Three weeks ago, Nvidia cofounder, president, and CEO Jensen Huang stood at the podium at the huge GTC Berlin AI Conference. Among his company's intriguing partners, he mentioned Israeli startup AI21 Labs, thereby providing great encouragement to its co-founders, chairman Yoav Shoham and CEO Ori Goshen, after a decade of the company's unsuccessful efforts to penetrate the market.
"Many people at the conference were shocked - we were there as a key partner of Nvidia, not one of the featured companies," says Goshen in his first interview following AI21 Labs' latest restructuring.
AI21 Labs was founded almost a decade ago by Mobileye founder Prof. Amnon Shashua; Prof. Yoav Shoham, a pioneering AI researcher at Stanford University; and young CEO Ori Goshen, long before ChatGPT ushered in the new AI era just four years ago. The Israeli startup was one of the few companies that dared to compete with Anthropic and OpenAI by developing a large multipurpose language model. AI21 Labs was supported by several leading Israeli AI specialists and prestigious investors, among them Google and Nvidia.
AI21 Labs had great ambitions, but the company has switched its business focus four times since it was founded in 2017. Most recently, it has announced a new sphere of activity: optimization and management of AI agents. The company was nearly acquired at one point. Meanwhile, last June, Shashua left AI21 Labs after founding doubleAI, an exciting new AI company, with other partners.
This time, Goshen, 41, and Shoham, 70, feel that they are on the right track. "Suddenly, we feel that we've found it," says Goshen. "You feel it in talks with customers and in the whole dynamic around you." Shoham adds, "We've never had anything like this. We've had responses before by researchers or reporters, but we never felt such energy from the market."
"An emotionally difficult meeting"
The search for a path forward reached a fateful crossroads last May. AI21 Labs laid off 60% of its staff - 110 of its 180 employees, including most of its marketing, sales, operations, and product workers, plus several AI developers. The remaining employees were in the company's core business. Goshen says that many of these employees chose to stay with the company despite receiving other job offers. The decision was taken after a large company (believed to be Nebius Group) contacted AI21 Labs and attempted to acquire it. After lengthy discussions, however, the offer did not go through; instead, the two companies maintained their commercial relations.
"It was a huge mess. There were all sorts of rumors flying around and things appeared in the press and were also seen by the employees," Shoham admits. "Then we had a company meeting that was emotionally difficult for all of us because we had to tell people who had contributed so much to us, 'What you heard is incorrect' and then we had to tell them, 'Two thirds of you must leave the company, but not because you did anything wrong - you did something good.'
"It was the worst hour I've ever experienced in my whole life. I had tears in my eyes. I started the meeting and let Ori (Goshen, A.G.) conduct most of the conversation. They told me to say that it was a complicated meeting, but it wasn't a complicated meeting - it was a very hard one. They say that there are five stages of grieving. We had to go through all five of them in a single hour.
"Fortunately, many people took it extremely well. It was done respectfully and transparently. People thanked us for the way we did it. We also helped the employees find jobs."
Since it was founded, AI21 Labs has been called the most promising AI company in Israel that eventually proved a disappointment. This was unquestionably a heavy burden to bear.
Shoham: "A lot of what you say is true. We'll see whether we're the leading AI company in Israel, but we're certainly no longer the 'new sexy guy in the neighborhood.' Many things have been written about us, some correct and some incorrect, but what we've built up until now is a kind of validation that everything we did has value and everything suddenly happening now is because of it."
In retrospect, was it a mistake to compete with Anthropic and develop your own language model?
Shoham: "Talking about 'what if' is theoretical and doesn't matter much, but since large language models are now at the core of AI, we developed a muscle that we wouldn't have had if we hadn’t developed our own language models. The infrastructure we developed in order to do this is enabling us to do what we're doing now. Developing the model requires an enterprise, and we built that enterprise."
Still, that gamble was very costly for you. You eventually had to downsize the company.
