This is the final part of a three-part series on Embodied Intelligence. Part 1 made the architectural case. Part 2 showed what the right architecture looks like in production. This part makes the investor case.
The capital is real. It is also concentrated in the wrong layer
In the two years between 2023 and 2025, the world committed more capital to datacentre AI infrastructure than to almost any other technology build-out in history. Data centre infrastructure capex grew 51% year-on-year to $455 billion in 2024, then rose another 53% year-on-year to $134 billion in the first quarter of 2025 alone, driven by hyperscaler demand for AI infrastructure. The four largest cloud operators namely AWS, Google, Meta, and Microsoft are expected to invest over $650 billion in AI infrastructure in 2026. Global data centre infrastructure spending is on course to surpass $1 trillion annually by 2030. To put that in physical terms: we are restarting nuclear power stations and building gigawatt-scale facilities to run these systems, something that I shared in the second piece.
None of this was wrong. The first wave of AI genuinely needed that infrastructure. You cannot train frontier models without it. The capital was necessary, and it built something real.
But necessary is not the same as where-to-deploy-next. And capital concentration at a single layer of a technology stack has a well-documented history of ending badly for the late entrants — even when the underlying technology is real. Goldman Sachs Chief Economist Jan Hatzius noted that AI investment contributed “basically zero” to measured US GDP growth in 2025, pointing out that much of the spending flows to imported equipment and has yet to show up in measurable productivity. The infrastructure is running ahead of the economic output it is supposed to enable. That gap does not mean AI is wrong. It means the next dollar of capital should not go where the last hundreds of billion went.
Man Group put it plainly: the companies that have lived through multiple semiconductor cycles, the ones who know what overcapacity looks like from the inside, do not believe the capex projections. They understand that GPU architectural generations turn over rapidly, that today’s training-heavy workload mix will not persist, and that the gap between capital deployment and revenue cannot be sustained indefinitely. My view is that the most likely unwind is not a crash. It is a plateau: a quiet flattening of the hyperscaler spending curve that could take 12 to 18 months for the market to fully accept. Crashes are obvious. Plateaus are expensive.
You have seen this before. In the late 1990s, US telecoms laid tens of millions of miles of fibre optic cable. The internet was real. The demand was real. But much of that fibre sat dark for years, and many of the companies that raised the most entered bankruptcy. The infrastructure was needed. The capital timing was catastrophic. The question now is which layer of the AI stack plays the role of that dark fibre and where the smart capital goes instead.
The semiconductor money chased the same idea and the same flaw
The infrastructure overbuild is the visible part of the story. The less visible part is what happened in the semiconductor funding market alongside it, and that is where the more instructive parallel lies.
The thesis that attracted the most chip capital over the past many years was some version of the same idea: NVIDIA has a stranglehold on AI compute, the margins are extraordinary, and therefore there is room to build an alternative. Investors backed that thesis at scale. AI and machine-learning semiconductor venture funding rose from $4.8 billion in 2024 to $8.4 billion in 2025, over $13 billion in two years, even as deal count fell, meaning the capital concentrated into a small cohort of NVIDIA challengers.
Look at where it went. Cerebras raised over $1 billion privately and then $5.55 billion at its IPO. Groq raised $750 million before entering a reported $20 billion NVIDIA transaction structured around licensing, assets, and talent rather than a conventional acquisition. Etched raised roughly $500 million at a $5 billion valuation for a transformer-only ASIC with its founder’s own words: “If transformers go away, our company will die.” MatX raised $500 million. d-Matrix raised $275 million at a $2 billion valuation, backed by Microsoft. SambaNova has accumulated over $1.5 billion. Fractile raised $220 million to run frontier models faster. Tenstorrent has raised more than $1.18 billion and is seeking $800 million more.
Every one of these is a different architecture. But they share the same target: serving large models, in the datacentre, faster and cheaper than NVIDIA. That is a real engineering problem. It is not a defensible investment thesis, for one reason almost nobody prices correctly.
NVIDIA’s moat is not just the GPU. It is CUDA build over two decades of developer tooling, libraries, and millions of researchers who write code that runs on NVIDIA hardware first. A new entrant that wins on silicon benchmarks still has to convince that entire community to rewrite its stack. Hardware without software and an ecosystem is not a moat. The highest-value semiconductor positions have always combined architecture with software, tooling, and customer relationships deep enough that switching has a real cost. NVIDIA did not win because the GPU was faster. It won because, by the time anyone noticed, the software ecosystem had become the actual product.
