Habana Labs

2016: a company set up to design two chips, not one

Habana Labs was founded in 2016 in Caesarea by Avigdor Willenz, who served as chairman of the board and was its first investor, together with David Dahan, who served as CEO, and Ran Halutz. The company raised about $120 million from private investors and venture capital funds, and besides its Caesarea headquarters it had development centers in Gdansk, Poland, and San Jose, California. In 2019, the year it was sold, it employed only about 150 people, a number worth keeping in mind when you read the purchase price below. The idea behind it was that the hardware the industry uses to run neural networks is not necessarily the right hardware for the job. A graphics processor was originally designed only for graphics computations and was adapted to deep learning only later. Habana started from the assumption that you could design a chip from scratch for the task and get better performance per unit of power. That bet was not obvious in 2016. Designing a new chip from scratch costs tens of millions of dollars before a single unit is sold, and the leading competitor already held both the market and a software system that almost every researcher in the world worked with. A small company that enters an arena like that does so on the assumption that it can design better, not cheaper.

Goya and Gaudi: two chips for two different stages of a model's life

Habana's interesting decision was not to build one dedicated chip but two, for two completely different stages in the life of an AI model. Goya, announced in September 2018, was meant for inference: running an already-trained model on new input. This is the operation that repeats millions of times a day once a product is on the market, so what matters is speed and cost per unit. According to the company, Goya could analyze 15,000 images per second, and it was designed to work with the common development frameworks such as TensorFlow and ONNX, so customers would not have to rewrite their models. Gaudi, announced in June 2019, was meant for the other side of the equation: training, that is, building the model itself. This stage needs far more computing power and happens far less often, so it calls for a completely different balance between speed, memory and communication between accelerators. The decision to split into two chips sounds like doubling the development cost, and that is exactly what it was. The logic was that a precise fit for each stage would give an advantage large enough to justify the double price.

Training and inference: why these are two problems at all

This distinction comes up across the whole field of AI chips, and it is worth understanding simply. Training is the process in which the model is built: it is fed huge amounts of examples, and it adjusts its internal parameters again and again until it succeeds at the task. This is a long, very expensive process that takes place in a data center and involves hundreds or thousands of accelerators that must communicate with each other at enormous bandwidth. Inference is what happens afterward: the model is already built, and new input is fed into it to get an answer. Each such operation is relatively small and fast, but it repeats endlessly. So the ideal hardware for each stage is different: for training, raw computing power and communication between chips matter; for inference, low cost, short latency and moderate power use matter. Many companies try to cover both needs with one chip; Habana chose two. The distinction is relevant even to people who are not interested in hardware, because it explains why the cost of training a model is talked about as a huge one-time event, and the cost of inference as an ongoing expense that comes with every single use. The two kinds of cost behave completely differently in business terms.

December 2019: Intel, $2 billion

In December 2019 the company was sold to Intel for $2 billion in cash. It was one of the largest chip deals ever made in Israel, and an especially unusual deal given Habana's size at the time. For Intel this was not buying a product but buying a way in: the AI accelerator market was almost entirely controlled by Nvidia, and Intel, which had dominated the general-purpose processor market for decades, found itself outside the most important growth area in the chip industry. In the years after the acquisition, the Gaudi family did indeed become its spearhead against the competitor, and in 2024 the third generation, Gaudi 3, was introduced. The acquisition was also seen as a milestone for the Israeli chip industry: it showed that giant corporations were willing to pay huge sums for a young Israeli hardware company, not only for software or cybersecurity companies. Two billion dollars for a company that employed about 150 people is a price that reflects positioning, not revenue: Intel paid for a team, for a design and for the chance to take part in the game, not for an existing profitable business.

What actually happened inside Intel

Here the story gets complicated, and it should be told without softening it. According to a Calcalist report, Gaudi 3 chips did not meet the revenue targets set for them, and Intel later decided not to market at all the next-generation chip that had been planned, Falcon Shores. At the same time the company emptied out: almost all the founders, managers and engineers left when the four-year retention periods set in the deal ended. This is a common arrangement in which the buyer ties part of the payment to key people staying. By 2025 Habana had stopped operating as a separate unit and the Caesarea offices were vacated. Intel continues to list Gaudi 3 products and purchase options on its website. The end of Habana as a separate unit is not the same as the end of all Gaudi products. A big acquisition, and a unit that stopped operating separately: this is one of the possible outcomes of an acquisition, and not the one usually talked about. Habana's story has since served as an Israeli test case for the question of what happens to a tech startup after a large corporation buys it, and what happens to the engineering team when the corporation changes direction. It is worth saying outright, because people tend to soften it: a large acquisition is an excellent event for the founders and investors, and it guarantees nothing about the fate of the technology itself. These are two completely separate questions.

Element Labs: same offices, same founders, a new venture

In October 2025 Calcalist reported that the same three founders, Avigdor Willenz, David Dahan and Ran Halutz, were returning to the same office complex in Caesarea with a new chip venture called Element Labs. It develops AI processors tuned for inference and meant for small, local data centers, that is, closer to the users. This choice is not random: it continues exactly the Goya side, inference, and not the Gaudi side. Willenz himself is a recurring figure in the Israeli chip industry. Before Habana he also founded Galileo, which was sold to Marvell, and Annapurna Labs, which was sold to Amazon. The pattern revealed here is worth noticing: engineering knowledge does not disappear when a unit stops operating separately; it moves on with the people. Some of the engineers who left Habana went to other Israeli AI chip companies, and some, including the founders themselves, simply started over. In that sense the story does not end with the closure but, at least for now, comes back to exactly where it started.