"Let's distinguish between two things. I think that it was clear that something had to change long before our major downsizing. It was obvious that selling models was just not good business. Earlier in the company's history, when we had already put out Jurassic, we decided against building a chat bot (i.e. not to issue a product in competition with ChatGPT or Claude, A.G.), not because we were unable to do it, but because it was a Sisyphean task. There are a great many cases that must be taken into account and it's also expensive to raise capital for it. We raised as much as we needed, not ten times as much.
"So we decided that this wasn't us. Perhaps we could have acted differently, but that was what we decided. As soon as we realized that selling models wasn't a good business for us, we knew that we also had to change the work mix, and the personnel mix as well."
The chat that changed everything
The original idea of Shoham, Goshen, and at the time also Shashua, was to develop AI and a product that is now taken for granted, an application named Wordtune that provided proofreading and rewrote texts. The program achieved initial success and the company completed a $7.5 million seed round in 2019 with participation from venture capital funds such as Pitango and TPY Capital. Shashua himself invested several million dollars. He later brought in Lip-Bu Tan, now CEO of Intel, but at the time known as a seasoned venture capital investor, whose venture capital fund, Walden Catalyst, led AI21 Labs' $20 million B financing round in November 2021 at a company valuation of $400 million.
ChatGPT, launched in late 2022, upset everyone's plans. OpenAI's chat, followed by those of Anthropic, Google, and Microsoft, offered for free what the Israel company had been selling, plus a lot more. AI21 Labs then adopted a different, more ambitious strategy: direct competition against these companies and developing language models from scratch. AI21 Labs' founders believed that companies and organizations were afraid to adopt AI technology because of inaccuracies and concern that information would leak out. Developing a new model based on technology that the company had already developed was designed to help AI21 Labs to reach markets that Anthropic and OpenAI were finding difficult to penetrate.
At this point, AI21 Labs launched models that received worldwide recognition: Jurassic, followed by Jamba. In the summer of 2023, Nvidia, Google, and Intel took part in a $200 million AI21 Labs financing round at a $1.4 billion company valuation.
It soon turned out, however, that French company Mistral AI, Canadian company Cohere, and dozens more companies, were also competing in the niche of models for organizations. When Chinese open free-of-charge models appeared, the race proved to be too competitive and expensive.
While its competitors were raising billions of dollars in a short time for the purpose of buying more and more graphic processors for development, AI21 used what it had and claimed that its models were focused. "Money is important, but it's not everything," Shoham said in the "Globes" Hatzolelet (The Submarine) podcast in March 2025, about a year before AI21 Labs began its wave of layoffs. "Our technology, which is extremely smart, doesn’t require that kind of wastefulness."
That very month, however, the company changed its focus for the third time by launching Maestro, a system that enabled companies to manage many models and the AI agents resulting from them, AI-based software capable of carrying out tasks and assignments. "Yes, this is the future of the industry," Shoham declared in that March 2025 interview. "The various models will always have some degree of inaccuracy. So what’s missing is an AI system that adds oversight and increases reliability."
A year later, AI21 again changed focus, and Maestro is no longer the main product in the new strategy that Nvidia endorsed last month. "The market is moving so fast and becoming so complicated," Shoham explains. "Maestro had many components, one of which is the thing that we're now putting on center stage - the process that is facilitating the constant improvement of AI agents."
"We studied the market’s needs"
At the GTC Berlin AI Conference, Huang spoke about the Israeli company for just under a minute. He talked about the new product that helps Alice, another Israeli company (which protects Anthropic or Meta against malicious use of their agents) train, improve, and upgrade its language models. With the help of AI21 Labs, Alice operated an internal language model that outperformed even Anthropic's models.
The Nvidia CEO thereby revealed to the world a new niche in the AI sector, called "post-training" or " lab-as-a service." In this process, companies take existing AI models and continue training them with the specific information and procedures of their customers. The result is an internal language model tailored to the customer that can be more resistant to cyberattacks and can also attain better performance than the off-the-shelf models available to everyone from Anthropic or OpenAI.
AI21 Labs achieves this result by using open language models, some of them available free-of-charge and others from Nvidia itself. The process is designed to improve the special models developed by its customers, among them AI companies Gong and Deel.
"We go to companies that want to control their AI, enable them to train their own models, and ensure that they are safe and calibrated for optimum performance, so that the potential is fulfilled. In effect, this is our specialty," Goshen explains.