That is precisely why ARM’s recent move is so revealing. For decades ARM was the most elegant business in semiconductors, licensing its architecture, collecting a royalty on more than 310 billion chips shipped, letting others handle the capital-intensive work of production. In March 2026, ARM broke that model and launched the AGI CPU, its first complete production datacentre chip, built on TSMC 3nm, with Meta, OpenAI, SAP, and Cloudflare among its launch partners and early customers. ARM projects this could scale from roughly $1 billion in chip revenue by 2028 toward $15 billion, transforming it from an asset-light licensor into a capital-intensive vendor competing much closer to the customers it historically enabled.
Why would ARM take that risk? Because it believes its ecosystem, the software, the developer base, the architectural gravity it spent thirty years building, gives it the right to win where pure-silicon challengers cannot. That is the correct read of why NVIDIA is hard to displace. Whether it is the right strategic bet for ARM is a separate question worth its own analysis. What it signals unambiguously is this: even one of the companies holding the most defensible ecosystem positions in the industry concluded that the datacentre layer is now a game of ecosystems, not chips alone. When the most patient licensor in semiconductors decides it must move closer to direct competition, you are watching a market where ecosystem control has become as important as silicon performance. That may prove to be a late-cycle signal. At minimum, it shows that pure chip differentiation is no longer enough.
A distinction the market keeps getting wrong
It is worth being precise about what “edge AI” means right now, because the term is doing a lot of work in funding decks.
Much of the capital flowing into what the market calls edge inference is funding a different problem entirely: running large models locally, on relatively high-power devices, without a cloud round trip. That is a meaningful engineering challenge. But it is still a large-model problem, built on the same architectural assumptions as the datacentre, just moved closer to the user. You can watch it happen in real time: Axelera began as a genuine edge company, computer vision, robotics, industrial inference, then its funding narrative expanded “from edge to datacentre” to chase generative AI and large language models. That is market gravity at work: companies that begin at the constrained, physics-first edge are pulled toward the better-funded, easier-to-explain large-model infrastructure market.
The sensor edge is a different problem altogether. It is not about running a smaller GPT on a device. It is about intelligence designed from the physics of the sensor signal up: event-driven, sub-milliwatt, always-on. The two markets have completely different competitive dynamics, moat structures, and capital requirements. Conflating them is one of the most common analytical errors in how this sector is evaluated, and it is precisely the error that leaves the genuine opportunity underpriced.
Where the constraint actually is
There are roughly 21 billion connected IoT devices in the world today. The number of individual sensors embedded within those devices is considerably higher: a single connected machine can carry dozens of sensors monitoring vibration, temperature, pressure, acoustics, and current simultaneously. Some industry estimates put the deployed sensor base above 50 billion when counting individual sensors rather than connected devices. By 2030, connected devices alone are forecast to reach 39 billion. The sensor count behind that number will be a multiple of it. Every one of those sensors sits at the edge of a physical environment and generates continuous signal that must be interpreted and acted. Most of that signal cannot go to the cloud. The latency is wrong, the power budget is wrong, the privacy constraints are wrong, and the economics of moving hundreds of billions continuous data streams through centralised infrastructure simply do not work.
This is the constraint. Not a software problem. Not a model problem. A physics problem. And physics problems in semiconductors create defensible IP in a way that model optimisation cannot replicate. The architecture required to sit at that constraint is event-driven, ultra-low-power, and designed from the ground up for the statistical nature of sensor signals: sparse, continuous, almost always silent until something meaningful happens.
That is fundamentally different from what wins in a datacentre. A GPU optimised for parallel matrix multiplication at kilowatt scale has nothing useful to offer a device that must run for three years on a coin cell and detect a fall in a bathroom. The physics are incompatible. That incompatibility is a moat, but a different kind from CUDA. It is rooted in architecture, power budgets, signal-processing expertise, embedded tooling, and design-in relationships rather than software lock-in alone.
Jevons is on the side of the sensor edge
DeepSeek arrived in early 2025 and briefly panicked the market. If models could be trained and run at a fraction of the cost, DeepSeek claimed to cut training costs by roughly 18x and inference costs by roughly 36x versus GPT-4o, the reasoning went that compute demand would fall. NVIDIA lost $600 billion in market value in a single day on that logic.
The logic was wrong, in a way that matters for sensor edge investors. The Jevons Paradox, a 160-year-old principle, holds that when a resource becomes more efficient to use, total consumption of it rises, not falls. When the steam engine became more efficient, coal use increased because more industries could afford it. Inference pricing has fallen at a median of roughly 50x per year; GPT-4-equivalent performance that cost $20 per million tokens in late 2022 now costs $0.40, while enterprise AI spending rose from $11.5 billion in 2024 to $37 billion in 2025.
