As the artificial intelligence landscape shifts rapidly from training massive foundational models to running them at scale, the economics of the industry are undergoing a profound transformation. Inference—the process of running trained AI models to generate real-time responses—is rapidly emerging on a path to becoming the largest market in the world. Industry observers and technologists suggest that if the broader vision of the AI era fulfills its promise, inference will soon underpin nearly every digital touchpoint across modern society. From powering automated storefronts and cinematic production pipelines to clinical medical encounters, complex video games, legal proceedings, and automated software engineering, the underlying compute requirements are staggering.
Recognizing this massive structural shift, the market for inference-specific systems has evolved from a speculative niche into a widespread industry consensus. Yet, when founders Gavin, Chris, and Rob first conceived of the venture back in 2022 while still residing in their dorm rooms at Harvard University, building dedicated hardware for inference was viewed as a deeply contrarian bet. While major incumbent technology firms and well-funded contemporaries focused heavily on general-purpose hardware or accelerated training architectures, the young founders recognized that inference required a completely distinct, first-principles approach. The result of that early conviction is Etched, which now stands uniquely positioned as the sole post-ChatGPT-era hardware startup to feature production-ready custom silicon slated to ship in commercial quantities by 2026.
Over the past few years, the team at Etched has quietly pioneered several critical research breakthroughs, including innovations in low-voltage inference and cluster-scale memory management. Each of these achievements represents a hard-won architectural choice designed to drive step-change improvements in both throughput and computational latency. As the fundamental unit of enterprise compute has progressively evolved away from individual chips toward larger server racks and ultimately interconnected clusters, Etched has deliberately engineered its products to match that physical reality. Rather than focusing solely on isolated microprocessors, the company has designed comprehensive, cluster-scale inference systems built to handle the immense demands of next-generation AI infrastructure.
Driven by these foundational hardware innovations, Etched’s custom inference system excels across multiple operational metrics, delivering exceptional throughput and interactivity over the full span of modern frontier models. This includes everything from large sparse Mixture-of-Experts architectures to dense transformer models and alternative neural network designs such as Mamba. However, believing in the theoretical necessity of custom silicon is only half the battle; executing on the manufacturing and design is an entirely different challenge. Developing novel computing hardware remains among the most difficult engineering feats in the world, particularly in an ecosystem where the underlying technological ground shifts constantly. AI models evolve continuously over the span of weeks—context lengths stretch dramatically, attention mechanisms are regularly reinvented, and the operational balance between dense and sparse parameters continues to shift.
Succeeding in this high-stakes environment requires a rare dual capability: the agility to iterate quickly enough to keep pace with rapid software model advancements, alongside the technical vision to push the absolute physical limits of hardware performance. Accomplishing this demands an exceptional kind of engineering team—one visionary enough to conceptualize the right architectural designs from scratch, young and dynamic enough to move with urgency, and proven enough to successfully tape out chips into actual production. Furthermore, building a sustainable enterprise in this sector requires an organizational structure designed to endure for decades rather than merely surviving a single chip tape-out cycle.
Stepping into the Etched corporate offices for the first time offers an unmistakable sense of intensity and focus. Whether visiting during early morning hours or late at night, the collective energy remains remarkably high and infectious. Founders Gavin and Rob exemplify a breed of technical outliers bold enough to design silicon from first principles while systematically knocking down every operational and technical obstacle in their path. The broader team they have assembled reflects that same relentless drive, mirroring the qualities that traditionally separate organizations that merely talk about hardware innovation from those that successfully manufacture and ship physical products.
This operational philosophy is visible across the company’s various facilities, including a live testing laboratory in San Jose, California, designed to host their initial server racks, and a dedicated strategic office established in Taiwan to colocate closely with core manufacturing suppliers and expedite testing phases. By continuously questioning every historical assumption regarding what is technologically possible—such as the viability of low-voltage inference—the company operates under a clear internal mantra: production is the product, and physical shipping is the only metric that truly matters.
The trajectory of the company’s progress over the past year has been remarkably rapid. Earlier this year, Etched successfully taped out its first-generation custom silicon chip at Taiwan Semiconductor Manufacturing Company, marking a significant corporate milestone. This achievement established Etched as the first post-ChatGPT-era enterprise to successfully execute a full-reticle A0 chip tape-out on TSMC’s advanced leading-edge semiconductor nodes. Following this manufacturing milestone, the engineering team achieved another notable feat by bringing up their first operational cluster of chips within a span of just 40 days. This cluster successfully executed inference workloads across a diverse array of frontier artificial intelligence models, achieving Pareto-dominant performance metrics across industry-standard throughput and interactivity curves. With these production milestones successfully met, Etched’s earliest enterprise customers are now gaining direct access to the hardware, allowing them to evaluate the raw speed, throughput, and expressiveness of the custom system firsthand.
While the company’s first-generation hardware represents a sophisticated achievement that is expected to perform strongly within the commercial market, industry backers emphasize that the true value lies deeper within the enterprise. Investors note that Etched has successfully constructed the foundational apparatus required to continuously build subsequent iterations of advanced hardware—effectively creating the machine that builds the machine. This institutional backing highlights a team possessing both the design taste and the relentless operational execution required to ship successive generations of increasingly ambitious inference machines at an accelerated pace.
In a macroeconomic environment where the global intelligence race remains fundamentally constrained by available compute capacity, pushing the boundaries of physical engineering to maximize intelligence output per watt represents both an engaging operational challenge and a critical mission for technological advancement. To support this ongoing expansion and the commercial rollout of its custom silicon architecture, Etched has officially announced the closing of a $300 million Series C financing round. The investment values the company at $10 billion on a pre-money basis, with the funding round led by major financial institutions and strategic partners, including Jane Street, Andreessen Horowitz, Diffusion, and SK Hynix.
Comments (0)