Environmental remediation is a large, lucrative market defined by fragmentation and dominated by a service-heavy model (see appendix). To date, the market, which by some estimates delivers $20b+ in annual revenue, has largely resisted productization and technologically-forward approaches. 

We believe change here is inevitable and that a combination of technological momentum and novel business and financing models is creating substantial company-building opportunities for new players.

Today, there are over 450,000 brownfield sites in the US and nearly 3 million suspected contaminated sites in Europe. As a result, over $3 trillion in real estate value is essentially stranded. Traditional cleanup approaches are slow and costly, with the average federal cleanup taking 12-15 years, thanks to substantial regulatory overhead, labor constraints, and limited automation. 

Having had a front row seat to the way KoBold has grown, we see potential for a similarly data-driven business model that i) identifies high-value contaminated sites, ii) applies superior remediation technologies (via advanced chemistry or otherwise), and iii) captures the value differential between pre- and post-remediation states through, for example, real estate development or the redevelopment or repurposing of stranded assets (energy, resources, industrial equipment).

The why now is backed up by technological convergence that is helping to change the economics of brownfield transformation. Advanced sensing can precisely map contamination without extensive manual sampling, while innovative treatment methods can destroy persistent pollutants in days rather than years. Industrial robotics is increasingly becoming viable, which is particularly important in a market where labor constraints are substantial and worsening (thanks to an aging workforce and little infrastructure for backfilling the gap).

We imagine the ideal business model and capital stack will resemble KoBold’s hybrid approach: venture capital for technology development paired with infrastructure-like financing for asset acquisition and development. This bridges fast-moving tech-enabled discovery with slower land and asset development cycles while allowing value capture from both services and asset appreciation.

We see remediation as a wedge to a broader opportunity, which we have referred to in the past as a “stranded asset exploration company”. Companies will likely start by focusing narrowly on specific geographies, pollutants, or site types (mines, agriculture, waterfronts, etc.) before expanding.

PFAS in dirt or water, for example, represents perhaps the most immediately addressable and culturally salient initial opportunity areas. However, there are a wide variety of potential starting points. 

• Urban brownfields in high-demand real estate markets

• PFAS-contaminated military bases and airports

• Former manufacturing facilities with prime locations

• Mine tailings containing recoverable metals

• Waterfront properties with sediment contamination

• Agricultural land impacted by contaminants near expanding suburbs

• Contaminated aquifers in water-stressed regions

Companies like Valinor and biotech precursors like BridgeBio present another structural model to consider for such a company. Our understanding of how the “go to market” model works in remediation, particularly around remediation of government-owned and monitored sites, is that it is highly relationship-based, even more so than other government procurement processes.

Thus, if a company can centralize those relationships and the playbook for winning remediation mandates, they can scale that across a range of vertical focuses that can be staffed and financed in ways that are right-sized for a given opportunity. There are also centralizable technical capabilities that can be developed, from advanced chemistry to robotics.

There is substantial policy uncertainty surrounding the opportunity. The EPA and the environmental market overall are hot-button political issues where resourcing could be shifted or shut down with little warning. However, we think the amount of trapped value in the form of real estate, industrial equipment, and resources (minerals, energy, etc.) that can be unlocked through better data and automation will make this opportunity resilient against policy shifts. 

One way to think about the slowdown in scientific discovery is that it is operating at the wrong level of resolution. Our current research ecosystem – constrained by peer review, “publish or perish” productivity incentives, and hyper-specialization – has systematically dismantled the scientific freedom that enabled the researchers behind the 20th century's most transformative discoveries – what Donald Braben termed the “Max Planck Club”.

This creates massive opportunity costs to society and signals a significant company-building opportunity. 

We see two emerging pathways to recalibrate this resolution:

1. Individual Amplification: Tools that enable exceptional researchers to transcend institutional constraints and operate like 1-person research institutes. AI-augmented scientific workflows could dramatically expand what a single scientist can accomplish across literature review, experiment design, and analysis

2. Mass Collaboration: Platforms that enable unprecedented collective intelligence by breaking down information silos between disciplines and creating new coordination mechanisms for distributed discovery.

We are most interested in pathways focused on one or more key levers below:

Funding & Incentive Redesign: Models that reward researcher autonomy, risk-taking, and open collaboration rather than incremental, pre-specified outcomes. This includes both traditional capital allocation mechanisms and novel approaches like prediction markets for scientific outcomes.

Human-AI Scientific Workflows: Tools that seamlessly combine what AI is good at (e.g., literature review, document creation, etc.) with human scientific intuition (e.g., better questions), 

Proprietary Data Assets: Platforms that generate unique, structured scientific data through novel experimental methods or researcher collaboration.

While many of the hybrid non-profit/for-profit research institutions that emerged in 2021-22 (the various attempts at building OpenAI for Chemistry or Biology) have failed to gain traction, we remain interested in thoughtful full-stack approaches, particularly given the capital availability that now exists again such “compound” company building models. The most compelling version might integrate all these elements—new funding models, workflow tools, and collaboration platforms—into coherent systems that reshape how science is produced and validated.

While we are already invested in Sakana and recognize that many solutions may fit well within the attention radius (if not the current technical blast radius) of large AI labs, given the philosophical proximity of science and general intelligence, we also believe the why now argument is compelling as trust in academic insutitutions is at an all time low and the geopolitical necessity of scientific breakthroughs at a generational high-point.

We are entering an era where the notion of “markets in everything” – long an economist blogger trope – is happening at never-before-seen speed and scale. Prediction markets are one obvious instantiation of this trend. 

But the opportunity is more expansive than that. Massive amounts of (real-world) data coupled with rapidly improving processing and sensemaking capabilities create the foundation for understanding and pricing previously opaque risks, as we have written about in multiple memos: new standards, insurance products, or ways to price externalities. We expect this trend to accelerate as AI compresses the time between cause and detection. Information asymmetries will close faster, enabling players to identify and act on risks sooner. We especially see an opportunity in areas where AI does not only help detect these information advantages (unless there is unique access to data, these are only temporary), but can help act on those in much more efficient and effective ways, shifting the expected ROI of specific opportunities.

One example we are excited about in this context is detecting fraud and corporate misconduct related to harmful products. Instead of waiting years for harmful products to be flagged by regulators, platforms can mine medical and consumer data to spot dangers early. Combining these insights, AI-enabled reduction in litigation costs, and increases in litigation success rate (e.g., as a result of early issue detection and case abandonment), the ROI associated with contingency fees can become increasingly attractive. The key question for such a business is how it can establish itself as the primary consumer platform for litigation that enables it to maximize its volume and acquire proprietary data at scale.

Another example is that this could give rise to a new class of distressed debt investors. AI streamlines distressed investing by quickly analyzing legal documents and covenants to pinpoint optimal positions in the capital stack while leveraging historical bankruptcy data to forecast restructuring outcomes under different legal regimes. Additionally, AI-driven litigation analytics and scenario modeling empower investors to decide between litigation and out-of-court restructuring to maximize recoveries strategically.

