Cerebras Systems went public on Nasdaq this month at a $100 billion valuation. The more consequential story is how it got there.
On May 14, 2026, the opening bell at Nasdaq marked one of the most closely watched debuts in years. AI chip maker Cerebras Systems (ticker: CBRS) priced its initial public offering at $185 per share, surged to an intraday high of $386, and closed at $311 — a first-day gain of 68%.
At its peak, the company's market capitalization briefly exceeded $100 billion, making it the largest U.S. technology IPO since Uber in 2019 and, by most measures, the largest pure-play AI IPO on record.
The debut sent ripples through Silicon Valley. Since then, the stock has pulled back to around $290. ARK Invest founder Cathie Wood has already moved to add the stock to her portfolio. Yet the more analytically significant story behind Cerebras is not its market debut — it is how a five-person startup persuaded TSMC, the world's most advanced contract chipmaker, to help engineer what is now the largest single AI chip ever produced.
The "Big Chip" Bet: Going Against the Grain
While the global semiconductor industry has spent decades pursuing miniaturization — from 7-nanometer nodes down to 2nm — Cerebras pursued the opposite logic: take an entire 300mm silicon wafer and fabricate it as a single, monolithic chip.
The result is the Wafer-Scale Engine 3 (WSE-3), measuring 46,225 mm² — more than 50 times the die area of Nvidia's largest GPU. It contains 4 trillion transistors, 900,000 AI-optimized cores, 44GB of on-chip SRAM, and a memory bandwidth of 21 petabytes per second.
The architectural logic is straightforward: by eliminating inter-chip communication — a well-documented bottleneck in conventional GPU clusters — this dinner-plate-sized chip reduces latency and power consumption while simplifying scaling for large AI model training and real-time inference.
This is not a prototype. The WSE-3 is in production and shipping to customers. OpenAI, Amazon Web Services (AWS), and Meta have all placed orders. Cerebras has reported a backlog of $24.6 billion in cumulative orders, with a single OpenAI contract accounting for more than $20 billion of that total.
Five Engineers, Thirty TSMC Specialists
In 2015 and 2016, Cerebras was a company of five people with an idea that industry veterans widely dismissed. No commercial producer had ever successfully manufactured a full-wafer single chip at scale. Yield loss, power delivery, thermal management, and testing each presented what engineers at the time considered insurmountable obstacles.
Co-founder Jean-Philippe Fricker later recalled the team's first direct engagement with TSMC. The Cerebras engineers flew to Taiwan, entered a conference room, and faced 30 of the foundry's most senior specialists. "They looked at us like we were telling a fairy tale," Fricker said.
Rather than leading with vision, the Cerebras team decomposed the problem into discrete, executable steps. They proposed modifying scribe lines — the narrow gaps between die on a wafer — and routing millions of short interconnects across them, effectively stitching 84 reticle dies into a single logical unit. They introduced keep-out zones to eliminate conventional test structures, developed fault-tolerant core architectures that could route around defects, and engineered custom power delivery systems, liquid cooling enclosures, and specialized photomasks.
TSMC's engineers, according to accounts from both sides, responded to the technical specificity of the proposal. What began with the WSE-1 on a 16nm process node has since grown into nearly a decade of deep co-development. TSMC not only adapted production lines for an then-unknown startup — it subsequently identified wafer-scale integration as a strategic direction, with plans to commercialize System-on-Wafer (SoW) technology more broadly.
In 2025, TSMC North America Vice President of Business Management Lucas Tsai stated publicly at an industry conference: "We are proud to work with industry innovators like Cerebras — from pioneering startups to leaders, we support customers of all sizes in turning transformative ideas into reality."
Why Would TSMC Take a Chance on a Five-Person Company?
The answer lies in what Cerebras was solving. Memory bandwidth and interconnect efficiency represent two of the most persistent structural constraints in large-scale AI computation. For TSMC, the partnership was not purely commercial — it was an opportunity to accumulate process knowledge relevant to next-generation advanced packaging and integration techniques.
Critically, Cerebras achieved performance that surpassed conventional GPU clusters on a 5nm node, without competing for TSMC's most constrained 3nm or 2nm capacity. This gave TSMC's more mature process nodes a high-value application that did not cannibalize its leading-edge customer base.
For Taiwan's broader semiconductor supply chain, the dynamic represents a structural advantage. Cerebras is almost entirely dependent on TSMC for fabrication. Its commercial success directly amplifies TSMC's strategic position in AI infrastructure — at a moment when Nvidia's dominance in GPU-plus-CoWoS packaging is drawing both regulatory scrutiny and competitive responses globally. Cerebras's monolithic wafer-scale approach opens a distinct architectural track, adding a second growth vector to Taiwan's chip ecosystem.
From Dismissed Idea to a $100 Billion Debut
Cerebras's success is not merely a stock market victory — it is a model for what becomes possible when a small startup and a manufacturing giant innovate together. It demonstrates that genuine hardware differentiation, backed by real customer contracts and shipping product, can still command extreme valuation premiums even in a capital environment increasingly skeptical of undifferentiated AI hardware plays.
As OpenAI, Anthropic, SpaceX, and other high-profile private technology companies prepare for their own public market debuts, Cerebras has established a clear reference point: a differentiated product with real customers and shipping revenue can access public capital on its own terms.
The broader question — whether wafer-scale integration can evolve from a niche architectural bet into a mainstream infrastructure standard — remains open. What is not in dispute is that a five-person team's willingness to bring a technically rigorous, step-by-step proposal to TSMC's engineers initiated a partnership that now sits at the intersection of AI infrastructure, semiconductor supply chain resilience, and cross-Pacific industrial collaboration.
This dinner-plate-sized chip carries not only 900,000 AI cores, but also a story of audacity, perseverance, and cross-border collaboration.
From a wafer fabrication facility in Hsinchu, Taiwan, to the Nasdaq opening bell in New York, the chapter that Cerebras and TSMC have written together will be recorded in the history of semiconductors and AI. When future accounts ask who truly drove the revolution in AI computing, this story of Taiwan-U.S. collaboration will stand as one of its most defining chapters.
