In December 2020, Amazon announced in a fanfare that it will use Gaudi chips from small Israeli startup Havana Labs to train large language models (LLMS) in the cloud. Generation AI was not yet a home term at the time, but it was a major development in ecosystems. Intel, which acquired Havana the previous year for $2 billion, boasted about the event, calling it “the first crack in Nvidia’s advantage in GPUs.” At the time, Intel and Amazon were dominant players in the market, but Nvidia was mostly known to gamers who praised the graphics capabilities of its processor. But Inter’s concerns about his younger, ambitious competitor were already clear. A few years later, it was hard to imagine Nvidia going to trade at a valuation of $3.5 trillion, as the world’s largest company, but Intel is worth just $80 billion.
Last week, in addition to reporting disappointing financial results, Intel announced that the next-generation Habana processor, Falcon Shores, will not be sold commercially as it received negative feedback from customers. This follows a previous announcement that Gaudi failed to meet its expectations of reaching $500 million in revenue by 2024. As a result, Intel has decided not to develop a different generation of Gaudi beyond the existing Gaudi 3. Habana Labs’ failure failed as another entry in a long history of failing to acquire Intel. This result may not be surprising to Intel, but it is a rare event in the track record of Avigdor Willenz, the Israeli entrepreneur behind Havana.
1 View the gallery
Habana Labs founders Avigdor Willenz and Intel Offices.
(Photo: Eyeal Toueg and Nir Keidar))
Considered a genius in the semiconductor world, Willentz has a long history of successful exits. His achievements include Galileo’s $2.7 billion sale to Marvell in 2000. He later sold Annapurna Labs to Amazon, selling a relatively modest $350 million. More recently, he is one of the founders and early investors of Astera Labs in the US, published at a valuation of $5.5 billion last year and is now worth $16 billion. Given his record, Habana Labs’ failures are rare failures, raising questions: what went wrong? The simple answer is probably the struggle between itself and its well documented acquisition.
Even before the negative announcements about Gaudi and Falcon Shores, Habana Labs was already falling apart. By 2024, almost all of its founders, managers and engineers had left Intel and most of them left shortly after its four-year retention period ended. Many were not waiting for Intel’s voluntary retirement plans. Today, most of Habana’s original hardware engineers are not left at Intel. Recognizing the challenges involved in integrating acquisitions, Intel had previously experimented with maintaining the acquisitions as separate units. This approach was applied to Mobileye until it determined that the structure was ineffective and that it was running as a public company. Habana was held as a separate unit until last year, when it was absorbed into existing Intel operations.
At first, I was as optimistic as usual. Unlike some of its past failures, it correctly identified AI as the future, as if it were missing out on the mobile revolution. Five years ago, the company predicted that by 2024 it would generate revenue from its AI processors at $25 billion. This is a goal that is far from reality. As part of the AI push, Intel has acquired startup Nervana for $400 million. Nervana specialised in processors for training language models, but Intel quickly realized that it was betting on the wrong technology. In 2017, seeing Nvidia’s rapid advances in GPUs, Intel recruited former AMD executive Raja Koduri to lead the AI efforts. However, customer feedback was unfortunately disappointing. Intel’s processors were insufficient and could not meet market demand.
To demonstrate its commitment to AI, Intel sought dramatic moves. They tried to acquire Melanox, but refused to pay the $7 billion that Nvidia ultimately offered in a deal that surprised the industry. Nvidia CEO Jensen Huang had a clear vision for the strategic importance of Mellanox, but Intel hesitated. After the loss, Intel turned to the second name its customers were recommending for Habana Labs, a competitive AI processor. Habana, founded in 2016 by David Dahan and running Halutz, a veteran of Primesense, the Israeli company behind Apple’s facial recognition technology, already had a functional AI training processor, Gaudi . Unlike Nvidia’s GPU, Gaudi is based on a different architecture, with Amazon later claiming it was 40% more efficient than Nvidia’s offering. At the time, Amazon was rumored to be interested in buying Havana, but Inter, hoping to recover from the Melanox fiasco, was willing to pay more. Intel completed its acquisition in 2020 amid the height of the Covid-19 pandemic, and simultaneously shut down Nervana.
Nvidia completed the acquisition of Mellanox almost simultaneously, but the results of the two Israeli companies diverged significantly. Nvidia has successfully integrated Mellanox. Mellanox currently generates more than $13 billion in annual revenue and employs thousands of people in Israel. Meanwhile, Habana Labs was surprised under Intel’s leadership.
A source who spoke to Calculist about Havana’s downfall unanimously agrees that Intel mismanaged the acquisition by pursuing multiple competing AI strategies without a complete commitment. Intel believes that it has acquired Havana just to hide the nervana failure and signal to investors that it is investing in AI, with the intention of challenging Nvidia through Habana. This was evident in Intel’s organizational structure. Havana was not placed under Koduri’s GPU division (AXG), but instead under the Data Platform Group (DPG). Koduri, a well-known Intel employer, reportedly had a significant influence and received a large portion of his development and marketing budget.
“From the moment Havana acquisition was completed, people within Intel couldn’t understand why they run both Havana and the GPU division, which are developing competing architectures,” said a former Intel executive. Ta. The former Havana employee also described Intel’s bureaucratic inefficiency as a major obstacle. “In Havana, we were able to make decisions through a five-minute corridor conversation. At Intel, the same decision requires three meetings with dozens of participants, and nothing has been made. That’s what it was,” recalled the former Havana employee.
Until 2022, Intel has pursued both strategies simultaneously, developing a competing GPU, the Ponte Vecchio, while selling Gaudi processors. However, with the rise of ChatGPT and other generator AI models, Nvidia’s advantage cannot be denied, and Intel has once again received negative customer feedback. Ponte Vecchio was cancelled in 2024, just two years after its launch, and later that year, Intel announced that it would not develop a new Gaudi generation. Following the departure of Koduri in 2022, Intel has integrated Habana and its GPU division to develop a new AI processor, Falcon Shores, a hybrid chip that combines GPU (like NVIDIA) and CPU (Intel’s Specialty). sought to integrate efforts to develop. In Havana, the move was filled with skepticism and gentle humor. “All of a sudden they remembered us,” he joked.
Currently, we see that Falcon Shores is unable to meet expectations and is only used for internal Intel testing. The company is shifting its focus to the new chip, Jaguar Shores, and is trying to jump over several generations to catch up with nvidia. However, industry experts believe it is nearly impossible to fill the gap. Nvidia has spent the last 20 years refined AI processors, but Intel recently arrived late like a dull hippo in a sprint contest.
Meanwhile, the former Havana team, along with Willentz, has already moved from its new Tel Aviv office to the next AI venture. They can’t complain much – Intel was paid in cash, and estimates suggest that 100 employees from Willentz, founders and startup received roughly $1 billion in bulk It’s there. But despite the economic downturn, Havana’s failure is a rare scar on Willentz’s record and, more importantly, an opportunity missed for Israel. A better execution would have allowed Havana to be the foundation of Intel’s AI strategy, rather than another victim of its problematic acquisition history.
Intel says: “Our AI opportunities are focused on solving the most pressing challenges for our customers, particularly reducing computational costs and increasing efficiency. We provide the most compelling total cost of ownership across this continuum. With the goal of doing core and existing assets.
“Following customer feedback and market dynamics, we plan to use Falcon Shores as an internal test chip and concentrate resources to bring Jaguar Shore to the market instead. Feed directly on Jaguar’s coast. I’ll give you.”
