In the summer of 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton demonstrated that two consumer NVIDIA GPUs could train a convolutional neural network that outperformed every hand-crafted computer vision system ever built. It was, in retrospect, the moment the modern AI era began. But it was also the beginning of an architectural dependency that the AI industry has been racing to extend, and is now running out of runway to maintain.
The GPU's dominance of AI computation is not an accident. It is the product of a fortunate architectural alignment: GPUs were designed for massively parallel floating-point arithmetic to render 3D graphics, and training neural networks turns out to require massively parallel floating-point arithmetic. For more than a decade, NVIDIA rode this alignment to a market position that the company is worth more than most countries. But the alignment is straining. The physics of silicon manufacturing, the economics of power delivery and cooling, and the specific computational demands of next-generation AI workloads are collectively pushing the industry toward a hardware transition that will be as significant as the move from CPUs to GPUs twelve years ago.
At Neuron Factory, we are investing in the companies that will define what comes next. This piece is our map of the transition — where we are, where we are going, and what it means for investors and founders building in this space.
Wave One: The GPU Era (2012–Present)
The GPU era was defined by a simple formula: more memory bandwidth plus more FLOPS equals better AI. NVIDIA's genius was recognizing this before anyone else and building the software stack — CUDA, cuDNN, the NCCL communication library, the Tensor Core architecture — that made GPUs not just technically superior to alternatives but practically irreplaceable. By the time competitors recognized the opportunity, NVIDIA had built a software moat that its hardware advantage alone could not have created.
The GPU era is not over. NVIDIA's H100 and B200 accelerators are extraordinary pieces of engineering, and they will be the dominant training platforms for large language models and multimodal AI systems for at least the next three to five years. But the GPU era is clearly peaking. The signals are unmistakable:
- Power consumption per rack has exceeded the capacity of conventional data center electrical and cooling infrastructure, forcing the industry to build custom facilities at enormous capital cost.
- The cost of training frontier models has grown faster than the revenue models that can justify it, creating pressure on AI companies to find more efficient inference approaches.
- NVIDIA's gross margins — currently above 70% — are so extraordinary that every major technology company is investing aggressively in alternatives. Google, Amazon, Microsoft, Meta, and Apple all have active AI chip programs designed to reduce NVIDIA dependency.
- The semiconductor process technology roadmap is approaching physical limits. Moore's Law has effectively ended for traditional transistor scaling, and the performance gains from moving to 3nm and 2nm processes are diminishing rapidly relative to their cost.
Wave Two: The ASIC Era (2020–2030)
The second wave of AI hardware is already underway. Application-specific integrated circuits — chips designed for a specific class of AI workload rather than general-purpose computation — are capturing an increasing share of AI inference and, increasingly, training workloads.
What ASICs Do Better
By eliminating general-purpose capabilities that are irrelevant to their target workload, ASICs can achieve dramatically better energy efficiency and cost per inference than GPUs. Google's TPU v4 achieves competitive performance with NVIDIA's A100 at roughly 40% of the power consumption for transformer inference workloads. Amazon's Trainium and Inferentia chips are enabling AWS customers to run LLM inference at 30-50% lower cost than equivalent GPU instances. These are not small differences — at the scale of billions of inference calls per day, they translate to hundreds of millions of dollars in annual operating savings.
The ASIC wave is creating a rich ecosystem of opportunities beyond the hyperscaler in-house programs. Startups designing inference ASICs for specific vertical applications — healthcare, financial services, industrial automation — can achieve cost and efficiency advantages that NVIDIA's general-purpose architecture cannot match. The software challenge is real: building a competitive CUDA ecosystem on a new chip architecture requires years of developer toolchain investment. But for workloads that can be served by a fixed model architecture with predictable input distributions, the ASIC economic case is compelling.
Where ASIC Companies Face Risk
The principal risk for ASIC companies is architectural obsolescence. AI model architectures are evolving rapidly — the attention mechanisms that transformer-based models rely on look different from the CNN architectures that dominated five years ago, and there is no reason to believe the rate of architectural change will slow. An ASIC optimized for current transformer architectures may face significant competitive pressure when the next architectural paradigm emerges. The ASIC companies with the best long-term prospects are those that have designed sufficient architectural flexibility into their chips to accommodate future workload changes without requiring a complete respin.
Wave Three: The Neuromorphic Era (2025–2035)
The third wave is the one we are positioned at the leading edge of. Neuromorphic computing architectures — chips that process information through sparse, event-driven spike patterns rather than continuous floating-point arithmetic — offer a fundamentally different value proposition from both GPUs and ASICs. They are not primarily about higher peak performance. They are about dramatically lower energy consumption for a specific class of tasks, particularly those that involve processing streams of sensory data in real-time with tight power constraints.
The Core Neuromorphic Advantage
A spiking neural network running on a neuromorphic chip consumes energy approximately proportional to the number of spikes it processes — which is a small fraction of total neurons on any given input. In biological neural networks, this sparsity is the primary mechanism through which the brain achieves extraordinary computational efficiency at 20 watts. In silicon implementations, benchmarks from Intel's Loihi 2 and IBM's NorthPole show 10x to 100x energy efficiency advantages over GPU-equivalent inference for specific workloads, with the largest advantages in event-based sensing applications.