The rise of foundation models has transformed how we handle unstructured data like text and images. Yet many organizations still struggle to extract value from their structured data (tables, databases, time series). A study by MIT and Databricks found that only 13% of large enterprises are driving material business value from data and machine learning strategies – a problem unlikely to be eased as data volume grows exponentially, regulations expand, and talent shortages persist. 

Every industry runs on structured data. Faster, more accurate analytics directly impact operational decisions and bottom-line results, and we believe companies laser-focused on delivering this improvement have the potential to create and capture substantial value.

The current landscape around structured data is – and always has been – defined by bottlenecks: data integration across disparate systems, manual feature engineering requiring domain expertise, model selection involving extensive trial-and-error, and complex workflow integration. Foundation models address these challenges by learning from vast datasets to automate these processes end-to-end.

Recent breakthroughs like TabPFN demonstrate that models trained on millions of synthetic datasets can achieve zero-shot predictions on new tables, consistently outperforming traditional methods. These models eliminate the need to train from scratch for each new problem, creating a paradigm shift in how organizations approach analytics. The combination of these algorithmic advances with multimodal capabilities, evolving business models focused on capturing outcome-based value, and persistent talent constraints is accelerating adoption.

Despite the generalizability of many emerging approaches, we believe the complexity of the workflows required to make use of the data and insight necessitates a sharp focus on building industry-focused solutions (with, of course, a strong vision for sequencing into further segments). 

Better availability and assessment of structured data is, in fact, a core element of our “action layer” thesis across critical industries. 

• Insurance underwriting benefits from automated risk scoring, incorporating structured profiles alongside multimodal data like claim photos.

• Life sciences, particularly clinical trials, can leverage synthetic control groups from historical trial data, reducing required patient numbers and accelerating drug development timelines.

• Supply chain operations leverage time-series foundation models for demand forecasting, recognizing patterns across products and locations by transferring knowledge from similar historical patterns.

• Manufacturing facilities deploy models trained on sensor data from thousands of machines to predict failures before they occur—even for equipment that has never failed before—enabling proactive maintenance and reducing costly downtime.

• Financial services models can process years of transaction data to identify patterns across thousands of variables, enabling faster, more accurate risk assessment and fraud detection.

Across the model landscape (for structured and unstructured data), we continue to be open to different arguments on where sustainable value will accrue in this ecosystem – i.e., where does the proverbial LLM blast radius end? 

While specialized foundation models for structured data have clear advantages over general-purpose models today, the pace of competition makes a sharper focus on structured data inevitable. Thus, durable advantages may ultimately lie not in the models themselves, but in the surrounding elements of the stack – proprietary data access, workflow integration, and physical world action layers – the mechanisms that translate insight and predictions into real-world operational outcomes.

Our ongoing exploration of the hardware development stack, from AI-driven CAE to PCB automation, has consistently pointed us toward a fundamental challenge: the immense complexity of coordinating diverse tools, stakeholders, and workflows across the hardware development lifecycle. While individual design tools have evolved, the job of orchestrating these pieces – managing requirements, test data, manufacturing handoffs, and team collaboration – remains a major bottleneck.

As Western economies pour unprecedented capital into hardware innovation across aerospace, energy, and defense, an entirely new class of hardware companies is emerging. And they are building with fundamentally different expectations around tooling and development speed. The incumbent hardware solution stack fails to meet these heightened expectations – it is fragmented across systems, heavily manual, and lacks real-time visibility. 

As a result, we have seen many emerging hardware companies rolling their own solutions to solve internal and external orchestration across the development lifecycle. Stoke Space’s Fusion, an internal tool that they externalized, is one such effort. This trend, which we have seen inside of several other companies, signals both the severity of existing tooling gaps and validates demand for better solutions.

As such, we see a Stripe-like opportunity to enable and capture a portion of the value created by this new class of companies through the type of critical, but boring, infrastructure that we have deemed “plumbers for reindustrialization” in other research.

We see three primary areas of opportunity for new companies at the orchestration layer:

Test Data & Observability: The proliferation of sensors and testing equipment has created data noise that existing tools can't handle effectively. Real-time analysis of test data, coupled with AI for anomaly detection and optimization – DevOps-like telemetry and monitoring – could transform validation processes that historically relied on manual review and tribal knowledge.

Unified Data & Collaboration Hubs (Next-Gen PLM): The shift toward distributed engineering teams and expansive supply chains (e.g. via outsourcing) has exposed the limitations of current tools. Engineers spend a material amount of their time on non-value-added work like converting files, updating documents, or searching for the latest designs. Modern cloud-based hubs that unify product data (requirements, CAD, tests) could dramatically improve productivity.

Manufacturing Orchestration: The gap between design and manufacturing is a major bottleneck. Tools that automate the translation of designs into manufacturing instructions and provide real-time feedback on manufacturability could significantly reduce iteration cycles and costs.

New platforms built specifically for these emerging workflows – distributed by default, data-intensive by design, and automation-ready from the start – are naturally advantaged.

From a go-to-market perspective, focusing on emerging hardware companies helps orchestration companies avoid legacy processes and tooling and instead focus on shaping modern development workflows. These companies are building complex hardware under intense time (and capital) pressure – they need tools that can keep pace. As these tools prove their value to early adopters, they can expand both vertically (serving larger enterprises) and horizontally (connecting more parts of the development process). 

However, this means our thesis relies on this new class of hardware companies being a durable and scalable customer base. If the end game is dozens of sub-scale acquisitions and a select few successes – leaving today’s incumbent hardware OEMs as the primary market – the entrenchment of existing tooling and orchestration companies (from Siemens to Jama to PTC) will be harder to break.

Similar to what we have concluded in our research into AI-driven CAE, success doesn’t require displacing incumbent tools outright. Rather than competing head-on with entrenched CAD/CAE/PLM systems, new platforms can focus on making these tools work better together – becoming the connective tissue that coordinates modern hardware development. Once established as coordination layers, these platforms position themselves to expand their footprint over time.

The PLM and hardware development tooling market can already be measured in the tens of billions, but we believe the truly transformative companies will win by expanding the market and helping hardware companies iterate and build at the speed of software. This acceleration creates a powerful flywheel: faster development cycles enable more products, which drives increased tool usage and data generation, further improving development speed. Just as software development tools expanded their market by enabling faster iteration cycles, we believe the winners in hardware orchestration will grow the market by unlocking new levels of development velocity.

The risks are real – long sales cycles, integration complexity, and regulatory requirements in sectors like aerospace and defense. But we believe the confluence of market demand (driven by reindustrialization), technological convergence, and incumbent blindspots create a unique opportunity for new platforms to emerge.

Autodesk is a $65B company with 90% gross margins and earnings growth of 10%+ annually over the past decade. It is, in the views of many practitioners in the ecosystem, a monopoly in the worst sense of the word – extractive price increases paired with degrading product quality, closed and proprietary standards that lock in customers, and a lack of feature-level evolution to meet the needs of architects, engineers, designers, and contractors.

