When Ofer Shacham left his role leading chip development for Meta’s metaverse AI glasses, he joined forces with two other senior executives, Shahriar (Sha) Rabii and Masumi Reynders, to found Majestic Labs, a semiconductor startup taking on Nvidia from an unexpected direction: developing a graphics processor designed to outperform Nvidia’s chips in AI processing efficiency.
Nvidia’s famous GPU actually has a relatively limited onboard memory, making processing difficult when running very large language models (LLMs). Shacham’s new venture seeks to change that. "We want to bring the AI message to everyone, not just the cloud giants," he tells "Globes" from the company’s recently opened offices in Ra’anana.
Majestic Labs raised over $100 million in November 2025. About a month ago, it unveiled its products in "Globes": a graphics processor called Ignite based on intellectual property from British company ARM, the open source RISC-V platform, a proprietary technology developed by Majestic, and Prometheus, a server densely packed with memory modules that give each chip access to extensive memory resources for computation, facilitating more efficient AI-processing, and positioning it as a potential challenger to Nvidia.
To appeal to software developers already accustomed to working with Nvidia-based systems, Majestic enables AI engineers to develop applications using the popular PyTorch programming framework and to run them on OpenAI’s operating system, which has in recent years emerged as Nvidia’s main rival within the AI ecosystem. According to Majestic’s plans, the servers and chips will be launched in mid-2027 and will enter production in the coming months.
Majestic Labs’ server is still under development, but according to industry estimates, the company has already received orders worth several hundred million dollars on paper for its servers, which are expected to launch next year. These servers are said to be capable of performing up to 50 times more AI processing operations, or "tokens," per megawatt of power consumption. "I have a customer currently building a data center with a power capacity of 500 megawatts. The price of the server is less important to him than the number of tokens he can generate per megawatt, and I can deliver up to 50 times the industry standard," Shacham says.
"This isn’t a market with lots of customers buying small amounts. It’s a market with just a few customers - each one spending hundreds of millions of dollars. We estimate that dozens of customers will start making orders in the hundreds of millions of dollars as early as next year," he adds.
Nvidia's problem
Shacham is saying what many in the industry think but are afraid to say out loud - that buying more and more Nvidia servers and chips will lead to an energy crisis. "The most powerful LLMs, like Gemini, Claude or GPT, are struggling to operate with the limited memory of each GPU. To compensate, companies have to buy more and more processors and servers, most of which remain underutilized most of the time because of their memory constraints, or what's known in the industry as the 'memory wall' - the processors sit idle for much of the time, consuming energy while waiting for data to be fed into them," says Shacham.
For the five cloud giants alone - Amazon, Google, Microsoft, Meta, and Tesla - this problem dictates huge capital expenditures, which amounted to $443 billion last year and are already forecast to rise to $700 billion this year. "And with the increasing number of edge models being released by companies like Anthropic, OpenAI, and Gemini, companies are being forced to buy more and more servers to accomodate this enormous processing power."
The three entrepreneurs are well-known in Silicon Valley and are believed to have good relations already with some of the largest US-based AI model companies, some of the largest cloud providers, and especially the neo-cloud companies that build AI server farms and lease processing services to the cloud giants and other companies. These may be the first to adopt Majestic's technology.
A different breed
Shacham and his American partners are a different breed in the Israeli startup scene; each has at least 20 years of career experience managing hardware labs at tech giants. Shacham, who until the beginning of the previous decade worked in IBM's hardware team in Israel, just wanted to study to complete his MA in electrical engineering at Stanford. Those studies led to a doctorate and took him to the university's chip design laboratory.
When his supervisor, Prof. Mark Horowitz, left for a sabbatical year, Shacham, then in his early thirties, had the opportunity to run the lab where he served an intriguing client: the Defense Advanced Research Projects Agency (DARPA), the R&D branch of the US Department of Defense. DARPA had come to Stanford with a request to develop highly efficient innovative processors that could perform intelligent image processing in mobile edge devices. "That was before the AI era. Back then, we were talking about neural networks, machine learning, and computer vision," Shacham says.
In parallel to his work at the lab, Shacham founded his own company that applied some of the principles he had learned to develop efficient chips. Within a short time, he signed an interesting customer: Google, which in 2011, after acquiring Android, began to examine the subject closely. The person who managed that area at Google, Sha Rabii, acquired Shacham's company, became his direct manager, and over the years moved with Shacham, first to Meta and then to co-found Majestic Labs with him.
In 2013, Shaham and Rabii began leading the development of the first AI processors for Google’s new smartphones, which would later be released under the Pixel brand name. The two were joined by a lawyer named Masumi Reynders, who turned out to be a commercial talent. Reynders became the third member of Shaham and Rabii’s team, serving as a senior product manager across their various activities, and Majestic's third co-founder.
"We did things that Google didn't believe could be done. I showed our roadmap to one of the senior managers at Chrome - he looked at me and said, 'Are you going to put all that on a phone?' It's totally crazy, but I like doing crazy things - just hire five more people and do it. We started with five people, and when I left, there were 300 of us. Today, when I see Waymo's autonomous cars, I know that part of the LiDAR radar chip was developed in our lab. I know that some of Gmail’s security features run through chips we designed, and that every YouTube video passes through software we built for video compression on those chips."