The Software Stack Challenge
The primary barrier to neuromorphic computing adoption today is not hardware performance — it is the software stack. Training algorithms for spiking neural networks are less mature than the backpropagation methods used for conventional ANNs. Developer toolchains for neuromorphic hardware are primitive compared to CUDA. The ecosystem of pre-trained SNN models is tiny relative to the Hugging Face model library for transformer architectures.
This software gap is the central challenge that neuromorphic computing companies must address in the next five years. The companies in our portfolio and pipeline that we are most excited about are those that are investing as aggressively in their software stack as their hardware — building SNN compiler toolchains, training frameworks optimized for spike-based learning rules, and developer APIs that abstract away the complexity of the neuromorphic architecture for application-level engineers.
Application Domains Where Wave Three Wins Today
The neuromorphic era will not arrive uniformly. It will arrive application by application, starting with the domains where the energy efficiency advantage is most directly economically valuable and where the software complexity is most manageable. Our assessment of the application priority order:
- Event-based vision and sensor fusion: The market is ready now. Automotive LIDAR/radar fusion, industrial machine vision, and security monitoring are all processing streams of sensory events in real-time under strict power budgets. Neuromorphic architectures have demonstrated commercially compelling advantages in these applications with existing technology.
- Medical device embedded AI: Wearable cardiac monitors, neural prosthetics, continuous metabolic monitoring — all of these require AI inference on sub-watt power budgets. The regulatory timeline is long, but the commercial opportunity is enormous and the technical fit with neuromorphic architectures is precise.
- Autonomous robotics: Mobile robots in logistics, agriculture, and construction need local intelligence for navigation and manipulation under strict power constraints. This application domain is 2-3 years from significant commercial scale but is developing rapidly.
- Data center inference acceleration: For specific workload categories (time-series forecasting, anomaly detection, reinforcement learning simulation), neuromorphic co-processors will achieve TCO advantages in data center deployments within 3-5 years.
What the Three-Wave Map Means for Investors
The most important insight from this framework is that AI infrastructure investing is not monolithic. The investment opportunities in Wave One (GPU supply chain, data center infrastructure, CUDA ecosystem software), Wave Two (inference ASIC design, model compression and optimization, serving infrastructure), and Wave Three (neuromorphic hardware, SNN training frameworks, event-based sensing applications) are almost entirely non-overlapping. Each wave requires different technical evaluation expertise, different go-to-market understanding, and different time horizons.
At Neuron Factory, we have made the deliberate decision to focus our investment activity at the leading edge of Wave Three, while maintaining our technical coverage of Wave Two through our advisory relationships. We believe Wave Three is where the highest-conviction venture opportunities exist today — not because the technology is most mature, but because the founding team quality is highest (the researchers transitioning from Wave Three academic programs to company formation represent an extraordinary cohort), the competitive dynamic is most favorable (the space is underinvested relative to its commercial potential), and the timing is right (the first commercial applications are de-risking the technology category for the broader investment community).
What the Three-Wave Map Means for Founders
If you are a founder building in AI infrastructure, understanding where your technology sits in this wave map is essential for your fundraising strategy, your hiring plan, and your customer development approach. The investors, customers, and team members who are most relevant to Wave One companies are almost entirely different from those relevant to Wave Three companies.
For Wave Three founders specifically, our advice is to resist the pressure to generalize. The neuromorphic companies most likely to succeed in the next five years are those that dominate a specific application domain — not those that pitch themselves as general-purpose alternatives to NVIDIA. Pick your beachhead, win it decisively, and build from strength. The platform opportunity will come, but it comes after you have established undeniable proof points in a specific market, not before.
"Every major platform technology in computing history — from the mainframe to the PC to the GPU to the mobile chip — started with a specific application where it was undeniably, categorically better than the incumbent. Neuromorphic computing will follow the same pattern. The question is not whether it will happen. The question is which application creates the first undeniable proof point, and which company owns it." — Dr. James Okonkwo, Neuron Factory
Our Current Investment Focus
We are actively evaluating companies in the following segments of the AI infrastructure transition:
- Event-based sensing + neuromorphic processing: Companies building end-to-end systems that combine event-based cameras, LIDAR, or radar with neuromorphic processing pipelines for automotive or industrial applications.
- SNN training tools: Companies building developer tools that make it significantly easier to design, train, and deploy spiking neural networks without requiring deep expertise in neuromorphic circuit design.
- In-memory computing: Companies developing processing-in-memory architectures for specific AI workloads — particularly matrix-vector multiplication in transformer inference — where the memory wall is the primary bottleneck.
- Analog AI accelerators: Companies using analog computing elements to implement neural network inference with dramatically better energy efficiency than digital approaches for specific workload profiles.
If you are building in any of these spaces, we would like to hear from you. The transition from Wave Two to Wave Three is underway, and the companies founded in the 2024-2027 window will define the infrastructure of AI for the next generation. We want to be part of building it.