, AI, and data analytics roles at J.P. Morgan and Citi. What specific gap did you see in how enterprises understood agricultural and soft commodity risk that convinced you Treefera needed to exist?

Nature-based commodities represent $2.1 trillion of world trade, yet the data underpinning risk assessment for those assets was manual, delayed and structurally inaccurate. Working with risk models at J.P. Morgan and Citi, I saw pricing and exposure decisions that depended on data originating at the first mile of agricultural supply chains: collected by hand, prone to omission and often weeks or months behind reality. Sixty percent of supply risk originates at the first mile, before commodities ever reach a port or an exchange, yet that was precisely where visibility was thinnest.

What convinced me to build Treefera was the convergence of two things: the scale of the problem and the arrival of tools capable of addressing it. Satellite resolution and coverage had reached a point where you could observe crop conditions at field level across major production regions at minimal cost. AI had matured to the point where it could turn that raw signal into something financially interpretable. No one was connecting those dots in a rigorous, science-first way. Enterprises were still pricing risk based on government reports that lagged ground conditions by months, national averages that masked local variation and linear models that could not handle climate volatility. The gap between what was knowable and what was known was enormous. That gap is where Treefera sits.

Treefera is often described as an AI-powered intelligence platform, but your approach is explicitly not centered on LLMs. How do you explain the difference between predictive AI for supply chains and the generative AI systems that currently dominate the conversation?

Generative AI and large language models solve a fundamentally different problem. They are productivity tools: extraordinarily useful for simplifying repetitive tasks, like drafting and summarizing. The commercial challenge for those systems is adoption, getting people to change how they work. That is a market education problem, not a scientific one.

Treefera uses AI to solve scientific problems with financial-grade accuracy requirements. Our deep learning models are trained to interpret satellite imagery, climate signals and crop biology to produce yield and production area forecasts accurate enough to inform capital allocation decisions. The question we are answering is not “what does this document say?” It is “what will this region yield in three months, and how confident should you be?” Those are not the same class of problem and they do not require the same class of model. LLMs are optimized for language; our models are optimized for physical world interpretation. Confusing the two leads to applying the wrong class of tool to a problem it was not built for.

Many AI companies argue that better performance requires more compute, larger models, and greater access to GPUs. Treefera appears to challenge that assumption. What does “frugal computing” mean in practice, and why is it important for applied AI?

The prevailing assumption in AI is that scale equals performance: more parameters, more GPUs, more cloud infrastructure. For applied AI in domain-specific contexts, that assumption is wrong, and it produces unnecessary cost and energy waste.

Frugal computing, in practice, means three things for us. First, we decouple compute from time. Most of our processing tasks do not need to happen at a specific second. Rather than running always-on infrastructure, we identify periods of excess network capacity and borrow compute during those windows.

Second, we decentralize workloads. Rather than routing everything through a single cloud hub, we distribute across a network of available nodes, including blockchain infrastructure, which carries significant idle capacity at certain periods. If one node becomes inefficient, tasks reroute dynamically.

Third, we right-size the hardware. We use NVIDIA AG6 rather than top-tier chips, where the performance is equivalent for our workloads at a fraction of the energy and cost. The reason this matters beyond efficiency is precision. Frugal computing forces discipline about what computation you actually need. That produces leaner, more interpretable models – the kind financial and operational decision-makers can actually use. They do not need a larger model; they need a more accurate answer.

Your platform reportedly delivers predictive outputs on crop yield, land use, and supply risk without relying on always-on cloud infrastructure. How do you decouple compute from time while still delivering commercially useful, near real-time intelligence?

Near real-time intelligence and continuous computation are not the same thing. Our customers need weekly updated forecasts; they do not need the compute that generates those forecasts to be running continuously.

We map our processing cycles to the natural rhythm of the data. Satellite imagery arrives on a cadence. Weather inputs update on a cadence. The analytical questions our models answer are also cadence-driven: what is the yield trajectory for this region, what has changed this week in planted area. So we pre-schedule compute jobs to run against those data windows, borrowing capacity from distributed infrastructure during low-demand periods. The output to the customer is a weekly data feed that is current, actionable and model-ready. The infrastructure behind it only runs when there is something meaningful to process. That architecture is also more resilient. A distributed workload that routes around failed nodes is more reliable than a single always-on server with a single point of failure.

Agricultural and soft commodity markets still rely heavily on delayed, survey-based, or fragmented data. How does AI change the way companies, banks, insurers, and traders can assess risk before it becomes visible in official reports?

The structural problem with survey-based reporting is that it is backward-looking by design. By the time a government agency publishes a supply estimate, the physical conditions underpinning it are weeks or months old. Markets that move on that data are reacting to history.

AI changes that by shifting the information source from surveys to direct observation. Satellite imagery, climate signals and crop development data are available now, not in six weeks when a report is compiled and published. What our models do is translate that physical data into the financial language that traders, underwriters and analysts actually use: yield forecasts with stated uncertainty intervals, production area estimates with weekly updates, stress scores that quantify risk at origin before it surfaces in prices.

