Since 2021, we at Two Small Fish Ventures have been investing in deep technology startups across the AI tech stack, including semiconductor technology. Then, in mid 2025, we shared the evolution of our investment thesis on the five interconnected layers shaping the next frontier of AI: vertical AI platforms, physical AI, AI Infrastructure, advanced computing hardware, and smart energy. You can check the blog post here.
A few months later, Jensen Huang stood on stage at CSIS in December 2025 and walked through essentially the same framework.
Great minds think alike, right, Jensen?
For those of us who understand technology deeply, we naturally see the same thing: AI doesn’t just change applications. It touches the full stack from the chip to the grid. And that’s exactly where we’ve been placing our bets.
Since then, I’ve been invited to speak at multiple events about where the frontier is moving. Founders asking about bottlenecks. Investors trying to map where to allocate capital. Operators trying to understand what’s actually happening inside each layer. So here, let me break down how we think about the five layers, and where some of the opportunities will be.
What’s unfolding across these five layers shows that hard problems remain unsolved, and real tensions exist in how entrepreneurs are approaching them. That’s exactly what makes this moment interesting.
Vertical AI Platforms: The very obvious move in the early days
When the world found out about foundation models with the launch of ChatGPT in November 2022, it was so obvious for many builders to bet on vertical agentic systems. Since then, we have seen domain-specific agentic systems for law firms, doctors, manufacturers, supply chain, etc. Everyone got excited. Everyone started building.
Many of the approaches are really about, “this company has data, let’s throw a model at it.” The result is usually incremental automation, shaving percentages off existing processes. But many enterprises have since then reported that their investments are nowhere near the step-function shift. It became clear that unlocking that exponential value requires deep domain expertise and purpose-built agentic systems.
At the same time, building AI-powered software has never been cheaper or faster. Enterprises are moving more towards building the solution internally. The case that Kirkland & Ellis committed $500 million to build their own legal infrastructure rather than adopt a third-party solution is an example of how some enterprises approach AI adoption. After all, they want to control data, workflows, and their institutional knowledge.
So now entrepreneurs are navigating two challenges at once. Convincing enterprises to buy rather than build, while standing out in a suddenly crowded field where everyone is shipping similar solutions at the same time.
However, we believe the founders worth watching are the ones with strong domain expertise. They understand and are willing to tackle hard problems and the old, archaic systems holding industries back. Enterprise data, privacy, risk management, compliance, all of it, while building agentic systems that actually move the needle. That combination is rare, and frankly, that’s exactly what this market needs more of.
Physical AI: Bringing Intelligence into the Physical World
Robotics has existed for decades. Canada built the Canadarm for NASA back in the 1970s. Walk into any old manufacturing plant, and you’ll find machinery programmed to do the same repetitive tasks, day in and day out.
What’s different now is that machines can actually learn from their environment. Intelligence systems that use multi-modal data such as visual data, motion data, sensor data, and other forms of data create robotic systems that are more adaptive, more versatile, and more affordable to deploy.
As we said a few years ago, physical AI is accelerated by the convergence of hardware and software. Take Haply, one of our portfolio companies. They combined touch sensors with software that transforms raw sensor data into force data that robots can understand. That’s the kind of insight that only emerges when hardware and software are designed together from the ground up.
Physical AI is more than just robots and humanoids. It’s the entire ecosystem required to deploy intelligent systems safely in the real world. Issues such as fail-safe mechanisms, safety frameworks, governance and operability all require technological innovations. When a robot makes a mistake in a factory, the question is whether the environment around it stays safe and what the human-in-the-loop process looks like.
AI Infrastructure: The Necessary Layer that Makes AI Systems Trustworthy
While everyone’s excited about vertical AI systems and smart robots, wide adoption of AI is still not happening. Why? It’s because while AI systems are powerful, they are not necessarily trustworthy or frictionless when it comes to adoption. A big part of that comes down to the software stack that makes AI safe, observable, and trustworthy enough to deploy at scale.
Most agentic systems are being built without a clear way to benchmark whether they’re working better than the processes they’re replacing. Explainability at the model layer is one thing. Auditability across an entire agentic workflow is a completely different problem. Yscope, one of our portfolio companies, is tackling distributed data logging across complex enterprise environments.
Hyperscalers are starting to address this through their enterprise teams. OpenAI and Anthropic are building up their FDE teams to help enterprises adopt their products. They are really gathering the know-how on the tooling needed for agentic observability, benchmarking, and auditability, etc. The opportunities in this layer are about creating trustworthy layers between models and agents, and between agents and humans. That’s the white space.
Advanced Computing Hardware: The Bottlenecks Become Obvious
The compute bill is brutal. Uber burned through their entire annual AI token budget in four months. Microsoft switched to DeepSeek models because cloud inference costs were unsustainable. The direction becomes obvious: compute has to be distributed.
The hyperscalers are already moving that way. NVIDIA and Microsoft are bringing agentic capabilities to the desktop. AMD is enabling inference on edge devices, and Google is opening TPU access. Tesla is rumoured to be building container-scale data centers as mesh networks.
We’ve been betting on this for a few years. Blumind on neuromorphic edge compute is one example of where we see the hardware layer heading. The white space runs from wearables to edge computing devices, all the way to semiconductor architecture designed specifically for distributed AI workloads. And further down the timeline, quantum computing adds another dimension entirely.
Smart Energy: Everything Draws Power
AI’s appetite for power is well documented. What gets far less attention is the intelligence layer needed on top of the existing grid infrastructure. Starting right at the endpoint.
Current energy infrastructure is entirely reactive. Someone plugs in a device, draws power. That’s it. No optimization. No intelligence. The grid won’t change for decades. But endpoints can get smart right now.
FUTURi, one of our portfolio companies, is building smart systems that optimize residential energy consumption in real time. Unlike Smart meters, which have existed for years, their systems collect data and apply intelligence to it. As this kind of endpoint intelligence matures, the next layer will naturally emerge: district-level coordination aggregating signals from thousands of endpoints to optimize local grid management.
Looking Beyond the Five Layers of AI
These five layers in the technology stack of AI are moving fast, and they are interconnected.
However, the talent to drive adoption in the enterprise, maintain their system, and stay in the operation loop is a completely different breed of our existing workforce. We need people who understand agentic systems deeply, can architect distributed infrastructure, debug multi-model workflows, and keep human judgment embedded throughout.
As mentioned before, OpenAI and Anthropic have started embedding Forward Deployment Engineers directly inside enterprise customers. That’s not a sales function. That’s a signal of what’s coming. New categories of roles are being created that didn’t exist just a few years ago.
This is one of the most exciting moments to be building and investing across these five layers. The problems are hard, the stakes are high, and the talent and capital flowing in are unlike anything we’ve seen before.
I’ll be diving deeper into each layer over time, sharing more of what I’m seeing and where I think the most interesting opportunities are taking shape. Stay tuned.
