NVIDIA, the $5.5 trillion market cap giant dominating the AI chip industry, is facing a formidable new rival: Cerebras. The AI chip startup is poised to launch its IPO on the Nasdaq, marking one of the year’s largest global public offerings. Priced at $189 per share, Cerebras aims to raise $5.55 billion, valuing the company at $56.4 billion (approximately 380 billion RMB). This blockbuster IPO—also likely the largest in the first half of 2026—has thrust Cerebras into the spotlight, as it seeks to disrupt NVIDIA’s long-standing dominance in AI compute hardware.
Cerebras’ Game-Changing Core: Wafer-Scale Engine (WSE)
Cerebras’ rise hinges on its radical technical innovation: the Wafer-Scale Engine (WSE), the largest chip ever built by humans. Unlike NVIDIA and AMD, which cut 12-inch silicon wafers into small individual GPU dies before packaging, Cerebras flips the script. It repurposes an entire 12-inch wafer as a single "giant chip" for AI inference, currently in its third iteration (WSE-3).
To put the scale into perspective: a traditional GPU is roughly the size of a postage stamp, while Cerebras’ WSE is as large as a dinner plate. Packed with 4 trillion transistors and over 900,000 AI compute cores, the WSE consolidates all compute, memory, and network communication onto one wafer. This solves a critical bottleneck plaguing modern large AI models: GPU cluster inefficiency. Training state-of-the-art models today requires thousands, even tens of thousands of GPUs linked by high-speed networks—but more GPUs mean crippling communication latency. Many large models stall not from lack of computing power, but from data bottlenecks between GPUs.
Cerebras’ unorthodox approach eliminates this problem entirely. By housing all processing on a single wafer, data no longer needs to shuttle between scattered GPUs. The result is lower latency, reduced power consumption, faster training and inference speeds, and seamless scalability for ultra-large models. This architecture represents a stark alternative to NVIDIA’s GPU-centric ecosystem, earning Cerebras the title of the most disruptive player outside NVIDIA’s dominance.
Founder’s Bold Vision: Defying Industry Norms
Cerebras was founded in 2016 by Andrew Feldman, a serial entrepreneur with a track record in low-power server technology. Before founding Cerebras, Feldman co-founded SeaMicro, which was acquired by AMD for $334 million in 2012. His pivot to AI chips stemmed from a key observation: the rising cost of data movement in deep learning would eventually render the traditional CPU/GPU "small-chip assembly" model obsolete.
Feldman set out to do what many in the industry deemed impossible: build a fully functional wafer-scale chip. Critics argued the idea was reckless—any defect on the wafer would render the entire chip useless, which is why wafers are always cut into smaller dies. Feldman’s counterintuitive solution was to design the system to bypass defects, operating only on the wafer’s functional regions. In 2019, Cerebras launched its first-generation WSE, proving the concept viable and revolutionizing AI hardware design.
Explosive Growth & $24.6B Order Backlog
Cerebras’ trajectory has been nothing short of spectacular, driven by the booming demand for AI inference hardware. The company’s revenue skyrocketed from $24.6 million in 2022 to $510 million in 2025—a 19x increase over four years, with a 76% year-over-year jump in 2025. Even more notably, Cerebras turned profitable in 2025, posting a net profit of $238 million, reversing a $482 million net loss in 2024. Such growth is rare among AI hardware startups.
A defining metric fueling investor enthusiasm is Cerebras’ staggering $24.6 billion order backlog—a mind-blowing figure for a company with $510 million in annual revenue. This backlog underscores a seismic shift in the AI industry: buyers are now pre-purchasing AI compute capacity rather than just buying chips.
Key Customers Powering Growth
Cerebras’ revenue and order book are highly concentrated in a small group of ultra-large clients:
- OpenAI: The single largest customer, with a $750MW AI compute agreement valued at over $10 billion (potentially exceeding $20 billion through 2028). OpenAI CEO Sam Altman is an early Cerebras investor, and the company may invest $1 billion in Cerebras’ data center buildout, securing up to a 10% stake.
- G42: A UAE sovereign wealth fund-backed tech firm, one of Cerebras’ top clients, with multi-billion-dollar orders in the backlog.
- MBZUAI: The UAE’s Mohamed bin Zayed University of Artificial Intelligence, accounting for 62% of Cerebras’ 2025 revenue. Combined with G42, Middle Eastern clients contributed 86% of 2025 revenue.
Cerebras does not sell standard chips—it delivers full AI supercomputer systems, including WSE chips, servers, networking, software, and end-to-end data center deployment. This business model targets national AI initiatives, supercomputing centers, and large model developers—perfect for countries like the UAE, which aims to build sovereign AI infrastructure independent of Western tech giants.
Can Cerebras Truly Challenge NVIDIA?
For all its momentum, Cerebras is far from dethroning NVIDIA. The core issue is not raw chip performance, but NVIDIA’s unassailable ecosystem dominance.
1. CUDA Software Moat
NVIDIA’s greatest strength is not its GPUs, but the CUDA software ecosystem. Nearly all global AI frameworks, training systems, inference tools, and engineering libraries are built around CUDA. Enterprises do not just buy NVIDIA GPUs—they rely on CUDA’s compatibility and tooling. Cerebras’ software ecosystem remains vastly underdeveloped compared to CUDA, creating a high barrier to mainstream adoption.
2. Concentrated Customer Base
Cerebras’ revenue is heavily reliant on a handful of mega-clients, making it a project-based firm rather than a platform-wide provider like NVIDIA. This lack of broad market penetration limits its ability to scale beyond niche large-scale AI projects.
3. No One-Size-Fits-All Hardware
AI hardware has no universal solution. Research from Harvard and other institutions confirms different workloads require specialized hardware. Cerebras excels at large-scale inference, but NVIDIA GPUs dominate training, small-scale inference, and edge AI. Competitors like Groq, TPU, and Gaudi also carve out their niches—no single chip can replace all use cases.
4. Supply Chain Dependence
NVIDIA controls critical AI hardware supply chains, with deep ties to TSMC’s advanced CoWoS packaging, Micron’s memory, and SK Hynix’s HBM. Even if Cerebras’ architecture succeeds, it still relies on the same constrained supply chain as NVIDIA.
5. Customized Business Model
NVIDIA sells standardized, mass-produced GPUs. Cerebras builds highly customized supercomputer systems—slower to deploy, more capital-intensive, and tied to long-term client contracts. This model limits scalability compared to NVIDIA’s global, off-the-shelf product line.
Conclusion
Cerebras’ upcoming IPO marks a pivotal moment for AI hardware innovation. Its wafer-scale architecture addresses critical GPU cluster bottlenecks, and its $24.6 billion order backlog—led by OpenAI—validates its value for large-scale AI inference. However, challenging NVIDIA remains a distant goal. NVIDIA’s CUDA ecosystem, supply chain control, and broad market reach create insurmountable barriers for niche players like Cerebras.
Cerebras will likely emerge as a key niche leader in ultra-large AI inference, serving national AI projects and big model developers. It will not replace NVIDIA, but it will diversify the AI hardware landscape, offering a powerful alternative for workloads where GPU clusters fall short. As AI demand continues to surge, the industry will increasingly embrace multiple specialized hardware architectures—and Cerebras is poised to claim its share of the market.
