2025 has been, by most objective measures, the year that deep-tech AI crossed from promising category to established asset class. Total global investment in AI hardware, AI-for-science, and AI infrastructure reached approximately $47 billion in the first three quarters of the year, a 62% increase over the same period in 2024. More significantly, the composition of that capital has shifted: a growing share is flowing to companies with genuine technical differentiation — proprietary hardware, novel training paradigms, defensible data assets — rather than application layer businesses built on commodity API access.
At Neuron Factory, we view this shift as validation of the thesis we articulated at our founding: the most durable value in the AI economy will be created at the infrastructure and hardware layer, not the application layer. The companies we are most excited about are not building the tenth AI writing tool or the fifteenth AI customer service platform. They are building the picks and shovels, the foundational science, and the enabling infrastructure for the entire AI ecosystem.
Funding Trends by Domain
Looking at investment activity across the six domains central to our investment thesis, several patterns are notable.
AI Hardware and Custom Silicon: Investment in AI-specific semiconductor companies reached $8.2 billion globally in the first three quarters of 2025, driven by the continued insatiable demand for inference compute and growing interest in alternatives to the GPU for specific workload categories. Neuromorphic and analog compute startups saw particular acceleration, with aggregate seed and Series A funding up 180% year-over-year. We expect this trend to continue as the energy cost of AI at scale becomes a genuine operational constraint for hyperscalers.
Foundational AI Systems: Investment in AI infrastructure — training frameworks, data curation tools, evaluation platforms, and synthetic data generation — reached $12.4 billion in the period. The most interesting activity is in the post-transformer architecture space, where a cohort of well-capitalized startups are exploring state space models, sparse mixture-of-experts architectures, and novel attention mechanisms that offer better efficiency at inference time.
AI for Science: Drug discovery AI and materials science AI continue to attract substantial capital, with aggregate investment of $7.1 billion in the period. The most consequential development of 2025 in this domain is the emergence of AI systems capable of designing novel molecular structures with specified properties from first principles — a capability that could compress drug discovery timelines from an average of twelve years to under three for certain target classes.
Quantum-AI Convergence: Still early-stage as a commercial category, but the investment signal is strong: $2.8 billion in aggregate funding for companies working at the intersection of quantum computing and machine learning. The most credible near-term opportunity is in quantum optimization for training data selection and hyperparameter search, where quantum advantage may be achievable on near-term hardware.
AI Safety and Alignment: A category that was almost exclusively academic two years ago now attracts substantial commercial investment: $1.6 billion in the period, with significant contributions from both dedicated safety-focused funds and major technology companies building internal research teams. We view this as an important and underappreciated opportunity category, and it is an area where Neuron Factory is actively evaluating investments.
Autonomous Systems: The most mature category in terms of capital deployment, with $15.9 billion invested in the period across robotics, autonomous vehicles, and intelligent agent platforms. The key trend is the shift from research to deployment: a growing number of autonomous systems companies are moving from controlled pilots to production-scale commercial operations.
Most Consequential Research Advances of 2025
Several research developments published or deployed in 2025 have materially shifted our understanding of what is technically possible, and by extension, what we think is worth investing in.
The first is a result from the University of Toronto demonstrating that spiking neural networks trained with a novel surrogate gradient method can achieve accuracy within 2% of state-of-the-art conventional networks on standard computer vision benchmarks while consuming 85% less energy. If this result generalizes — and early replications suggest it does — it substantially strengthens the commercial case for neuromorphic hardware in edge AI applications.
The second is a series of papers from DeepMind, the Allen Institute, and three academic groups demonstrating that large language models trained on scientific literature can generate valid novel hypotheses for experimental testing at rates that meaningfully exceed expert human performance in constrained domains including protein-ligand binding prediction and materials property optimization. This is the research foundation for the next generation of AI-for-science companies.
The third is a practical demonstration by a government-affiliated research laboratory of a working quantum key distribution network operating continuously over 140 kilometers of installed fiber with commercially viable key generation rates. This result — not yet published but privately communicated by sources we consider highly credible — suggests that metropolitan quantum networking is a commercial reality within a much shorter timeframe than mainstream industry forecasts suggest.
Five Vectors We Are Watching
Looking ahead to the next three to five years, we are particularly focused on five technology vectors where we believe the most important new companies will emerge.
Physical AI: The convergence of embodied robotics, neuromorphic sensing, and foundation models trained on physical interaction data. Companies that can close the sim-to-real gap and build robots that reliably handle the unpredictability of the physical world will create extraordinary value in manufacturing, logistics, construction, and healthcare.
AI-Native Networks: The next generation of telecommunications infrastructure will be designed from the ground up for AI workloads rather than retrofitted to handle them. This includes AI-optimized routing, intelligent spectrum management, and in-network compute architectures that bring inference close to data sources.
Biological Intelligence Augmentation: Non-invasive and minimally invasive brain-computer interfaces are approaching a level of fidelity that enables genuinely new classes of human-computer interaction. We are watching this space extremely carefully and believe the first commercial breakthrough applications — likely in assistive technology and professional augmentation — are less than three years away.
Synthetic Scientific Discovery: AI systems that can autonomously design experiments, execute them through robotic laboratory automation, analyze results, and generate the next hypothesis. The integration of language models with automated wet labs is already producing publishable results in a small number of leading research institutions.
Distributed Intelligence Infrastructure: The compute, networking, and storage infrastructure required to run AI workloads at the distributed edge — where latency constraints, energy budgets, and connectivity limitations make centralized cloud AI impractical. This is a fundamentally new infrastructure problem, and it will require fundamentally new solutions.
A Note on Our Portfolio
Neuron Factory made five seed investments in the first half of 2025, spanning neural interface technology, quantum networking, neuromorphic computing, AI perception, and autonomous logistics. We are proud of the technical caliber of every founding team in our portfolio, and we are even prouder of the commercial progress each company has made in the months since we invested.
If you are building in any of the domains described in this report and are looking for a seed investor who will engage seriously with your science, we want to hear from you. Our pipeline is always open.