But Autodesk is just a symptom of a deeper problem. The entire AEC technology stack has evolved to reinforce silos rather than break them down. Each specialty has its own tools, workflows, and data formats, creating artificial barriers between disciplines that naturally need to collaborate. The result is an industry that remains extremely inefficient – construction productivity has historically grown at under 1% annually despite billions spent on software.

Perhaps counterintuitively because of the stranglehold Autodesk (and other deeply ingrained products) holds, our early hypothesis is that focusing on design is the perfect wedge to transform this massive industry. Every project's design phase naturally brings together architects, engineers, contractors, and owners – each making decisions that cascade through the entire project lifecycle. This, in turn, creates the possibility to develop network effects (the type enjoyed by Autodesk) once at scale.

The question, then, is what creates the wedge for companies to build distribution in the first place and kick off the network effects flywheel – something that has been a challenge for new entrants, as evidenced by the lack of massive venture-backed outcomes to date. We believe several converging technologies are coming together to massively reduce the costs of experimentation, lower the barriers to real-time design collaboration between parties (minimizing the cascading delays that begin at the design phase), and expand the creative canvas of design possibilities.

• WebGL and cloud-native architectures finally enable true browser-based 3D modeling at scale. Just as Figma used these technologies to make design collaborative, new platforms are rebuilding BIM from first principles for seamless multi-user collaboration

• Advances in physics-based simulation and generative AI allow instant validation of design decisions - emerging tools can compress structural engineering workflows from weeks to minutes and automatically optimize building systems for performance.

• New platforms are bridging design and construction by translating BIM models directly into fabrication instructions, creating the potential to significantly reduce MEP installation costs.

We see three approaches emerging to leverage these technologies and begin embedding them into multi-stakeholder workflows:

1. Next-gen cloud BIM platforms (e.g., Motif, Arcol): Browser-first collaborative design tools – i.e. "Figma for buildings". Here, we believe companies can build momentum through counter positioning – API-first models that prioritize open document and data standards.

2. AI-powered point solutions (e.g., Genia, Qbiq): Focused tools that dramatically accelerate specific workflows. Genia automates structural engineering analysis and optimization, while Qbiq uses AI to generate space planning options for real estate teams.

3. Design-to-fabrication platforms (e.g., Stratus): Bridging the gap between design and construction by automatically translating BIM models into fabrication-ready instructions. Stratus has shown particular success in MEP, where better coordination can significantly reduce installation costs.

The path to end-to-end orchestration will follow a clear sequence: Start by connecting architects and engineers through real-time design collaboration. Then extend to contractors, automatically translating designs into construction planning. As the platform becomes the system of record for design and planning decisions, it can naturally expand into procurement, payments, and project financing - using its unique data position to reduce risk and unlock better financial products. Eventually, these platforms could have a shot at orchestrating the entire building lifecycle - from initial concept through operations and maintenance.

Most importantly, these platforms will enable fundamental shifts in business models and incentives. Today's hourly billing and fixed-fee structures actually discourage efficiency - architects and engineers are paid for time and deliverables, not outcomes. But platforms that can measure and validate impact could enable new performance-based pricing models. Early adopters might start with simple metrics like design iteration speed or coordination time saved. Over time, as platforms gather more data across the building lifecycle, they could facilitate true outcome-based contracts where designers and engineers share in the value they create through better, faster, more efficient projects.

However, the trajectory of the geospatial intelligence market has largely been a story of massive perceived potential and disappointing results for builders, investors, and customers. While the use cases have been evident for decades, commercial value at scale has been slow to materialize and the net ROI of most earth observation companies has likely been negative. Adoption has been broad, but shallow – few commercial customers spend more than $1m per year on data and value-added services related to geospatial intelligence. Leaders on the upstream data collection part of the value chain (like Airbus and Maxar) still rely on government customers for a majority of their business while companies like Planet Labs still struggle to project commercial demand from quarter to quarter, indicating a lack of urgency to the data and analysis being offered. 

Solving the bottlenecks around geospatial intelligence that have kept deep commercial adoption out of reach – namely expensive data acquisition costs (for high fidelity data), fragmented data accessibility, and a lack of connectivity from data to core enterprise/industrial workflows has substantial implications for economic growth and human flourishing. The World Economic Forum projects that geospatial intelligence, as a platform technology, has the potential to drive more than $700 billion in economic value annually by 2030. A vast majority of this value will be created in critical physical industries – transforming land use, mitigating natural disasters, transforming how we build and maintain infrastructure, reducing greenhouse gasses, and addressing security and safety issues more proactively. 

Why is this interesting? 

We believe these bottlenecks are finally beginning to fall thanks to two converging factors – technological step-changes and the emergence of urgent buyers for the key technological building blocks that will make cheap, precise, and actionable geospatial data possible. 

• Launch costs have fallen 80-90%, enabling massive sensor deployment. While it took nearly 60 years to put 2,000 satellites in orbit, we launched 3,000 in 2023 alone

• Next-generation sensors are achieving unprecedented coverage and precision. Today's systems can detect not just the presence of objects but their composition and behavior from hundreds of kilometers away, at sub-meter resolution

• AI and compute advances have collapsed processing times and made it possible to for non-specialists to make sense of multi-modal data – what took human analysts years now often takes minutes

The demand side pull, while still not fully materialized, is equally as important and developing quickly:

• Insurance companies – and the entire insurance model – face existential pressure from climate-driven catastrophic losses (and regulatory intervention). Beyond risk assessment capabilities, improved, more transparent/accessible tooling can help to rebuild trust in this important segment of the financial system. 

• Autonomous systems (and with it shorter decision-making windows) are increasingly factoring into defense and intelligence operations, putting a premium on breaking down the current data silos to develop advantaged (precise and real-time) sensemaking capabilities.

• As we have observed through KoBold, the energy transition is creating entirely new customer segments (and forcing agility from large incumbents) focused on critical mineral discovery, methane detection, and other resource categories like water or forestry. 

• Infrastructure operators, utilities, and construction firms are scrambling to maintain the trillions of dollars of assets needed to reindustrialize, electrify, and – more critically – simply keep the modern way of life (e.g. clean water) running. Proposed initiatives like The Stargate Project create another major tailwind for the geospatial intelligence market. Above are just the handful of use cases we have been most exposed to through our investments and research. Like most great platform technologies, though, we believe many of the most valuable applications will be emergent. Thus, as we look at investments in the category, we are most interested in companies positioned to surf rather than compete with the dual blast radii of LLMs and Space Launch.

Which areas are most investible? 

Sensor Advantage / Infrastructure → While much of the sensor stack is being commoditized, competition at the powerful world model level (e.g. Niantic’s LGM) will drive demand for truly differentiated imaging and sensor suites. High precision, platform agnostic, high bandwidth, and real-time hyperspectral imaging stand out.