Not worried about the competition
After years of working on chips for peripheral devices like smartphones, glasses, and VR helmets, they moved on to work on a different type of device: an AI processing server. "We realized the real problem was where you need to run language models that are growing at a staggering pace. Since around 2020, they've been scaling exponentially - from 100 million parameters to a billion, then to 100 billion, and now toward a trillion parameters. At the same time, the hardware needed to run these models simply hasn't been advancing at the same pace.
"It's the memory limitations of existing hardware. Until recent years, there was no software that forced you to buy a $5 million server rack. Today, every new model release requires you to keep increasing your investment in hardware. That’s not sustainable, and it calls for a new architecture."
Shacham isn't worried about the growing competition with new chip companies like Cerebras Systems, which successfully went public last month after launching a chip to power AI agents with a high memory rate, or the resurgence of Intel and AMD thanks to the growing demand for their core processors in AI processing facilities. "I don't see competition as a problem," he says, almost dismissively. "On the contrary." In his view, the fact that Nvidia still owns about 90% of the market only illustrates how open the arena is, and how much room there is for other players to offer other approaches to solving other problems.
"The success of companies like Cerebras proves to data centers and huge customers that there is more than one way to build an AI infrastructure. This world is so versatile and so vast, that different solutions exist for different problems," he says.
And this, in his view, is exactly the point: the market is not shifting into a world of a single replacement technology, but into an era of heterogeneous computing. A world in which several types of processors operate within the same data center, each optimized for different tasks. Some kinds of workloads still fit Nvidia’s architecture very well; others will be better suited to new architectures.
As far as he is concerned, the scale of the opportunity has long since gone beyond all reasonable projections. When the company launched in 2023, he notes, the market was forecast to reach $80 billion in 2024, and half a trillion dollars by the end of the decade. In fact, he says, the market has already raced far beyond that. "It's a crazy market, and we're in it," he says.
When the five largest companies alone are pouring massive sums into infrastructure, it no longer really makes sense to assume that a single player will provide all the solutions. On the contrary: every successful new company pushes the market forward, proves what's possible, and expands the legitimacy of alternative architectures.
The memory wall
Within this picture, Shacham positions his company in a very specific place - not in the race for flashy benchmark speeds, but at the heart of one of AI's most painful bottlenecks: the memory wall. He explains that the inference stack is effectively split into two distinct parts: on one side, processing input - prompts, documents, and context - and on the other side, generating tokens. Each stage requires different resources. "One needs more compute, the other needs more memory," he says. In his view, the future does not belong to a single machine that does everything, but to systems that can properly divide the work and get all components to communicate efficiently.
This is where, he argues, the company’s advantage lies: "We're very focused on the memory wall problem and how to dismantle it." Instead of focusing solely on peak computational speed, the company is targeting large, long, and complex models - precisely the cases where the real bottleneck is not raw computing power, but the ability to feed, store, and move vast amounts of information efficiently.
The business logic, for him, is very clear: In a world in which data centers are constrained by energy limits, success will be measured not only by speed but also by efficiency; not who runs fastest, but who serves the most users within the same power envelope. "If I have someone with a 100-megawatt server farm, and I can deliver 50 times more users within the energy framework of the grid, he'll be happy," he says. The bottom line: " We're not focused on the fastest computations, but on the largest number of customers at the highest efficiency."
The two American founders - Rabii and Reynders - are located in the heart of Silicon Valley, while Shacham runs the branch located in Ra'anana's sleek new industrial park. "We're trying to get the best of both worlds," he says. "There's incredible software and hardware talent in Israel, and at the same time the company has a very American DNA that's closely connected to everything happening in Silicon Valley, just an hour's drive from Google, Meta, OpenAI, Anthropic, and Nvidia.
"We live 'round the clock - we finish a project in Israel, then hand it off to employees in California who send it back to Israel at the end of their day. And if there’s a customer who needs to meet us, there's always someone in Silicon Valley who can take it on themselves," Shacham says. "It’s not about building the biggest company, but the most efficient one that runs the fastest and builds the best product. But there are challenges, like flight cancellations that sometimes make things harder, and you end up closing deals over Zoom instead of face-to-face. The weak dollar exchange rate also makes employees very expensive. Does that mean we'll hire fewer people here? Not necessarily. We'll hire wherever we find the best talent, but without a doubt, Israel’s advantage in employment costs is eroding."
Year founded: 2023
Capital raised: Over $100 million
Employees: 50, 25 in Israel
Investors: Lux Capital, Bow Wave, Hetz Ventures, Grove Ventures, TAL Ventures, Upfront Ventures, Aidenlair Capital, SBI, QP Ventures.
Published by Globes, Israel business news - en.globes.co.il - on June 17, 2026.
© Copyright of Globes Publisher Itonut (1983) Ltd., 2026.