AI just snuck into our team photo. Generated by ChatGPT 1.5. It’s not quite there yet, but 2026 could change that.
But one thing AI will not change: you are only as good as the people you surround yourself with (right, Einstein?). And we could not agree more.
To our LPs, GPs, and founders, thank you. As we close out the year with 60+ GP relationships and 120+ funds, we are reminded that these are not just numbers. They are partnerships built on trust, and we feel privileged for it.
Wishing you all a wonderful holiday season!
2025 Set the Limits. 2026 Will Test Them.
Why Infrastructure and Industrial Execution Now Define Venture Outcomes
AI just snuck into our team photo. Generated by ChatGPT 1.5. It’s not quite there yet, but 2026 could change that.
But one thing AI will not change: you are only as good as the people you surround yourself with (right, Einstein?). And we could not agree more.
To our LPs, GPs, and founders, thank you. As we close out the year with 60+ GP relationships and 120+ funds, we are reminded that these are not just numbers. They are partnerships built on trust, and we feel privileged for it.
Wishing you all a wonderful holiday season!
2025 Set the Limits. 2026 Will Test Them.
Why Infrastructure and Industrial Execution Now Define Venture Outcomes
By Reference Capital
Dear Partners & Friends,
This December edition closes a year that, in hindsight, followed a coherent arc driven less by theme than by constraint. We began with power grids and data centers, as the AI boom exposed energy and compute as hard limits. Attention then expanded to sensors and factories, where the physical systems feeding and executing intelligence emerged as the next bottleneck. Capital markets and geopolitics entered the frame as forces shaping who can build, scale, and control these systems. By year’s end, space infrastructure and defense stood out as areas where constraint, sovereignty, and physical reality are inseparable. Seen this way, 2025 marked the moment venture capital began to question whether software alone could move the world, and increasingly aligned capital with technologies capable of reshaping entire economies.
In that context, software did not lose relevance. It moved down the stack, becoming the coordination layer governing energy systems, machines, supply chains, and security infrastructure. Our research tracked this transition not by category, but by observing where technical limits translated into strategic and investment decisions.
2025 in Review: Themes That Compounded
If there was one pattern that repeated all year, it was this: the next wave of technology progress is being shaped less by software innovation and more by physical constraints. Power, sensors, factories, supply chains, security, these are not side topics anymore. They are the enabling layer. And the year’s work kept circling back to the same conclusion: whoever controls infrastructure controls the pace.
Infrastructure became the binding constraint
We started the year by looking at electricity grids. That work made something very concrete: the core limit beneath the AI boom is not energy in the abstract, it’s the infrastructure that delivers it. Transmission congestion, distribution bottlenecks, and slow permitting are now first-order constraints on where and how quickly compute can scale.
What sits underneath those issues is coordination. The grid was built for centralized, one-way power flow. It is not designed for distributed generation, volatile demand, and compute loads that increasingly look like industrial plants. Generation capacity still matters, but without the ability to move and deliver power, it often stays stranded.
That dynamic pushed investment in two directions at once. On one side: reinforcing centralized transmission and distribution. On the other: accelerating local energy production and storage that can bypass bottlenecks entirely. It also reshaped the strategic frame: energy systems built closer to demand tend to be more resilient, less exposed to single points of failure, and less dependent on foreign suppliers. In practice, control over energy infrastructure is becoming a prerequisite for AI leadership. Nations that solve this will set the pace; those that don’t will be constrained not by talent or capital, but by electrons.
From power to sensing: making the physical world programmable
Once you accept that AI is constrained by physical systems, the next question becomes: what does it actually run on, beyond compute? After large language models effectively consumed the world’s supply of digital text, attention naturally shifted toward the physical world, the next frontier for data.
That is where our work on decentralized sensor networks fit. Across fleets, satellites, industrial arrays, and community-powered networks, the ability to generate real-time, verifiable physical data moved from a technical consideration to a strategic control point. This is the beginning of physical observability at scale: making the real world legible to machines.
And as AI systems increasingly act in the world, the value migrates. It accrues less to those who merely store data, and more to those who generate proprietary data and control the layer where work happens. In that world, the database becomes plumbing. The advantage belongs to platforms that own both sensing and execution.
Execution became the next frontier: industrial production returns
Once data can be captured and interpreted, execution becomes the next bottleneck. Mid-year research brought us back to production itself. Advanced manufacturing and robotics are finally crossing from experimentation into scaled deployment.
The recurring failure mode is the same: many companies can prototype; far fewer can scale. The winners are increasingly those who control the full system. That was visible in early proof points across aerospace, medtech, defense, and construction. The deeper insight, though, was that systems matter more than machines. Competitive advantage now sits in vertically integrated workflows, software, sensors, materials, and production tightly linked in closed loops.
At the frontier, AI-driven materials discovery and generative design are compressing the time and capital needed to move from concept to qualified parts. That shift matters because it weakens what used to be durable moats: massive factories and balance-sheet CapEx. New entrants can challenge incumbents by moving faster through design-to-production cycles, not by outspending them on physical footprint.
It’s also worth being precise about what automation changes, and what it does not. Advanced manufacturing will reduce dependence on skilled production labor over time, but the bottleneck shifts rather than disappears. Engineers, technicians, and systems integrators who can design, deploy, and maintain these systems remain scarce, and no amount of robotic capacity substitutes for the people who make it work.