Data Fusion → As launch (and other sub-orbital geospatial sensor deployment) grows exponentially, data generation will scale along with it. If the status quo holds, silos and the need for bespoke solutions will only worsen. There is a Snowflake-scale opportunity to build data warehousing and piping for multi-modal geospatial data.

Geospatial Data as an Industry Transformation Wedge → Similar to Gecko in robotics, we believe the most valuable geospatial companies won’t be thought of as geospatial companies when all is said and done. Instead, we see major opportunities to use geospatial data as a wedge to build the workflows and intelligence engines that transform physical industries.

Energy demand is rising for the first time in over a decade thanks to rapid electrification, reshoring of manufacturing, and perhaps most notably, AI. This demand is being driven top-down via policymakers and bottom-up from the private sector. Regulations like the IRA and CHIPs Act have driven significant growth in new manufacturing construction. Meanwhile, energy constraints have overtaken GPU availability as the core bottleneck to scaling AI for companies like Microsoft, Google, Meta, and OpenAI. 

The willingness of big tech companies to spend whatever is necessary to access energy in the name of AI has led to amusing re-estimations of future data center energy demand every few months.

“[Our expectation of] 83GW is up from ~56GW from the prior September 2023 modeling. Overall McKinsey now forecasts US data center energy consumption in terawatt hours (TWh), rising to 606TWh in 2030, representing 12% of total US power demand. Critically, this is up from ~400TWh in the September 2023 modeling refresh. This is relative to 147TWh in 2023 and 4% of overall US power demand.”

Meeting this energy demand, whether in service climate objectives or geopolitical, energy, and technological sovereignty priorities, is of existential concern to economies around the world. As the saying goes, there is no such thing as an energy-poor rich country. Europe, in a trend that has continued since Russia invaded Ukraine, continues to struggle to meet the energy-related needs of its industrial champions. This has pushed them in droves to the US and other geographies, putting the continent’s competitiveness, productivity, and growth at risk. 

Energy abundance generally and response to data center demand specifically hinges on three important pillars: new power production, better transmission and distribution, and more efficient utilization.

As highlighted in other research, owning and operating physical assets can provide companies with a tremendous moat and allow them to capture more of the value they create. For this reason, companies focused on new power generation or physical infrastructure related to better transmission and distribution are interesting. However, such opportunities are often held back by factors like permitting that are outside their immediate control. 

Efficiency, on the other hand, is a problem best addressed by software and AI. This is particularly true for commercial and industrial buildings, which account for ~ 20% of final energy use (and rising thanks to the growth factors highlighted above). In some places, like Ireland, data center use alone promises to consume nearly one-third of grid capacity in the near future. As energy costs become a more substantial profitability factor and increased consumption puts pressure on sustainability objectives, better solutions for commercial and industrial energy efficiency represent one of the biggest opportunities of the next several decades.

Many of these operations have concrete optimization functions with goals and constraints. However, in many cases, the degrees of complexity of the world are too large for humans to grasp. Therefore, we fail to set up appropriate optimization functions and systems around them, leading to systems far from their global optima. That’s where we see massive opportunities for reinforcement learning. Advanced RL has enabled us to address areas previously unfeasible to optimize for due to their levels of complexity. 

Managing the energy usage of highly energy-intensive operations (e.g., data centers, cooling facilities, etc.) fits these criteria. RL models are capable of driving significant performance improvements autonomously, saving substantial energy and cost. The team behind Phaidra, one company that applies these models, was started by a team of Google employees who deployed these methodologies at Google data centers and saw up to 40% energy savings. They recently announced that they could drive energy savings of 16% at Pfizer’s data centers. Meta has published similar efforts. 

One of the key questions is whether there is enough quality data from sensors to support these plans and whether there is enough digitization of the physical world (and of its controls) for models to drive actions in the physical world. It is likely reasonable to assume that digitization has penetrated well enough for us to have a reasonable granularity and actionability, but the assumption is that the more data and the more actionability, the better. 

This field sits at the intersection of two areas that are core to our broader AI theses: 

1. Massive economic will happen in the physical world.

2. The most interesting AI use cases are in areas where AI helps us develop an understanding of the world where complexity is so high that we previously could not. In previous writing, we have referred to these as “expansion” use cases. 

Moreover, similar to KoBold, we expect that building a company in this space will require hiring world-class people across various fields: 1) AI/ML, 2) Software, and 3) Optimization of niche systems. We believe that for companies, being able to build a company that combines these three talent sources will build up substantial talent moats.

As LLMs can create software cheaply and agents become skilled at connecting user experiences in novel ways, companies are starting to push ideas around self-assembling/spontaneous software. We believe that, enabled by LLMs, a new paradigm could be on the horizon that increasingly merges the creation and consumption of software and makes a longstanding vision a reality.

We have previously written about this in the context of business software (see here), but we see an equally interesting opportunity in pro/consumer software and applications. It is important to stress that this is an incredibly nascent area with more questions than answers. 

The few questions we have are: 

1. Where does this happen? For these experiences to feel genuinely magical and software to feel spontaneous, LLMs must have the maximum context of a user's digital expertise, data, and usage patterns across applications. The most likely place for this to live is within a user’s operating system. Assuming operating systems are too slow to adopt, apps will likely emerge. However, it is unclear how long their staying power will be and how useful these will be if the tools and experience they create/enable are removed and not interconnected with default operating systems. In that case, the default place where these things live could be the interfaces of the large LLM providers. Claude has taken steps in that direction. 

2. How do these systems' privacy mechanisms work? As described above, they require a lot of context to feel magical. The question is how this context is handled privately. Some approaches mitigate risk, such as private cloud enclaves, but there could be a world where these kinds of applications only start taking off when models have 1) memory and 2) can run on consumer devices (e.g., phones and PCs).

3. What do monetization and business models look like here? It is unclear how much users will pay for custom software tools, especially if this requires work/creating tools. Only 30% of Android users customize their OS, and the current app paradigm has not trained people that they need to pay for utility-type services (this is the result of a combination of tools as a way to lock in as well as ad-supported models). In a world where apps become cheaper to produce and likely more abundant (due to the same dynamics discussed here), it is unclear whether most users will not just use apps that are increasingly available for niche use cases until software becomes completely self-assembling, assuming users every intent ahead of time. 

If we find good answers to these questions, we will be excited about this space and its potential.  

• These models will support multiple aspects of development: cognitive growth through personalized learning experiences, social-emotional development via interactive conversations, and creativity through imaginative play and storytelling. AI-enabled robot toys have significantly improved the social skills of children with behavioral development disorders (here).

• Toys can develop highly personalized interactions based on each child's unique patterns, pace, and interests, enabling children to lean into their curiosity. Moreover, by following their interests with relevant follow-up questions, they create an engaging cycle of discovery and learning.

• These interactive toys help children become comfortable with AI technology early, preparing them for a future where human-AI collaboration will be commonplace.

