Majestic Labs has emerged from stealth with $100 million in funding, a $10 million seed round, and $90 million Series A round. The company's cofounder and CEO, Ofer Schacham, is an Israeli who lived in Silicon Valley for many years but recently returned to Israel.
Schacham was talking about AI long before ChatGPT burst onto the scene three years ago. Twenty years ago, he was asked by the US Defense Advanced Research Projects Agency (DARPA) to develop AI processors for military needs after completing his doctorate at Stanford.
Google hired Schacham, who was responsible for developing the computer vision chips for Pixel devices, and later became head of AI processors for Google's edge devices. Mark Zuckerberg brought him in to set up Meta's chip lab to develop its metaverse headsets, and for years, he managed the entire chip division from Israel.
After being the "chip czar" for Sergey Brin and, later, Mark Zuckerberg for most of his career, Schacham is now his own man at Majestic Labs, which he founded with his former colleagues at Meta and Google, Sha Rabii and Masumi Reynders.
Majestic Labs has been founded to challenge Nvidia in its soft underbelly. The inefficiency of its memory components, an issue that severely limits the computational capacity of each individual processor, dictates that companies interested in training complex language models buy increasingly expensive graphics processors.
The same components are so critical for training AI models. Nvidia itself benefits from this inefficiency because it sells more processors.
In the meantime, the problem remains unsolved and is even burdening Nvidia and OpenAI. Altman's Law, a rule of thumb somewhat analogous to Moore's Law in the world of chips, states that the cost of AI processing is proportional to the square of the number of its users.
When the number of AI queries increases dramatically, whether in chat engines, video creation, or document creation, the industry reaches a literal boiling point, where not only costs rise, but also the temperature in data centers, which in turn slows down the chips.
One of the main bottlenecks of the graphics processors that dictate this, whether they are Nvidia, AMD or Google processors, is the limited active memory on each chip, up to 192 gigabytes of memory in each. It requires the purchase of many expensive chips, and Majestic Labs is solving this problem.
Majestic does not offer its own graphics processor, as do more successful competitors like AMD, or less successful ones like Intel, SambaNova, Cerebras, and Groq, but replaces Nvidia's AI server with its own server.
It is a very ambitious vision that includes not only a dedicated processor developed in Israel, but also a new server structure that connects to each graphics processor huge amounts of memory with higher bandwidths than anything currently seen on the market.
The result is equivalent to connecting up to 1,000 times more memory to each graphics processor than on Nvidia, AMD, or Google AI servers, according to the company. It does all this from a modest office in a building on the border of Kfar Saba and Hod Hasharon. The company has 50 employees, half of whom are in Israel.
Majestic Labs has declined to say at what valuation the funds were raised, but "Globes" estimates the company's valuation at $300-400 million. The seed round was led by Lux Capital, and Bow Wave Capital led the Series A round.
Other investors include SBI, Upfront, Grove Ventures, Hetz Ventures, QP Ventures, Aidenlair Global, and TAL Ventures, the only Israeli fund in the larger $90 million financing round.
Increase memory accessibility
Schacham says, "The industry is trying to run language models with trillions of parameters in aggregate and all this on processors whose maximum memory capacity does not exceed 192 gigabytes."
He explain, "At home, our AI uses are no longer satisfied with scanning a QR code and decoding it: we ask the chat several questions a day, upload a 150-page document and ask it to translate it into Chinese and create images from scratch using verbal descriptions, so the memory requirement is estimated in terabytes of data on processors that can work in gigabytes."
By rebuilding the server, the company claims, Majestic Labs can increase memory accessibility by up to 1,000 times compared with existing solutions, improve the server's computing performance by more than 50 times, and actually save 10 or 20 Nvidia servers for every server it sells.
The experience of the three founders, who Zuckerberg himself hired to run Meta's chip lab, their doctorates, and the hundreds of patents they have registered in the US over the years.
Their industry connections play a major role, allowing them to speak with many potential customers and develop servers that will meet their future needs in 2027. "We know how to work and understand their product requirements," says Schacham.
Majestic Labs is not the only startup that intends to compete with Nvidia. It is challenged by US companies that have developed dedicated AI chips such as SambaNova, Cerebras, Groq, and Tenstorrent, as well as Israel's NextSilicon.
"The company was founded for the AI era," says Schacham, without referring to the rivals by name. "And while many companies try to come from a hardware or 'hardened' solution. As chip giants launch another solution that doesn't address the memory problem, the models continue to grow and evolve, so it's important that the solution be flexible and programmable."
Schacham explains that "One of the lessons I learned the hard way as a chip development manager, as someone who spent 12 years at Google and Meta, is that software is king. The bottleneck in training large models can probably be solved at the system level, not by launching this or that chip."
Is there a bubble?
Schacham doesn't believe there is a bubble in the AI industry. Still, he distinguishes between companies that will become bigger and those that will disappear: "As early as 2014, I drove the first prototypes of Google's autonomous cars, and when I visited their data centers and saw how with AI efficiency increased by 20%, I understood where the future was going," he says.
"Without a doubt, every field that AI touches will change rapidly. I saw it from the cockpit at Google, and I saw it in the areas of image processing."
He continues: "But is there a bubble? AI is growing at a dizzying pace, and some companies are justifiably valued at ever-increasing levels, but others are overvalued and won't survive. AI will be a greater revolution than all the other revolutions that preceded it, but not every company that enters will survive. Is Google's and Meta's spending on capital equipment for AI unjustified? I don't think so. Tell me how many times you use ChatGPT or try an AI engine to renovate images."