Cerebras

Cerebras Systems is an American artificial intelligence hardware company headquartered in Sunnyvale, California, founded in 2016 by Andrew Feldman, Sean Lie, Gary Lauterbach, Jean-Philippe Fricker, and Michael James, all veterans of the SeaMicro microserver startup that AMD acquired in 2012. It designs the Wafer-Scale Engine (WSE), the largest single chip ever produced, with the third-generation WSE-3 (released March 2024) integrating 4 trillion transistors and 900,000 AI-optimized cores on a single 46,225-square-millimeter die, and operates Cerebras Inference, one of the fastest commercial AI inference services for open-weights large language models. As of April 2026, Cerebras has filed for an IPO on the Nasdaq, has commercial deployments with G42 (the principal customer through the Condor Galaxy AI supercomputer build-out), the US Department of Energy national labs, and other enterprise and research customers, and is one of the principal commercial alternatives to NVIDIA-based AI compute.

At a glance

• Founded: 2016 in Sunnyvale, California, by Andrew Feldman, Sean Lie, Gary Lauterbach, Jean-Philippe Fricker, and Michael James.
• Status: Private. Filed for an initial public offering on the Nasdaq in October 2024 (subsequently paused pending CFIUS review of the G42 customer relationship); public-listing timeline remains uncertain through 2026.
• Funding: Approximately $720 million-plus cumulative private capital. Backers include Foundation Capital, Eclipse Ventures, Benchmark, Sequoia Capital, Coatue, Eclipse, and other investors. Latest reported pre-IPO valuation around $4 billion.
• CEO: Andrew Feldman, Co-Founder and Chief Executive Officer. Technology entrepreneur; former CEO of SeaMicro.
• Other notable leadership: Sean Lie, Co-Founder and Chief Hardware Architect. Julie Choi, Chief Marketing Officer. Dhiraj Mallick, Senior Vice President of Engineering.
• Open weights: No (Cerebras designs hardware rather than producing foundation models). Cerebras Inference, the commercial inference service, hosts open-weights models from Meta AI / FAIR (Llama family), Mistral, DeepSeek, and other open-weights providers.
• Flagship products: Wafer-Scale Engine (WSE-1, WSE-2, WSE-3), CS-2 and CS-3 systems (the rack-scale platforms housing each WSE), Cerebras Inference (commercial inference service for open-weights models), and the Condor Galaxy AI supercomputer build-out (the G42 partnership announced 2023).

Origins

Cerebras was founded in 2016 by a five-person founding team, all veterans of the SeaMicro microserver startup that AMD had acquired in March 2012 for approximately $334 million. Andrew Feldman (CEO), Sean Lie (Chief Hardware Architect), Gary Lauterbach, Jean-Philippe Fricker, and Michael James had departed AMD by 2015 with a thesis that the transition to deep-learning workloads called for a fundamentally different chip-design approach optimized for the data-flow characteristics of neural-network training and inference.

The 2016 to 2019 founding period was in stealth, with the company building out the wafer-scale chip-design and engineering capability across silicon, packaging, power, cooling, and software layers. The first-generation WSE-1 was unveiled publicly in August 2019 with the CS-1 system at the Hot Chips conference, integrating 1.2 trillion transistors and 400,000 cores on a single 46,225-square-millimeter die. This was the largest single chip ever produced and substantially larger than any contemporary GPU.

The 2021 to 2024 period saw Cerebras' product iteration. WSE-2 (April 2021) doubled the transistor count to 2.6 trillion. WSE-3 (March 2024) reached 4 trillion transistors and 900,000 AI-optimized cores. The CS-2 and CS-3 rack-scale systems anchored Cerebras' commercial deployment, with the Condor Galaxy AI supercomputer build-out announced in 2023 in partnership with G42.

The 2024 release of Cerebras Inference (August 2024) was the company's most consequential commercial transition. The service hosts open-weights models from Meta AI / FAIR, Mistral, DeepSeek, and other providers, with faster token-generation latency than NVIDIA-based inference services. Industry coverage has consistently characterized Cerebras Inference as the fastest commercial inference service for open-weights large language models through 2024 to 2026.

The October 2024 IPO filing on the Nasdaq was the company's planned commercial-public-listing transition. The IPO was subsequently paused pending CFIUS review of the G42 customer relationship, given G42's UAE state backing and the April 2024 Microsoft investment that had reshaped G42's commercial-customer compliance posture. The continued IPO timeline remains uncertain through 2026.