Space followed the same logic. It is no longer defined by exploration alone; it is becoming infrastructure: launch, communications, in-orbit logistics, and eventually off-planet production. Here too, whoever controls infrastructure controls the pace of innovation.
By autumn, the threads converged: security as infrastructure
As the year progressed, geopolitics turned what could have been separate conversations, energy independence, industrial capacity, resilient supply chains, into one integrated picture. Space followed the same logic: no longer defined by exploration alone, it has become a race to build and control critical infrastructure, from launch and communications to in-orbit systems that increasingly underpin economic and military power.
Defense technology, once largely avoided by venture capital, has become a frontier shaped by autonomy, advanced materials, and AI-enabled command systems. Procurement reform and the rise of dual-use startups are pulling private capital directly into national security ecosystems.
By then, the conclusion was hard to avoid: security is infrastructure, and infrastructure is investable again.
Capital and liquidity: the market stayed split
Underneath the industrial and geopolitical shifts, the financial aftereffects of the 2021–2022 cycle continued to play out unevenly. Many companies that failed to grow into their valuations are now stuck in a long middle period, walking dead outcomes, or as we’ve called them, zombicorns. Liquidity remains constrained for that cohort, pushing outcomes toward restructurings, secondaries, and extended holding periods rather than clean resolutions.
At the same time, this isn’t a uniform reset. In areas like AI, and select industrial and defense segments, capital concentration has accelerated. The result is not a broad return to discipline, but a bifurcated market where scarcity and excess coexist, leading to very different paths to scale, pricing, and liquidity depending on where you sit.
That environment shaped our work on institutional venture capital later in the year. In a market this split, the difference between durable firms and transient ones comes down to governance, process, and structural rigor, capabilities that matter across cycles, not just in one valuation regime.
China: field insights from a parallel industrial system
Many of these dynamics are already clearer in China’s innovation ecosystem, where energy, manufacturing, capital, and policy are more tightly coordinated than loosely coupled.
Conversations in Shanghai and Beijing materially reshaped our understanding. What stood out was not imitation, but coordination at scale. China’s dominance across the electric stack, batteries, power electronics, EVs, and critical materials, now looks structural. Manufacturing scale compounds across sectors, supported by aligned industrial policy and capital allocation.
For allocators, an equally important shift is that capital markets have localized. The majority of new funds are RMB-denominated, with provincial governments acting as anchor LPs. USD participation has consolidated sharply, changing how foreign capital interfaces with the world’s second-largest economy.
The takeaway for global investors isn’t a simple yes/no on exposure. It’s recognizing that China represents a parallel industrial system that shapes pricing, supply chains, and competitive dynamics across energy, AI hardware, robotics, and biotech, regardless of where capital is ultimately deployed.
Looking ahead to 2026: what we’re watching
Against this backdrop, the transition underway is unlikely to reverse in 2026. If anything, it should accelerate. Here are the dynamics we think will matter most:
- Capacity becomes the filter. The gap between teams that can prototype and teams that can build at scale will define winners. Capital increasingly flows to companies that control manufacturing, supply chains, or workforce pipelines, not just IP.
- Model commoditization accelerates. Foundation models consolidate into a handful of utilities. Defensibility shifts to proprietary data, embedded workflows, and regulatory moats. Pure model plays become difficult to underwrite except at the frontier.
- AI moves down the stack. Compute efficiency, inference optimization, specialized chips, and energy integration will dominate the next phase. Model races give way to infrastructure economics. The shift is especially clear with agent-native infrastructure: enterprise backends were built for humans clicking through workflows, predictable, low-volume, one request at a time. Autonomous agents behave differently. A single task can spawn thousands of sub-queries in milliseconds, overwhelming systems designed for a gentler pace. The winners will be architected for this from day one: coordination, state management, and massive parallel execution as default assumptions.
- Security and trust become non-negotiable. Agents acting across systems at machine speed create new attack surfaces, credential sprawl, prompt injection, unaudited tool calls. The companies that enable agentic deployment at scale will be the ones solving authentication, permissioning, and auditability as core architecture, not bolt-on features.
- Systems of record lose primacy. For decades, enterprise software derived power from being the place where data lived. AI changes the equation. When models can read, reason, and act across operational data directly, the database becomes plumbing. Value shifts to whoever controls the layer where work actually happens. Incumbents built moats around records; the new advantage belongs to platforms that own execution.
- Skilled labor stays the binding constraint. The manufacturing job shortfall doesn’t resolve in 2026. This sustains tailwinds for automation, workforce training, and companies that can operate with leaner teams.
- Industrial and defense tech remain structurally advantaged. Public budgets and strategic procurement anchor demand, though long sales cycles and execution complexity require patient, technically fluent capital.
- Energy stays the quiet bottleneck. Grid reinforcement, long-duration storage, and flexible demand underpin everything from AI to electrification.
Even if macro conditions improve, capital is likely to remain selective, rewarding teams with strong execution and clear unit economics.
Closing thoughts
As 2025 ends, the question is no longer whether a new industrial era is coming. It’s how decisively it is already reshaping venture markets, startup strategy, and capital allocation. The most transformative technologies of the next decade, AI, robotics, energy, defense, and space, depend on rebuilding physical systems at scale.
At Reference Capital, our role remains consistent: map these transitions, interpret where capital, policy, and technology intersect, and support investors navigating a market where software, hardware, and geopolitics are increasingly inseparable.
The Reference Capital Team
