There are certain stories that force us to stop and reconsider what we believe is possible, and the story of Etched may be one of them.
Three Harvard dropouts, all only 24 years old, have built a semiconductor company that is now being valued at approximately $21 billion, and what makes the story particularly remarkable is that in a matter of 4 months they moved from an idea on paper to working silicon, a fully integrated computing system, a major customer and more than $1 billion in orders.
In an industry where experience, capital and patience have traditionally been prerequisites for success, three young entrepreneurs have challenged an assumption that has governed semiconductor development for decades: that hardware has to move slowly.
The company was founded by Gavin Uberti, Chris Zhu and Robert Wachen, with Uberti and Zhu having met at Harvard in Math 55, one of the university’s most demanding mathematics courses.
All three are 24 years old, and when they initially approached investors, their age was viewed less as an asset than as a liability. Silicon Valley has become accustomed to young entrepreneurs disrupting software, social media and consumer technology, but semiconductors are different because the consequences of getting something wrong are measured not in lines of code but in millions of dollars, years of engineering and sometimes the complete failure of a company.
The skepticism was therefore understandable. Designing an advanced semiconductor requires knowledge that is accumulated over decades, and the process involves architecture, chip design, , testing, memory, networking, power management and the physical infrastructure required to make all of those pieces work together. A software company can build a prototype on a laptop and change it overnight, while a semiconductor company can spend hundreds of millions of dollars before it even knows whether the product will work as intended.
READ: Sreedhar Potarazu | Who wrote this? The crisis of authorship — Vedas, Shakespeare and now, AI (August 18, 2026)
Etched appears to have looked at that reality and asked a different question: what if the semiconductor industry itself could be approached with the urgency of software?
Production is the product
That philosophy is captured in the company’s motto, “Production is the product,” because the founders did not view the chip as the finished product and manufacturing as something that would happen later.
They treated the entire process of moving from an architectural concept to a functioning computing system as part of the product itself, investing tens of millions of dollars in diagnostic equipment, developing software and custom server racks before the first silicon had even been fabricated, building a 2-megawatt data center inside its offices and taking greater control over portions of its supply chain and testing process.
The result was an extraordinary acceleration of the traditional hardware development cycle. After receiving its test chips from Taiwan Semiconductor Manufacturing Company, Etched says it took only 44 days to get them running AI inference workloads, a process that can reportedly take six months or more in the traditional semiconductor development cycle.
Whether every comparison ultimately survives independent benchmarking is something the market will have to determine, but the speed itself illustrates the philosophy behind the company: do not treat production as the final step in developing the technology; make production part of the development process from the beginning.
What makes Etched particularly interesting, however, is not simply how quickly it builds chips but what problem it believes the chips should solve.
Much of the first wave of artificial intelligence has been defined by training. Training is the process through which a model learns from enormous amounts of data, requiring massive amounts of computational power to adjust billions or even trillions of parameters until the system becomes capable of understanding language, recognizing images, writing code or performing other increasingly sophisticated tasks.
Nvidia became the dominant force in this environment because its GPUs are exceptionally well suited to the parallel mathematical operations required to train these models.
But once the model has been trained, the economics of artificial intelligence change because the model has to be used, and that process is called inference.
Inference
Every time someone asks a question, generates an image, analyzes a document, writes software, interprets a medical image or asks an AI agent to perform a task, the system has to run the model and generate a response.
As artificial intelligence moves from an occasional application into something that operates continuously across medicine, finance, education, manufacturing, robotics and virtually every other industry, inference could ultimately represent an enormous share of the world’s computing requirements.
This is where Etched is making its bet. Rather than building a general-purpose processor designed to perform a broad range of computing tasks, the company is developing specialized hardware designed specifically around AI inference.
The idea is not entirely new because computing has always benefited from specialization, but the scale and importance of AI create an opportunity to rethink the architecture of the computing system itself.
That brings us to one of the most interesting concepts in Etched’s architecture, which the company describes as “cluster-scale memory.”
Cluster-scale memory
The easiest way to understand the concept is to think about what happens when you take a problem that is too large for one person and divide it among a hundred people.
