AI's Money Is Moving Down-Stack: Why Silicon and Sensors Are the New Frontier

For two years the AI story has been told at the top of the stack. Bigger models, longer context windows, and the multi-billion-dollar rounds that fund them have owned the headlines. This week the Race to AGI deal tracker told a quieter and, arguably, more interesting story: a run of smaller investments landing squarely in the hardware and physical layers beneath the models.

Look at the shape of it. Oxmiq Labs raised a $35M Series A to commercialize OxCore, a licensable AI GPU architecture — a design chipmakers can license and customize rather than engineering an accelerator from scratch. SmartCore drew angel funding to scale silicon-photonics optical I/O chiplets, the interconnect that shuttles data between chips inside a data center. Luxonis raised a $14M Series A to expand its OAK cameras and OAK4 perception platform for robotics and industrial "physical AI." And Rockwell Automation and Cisco formed a partnership to co-develop software-defined manufacturing for factories. Individually, none is a headline number. Together, they trace a thesis.

The thesis is that capital is beginning to move down-stack — away from the application and model layers, where the marginal dollar buys less and less differentiation, and toward the compute substrate and the physical edge, where supply is scarce and moats are durable. It is the least glamorous part of the AI economy, and increasingly the part that decides who can build at all.

Start with silicon. The defining constraint of the current cycle is not ideas but accelerators, and that constraint has a name: Nvidia. Every lab that wants to train or serve a frontier model competes for the same supply. Oxmiq's pitch — a proven GPU core you license and adapt, rather than a chip you design over a multi-year cycle — is a direct response to that bottleneck. It is, in effect, the Arm playbook aimed at AI silicon: don't sell the chip, sell the architecture and let a hundred customers build around it. Whether OxCore specifically succeeds is unknowable today. But the logic behind funding it is sound, and it will not be the last bet of its kind.

SmartCore points at the same wall from a different angle. As clusters scale into the tens of thousands of GPUs, the limiting factor stops being the compute in each chip and becomes the bandwidth between them. Silicon photonics — moving data as light rather than electrons across copper — is one of the few credible paths to keeping interconnect from throttling the whole system. Funding an optical-I/O chiplet startup is a bet that the next efficiency gains come from plumbing, not parameters. That is a very different worldview from the one that dominated 2024.

Then there is the physical layer, and this is where the week's deals get genuinely interesting. Luxonis builds the perception hardware — cameras and on-device vision — that lets a robot or a machine actually see and act in the world. Rockwell and Cisco are wiring AI into factory floors. Both are wagers on the same proposition: that a large share of AI's next decade of value gets created not in a browser tab but on a shop floor, a warehouse aisle, or a robot's chassis. The application layer for text is crowded and commoditizing fast. The perception layer for machines is comparatively wide open, and it is capital-intensive in a way that keeps the tourists out.

It would be a mistake to read this as the top of the stack going quiet. The same week produced the loudest possible counterexample: Kuaishou is spinning Kling AI out as an independent video-generation company, backed by up to $3 billion from Tencent, Alibaba and Baidu, with an IPO mandated by 2031. That is application-layer money at a scale the hardware rounds cannot touch. But note the structure — three rival platforms jointly underwriting one company rather than each building in-house. Even at the top of the stack, the market is consolidating bets rather than multiplying them. And the compute those video models will run on has to come from somewhere: witness SpaceX's multi-year lease reserving Nvidia Blackwell capacity for Reflection AI. The application layer's ambitions are exactly what make the layers beneath it valuable.

For anyone trying to read the field, the practical signal is this. The headline metric of AI investment — the size of the largest model round — is becoming a poor proxy for where the durable value is accreting. The broader trend data increasingly shows two economies running in parallel: a crowded, fast-commoditizing race at the application layer, and a quieter, more defensible buildout of the silicon, interconnect, and perception hardware that everything else depends on. The talent-and-technology maneuvering around companies like Groq is part of the same story: the scarce resource is no longer the model, it is the ability to run one economically.

The hedge worth keeping is that down-stack bets are slow. Silicon photonics has been "about to matter" for the better part of a decade; robotics perception startups have a long history of raising well and shipping late. A $14M or $35M round is a hypothesis, not a verdict, and most will not pay off. But the direction of the capital is a leading indicator that is hard to fake, and right now it is pointing at the ground floor of the stack. The next time a frontier lab announces a record training run, the more revealing question may not be how large the model is — but who made the chips, the light, and the sensors that let it exist at all. You can watch that shift accrue, deal by deal, in the Race to AGI deal tracker.