• Parents benefit from built-in progress tracking that monitors developmental milestones from a developmental psychology perspective. Moreover, parents can set up the models so that they incorporate their unique set of values and teach them to the children. Therefore, these toys can extend parents’ role model function. 

Naturally, we believe that data privacy and safety are crucial to these toys being beneficial and successful. Therefore, we believe them to have the following properties: 

• Ideally, proprietary small models that run on the device. 

• If not on the device, high guardrails on the outputs provided by large models. 

• Ideally, combined with models that understand children's social development and interactions, they tailor behavior to specific situations and needs (e.g., no constant speaking, ability to let children play by themselves if necessary). 

• Private processing of data (e.g., in secure, private cloud enclaves) with full access for parents to delete data. 

• Hardware form factors that combine low cost with maximum utility (tradeoff between mobility, ruggedness, and ease of use). Voice is likely the most important form factor and default mode of communication. A camera is likely ideal but challenging from a privacy perspective. 

• If possible, safely and privately, the embodied form factors and associated data provide interesting optionality to train novel models. 

We see an interesting opportunity for a commercial player to emerge here. Specifically, we see an opportunity to build an operating system that meets the standards above and enables owners of IP and distribution to build on top. In addition, we see significant opportunities to extend this platform in other areas, such as elderly care.

Over the last 20 years, "game tape" and statistical sophistication have revolutionized athletics (and other highly quantifiable professions like enterprise software sales) by enabling precise feedback loops, accountability, and recognition. AI is now driving a similar transformation of the broader professional universe where the relationship between inputs and outputs is often harder to grasp. Professionals have always valued mentorship and coaching. But access has historically been limited by cost and scale (hence “executive” rather than “employee” coaching. AI promises to democratize this type of performance enhancement (and an organization’s ability to measure it) in the same way that companies like Synthesis address Bloom's Two Sigma problem in education. 

Our hypothesis is that “constellations of (AI) experts” – deployed across every facet of professional development and organizational performance – will become as fundamental to career success as mentors and coaches are to elite athletes today. Several converging catalysts are making this possible. 

• The mass market deployment of co-pilots and proto-agents has rapidly normalized AI-human collaboration. More than 60% of physicians now use LLMs to check drug interactions and support diagnosis – early evidence of adoption for high-leverage decision support. 47% of GenZ employees say ChatGPT gives better career advice than their boss – signaling dissatisfaction among young workers with the status quo.

• The proliferation of audio/video capture in professional settings generates rich data to help these systems better understand and improve performance. People increasingly operate with the assumption that every call is recorded, while younger employees regularly go viral for sharing layoff videos online. 

• The economics of AI are reshaping both organizational structures and individual incentives. Companies are shifting from fixed to variable cost models, flexing labor (human and agent) up and down based on demand.  This, in turn, is shifting how workers are measured and compensated. As a result, professionals must proactively adapt to succeed in this new paradigm where human judgment and AI capabilities become increasingly intertwined.

We see several areas where the “constellations of AI experts” will be professionally valuable. In each of these categories, we expect the most successful platforms will combine automated interactions, human experts in the loop, and content/validation that come together to create holistic systems of improvement. 

• Organization-wide solutions that integrate deeply with company context to provide AI-powered coaching and feedback loops. While employees have shown a willingness to trade privacy for better tools, trust and security guardrails are essential. 

• Individual-focused platforms that grow with professionals throughout their careers, combining performance enhancement with credential creation in an increasingly fluid labor market. 

• Solutions for high-turnover industries that capture and distribute best practices to improve training efficiency and retention (e.g. frontline audio-first interfaces)

• SMB owner enablement systems in areas like the skilled trades and family medicine, to make it possible to i) capture and transmit tacit knowledge (streamlining hiring/training while increasing terminal value) and ii) help operators compete without needing to bring in expensive consultants or PE expertise. 

These are, to be clear, highly divergent use cases that necessitate different product philosophies, business models, and competencies from the companies building solutions. However, they share important characteristics, namely that they all represent opportunities to use AI and better data to make professional tacit knowledge, action, and context visible and measurable, unlocking precise intervention to help individuals (and by extension teams and companies) grow into their potential. 

For companies deploying AI in legacy industries, a key question is whether to enable incumbents by selling them solutions or compete with them by taking a more full-stack approach. The trade-offs between these two models is something we have started to explore through our AI-enabled services analysis and this piece on when to compete with and when to sell to incumbents in an industry.

Recently, several firms have shared public theses on the opportunity for emerging AI companies (or vertical market software companies) to capture additional value in a given value chain by fully integrating via the acquisition of assets – as opposed to selling solutions to incumbents or taking a more organic (build instead of buy) approach to going full stack.

Slow, which helped finance the $1.6b acquisition of parking operator SP+ by software company Metropolis, calls the model “Growth Buyouts”. Equal Ventures, which recently opined on the opportunity for such a model in insurance calls it “tech-enabled consolidation”. Vertical market software investor Tidemark calls the approach “tech-enabled vertical roll-ups”. Re:Build Manufacturing calls its technology-driven manufacturing roll-up model an “American Kieretsu”. 

Our current hypothesis is that while the AI-enabled acquisition of services businesses (with venture dollars) may not be wise, there is a significant opportunity for emerging AI, software, and robotics companies to capture more value and develop value chain control by acquiring legacy assets in physical industries. 

For decades, venture capitalists have been involved in what seems like Sysyphosian tasks: digitizing businesses that operate assets in the real world. For many reasons, from software shortcomings to incentive problems and structural challenges in GTM to a lack of skilled labor on the consumer side. We see a trend for novel ML models to solve the first of these problems by being able to operate assets end-to-end without much human input. Yet the latter challenges remain. Therefore, AI-native companies addressing these problems are prone to leave value on the table and, due to slow adoption, likely are slower at training and developing their models, preceding a lot of additional value. Therefore, AI-enabled asset ownership represents one path to achieve this. 

Sequence matters for companies that go down this path. Companies should prove they can build technology and deliver ROI (for early customers or via a smaller-scale organic full-stack approach) before embarking on buying distribution via M&A. The only cases where early M&A can be attractive are cases where smaller targets that are structurally very similar to large-scale targets in the market can be acquired for amounts smaller than traditional GTM. Initially, these businesses have venture risk profiles, and only after the second or third large acquisition should they be derisked and become predictable/repeatable enough for investors with a lower cost of capital – Infra, PE, etc. – to consider participating. By reaching this point, venture investors will have seen highly attractive returns. 

Initial Hypothesis on Key Conditions

Below is an initial hypothesis for when it makes sense for a company to vertically integrate via acquisition as opposed to doing so organically or remaining a software/technology vendor to a given industry:

• The company must have a demonstrated “production advantage,”; i.e., a clear technological or product edge that creates compounding value in an industry. Companies leveraging exclusively off-the-shelf technology likely lack the differentiation to deliver venture-scale outcomes even with strong operational execution and financial engineering. If a PE fund working with Accenture can go after an opportunity or if human labor is cheaper on an efficiency-adjusted basis it is unlikely to be a VC case. If solving the problems requires a combination of world-class technologists AND operators, this becomes an interesting opportunity for venture-style risks and outcomes. 
  