READ: Sreedhar Potarazu | Almost a third of America does not have a primary care doctor (August 16, 2026)
At first, it seems obvious that the hundred people should be able to solve the problem much faster than one person, but that assumption only holds if they can communicate efficiently. If each person has access to only part of the information and has to constantly stop working to ask someone else for information, the advantage of having a hundred people begins to disappear.
The same principle applies to AI computing. Modern AI models are so large that their information can be distributed across multiple processors and memory systems, which means that the processors have to communicate with one another constantly. The processors may be extraordinarily fast at performing calculations, but if they spend too much time waiting for information to move between them, the computational power is not being fully utilized.
Consider a restaurant with one chef who has every ingredient within reach. Now imagine that the restaurant has fifty chefs, but each chef has a separate refrigerator located on the other side of the building. The kitchen has fifty highly capable chefs, but much of their time is spent walking back and forth to obtain the ingredients they need. If you redesigned the kitchen so that every chef could access the ingredients almost instantaneously, the same fifty chefs could operate much more like one highly coordinated kitchen.
That is the basic intuition behind cluster-scale memory. Instead of treating each processor as an isolated computing unit, the architecture attempts to allow multiple processors to communicate and access information so efficiently that the entire collection begins to function more like one enormous computational system.
The future of computing may increasingly be determined not simply by how quickly a processor can calculate but by how quickly the entire system can move information.
Etched says that certain communication tasks that take approximately 4,000 nanoseconds on Nvidia’s Blackwell architecture can be performed in approximately 700 nanoseconds using its architecture and custom interconnections. Those figures are company claims and ultimately need to be evaluated through independent testing and real-world workloads, but they point to a fundamental problem that is becoming increasingly important as AI models grow larger: sometimes the bottleneck is not computation itself but the movement of information required to perform.
Etched is taking a more focused approach by concentrating on inference and building hardware around the characteristics of the workload. The advantage of specialization is that the processor does not have to be everything to everyone. A Formula One car is not designed to carry a family to the grocery store, drive through a snowstorm and transport furniture; it is designed to perform one task extraordinarily well.
Etched is essentially betting that AI inference has become important enough to justify the computing equivalent of a Formula One car.
Approximately 15% of Etched’s roughly 400 employees reportedly previously worked at Nvidia, while the company has recruited senior engineers with decades of experience in chip design and data center systems. Some of those engineers spent more than twenty years at Nvidia, bringing with them the very experience that investors initially believed the young founders lacked.
That combination may be one of the most important lessons in the entire story. The founders brought a willingness to question the assumptions of an established industry and an urgency that may be difficult to create inside a large organization, while the engineers they recruited brought the institutional knowledge necessary to turn those ideas into working silicon.
The choice of Jane Street as the company’s first major customer is equally revealing. Quantitative trading firms operate in an environment in which extremely small differences in computational speed can have economic consequences, making them natural early customers for specialized inference hardware.
READ: Sreedhar Potarazu | If AI gives professional advice, does it need a license? (August 9, 2026)
Jane Street has not only become Etched’s first customer but is also leading the company’s new $700 million financing round, giving the technology something that every semiconductor startup ultimately needs: a sophisticated customer willing to put the product into the real world.
The $21 billion valuation, however, should not be confused with proof that Etched has already won.
Semiconductor history is filled with companies that developed extraordinary technology but failed to translate that technology into sustainable economics. Nvidia remains an extraordinarily powerful competitor, and Google, Amazon, AMD, Microsoft and numerous startups are developing their own AI architectures.
The real test for Etched will be whether its architectural advantages survive independent testing, whether its systems can operate reliably at scale and whether the economics of specialized inference justify the enormous investment required to build and deploy them.
But even if the ultimate outcome is uncertain, the idea behind Etched deserves attention because it reflects a much larger transformation taking place in artificial intelligence.
The most interesting question is therefore not whether three 24-year-olds were able to build a $21 billion company.
It is whether they have identified a fundamental change in the architecture of computing.
That is the promise behind cluster-scale memory, and it is also the reason Etched deserves to be watched closely.
The valuation may change, the technology may evolve and competitors may eventually prove that another approach is better. But the question these three young founders are asking is likely to remain one of the central questions of the AI era: What happens when we stop thinking about the computer as a collection of chips and start thinking about it as one enormous interconnected system?
Perhaps the most important innovations in technology begin exactly that way, not with an answer, but with the courage to question an assumption everyone else has accepted.