In 2022, our US Corn models flagged a downward revision five weeks before USDA published theirs. In January 2025, our models surfaced a stress score of 0.76 across Ghana’s cocoa belt; COCOBOD did not revise its season forecast until June. The information advantage is not marginal; it is structural. Enterprises still waiting for official reports to make supply and pricing decisions are operating with a lag that their counterparties may not share.

The “first mile” of supply chains has historically been one of the least transparent areas for global enterprises. Why is first-mile visibility becoming so critical now, especially as climate volatility, regulation, and geopolitical uncertainty increase?

Three forces are converging and each is individually material.

Climate volatility is increasing the frequency and severity of production shocks at origin. The same weather event now carries greater supply impact because crop systems are more stressed and extreme events are less predictable from historical averages. Linear risk models built on historical norms are structurally ill-equipped to handle that. You need real-time physical observation to see what is actually happening in the field.

Regulation is creating a direct liability link between what happens at the first mile and what an enterprise can sell or finance. EUDR, CSRD, TCFD: these frameworks require enterprises to know, with evidence, where their commodities come from and what the conditions at origin were. “We trusted the supplier” is no longer a defensible position. That liability is pushing traceability and provenance from a procurement preference to a legal requirement.

Geopolitical disruption has made single-origin dependency a boardroom risk. When one region accounts for a dominant share of a commodity’s global supply and that region becomes politically or physically unreliable, enterprises without first-mile visibility have no early warning mechanism. They find out when the market has already repriced.

There is also a broader shift happening in the data ecosystem itself. Google Earth AI’s recent launch of pan-tropical commodity maps – annual 10-meter crop footprints for cocoa, coffee, oil palm and rubber, released as open data – is a useful indicator of where things are heading. The physical world is increasingly legible from space, and demand for supply chain transparency is now mainstream enough to attract big-tech investment at scale. Treefera welcomes that. A richer foundational data layer raises the floor for the whole market and creates shared awareness that better information is not just possible, it is available.

What open discovery maps cannot do is close the intelligence gap. Knowing where crops are planted is not the same as knowing how this season is developing, what conditions at origin mean for your supply exposure or where your portfolio is at risk. The translation from observation to financial-grade insight is what Treefera is built for.

First-mile ignorance used to be commercially tolerable when the world was more predictable. It no longer is.

Your customer base includes major organizations such as JP Morgan, Microsoft, Bayer, and Anew. What are the most common problems enterprises are trying to solve with Treefera: compliance, supply risk, forecasting, sustainability, procurement, or something else?

The core problem enterprises bring to us is a version of the same thing: they have significant financial exposure to what happens at the first mile of agricultural supply chains and no reliable mechanism to see it before it costs them. The specific framing differs by sector.

For traders and commodity-exposed financial institutions, the question is information advantage: seeing supply shifts before they appear in prices or official data. For agricultural lenders and insurers, it is risk assessment; they are underwriting or financing operations whose performance is directly governed by conditions they cannot observe. For corporates with sustainability or compliance obligations, the question is evidence: proving, with defensible data, that their supply chains meet the standards regulators and counterparties now require.

The traditional answer – trust the supplier, wait for the government report, buy the consensus estimate – is no longer adequate. The precision and speed they need do not exist in the public data ecosystem. It exists at the first mile.

Treefera has reported 6x year-over-year growth, zero churn, and an oversubscribed Series B within two years. What does that level of adoption suggest about enterprise demand for AI systems that are precise, efficient, and operationally grounded rather than simply large-scale?

Zero churn is the most telling signal. Revenue growth can reflect commercial execution; zero churn reflects product-market fit. Customers who have used the data for a full season, back tested it against their own models and made decisions based on it, and then renewed, are telling you that the signal is real and that it is changing how they operate.

It also points to significant unmet enterprise demand for AI that is precise, auditable and operationally integrable – demand underserved by a landscape heavily weighted toward generative tools and large-scale general models. Supply chain and risk professionals need a forecast with a stated uncertainty interval that they can stand behind in a risk committee. When that bar is met – domain-specific data, financial-grade accuracy, transparent methodology – enterprises prioritize it. The oversubscribed fundraise reflects investor recognition of the same dynamic: the market is large, the problem is structural and the existing data infrastructure was not built to solve it.

Looking ahead, do you believe the next phase of applied AI will be defined less by model scale and more by operational efficiency, domain-specific data, and measurable business outcomes?

Yes, and the evidence for it is already visible.

Large-scale AI is beginning to hit the limits of what raw scale can solve. Adding parameters does not improve a crop yield forecast if the underlying data is coarse, delayed or geographically misaligned. The marginal value of compute is diminishing when the bottleneck is data quality and domain precision, not model size.

The next phase will be defined by domain-specific training data, right-sized models and verifiable outputs. In sectors like agriculture, finance and supply chain, where decisions carry financial and operational consequence, the question has never been “how large is the model?” It has been “how reliable is the answer and how quickly can I act on it?” Scale alone cannot answer that. The companies that will lead in applied AI over the next five years will have built proprietary data pipelines into the physical world, trained models on that data with appropriate scientific rigor and demonstrated measurable accuracy in live conditions. The technology is increasingly a commodity; the data and the domain expertise are not.

Thank you for the great interview, readers who wish to learn more should visit Treefera.