• Customers have proven structurally unable to adopt and deploy a company’s solution to its most productive extent. Alternatively, they seem unwilling to pay for the full value of it. This can be due to various reasons, from lack of scale to incentives leading to stasis in a given market (“if my competitor doesn’t innovate, I don’t have to”). We should be able to pinpoint a structural issue – and generally point to evidence from a company’s own experience –  with a given market/customer base to be certain the ineffectiveness is not a product issue. 
  

• Building on the previous criteria, companies spearheading the buy-out strategy should be building technology that changes the core way an asset is operated, transforming the economics of the business/industry. Most likely that is where existing operators are (somewhat paradoxly) least incentivized to adopt technological disruption. This is what makes the Metropolis acquisition of SP+ such a compelling application of this approach. SP+ has 3,000+ parking locations around the world where the core customer experience (paying for parking) can be largely automated. While the “work around the work” (maintenance, security, etc.) still requires people, the ROI around the primary transaction is much easier to understand than situations where the AI solution is helping people deliver the primary solution more efficiently (e.g. home services models, legal services, etc.). 
  

• Likely, there is a sweet spot around the level of complexity that goes into operating an asset that makes it a fit for AI-enabled acquisition. Complexity can either stem from the core value proposition being complex, several offerings being performed at the same asset leading to compounded complexity or the “work around the work” being significant (e.g., for regulatory reasons). Too little complexity at the core value proposition becomes a PE case; too much and the operational overhead reduces the leverage associated with improving the margins of the core asset. Ideally, the complexity/problems across holdings within the same space should be the same (e.g., parking lots), and skills easily transferable. We should be able to pinpoint these levels of complexity and identify assets/problems where they meet the sweet spot. 
  

• The category a company is operating needs to have acquisition targets that are operating at scale (ideally businesses worth USD 1B+ with additional value creation in the several hundred million – further analysis on this needed). Buying assets operating at scale that can be fully optimized and automated via AI is substantially more attractive than rolling up locally-delivered services businesses. Again, this is what makes the SP+ acquisition so attractive, SP+ has 3,000+ parking locations around the world that likely are all run very similarly. Ideally, solutions deliver not only cost savings but also growth opportunities. We are also interested in companies with a view on how the integration of software and legacy assets will unlock increasing ecosystem control and turn the business into an industry operating system. 

• Companies must have advantaged access to talent across functions. It is rare for a founder or founding team to understand “what great looks like” in areas where they have not had direct experience. A team of software engineers is likely unfamiliar with what makes a great industrial CFO or service-business COO. As a result, we may expect the pool of founders well-equipped to build such a business to be small. We have seen this play out at companies like KoBold Metals, which combine highly scientific founding teams with business acumen. 

These criteria still don’t fully answer why/when it is better to grow a full stack solution via acquisition rather than a more organic approach. One primary reason a company would choose to grow via acquisition is if the geographic footprint and surrounding “infrastructure” of an industry will look largely similar in the future as it does today. In such cases, the speed of distribution created by acquisition is enough of an advantage to overcome the accompanying cultural and organizational complexity that could be mitigated with a more organic strategy.

To use the Metropolis example, should we expect the footprint of the parking industry to be more or less the same in 10 years as it is today? While autonomous vehicles may make some impact on the margin during that time period, the inertia of the built environment probably means we should expect the flow of traffic and parking to remain relatively the same (airports, stadiums, commercial centers, etc). 

A counter-example is H2 Green Steel, which has raised multi-$B to produce steel with 95% lower emissions than traditional steelmaking. Because of the fact that the company’s steel production depended on access to ample clean energy, the company couldn’t just acquire and transform underperforming steel facilities despite the similarity in equipment, permitting, and human capital needs. Thus, to transform the industry around their vision, the company was forced to build a more organic approach. 

Companies also might pursue a buy instead of build strategy when the technology can be easily integrated with existing assets and infrastructure, substantially reducing time to value for a given solution. 

There are likely several other criteria in support of (and against) the strategy of vertically integrating via acquisition which need to be explored in further depth. 

While several initiatives (e.g., NREL, IEA) by various government agencies and NGOs to address the abovementioned challenges have been underway, none have impacted the market where easy open data access has been achieved. 

Therefore, we see a unique opportunity in a dedicated effort to aggregate the various data sources and make them available in a standardized format via API and application. The availability of such data can be the underpinning for a wide array of new applications and use cases that require such data (e.g., applying Reinforcement Learning-based optimization to the grid) and can be substantially improved if such data exists. In short, we see an exciting opportunity for the company that can aggregate and maintain the highest quality grid data to be the nexus of an emerging ecosystem.

We see an exciting opportunity for AI-enabled marketplaces to emerge. While there are a lot of opportunities for AI to enhance marketplaces (see good NfX write-up here), we are especially interested in situations where AI-enabled processes are in a reinforcing interplay with data advantages that lead to a sustained higher value proposition (i.e., better matching) in the marketplace (see graph below).

As outlined above, there are two interconnected feedback loops at play: 

1. Using LLMs and VLMs to collect the right proprietary data at scale (i.e., conduct interviews, ingest large documents, understand client requirements using document upload, etc.).

2. Use fine-tuned LLMs/VLMs + other ML models to understand demand and supply better, identify actions that reduce uncertainty around matching probability (e.g., follow-up questions) and to carry these actions in service of enabling more cost-effective/higher-value matching.  

We expect that businesses creating sustained value in this space to meet the following criteria:

1. LLMs, VLMs, and other models can perform tasks to an acceptable degree (i.e., they meet a bare minimum threshold) – both on the proprietary data collection and matching side.

2. Large amounts of unstructured data and feedback loops are useful for fine-tuning models that directly unlock economic value.

3. Nobody has collected data relevant for training/finetuning these models at scale as there has been no economic/technological incentive to do so.   

4. There are ways to create highly frictionless form factors using 1.)  that allow users to interact with these platforms seamlessly and in highly personalized ways to collect large amounts of data. 

5. Initial data and model advantages can be sustained and turned into lasting moats with little risk of second movers and other market participants (e.g., incumbents with large distribution) being able to catch up. 

We see opportunities in various areas, from HR to Traveling to healthcare provider (e.g., psychologist) matching. Especially in scenarios where a lack of information leads to low matching rates. A few cases:

Case Study 1: Staffing

Staffing is historically incredibly time-consuming, requiring a deep understanding of a candidate’s capabilities and the job requirement assessment. This is very hard to scale as quality assessment usually requires 1) reviewing materials, 2) conducting interviews to dig deeper + review these, and 3) feedback cycles to understand what type of candidates the demand side actually wants (stated vs. revealed preferences). This leads to many staffing marketplaces doing a bad job of vetting demand or being very expensive, with matching rates reflecting this. 