Jevons works at every layer, cheaper inference drives more datacentre demand too. But the asymmetry is the point. Datacentre capacity is being built years ahead of demonstrated demand. Sensor edge capacity does not yet exist at the scale the physical world will require. Every efficiency gain makes it feasible to push intelligence into a lower-cost, lower-power device, opening a new tier of the physical world that was not economically accessible before. The sensor edge market does not shrink as AI gets cheaper. It expands. It is not saturated at hundreds of billion sensors. It is at the beginning.
The market size objection and why it asks the wrong question
The sharpest challenge an investor will raise to the sensor edge thesis is market size. The comparison is not flattering at first glance. NVIDIA alone generated $215.9 billion in revenue in fiscal year 2026. The broader edge AI chip market sits at roughly $3.7 billion today, growing to $11.5 billion by 2031. The sensor-native segment is smaller today, but strategically more interesting because it sits closest to the physical constraint. The broader MCU market it disrupts is approximately $36 billion, growing steadily but without drama. On raw TAM comparison, the sensor edge looks small. That comparison is directionally correct on today’s numbers and wrong on the logic.
The investment cycle argument has already been made in this piece, the datacentre is the most aggressively funded layer in semiconductor history, the return on the next dollar deployed there is declining, and capital rotates when that becomes consensus. What that rotation needs is somewhere defensible to go next. The neuromorphic segment within edge AI chips is already growing at 48.3% CAGR, the fastest of any segment in the stack, and the M&A multiples described earlier reflect strategic value that public market investors have not yet priced. The sensor edge is not a small market. It is an underpriced one at the early stage of a rotation.
The MCU market ships over 35 billion units annually, much of it at dollar-scale or sub-dollar ASPs: commodity silicon, thin margins, largely undifferentiated. That is the market sensor-native intelligence begins to disrupt. And the disruption does not merely grow that market. It redefines what a unit in that market can be worth.
When a microcontroller-class device becomes an intelligent edge node, sensing, deciding, acting without a cloud round trip, it stops being a commodity input and becomes a differentiated system component. In high-value sensing sockets, average selling price can move from commodity MCU economics to five, ten, or fifteen dollars because the value delivered is categorically different. That is not a universal uplift across every MCU socket. It is a selective but highly valuable ASP expansion across the sensing sockets where intelligence changes the function of the product. The TAM is not simply the current MCU market.It is what the MCU market becomes when redefined by the right kind of intelligence.
One final distinction that matters for how you underwrite this. At the datacentre, value above the chip accrues elsewhere, to hyperscalers, software platforms, application businesses entirely separate from the chip vendor. At the sensor edge, architecture, application software, and customer relationship can all sit within the same company. The design-in relationship generates revenue across product generations rather than subscription months that can be repriced overnight. That is a more defensible financial profile than the current MCU market reflects and one the right companies are already building toward.
The question is not whether the sensor edge TAM matches the datacentre today. It is whether the capital cycle has turned, the architecture is defensible, and the companies building here can own the constraint as it scales.
The smart money is already moving and saying so out loud
Strategic acquirers do not wait for a market to fully price. They move when they see capability they cannot build fast enough internally.
In February 2025, NXP agreed to acquire edge AI NPU startup Kinara for $307 million in cash. Kinara had raised $54 million across its entire life. NXP was not buying revenue. It was buying architecture, tooling, and time-to-market. That deal sat among a cluster of similar exits: GrAI Matter to Snap, Perceive to Amazon. Strategic M&A often precedes broader public-market recognition, because acquirers move when capability is scarce, not when the category is obvious.
The directional signal is also being said plainly from the most informed seat in the industry. At GTC 2025, Jensen Huang described the next wave of AI as robotics, physical systems that understand friction, inertia, cause and effect. By GTC 2026 it was no longer a roadmap: over 110 robots on the floor, autonomous-vehicle partnerships across the major manufacturers, and physical AI presented as production infrastructure. NVIDIA’s own messaging has shifted toward physical AI, robotics, autonomous systems, and embodied intelligence, which is notable because NVIDIA has been the clearest beneficiary of the datacentre AI cycle. When the dominant datacentre AI company points repeatedly at the physical world, that is worth taking seriously.
What to look for
If you are allocating capital, building partnerships, or deciding where to place a strategic bet, the filter is not which company has the largest model library or the most benchmark wins. The questions that identify durable positions are different and they fall into two distinct categories: what the technology is actually doing, and whether the commercial model reflects the reality of how this market works.
On architecture and technology.
Does the company own the constraint, or merely optimise around it? The most valuable semiconductor companies are not built around generic performance claims. They are built around hard constraints that customers cannot escape: power, latency, thermal envelope, form factor, reliability, or cost. In the sensor edge, the defining constraint is not TOPS. It is useful intelligence per microwatt, per millisecond, per sensor stream. A company that owns that constraint is building infrastructure. A company that merely compresses models to fit within it is building an optimisation layer.
