The billions required to build a frontier AI model or a hyperscale data center have convinced many founders that the AI boom belongs only to companies with deep pockets. That assumption is wrong. The infrastructure beneath those systems is riddled with bottlenecks in power, cooling, connectivity and materials science, and a company that solves even one of them can build a durable, highly valuable business without ever training a model of its own.
Where the real scarcity lies
Software built on top of today's large language models faces a harsh reality: capabilities that look distinctive this year can be absorbed into a competitor's product, or into the model itself, within months. Physical infrastructure behaves differently. The laws of physics governing heat dissipation, electrical resistance and signal loss do not change with each new model release. A company that develops a genuinely superior way to move data optically, deliver power efficiently, or manage thermal loads is protected by engineering difficulty that a rival cannot simply code around.
That difficulty runs from the electricity grid down to the transistor. Where a data center can be sited depends on grid capacity, local energy prices and regulatory approval, all of which are increasingly scarce in the regions best suited to AI deployment. Inside the facility, the amount of installed computing power that can actually be used depends on power delivery, cooling and interconnects. Processors left idle while waiting for data, or throttled because heat cannot be removed fast enough, represent wasted capital on an enormous scale. Materials and chip architecture set the outer limits of what is possible before any of that infrastructure is even built.
An engineering base built over decades
Israel's position in this landscape did not emerge overnight. Chip design, photonics, high-speed communications and advanced materials have been areas of sustained national strength for decades, producing engineers who understand both fundamental science and the operational demands of systems running at enormous scale. Many of them have worked inside Nvidia, Intel, Apple or Google, where they encountered the precise points where existing architectures hit their ceiling. When these engineers leave to start companies, they are rarely building another interface for an existing model. They are addressing the physical constraints that limit what every model, present and future, can do.
From laboratory result to qualified component
Technical superiority alone does not win contracts. No company building a major product will risk its launch on a component that performs well in controlled laboratory conditions but cannot be manufactured at scale or integrated at a workable cost. Manufacturing yield, long-term reliability and ease of system integration are as central to a technology's value as its raw performance. This makes early engagement with industrial buyers and design partners essential, since only real operating conditions can validate whether a component is ready for deployment. Industry gatherings such as November's DeepTech Week TLV, which bring chip designers together with the system builders who will eventually use their work, reflect how central this validation process has become.
That same rigor, once cleared, becomes a form of protection. A component that has been qualified and designed into a larger system is costly and risky to replace, since doing so requires redesign and renewed testing. The difficulty of winning the first sale is precisely what makes a proven supplier valuable for years afterward.
Thinking beyond the first customer
An initial customer gives a young company a concrete problem worth solving, and a reference point for everything that follows. But the technologies attracting the most durable investor interest are those whose underlying science applies across multiple systems and markets, not just the one that provided the first contract. Founders who secure early commercial traction without letting a single customer's specifications define their ceiling are the ones best positioned to grow alongside, rather than beneath, the infrastructure giants.
The companies spending the most on AI infrastructure will keep expanding what their systems can deliver, and they will need better materials, better connectivity and better power management to do it. Some of the largest returns in this cycle are likely to come not from the platforms themselves, but from the deeptech companies supplying the engineering those platforms depend on.