Let’s go through the criteria set up above to see whether a hiring marketplace is a good fit to become intelligent:

1. LLMs can already review and synthesize vast amounts of unstructured data (e.g., CVs, websites). They are capable of doing the same with job requirements. They are also capable of performing job interviews to a somewhat satisfactory level. 

2. Models and AI interviews can be finetuned based on desirable outcomes (e.g., matching of demand and supply), thereby adjusting their reviewing and interview capabilities. This can happen even in a customized way (e.g., custom), given that certain parties on the demand side are large enough to guarantee a certain “offtake.” Mercor wrote this on their blog:

3. This part is not so clear in the context of staffing. For one, there are a plethora of existing and new AI-enabled hiring tools that use AI-supported video (e.g., HireVue), and existing staffing platforms (e.g., Upworks) are rolling out video interviews, too. It is unclear to what extent these platforms might or might not have large amounts of a combination of unstructured data with hiring matches that they can use to train better models. Also, by sheer scale and distribution, these platforms should be able to generate plenty of data easily. 

4. In the segments of the economy where jobs are sought after, people are eager for the opportunity to be in the talent pool that is considered for specific jobs. In these cases, people are willing to share their data CVs and conduct AI interviews – especially if the process is smooth. Given that the demand side (aka the companies looking to hire from the talent pool) is reasonably attractive, the CAC associated with acquiring the supply and data (i.e., video interviews, CVs, etc.) should be fairly low. 
   
   As described above, while we don’t assume AI-based matchmaking to be perfect yet, we believe that AI can be used to support already increasingly efficient matching, enabling the development of a cash-flow-generating business model while data is collected and models improve.

5. Given the dynamics described under 3, it is unclear whether an HR marketplace with an initial data advantage can sustain this advantage. What if existing platforms like Upwork roll out AI-based video interviews and start training their models? With their existing brand and supply, they should be able to generate more data than any startup substantially faster, leading to better models, etc. If not, what is a relevant quantity of data to establish a platform as the winner? Will general LLMs acquire the capabilities of finetuned models as they get better and context windows improve?

In recent years, change has occurred at all levels – from the EIB’s updated views on defense technology and the European Commission’s first-ever Defence Industrial Strategy to the rise of an actual defense technology ecosystem in Europe for the first time, catalyzed by the momentum of lighthouse companies like Helsing, ambitious efforts like the €1b NATO Innovation Fund, and grassroots organizations like the European Defense Investor Network.

But expanding markets, increased capital flows, and narrative momentum don’t always imply attractive forward-looking returns. 

Despite the market’s growth, inertia, fragmentation, and protectionism rule the European defense market. While European defense spending has returned to Cold War levels, the continent still lacks urgency relative to geopolitical allies and rivals. The conflict in Ukraine has done little to unite European perspectives on the what, how, and who of building next-generation defense capabilities. The EU’s two largest economic and military powers – Germany and France – remain fundamentally split on the role of Europe in its own defense. This philosophical gap threatens to worsen the severe physical fragmentation of European defense forces – Europe operates 5x the number of vehicle platforms than the US. At the same time, the UK has increasingly shifted attention away from the continent towards the AUKUS coalition. 

The US defense technology ecosystem, far more developed than Europe, inspires little confidence in what lies ahead. Through September of 2023, venture-backed companies were awarded less than 1% of the $411 billion in Defense Department contracts awarded in the government’s fiscal year – only a slightly larger share than in 2010 when few startups were building military technology. And while companies like Anduril have shown that the path to scale is possible, the company’s success may end up making it the new technology distribution chokepoint instead of a bellwether for a thriving new defense ecosystem. 

These factors present significant obstacles to building and scaling European defense technology companies. They may also present unique opportunities for a highly targeted investment approach in the space, aimed at turning the market’s weaknesses (e.g. fragmentation) into strengths and riding key catalysts that may help emerging companies overcome the inertia and sub-optimal industry structure.

Catalysts Creating Opportunity

To believe in the opportunity to invest in emerging European defense technology companies in the face of the incumbent market structure, we need to see significant technological, economic, social, and policy-related shifts that are, critically, available to emerging entrants and not incumbents.

Europe has been forced to prioritize resilience and strategic autonomy for the first time in a century. This has restructured the continent's procurement regimes, capital markets, and attitudes. The simple availability of early-stage capital for emerging defense and security companies in Europe cannot be overstated. With dozens of early-stage funds now focused exclusively or significantly on the space and later-stage investors slowly showing up, financing a defense technology company in Europe to the point of scale is now possible. As the EIB and other capital markets institutions continue to evolve their view, we expect many of the capital markets barriers to financing defense and security companies across the company life cycle will begin to fall away.  

Procurement remains a significant challenge, but the fragmentation of Europe creates opportunities for emerging companies to get to market faster by targeting smaller, potentially more agile countries more inclined to adopt new solutions. Greece, for example, now spends 4% of GDP to support a tech-forward defense strategy while countries close to the front in Ukraine have been forced to move quickly to adopt new solutions. 

The “primitives” for rapid, capital-efficient defense technology development have come online, making it possible for companies to ride multiple technological tailwinds to build solutions that meet the complex needs of government customers. Decreasing costs of hardware, enabled by advanced manufacturing, better (software) tooling, and the acceleration of physical-world foundation models make it possible for companies to develop complex defense technology systems at a fraction of the cost of incumbents. AI systems are already operating successfully in significant tests (like dogfighting with fighter pilots) and on the battlefield in Ukraine, which should drive more receptiveness from (risk-averse) buyers and users. 

Lighthouse companies and talent ecosystems are emerging around defense and national security-focused technology for the first time in Europe. The US defense technology ecosystem built momentum on the back of breakthrough companies like SpaceX and Palantir. The same pattern is playing out in Europe, with companies like Helsing and The Exploration Company forming the foundation for a thriving defense-industrial ecosystem in Munich. While less developed in terms of defense and space-focused technology, places like Stockholm (energy) nd Paris (AI) have become beacons for talent in areas adjacent to national security. Meanwhile, London has captured much of the early-stage energy and likely represents a strong ecosystem to launch a defense technology company thanks to its physical proximity to Europe and cultural proximity to the US.  

The Ukraine conflict has presented a unique opportunity for companies to develop proof points and revenue, creating a “backdoor” for future contracts with Western European governments. It has also highlighted the future of warfare. Rapid acquisition and deployment processes in Ukraine have helped companies generate real revenue and test systems in live situations. While larger Western European governments have been slower to respond, and more likely to simply drive business to existing primes, the proof points being developed by emerging companies should help their cases in (eventually) accessing larger, longer-term programs. Technologically, the predominance of electronic warfare has given a fundamental advantage to agile companies that can iterate rapidly to stay ahead of Russian competition. 

