Ofer Shacham and the art of making room

A computer can have an enviable number of processors and an embarrassing amount of waiting to do. Ofer Shacham has built his latest company around that predicament. The processors want data. The memory system cannot deliver enough of it quickly enough. Buying another processor may simply give the first one some company.

At Majestic Labs, where he is co-founder and CEO, Shacham is trying to change that arrangement. It is a large undertaking with a familiar shape. His career has moved through university research, phone cameras, and wearable computing. Again and again, the work comes down to deciding how much useful computation a physical system can afford.

Long before the server proposals and executive titles, he worked as a diving instructor while studying at Tel Aviv University. He later moved through IBM and Stanford into custom silicon. His route into chip design had several turns still to come. At Stanford, even the job he thought he was taking changed beneath his feet.

A short job that lasted six years

At Stanford, Shacham joined work led by Mark Horowitz that was supposed to occupy a couple of quarters. His doctoral acknowledgments record the result: he was still there nearly six years later. Temporary employment can be a remarkably effective recruitment strategy.

His May 2011 dissertation examined chip generators: frameworks that could produce customized chips from reusable design knowledge. Specialized computing could save energy, but developing each new design carried substantial costs. The proposal was to preserve useful common structure while making room for application-specific choices.

Think of the difference between drawing every house from scratch and having an adaptable architectural system. The analogy has limits, but it captures the attraction: reuse the hard-won knowledge without forcing every occupant into the same floor plan.

The thesis introduced Genesis2, a framework for constructing such generators. Its acknowledgments also describe a crowded intellectual household: advisers, collaborators, friends, and family. Shacham dedicated the work to his grandfather Simcha. Even a document about automatic chip generation begins with people.

The research was part of a wider collaborative effort. Stanford’s floating-point-unit generator project lists Shacham among its contributors, alongside Horowitz, Sameh Galal, and others. The work explored energy and area trade-offs using a parameterized design built on Genesis2.

This is an unglamorous but consequential corner of computing. A designer must choose what to make flexible, what to fix, and how to test the results. Every extra option has a cost. A system that can do everything may spend too much of its energy preparing to do things nobody asked for.

The chip behind the camera roll

Shacham co-founded Chip Genesis out of the Stanford research and joined Google through its acquisition. His subsequent responsibilities included Pixel Visual Core, Pixel Neural Core, consumer silicon design, and work on security and Tensor chips. The university problem had found its way into products people carried around.

In October 2017, Shacham and Masumi Reynders co-authored Google’s introduction to Pixel Visual Core. It described a custom co-processor inside Pixel 2, designed to handle imaging and machine learning within a mobile power budget.

The details made the case for specialization. Google said HDR+ could run five times faster on Pixel Visual Core while using less than one-tenth the energy of the application processor. Those figures applied to that particular comparison. They illustrated how much the fit between a task and its hardware could matter.

Software was part of the design. The processor used tools including Halide and TensorFlow, with a compiler translating work for the hardware underneath. Efficiency depended on decisions made across the system, rather than on a chip sitting in splendid isolation.

By February 2018, Shacham was explaining the benefit in the language of everyday photography. Pixel Visual Core was being enabled for applications including Instagram, WhatsApp, and Snapchat. Better processing could reach pictures taken inside those apps.

For the user, this was a pleasantly undemanding transaction. Take a picture; receive a better result. Nobody needed to admire an arithmetic unit first. Hardware engineers sometimes achieve their desired outcome when the customer has no reason to think about hardware at all.

There is another detail in the old announcement worth keeping: Reynders’ byline. Years before she and Shacham became Majestic co-founders, their names already appeared together on a product explanation. Their later partnership had a working history.

A computer has to fit somewhere

In 2018, Shacham left Google for Facebook. Sergey Brin had offered him room to choose his next project at Google. The competing attraction was the engineering challenge of augmented-reality silicon.

The brief for the glasses included two requirements: wear them all day, and make them socially acceptable. A device intended for a face has a rather unforgiving review committee. Its owner gets a vote every time they put it on.

He went on to serve as Meta’s Vice President and Head of Silicon. His professional biography describes an organization built across hardware and software, working within strict limits on power, size, and performance. A wearable product makes those abstractions unusually concrete.

Weight changes how an object feels. Power requirements affect the battery. The battery affects the shape. A software ambition eventually acquires a physical dimension. These are useful reminders in a business where a diagram can make almost anything look conveniently rectangular.

By 2023, Shacham was ready to start another company. He joined Sha Rabii and Reynders to found Majestic Labs. The new venture brought colleagues with experience of building silicon organizations back into the same undertaking.

Giving memory a seat at the table

Majestic emerged from stealth in November 2025 with over $100 million in financing. Bow Wave Capital led its Series A; Lux Capital had led the seed round. The announcement described the founders coming together for a third time.

Rabii became president, Reynders chief operating officer, and Shacham CEO. Their backgrounds span silicon engineering, systems, operations, and commercial relationships. Building a server company requires all those conversations to happen before the finished machine can have a useful conversation with a customer.

“We are starting from the memory, making it a first class citizen.”

Ofer Shacham, November 2025

The premise is straightforward. An AI workload needs information stored somewhere and delivered to processors. Memory capacity determines what can fit; bandwidth affects how rapidly data can move. Increasing one does not automatically solve the other.

Majestic’s approach includes new silicon, server hardware, signaling, and software tools. Shacham has described the goal as a sufficiently large memory pipeline. It is a system proposal, with several pieces that have to work together.

In September 2026, the three founders published an essay setting out their argument. They described a pattern in which organizations add GPUs partly to obtain more attached memory, then pay for compute they cannot fully use. Longer contexts and more complex AI workloads, they argued, make the imbalance harder to ignore.

It is an argument about where to begin a design. Start with the workload and its appetite for data, then decide how the machine should be organized. That sequence has a recognizable connection to Shacham’s earlier research, even though a phone camera and a data center make very different demands.

What a server promises, and what it must prove

Majestic calls its server Prometheus and its processor Ignite. The company advertises up to 128 terabytes of shared memory per system, with Arm cores and RISC-V vector and tensor cores operating in a common memory space. Its product page lists PyTorch, vLLM, and Triton among supported tools.

These are company specifications and claims about a developing platform. They describe the machine Majestic is offering to build. Customer deployments will establish how those choices perform across real workloads, operating conditions, and costs.

Shacham has set out a wider aspiration: make advanced AI available beyond the largest cloud companies. In June 2026, Majestic’s reported launch target was mid-2027. That leaves the familiar hardware interval between explaining an architecture and delivering the equipment.

The practical questions are substantial. Can a customer run the desired model? Can the system keep it supplied with data? What happens to power consumption, software effort, and the bill? A server earns its place through those answers. An evocative name gets it only as far as the meeting invitation.

Back at the blackboard

Alongside Majestic, Shacham teaches VLSI at Tel Aviv University and advises graduate students. The return to a classroom gives his career a pleasing circularity: the researcher whose work became a startup now has fresh industrial problems to bring back to students.

He also talks publicly about the work. A 2025 conversation with Dov Moran explored the move from big-company leadership to building across Israel and California. A Semafor interview addressed AI memory. On the Hebrew-language Hard Reset podcast, the discussion ranged across chips and AI; the episode itself grew from a conversation outside a conference.

Across those settings, the recurring subject is how to make computing useful under constraints. The rooms have changed, from Stanford laboratories to product teams and startup offices. Shacham is still concerned with what fits, what moves, and what waits. For Majestic, the next chapter will be written by the machine.