Deep33 partner Yael Barsheshet joins Yoel Israel to discuss the infrastructure required for AI to scale, Israel’s deep-tech talent, and why she left a clear path in cybersecurity to pursue a different opportunity.
The potential of AI is no longer difficult to demonstrate. The challenge now is building the infrastructure required to deliver that potential efficiently and at scale.
For Yael Barsheshet, that means looking beneath the applications people use and examining the data centers, chips, memory, interconnects, cooling systems, energy sources, models, and deployment tools supporting them.
Yael is a partner at Deep33, a $180 million US-Israeli fund focused on Israeli founders. In her conversation with Yoel, she explained why the firm invests in the lower layers of the technology stack and where Israeli companies may contribute to the next stage of AI development.
Investing Across Three Coasts
Deep33 operates across Tel Aviv, New York, and California. Yael described this as a “three-coast” model that gives the team direct access to the Israeli technology ecosystem and two major US markets.
The structure comes with a personal cost. Partners must remain available across significant time differences, especially when an investment begins gaining momentum. The team relies heavily on direct communication through WhatsApp to stay connected.
That presence is particularly important for Israeli deep-tech companies. Their technologies may take longer to develop and sell than conventional software, and reaching the relevant customers in the United States can be difficult.
Deep33’s team aims to connect founders with organizations such as AI labs, hyperscalers, and NeoCloud providers. The fund also works with US deep-tech investors that recognize Israel’s potential but may not fully understand its founders or local ecosystem.
Why AI Needs Infrastructure Innovation
Deep33 began with the belief that significant value would be created in the infrastructure supporting AI.
AI adoption is expanding, but that growth is encountering barriers. Companies must consider the cost of running models, the amount of energy their systems consume, the available compute capacity, and whether deployments can produce an acceptable return.
Yael defines AI infrastructure broadly. It includes compute, memory, interconnects, cooling, and data-center energy efficiency. It also extends into models, orchestration, enterprise deployment, governance, and security at the model layer.
Each of these areas can affect whether an organization can deploy AI successfully at scale.
“We all now understand the potential of AI, but there are still bottlenecks,” Yael said. “We still need to make AI with a good ROI for the companies.”
Moving From Experimentation to ROI
Organizations are becoming more attentive to their AI spending. It is no longer enough to demonstrate that a model can produce an impressive result. Companies must compare the value they receive with their token and infrastructure costs.
Yael believes this shift will create opportunities for technologies that improve the economics of AI. The goal is to make models and deployments more efficient for enterprises, model developers, and infrastructure providers.
Energy and memory are among the most visible barriers. Both are necessary to increase AI capacity, and both are constrained.
Yael expects infrastructure innovation to address these limitations during the next five to 10 years, well before more distant concepts such as data centers in space become practical.
Why Energy Has Become an AI Problem
Large technology shifts can create a mismatch between energy supply and demand. AI is producing that imbalance now.
Data-center operators need land, power, and compute. When these resources are connected and available, demand frequently exceeds the available supply. Companies are therefore competing for greater access to both energy and computing capacity.
Over time, Yael expects energy production and compute efficiency to improve. In the immediate term, however, the imbalance remains a significant constraint.
Location will also become more important as AI activity moves from training models toward running them in production. Yael said inference infrastructure has value when it is distributed closer to the places where it is needed rather than concentrated entirely in one location.
What Is a NeoCloud?
A NeoCloud is a data-center provider focused specifically on AI workloads.
Yael explained that some NeoCloud companies previously operated infrastructure for blockchain before shifting toward AI. Their facilities are designed to supply the compute needed to generate AI tokens, especially as hyperscalers and AI labs face shortages.
Some NeoClouds primarily provide access to compute. More advanced providers attempt to move higher in the technology stack by improving performance, margins, and energy efficiency.
Their objective is to generate higher-quality AI output at a lower cost and with less power.
Yael sees an opportunity for Israeli companies in the layer between hardware and applications. Israel has technical talent capable of optimizing workloads and helping data centers produce more tokens while consuming fewer resources.
She said NeoCloud providers are already looking at Israel and attempting to enter or expand within the local market.
How Israel Developed Its Deep-Tech Talent
Israel’s deep-tech ecosystem did not appear suddenly. Yael pointed to companies such as Mellanox and Mobileye as examples that helped develop experienced technical talent.
Successful companies can produce future founders, employees, and investors who carry their experience into the next generation of startups. Cybersecurity demonstrates this cycle clearly, with employees frequently starting new companies after only a few years.
Deep tech follows a slower timeline because the products are more complex and often require greater amounts of capital and research. However, Yael believes Israel now has the talent and ecosystem experience needed to create more of these cycles beyond cybersecurity and software.
The military also plays an important role.
Yael served for 11 years and reached the rank of major in Unit 8200. She began her service in 2007, when cyber was still an emerging field and the terminology surrounding it was different.
“The army teaches you that nothing is impossible,” Yael said. “When there is a big problem and it needs to be solved within certain constraints, you will solve it no matter what.”
She described this as a mindset rather than a purely technical education. Teams are expected to work around limitations, adapt as opposing capabilities change, and find a way forward without accepting that a problem cannot be solved.
Leaving the Expected Cybersecurity Path
After leaving Unit 8200, Yael had opportunities to join early-stage cybersecurity companies founded by people she knew. Given her experience, continuing in cyber would have been the expected decision.
She chose a different direction.
Yael joined Zebra Medical Vision, which applied deep learning to medical imaging. She was drawn to the company’s ambition and the opportunity to work on technology that could influence how patients receive care.
At Zebra, she gained experience across data processing, model training, AI deployment, governance, and the practical requirements of turning a model into a product.
Yael acknowledged that moving away from cybersecurity may have meant leaving money on the table. She remains confident in the decision because it allowed her to pursue deep tech, hardware, and AI while working toward a broader impact.
Learning to Invest Through Experience
Yael later joined Hetz Ventures, which she described as her VC school.
Judah Taub, the firm’s managing partner, told her she needed to meet approximately 500 companies before she would begin developing the necessary judgment. According to Yael, there is no substitute for seeing founders and companies repeatedly.
Investors also need time to observe full cycles. Yael believes it can take four or five years to see how companies and founders respond to changing economic conditions, growth, and difficult periods.
Her investment in Majestic Labs, a semiconductor company addressing the AI memory wall, helped clarify where she wanted to focus. It directed her attention toward the lower levels of the stack, where technical bottlenecks can restrict everything built above them.
Supporting Israeli Deep-Tech Founders
Deep33 combines its infrastructure focus with practical support for portfolio companies.
Its Deep to Market program connects founders with potential customers and partners across the AI infrastructure ecosystem. The fund also follows a “double dip” approach, aiming to connect founders with grants and other nondilutive funding for every dollar it invests in equity.
That knowledge is particularly valuable in deep tech, where development is expensive and founders must understand government programs and alternative sources of capital.
For Yael, the opportunity sits at the intersection of Israel’s technical talent and a growing global need. AI applications may receive much of the attention, but their progress depends on solving the infrastructure constraints underneath them.