Key Insights

The following factors are the most significant in driving success for emerging European defense technology companies. These lessons are drawn from the company and our customer/expert interviews.

New defense companies are software-first, R&D-centric, and mission-driven. Incumbent defense contractors operate on a cost-plus business model, essentially building to the specifications laid out by government buyers and layering a “reasonable” margin on top (5% - 15%). As a result, large primes spend less than 3% of their budget on R&D and lack the incentive to innovate. On the other hand, companies like Anduril and Shield AI take on product development risk themselves and spend massively on R&D. 

And while the hardware these companies build tends to garner the most attention, the software and autonomy systems underlying the hardware make everything work. Anduril’s Lattice platform ties together all of the company’s hardware products, fusing sensor data and creating an autonomous operating picture. This software-defined operating model drives better margin structures (Anduril targets 40% gross margin vs. Lockheed and other primes under 20%), allowing them to continue fueling an R&D flywheel.

Fragmentation remains the most challenging aspect of European defense. It may also present the largest opportunity. Europe’s fragmentation challenge needs little additional explanation. There is not one unified military-industrial complex on the continent, there are 27. Each has a different view on supporting its own national champions, different relationships with EU member countries, and divergent views on buying from outside (usually the US). This has resulted in a complex web of disparate capabilities (weapons systems, vehicle platforms, and communication models) that limit rapid response and collaboration.

Understanding this, and realizing that it is likely beyond the reach of one company to solve from a hardware (let alone cultural) perspective, is key to uncovering where opportunities sit. Helsing, for example, has leveraged its positioning as a multi-domain AI backbone to build early leadership around this concept. As cheap drones, cameras, and other sensors proliferate, the opportunities to coordinate the complex data and operational picture, solving capability and collaboration gaps through more modularity and interoperability become larger.

Technology differentiation is table stakes. The most successful companies will possess a “secret” to navigating procurement. Despite the shroud of complexity surrounding defense procurement, success remains largely “champion-driven”, as Anduril CRO Matt Steckman recently remarked. Companies don’t win through better technology, they win by solving specific problems for people with influence in the buying process. Companies must simultaneously engage in long-term relationship building (including lobbying) to build trust with procurement influencers while developing relevant proof points in the field. One way of doing this, as Anduril demonstrated and emerging European players like Lambda Automata are attempting to replicate, is by viewing defense and security as a “conglomeration of micro markets” – which includes adjacent opportunity areas like public safety and border control.

Narrative momentum is highly rated but likely remains underrated by European founders. The traditional stereotypes of European vs. American founders seem to have played out in the early part of this new defense tech wave – from Anduril’s narrative mastery to the momentum of ecosystems like El Segundo to the sophisticated way some companies and investors have integrated themselves into the Washington decision-making systems. As in all markets, there is a reflexive nature to success in defense – the companies that win figure out how to craft a better story and more social proof to attract capital and talent in advance of fundamental traction. 

Distribution bottlenecks inherent in government and defense contracting are already contributing to market consolidation for emerging defense technology companies. Competing against defense primes means eventually competing in every domain they operate in. As software-first companies break through, the returns to scale and breadth might become even greater – platforms like Anduril’s Lattice get stronger as they consume more data and control more hardware assets in the field. Combined with the defense market’s natural bent towards consolidation, companies that can be “first to distribution” in a given area will be very hard to displace and will be strongly positioned to roll up interesting technology and talent, as Anduril has already started to do aggressively. (The sheer number of Anduril references in this document reflects its outsize and rapidly compounding success in this space!)

Emerging Investment Areas

There are several valuable defense market maps and landscapes worth evaluating to understand different ways of breaking up the market, perhaps the most comprehensive is this from Quiet Capital’s Michael Bloch: National Security and Defense Market.

To avoid rehashing those efforts, our focus has been on identifying emerging themes that span multiple segments of such maps, supported by converging market, technology, and geopolitical tailwinds. While not comprehensive, these themes align well with the catalysts and insights above and are where we have seen several of the most interesting companies in our review – the most interesting companies tend to touch multiple themes. 

Modularity and Interoperability → Leaning into the fragmented nature of European defense through solutions that aim to unite disparate operating systems and coordinate complex environments. While software capabilities will be the core connective tissue, hardware plays a big role as well. Cheaper, smaller, interoperable systems built to be easily adopted (both budget and technology-wise) can help accelerate initial deployment and provide companies with a platform from which to expand. 

Rapid Response → Building a more dynamic defense-industrial base by shortening time and cost to intervention across domains and operating areas. This ranges from faster kinetic capabilities (e.g. hypersonics and electronic warfare) to rapid manufacturing capabilities (e.g. Replicator) to faster deployment of machines and people (e.g. counter UAS swarms, labor coordination platforms) to systems that can be deployed (and as importantly, replaced) quickly. 

Multimodal Physical World Data and Intelligence → Wayve’s recent autonomous driving demonstrations showcased the speed at which multi-modal models are making their way into the physical domain. Along with the rapid decline of hardware costs, models that can reason more dynamically create interesting opportunities in defense, where operating environments are extremely fluid (i.e. not repetitive like pick and place, etc.) and thus pose problems for more rigid AI systems. Better simulation data will also continue to play an important role in preparing autonomous systems for live action. This represents a more horizontal theme and is thus something we might pursue a deeper dive into beyond defense.

Software for Hardware → The declining cost of hardware also creates room for better tooling, both at a collaboration/workflow level (i.e. Atlassian and GitHub for hardware builders) and at a design level (i.e. better CAD/EAD, “Figma for chips”, etc.). Fusion, a software platform developed and externalized by space launch company Stoke, highlights the need for better tooling to serve the hardware revolution. Enhanced IP and data security along with high levels of required precision for certain use cases may create specific opportunities in defense.  

Maritime Production, Security, and Infrastructure → Control over maritime infrastructure represents a significant geopolitical and economic advantage. Over the past decade, China has invested heavily in shipbuilding capacity. Today, a single shipyard in China has more production capacity than the entire US shipbuilding industry. However, the importance of maritime control goes beyond just shipbuilding. Undersea cables, for example, are the backbone of the global financial and communications systems. – over 95% of the world's communications are carried by a network of roughly 500 cables laid across the oceans. These represent critical vulnerabilities that need to be proactively protected through better surveillance, kinetic deterrence, and cybersecurity technologies. 

Combatting Digital Authoritarianism → Control of the digital economy is highly centralized, with cheaper data processing and engagement-centric business models (i.e. advertising) feeding the strength of a small number of powerful companies and institutions. This has led to democratic deterioration and a loss of trust in key institutions. It also creates a more straightforward surface area for attack and manipulation by adversaries – spanning consumer-focused influence campaigns to corporate IP theft. Technology that empowers sovereignty over assets and information, increases privacy, and enhances secure communication and collaboration represents a somewhat orthogonal, and bottom-up, approach to investing in defense and security as the go-to-market model may be dependent on large-scale government procurement.