The sums required to compete at the frontier of artificial intelligence, whether by training a new large language model or constructing a new data center, now run into the billions of dollars. For most startups, that figure alone closes the door. Yet inside the machinery that powers those giant systems sit thousands of narrower problems, in power delivery, cooling, connectivity and chip materials, where a single well-engineered solution can become indispensable without the company behind it ever needing to raise anything close to that scale of capital.
Why the application layer is a weaker bet
Software products built on top of today's AI models face a structural problem: the advantage rarely lasts. Model capabilities shift quickly, interfaces are easy to copy, and a clever wrapper around an existing system can be replicated by a competitor, or absorbed by the model provider itself, within a single product cycle. Physical infrastructure behaves differently. A breakthrough in optical connectivity, power electronics, or semiconductor materials rests on engineering and manufacturing knowledge that cannot be cloned by rewriting code. That barrier to entry is precisely what gives infrastructure-focused startups room to build lasting businesses.
The bottleneck runs from the grid to the chip
Expanding AI computing capacity is not only a matter of adding processors. Grid capacity, electricity costs and regulatory approval already determine where new data centers can physically be built. Inside those facilities, the ability to deliver power, remove heat and move data between chips determines how much of the installed hardware can actually be put to use. A processor starved of data because of limited interconnect bandwidth, or throttled because its heat cannot be dissipated fast enough, delivers only a fraction of its theoretical performance. Improvements at any of these layers, from power conversion to photonic interconnects, translate directly into more usable computing capacity for the same investment.
A deep bench of relevant engineering talent
Israel's technology sector has spent decades building expertise in exactly these areas: chip design, photonics, high-speed communications and advanced materials. Engineers who have worked inside companies such as Nvidia, Apple, Intel and Google often see firsthand where current architectures hit physical limits. When they leave to found startups, they are generally not attempting to build another model or another consumer application. They are addressing a specific constraint, a thermal limit, a bandwidth ceiling, a materials problem, that larger systems cannot ignore.
Proving reliability before claiming an advantage
Commercial success in this space depends on more than a strong result in a controlled lab environment. Industrial buyers will not stake a major product launch on a component unless it can be manufactured at scale, integrated into existing systems, and relied upon over years of operation. Yield, reliability and ease of integration are as much a part of the technology's value as its raw performance. That is why early engagement with system builders and industrial design partners matters so much, and why industry gatherings such as November's DeepTech Week TLV, which bring chip designers together with the engineers who build larger systems, serve a genuine commercial purpose rather than a ceremonial one.
Qualification as a competitive moat
Once a component has been tested and designed into a larger product, swapping it out later requires fresh validation and retesting, work that buyers are reluctant to repeat without good reason. That friction, often seen as an obstacle during a startup's first sale, becomes a long-term advantage once a supplier is established. The same rigor that makes initial adoption slow is what protects a qualified supplier from being displaced.
Building beyond the first customer
A first commercial deployment gives a young company a concrete problem to solve and a reference point for investors. But founders and their backers need to look further: whether the core technology applies across multiple systems and markets, not just the one that provided the initial opening. As AI infrastructure providers continue expanding what their systems can deliver, the startups that supply the science and engineering underneath that expansion, rather than the applications sitting on top of it, stand to capture an outsized share of the value created.