AI21 Labs guarantees to reduce the cost of using AI agents to one fiftieth in comparison with Anthropic or OpenAI after the "training" in every possible area, from customer service to finance. Among other things, it achieves this through the growing use of cheap open language models (among them those of Chinese company DeepSeek and Meta's Emma). Use of these models has grown from 15% three months ago to 80% at present.
Goshen and Shoham talk about the demand for their new product, which they say is unlike anything that ever happened before at AI21 Labs. For Shoham, the long and exhausting journey has paid off: "All that code that we wrote and the muscle that we developed was designed for the precise purpose for which we are now using it - optimization of AI beyond the models. Unwittingly, our long journey prepared us for this task."
AI21 Labs is currently regarded as one of the three main companies in the post-training market. Its competitors are San Francisco-based Applied Compute, which raised $350 million this week at a company value of $3.25 billion (its investors include super-investor Elad Gil and former basketball player Omri Casspi); and Thinking Machines (also San Francisco-based), cofounded by OpenAI cofounder Ermira "Mira" Murati, which has raised $2 billion at a company valuation of $10 billion.
AI21 Labs differentiates itself by offering value that includes processing services and training of the AI agents, but also a security package and management of these agents. Like other companies in the AI market, AI21 Labs has transitioned to a service model that includes placing employees on the customer's premises. According to Shoham, the model narrows the company's profit margin but increases its revenue.
You’ve been through so many incarnations - are you finally on the right track?
Shoham: "We have always been a deep technology company, but I think that we weren't paying precise enough attention to the market. We built exciting technology and in the early years of AI there was a kind of privilege in the market for carrying out experiments like a laboratory. Today there is already a ‘responsible adult’ market, and in the meanwhile we have learned to think about what the market needs. It turns out that the nine years we spent developing the platform and the knowledge have enabled us to meet a need that people like the CEO of Microsoft and the CEO of Palantir Technologies are also talking about."
Where did Amnon Shashua disappear to?
From its earliest days, the company's third and most important leg was Shashua, who led Mobileye to a $15.7 billion exit in a sale to Intel and later to an IPO. He regarded AI21 Labs as a groundbreaker in language models in the pre-transformer era (the operational method underlying the models now being sold, such as GPT Gemini and Claude). His name is still featured in press reports and is linked to AI21 Labs, but Shoham says that Shashua has not worked at the company for a long time (sources inform "Globes" that he stopped working there at least three years ago), despite his position as unpaid chairman - until he officially left the company last June.
Before Shashua left, Goshen and Shoham were co-CEOs; Shoham has now become chairman. Mentioning Shashua's name in conversation seems to arouse painful feelings in Goshen and Shoham: silent moments of embarrassment, a feeling of discomfort, excessive caution, and uneasy shifting in their chairs. "These are natural processes. An entrepreneur becomes less involved and less familiar with what is happening - and that's fine," Shoham eventually says. "He became more distant, and at a certain stage something no longer interests you, and then you continue on your way - it was healthy for everyone.
"Amnon was very involved in the early years and we very much appreciated it then and we appreciate it now, but in recent years, he hasn’t been involved. What he's now doing in his company - optimization of processes and scientific thinking - is not something that we do here."
In 2023, together with a number of academic researchers, two of whom were AI researchers at AI21 Labs, Shashua founded doubleAI, which recruited several of AI 21 Labs' investors. Last year, Nvidia participated in a investment round by Shashua's new company at a valuation of $1 billion. Shashua announced several achievements this year that position doubleAI as one of the most promising AI companies.
The choice of Israel and the disappointment
Goshen and Shoham are not without regrets. They say that one of their mistakes was their company's excessive presence in Israel (most of the marketing, sales, and business development staff). This contrasts with the currently prevailing situation in which all the most highly-regarded AI companies not only operate in California, but in San Francisco itself.
"San Francisco is now the capital of AI. The market is there, a presence there is important, and someone there sees how fast everything is moving," Goshen admits. "In retrospect, a presence there is critical, and I say this today to every entrepreneur I meet on every question of AI infrastructure."
