We closed Kvistlund Fund I in March 2021. This essay is an attempt to articulate, in public, what we believed when we raised it — the structural argument for why AI-native logistics software will restructure European supply chains over the next decade, and why we thought that thesis was early enough to be interesting at seed stage.
Writing investment theses in public is useful discipline. It forces specificity about what you actually believe and creates a record against which you can measure your judgment. In five years, the parts of this that were right will be obvious, and the parts that were wrong will be instructive.
The supply chain software stack was built for a different era
The core systems that run European supply chains — ERP, WMS, TMS — were architecturally designed in the 1990s and 2000s, when the core constraint was data management and workflow automation. These systems solved real problems: getting inventory records out of spreadsheets and into databases, automating purchase order generation, connecting warehouse operations to financial reporting. They were genuinely valuable, and their installed base is immense.
What these systems were not designed for is prediction. They record, they automate, and they report on what has happened. They don't model what is likely to happen or optimize decisions under uncertainty. Demand forecasting modules in legacy ERP systems run exponential smoothing on weekly sales history. Routing modules in legacy TMS platforms apply heuristic rules to freight allocation decisions. The "optimization" in most supply chain software is not optimization — it's structured workflow automation that humans have designed to produce reasonable outcomes in expected conditions.
The gap between what legacy software does and what is now technically possible is large and growing. Machine learning models trained on sufficient data can outperform heuristic rules on most supply chain decision problems — demand forecasting, route optimization, inventory positioning, freight procurement, warehouse task prioritization — by margins that translate directly into measurable operational cost reduction. The question in 2021 was not whether ML would eventually be the core technology in supply chain decision software. It was how fast, through which market entry points, and who would win.
Why the timing was right
Three structural factors converged in the 2019-2021 period to make this the right moment to back software-first operators building ML-native logistics products.
First, data availability. Supply chain operations generate enormous volumes of operational data — shipment events, inventory movements, order transactions, carrier performance records, demand signals. For most of the last two decades, this data was trapped in siloed on-premise systems with no API access and no cloud aggregation. The shift to cloud ERP, cloud WMS, and cloud TMS systems over the preceding five years had finally created the data accessibility necessary to train production ML models at scale. The data was there. The engineering infrastructure to use it was newly available.
Second, ML tooling maturity. The gap between "ML research project" and "ML production system" narrowed significantly between 2016 and 2021. Frameworks like PyTorch and TensorFlow matured. MLOps tooling for model training, versioning, deployment, and monitoring became available as managed services. A small team of ML engineers in 2021 could build and operate production ML systems that would have required a large specialized team in 2016. This made the build economics for ML-native logistics startups much more favorable.
Third, operator-founders entering the market. The wave of logistics operators who started their careers in digitizing supply chains — not the generation that built the legacy systems, but the generation that inherited them and spent 10 years fighting with their limitations — was hitting the age and experience level where starting a company made sense. These were engineers and operators who understood deeply what the legacy systems couldn't do, had spent years trying to work around those limitations, and had enough technical context to understand what ML could actually change. The founder quality in European logistics AI startups in 2020-2021 was materially different from the previous wave of logistics tech — more technical, more operationally experienced, more focused on the core prediction problem rather than the workflow layer.
Why European geography specifically
The US had most of the capital and much of the first wave of logistics tech infrastructure investment. But European supply chains have structural characteristics that make them a compelling geography for this thesis. The EU operates as a single market with heterogeneous operational contexts — 27 national markets with different regulations, languages, customs procedures, infrastructure quality, and carrier networks, all needing to interoperate. This complexity means that the demand for intelligent software to manage cross-border logistics is higher in Europe than in any comparable domestic US logistics network.
European logistics operators — particularly 3PLs, FMCG distributors, and automotive supply chain teams — have sophisticated operational requirements and procurement capabilities for supply chain software. The Scandinavian markets in particular have a tradition of operational excellence and technology adoption in supply chain that makes them natural early adopters of ML-native logistics products.
And European logistics AI startups in 2021 were dramatically underfunded relative to their US counterparts. The same quality of founding team, building the same quality of product, addressing the same scale of market, was raising at 30-50% lower valuations than comparable US companies. That funding gap created an opportunity for a seed fund with European network and domain expertise to access the best founders at attractive prices.
What we were wrong about
Three years into deploying Fund I, one thing we underestimated is how long the market cycle in enterprise logistics software can be. The 2021 market was characterized by strong tailwinds — supply chain disruption had made logistics optimization a top-of-mind board priority for large operators, and software budgets were expanding. When that environment shifted in 2022-2023, the enterprise sales cycles that had been running at 3-4 months extended to 9-12 months. Companies that had planned headcount and burn on faster revenue recognition needed to reset their timelines.
The thesis held. The timing calibration in our financial models was optimistic. That's a normal seed-stage calibration error, and it doesn't change the direction of the bet. Supply chain software is being rebuilt from the ML layer. The question was always when, not whether. The answer appears to be: over a decade, not over three years. We plan accordingly.