Does the architecture match the physics of the signal? A company compressing general-purpose models and calling it edge AI is a software-optimisation business. One designed from the sensor signal up is a semiconductor business with a defensible architectural moat. In Part 2 of the series, I described going from an 80-million-parameter model on GPUs to the kiloparameter range on a neuromorphic microcontroller at equivalent accuracy for the target sensing task. That was not generic compression. It was starting from the right constraint.
Can the architecture scale across sensor modalities? The sensor edge is not one market. It is many markets connected by a common signal problem: sparse, continuous, noisy physical data that must be interpreted under extreme constraints. The strongest architectures will not be limited to one demonstration. They will generalise across modalities, audio, radar, motion, vibration, touch, pressure, current, and vision-adjacent event streams, because the underlying computational problem is similar even when the end markets differ. Modality breadth is one of the clearest signs that a company is building a platform rather than a point solution.
On commercial model and strategic position.
Is the design-in strategic or peripheral? Not every design-in is equal. A chip used as a feature accelerator can be swapped. A chip that becomes part of the sensing, decision, and control loop of a product is far harder to displace. The best sensor edge companies will be designed into the product architecture, not bolted onto it, sitting close to the signal, close to the firmware, and close to the application logic. That position creates technical and commercial stickiness that compounds across product generations in a way that cloud API relationships, repriced quarterly, cannot replicate.
Does the company reduce customer risk? In the sensor edge, the buyer is rarely looking for a benchmark. They are trying to ship a product that must work for years in the field, under severe power, cost, safety, and reliability constraints. The winning supplier reduces risk across the full journey: sensor integration, model development, firmware, validation, production qualification, and lifecycle support. Performance opens the door. Risk reduction wins the design.
Is there application pull, not just technology push? The best signal is not a benchmark. It is a customer willing to redesign a product around the architecture because it enables something they could not otherwise ship: an always-on wearable, a battery-powered industrial monitor, a privacy-preserving occupancy sensor, a smart appliance that senses without streaming, a robot that reacts locally, a safety system that cannot wait for the cloud. When customers change their product roadmap around the technology, you are no longer looking at a component supplier. You are looking at category pull.
Is the company selling silicon or creating a platform? A chip wins the first design-in. A platform wins the second, third, and fourth. The highest-value positions combine the right architecture with enough application software to lower the customer’s integration burden with models, tools, reference applications, sensor interfaces, developer workflows, and enough abstraction that customers can move from proof-of-concept to production without rebuilding the entire stack. The hardware is the differentiator. The software lowers adoption friction, expands the market, and compounds the moat. NVIDIA understood this with CUDA twenty years ago. The sensor edge companies that understand it today are building something with the same compounding logic applied at the constraint layer where the physical world meets intelligence.
Does the company understand the path from design-in to deployment economics? Sensor edge companies should not be judged by SaaS timelines. Design-ins take time. Qualification takes time. Production ramps take time. But once designed in, the revenue profile is durable across product generations. The right question is not whether revenue appears instantly after a proof-of-concept. It is whether the company is converting evaluations into design-ins, design-ins into production programmes, and production programmes into platform relationships.
At the sensor edge, the model is only one part of the economics. The real equation includes battery life, bill of materials, sensor cost, connectivity cost, installation friction, false positives, maintenance, and field failure rates. A solution that saves a milliwatt can change the battery, the enclosure, the service model, and the customer’s total cost of ownership. That is why power is not a feature. It is a business model variable.
The companies worth backing will not look like datacentre AI companies made smaller. They will look like systems companies built around the physics of the edge: silicon where the constraint is hardest, software where adoption friction is highest, and customer relationships where product architecture becomes difficult to change.
The close
The datacentre decade produced extraordinary value and a great deal of necessary infrastructure. It also produced an overhang, and a wave of capital chasing NVIDIA at the one layer where NVIDIA is most defensible. Most of that capital will not earn what it expected — not because the technology is wrong, but because returns in any technology wave do not accrue to whoever raises the most. They accrue to whoever owns the constraint.
The sensor edge is not the next wave in the sense of something still forming on the horizon. The production deployments exist. The M&A is already in motion. The most informed players in the industry are pointing at it explicitly. What has not caught up is the capital — and that gap between where the evidence points and where the money sits is precisely where durable returns are built.
The shift from brute force to elegant architecture is not a philosophical preference. It is where the physics forces the industry to go. The datacentre had its decade. The question now is not whether the sensor edge has its own — it is whether you are positioned before that becomes obvious to everyone.
If you missed the earlier parts of this series, here are the links – Part 1 on the architectural case for Embodied Intelligence, and Part 2 on what production deployments actually look like.