At the same time, they deny various reports, such as a $300 million financing round shortly before the wave of layoffs and an attempted sale of the company to Nvidia. "A lot of things have been written about us and we have no idea where they came from," Shoham says.
They also fail to respond to the reports of an attempted acquisition by Dutch neocloud company Nebius Group, whose founders are Israelis. It is believed that such an attempt was made but was unsuccessful, which affected the timing and extent of the layoffs at AI21 Labs. Shoham and Goshen say laconically, "We said that Nebius was our commercial partner. We are enriching their platform with optimization of AI agents."
Did the decision not to raise capital harm you? French company Mistral AI has raised $7.5 billion at a $24.4 billion valuation and Canadian company Cohere has raised $1.6 billion at a $7 billion valuation. Unlike you, they have succeeded in growing and penetrating the enterprise market.
Shoham: "Most of Cohere's revenue is attributable to the Canadian government and its largest investor is also involved with the state. The same is true of Mistral AI. I'll give you an example: not long ago, I was invited to an event in the Élysée Palace. There were 200 people in the hall and in front were French President Emmanuel Macron seated next to Maurice Levy (chairman of the Supervisory Board of Publicis Groupe, A.G.). Next to him was Jensen Huang and Mistral AI founder Arthur Mensch. That just shows you how Macron is promoting the French AI company."
And no one in Israel is pushing you forward.
"Let's begin with the fact that nothing like this exists in Israel. This is the message I have been sending for years to the Israel Innovation Authority, the prime minister, and everyone else who needs to hear it. I realize that there are budgetary, political, and other constraints, and I also understand fairness and the requirement to conduct tenders. But if Israel wants to be competitive as a country, it has to gamble on specific companies and promote them."
Was a tie between the state and AI21 Labs on the table?
"It's something that I put on the table, but no one took it up."
In the summer of 2023, Shoham was appointed to head a scientific advisory committee for artificial intelligence attached to Telem (the National Infrastructure Forum for Research and Development), a government organization headed by the Israel Innovation Authority, which had the authority to promote a national artificial intelligence program - authority that has never been effectively exercised.
"Every country in the world has a national AI program, but we in Israel are special - we have three or four AI programs, all of them good," says Shoham. "They asked me to establish a scientific advisory committee and head it, and I did so. I think it exists, but I haven't been invited to it recently."
Despite his cynicism, Shoham regards the new National AI Directorate in the Prime Minister's Office as an encouraging sign. "This is a significant and satisfactory step with larger budgets than in the past. Until the budget is approved after the election, however, nothing will move. When it is approved, I think more serious measures will be taken to implement it," he says.
"Open Code is a tsunami"
If even you acknowledge that companies are switching to open free-of-charge models, what does the future hold for Anthropic and OpenAI?
Shoham: "I don't think that the closed models will disappear. They have groundbreaking capabilities that others don't have yet. For the bulk of users, however, open code is a tsunami. I think that one scenario, and I'm not speaking from knowledge, is that they will become more vertical companies working in specific areas, not merely offering the model. In the recent past, Anthropic has spoken about entering the fields of science, biology, and law."
In recent weeks, Google's share plunged 12% because of an unending increase in commitments to acquire AI equipment and a cash flow that is becoming negative. Elon Musk has predicted a dim future for neocloud shares, and Anthropic's prospectus revealed $900 billion in commitments for the coming years. When AI revenue and profits fail to match this pace, it’s an alarming trend that is liable to cause panic and take many stocks down with it.
Goshen: "I think that there are arguments on both sides. Is AI a bubble, or is it propelling the global economy? The growth rate is after all unprecedented. I think that the only relevant question is whether the growth rate will continue the way it has over the past year. If it doesn’t, we have a problem. At present, however, the market is gambling that it will continue and even accelerate.
"In this sense, the core of the gamble lies in the investments by the giants in capital equipment, which will soon total a trillion dollars. These companies were profitable companies that are allocating their entire free cash flow to this build out. There's a gamble here by the industry that this is the future. I think only time will tell."
Published by Globes, Israel business news - en.globes.co.il - on November 11, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.