
Palo Alto's $4.5 billion acquisition exposes a missed opportunity for Israeli AI startups
The cybersecurity giant is moving into the business of refining and adapting AI models, a field that Israeli investors say should have been a natural strength for the local tech industry. Several promising teams are now trying to close the gap.
Palo Alto Networks’ $4.5 billion acquisition of Applied Compute caught the market by surprise, not only because of its price tag but also because the target company operates outside the traditional cybersecurity sector. Rather than acquiring another security firm, the cyber giant is entering the field of AI infrastructure, specifically technologies that enable organizations to develop and refine AI models tailored to their needs.
The deal reflects a broader strategic shift among major technology companies. Rather than relying solely on models provided by companies such as OpenAI and Anthropic, businesses are increasingly seeking to develop their own capabilities and gain greater control over the AI systems that power their products and services. For Palo Alto Networks, this could mean developing models specialized for cybersecurity, improving their performance while reducing dependence on external AI providers.
Yonatan Mandelbaum, a partner at venture capital firm TLV Partners, says the acquisition is linked to the emergence of an industry focused on refining existing models and adapting them to organizations’ specific needs. A key differentiator, he says, is access to a model’s “weights,” the parameters established during training that determine how it operates.
“When using closed models, such as those from OpenAI or Anthropic, you lack access to their weights, making it impossible to make independent adjustments. In contrast, open models allow companies to take an existing model, refine it, and tailor it to their specific needs.”
He adds: “If you are a startup or an enterprise, like Palo Alto Networks, that wants to retain ownership of its organizational intelligence, you will likely seek the optimal balance between cost and performance: using the most affordable model possible while ensuring that it delivers performance comparable to the world’s most advanced models. An entire industry is emerging around this need.”
Adapting open models to an organization's requirements involves complex technological infrastructure, including data collection and preparation, model refinement and operational optimization. The process requires research and engineering expertise that most companies do not have in-house. As a result, an ecosystem of service providers has emerged, offering everything from training data to the infrastructure needed to run these models. Mandelbaum sees Palo Alto's acquisition of Applied Compute as a sign that post-training, the process of refining models after their initial training, is developing into a substantial market of its own.
Mandelbaum believes Israel had a significant opportunity in this field. “On paper, this is a category in which we ought to excel, precisely because of the caliber of talent available here. Yet, as is the case with many AI categories, we are somewhat late to the party. A clear illustration of this is the fact that Palo Alto Networks acquired Applied Compute rather than an Israeli team specializing in model optimization or adapting models for the cybersecurity domain. In my view, this was definitely a missed opportunity.”
However, he adds, “It is not too late. There are several exceptional teams in Israel currently tackling these challenges from various angles.”
Ori Goshen, co-founder and CEO of AI21, which develops AI models for enterprise use, also sees the deal as reflecting a fundamental shift in how companies approach AI: away from dependence on third-party providers and toward greater ownership and control of the models themselves.
“An organization that procures AI services pays a ‘double cost’: once for the actual use of the models, and a second time through the missed opportunity to enhance that AI over time. In a landscape where AI infrastructure is increasingly concentrated in the hands of a few, controlling the model, and its continuous improvement, becomes a key competitive advantage.”
According to Goshen, companies are gradually realizing that simply having access to AI is not enough. The real advantage lies in accumulating organizational knowledge, tailoring models to a company's specific operations and continuously improving them. Instead of repeatedly paying an external provider for each use, a company can build a technological asset that serves it over the long term.
Mandelbaum points to two major questions that remain unresolved in this market. The first is whether model improvement and model execution will eventually converge into a single field. Applied Compute itself generates a significant portion of its revenue from running the models it has helped train, while many AI infrastructure providers have begun offering model adaptation and improvement services.
The second question is how much customers actually care about the technology behind the process. “A client doesn't necessarily care about post-training, fine-tuning or model optimization. What they want is the best possible performance at the lowest possible cost. They don't really care which technique is used to achieve that result.” The challenge, he says, is convincing customers to choose one provider over another.
Ultimately, Palo Alto's acquisition of Applied Compute is more than another major technology deal. It signals that the AI race is entering a new phase. After the push to develop ever-larger models, companies are now competing to make those models more efficient, cost-effective and suited to the needs of individual organizations. For companies such as Palo Alto, mastering these technologies could become as important as their core cybersecurity products. For Israel's tech industry, the shift represents a significant opportunity, but also a warning: the market is developing rapidly, and there is no guarantee that Israeli companies will capture the position the country's talent might otherwise allow them to claim.
