Mission and strategy

Cerebras' stated mission is to accelerate the future of AI by building the world's fastest AI compute. The strategic premise is that wafer-scale chip design, with all 900,000 cores on a single die rather than distributed across thousands of GPUs, eliminates the memory-bandwidth and inter-chip communication bottlenecks that constrain GPU-based AI training and inference. Cerebras' commercial argument is that for workload classes (training of dense large language models, inference at token-generation speeds, scientific computing workloads), the wafer-scale architecture delivers fundamentally better performance per watt and per dollar than NVIDIA-based alternatives.

The strategy combines three threads. First, the Wafer-Scale Engine product iteration with the CS-2 and CS-3 rack-scale systems anchoring training and other compute-intensive workloads. Second, Cerebras Inference as a commercial inference service hosting open-weights models with faster token-generation latency. Third, the Condor Galaxy AI supercomputer build-out with G42, providing deployed compute capacity in the UAE and other geographies.

The competitive premise is that the 2022 to 2026 supply-chain constraint on NVIDIA H100 and H200 GPUs, combined with NVIDIA's pricing power as the dominant AI compute provider, opens commercial space for alternative AI compute architectures. Cerebras' wafer-scale approach distinguishes it from peer alternative-compute startups including Groq, SambaNova Systems, Tenstorrent, and other peers, each of which pursues different chip-architecture approaches.

Distribution channels include direct commercial sales of CS-3 systems to enterprise and government customers, the Cerebras Inference commercial inference service through API and developer platform access, and strategic-customer relationships including G42, the US Department of Energy national labs (including Argonne and Lawrence Livermore), and other enterprise and research customers.

Models and products

• Wafer-Scale Engine (WSE). The largest single chip ever produced. WSE-1 (August 2019, 1.2 trillion transistors), WSE-2 (April 2021, 2.6 trillion transistors), WSE-3 (March 2024, 4 trillion transistors and 900,000 cores).
• CS-2 and CS-3 systems. Rack-scale platforms housing each WSE, with the memory, power, cooling, and networking infrastructure for commercial deployment.
• Cerebras Inference. Commercial inference service hosting open-weights models from Meta, Mistral, DeepSeek, and other providers. Faster token-generation latency than GPU-based inference services.
• Condor Galaxy. Multi-site AI supercomputer build-out announced 2023 in partnership with G42. Deployed compute capacity in the UAE.

Distribution channels include direct commercial sales to enterprise and government customers, the Cerebras Inference API and developer platform, and strategic-customer relationships including G42, the US Department of Energy national labs, and other enterprise and research customers.

Benchmarks and standing

Cerebras' evaluation framework is hardware performance metrics (training throughput, inference token-generation latency, performance per watt, performance per dollar) rather than horizontal foundation-model leaderboards. Cerebras Inference has been consistently characterized in inference-latency industry coverage as the fastest commercial inference service for open-weights large language models, with token-generation latencies faster than Together AI, Fireworks AI, Anyscale, and other GPU-based inference services through 2024 to 2026.

The Condor Galaxy AI supercomputer build-out with G42 has been characterized in industry coverage as one of the principal AI supercomputer deployments outside the NVIDIA-based AI training infrastructure. The deployment in the UAE, alongside the G42 commercial-customer relationship, has anchored Cerebras' commercial credibility through 2023 to 2026.

The CS-3 system has been deployed at the US Department of Energy's Argonne National Laboratory and other national labs, with scientific-computing workload deployment for protein folding, materials science, and other application domains.

Leadership

As of April 2026, Cerebras' senior leadership includes:

• Andrew Feldman, Co-Founder and Chief Executive Officer. Technology entrepreneur; former CEO of SeaMicro.
• Sean Lie, Co-Founder and Chief Hardware Architect. Chip-design engineer; former senior engineer at SeaMicro.
• Julie Choi, Chief Marketing Officer.
• Dhiraj Mallick, Senior Vice President of Engineering.
• Senior engineering and product leadership across the WSE chip-design, CS-3 system engineering, and Cerebras Inference commercial-product organizations.

Continued senior engineering recruitment has supported the WSE-3 product transition and the Cerebras Inference commercial service launch through 2024 to 2026.

Funding and backers

• Cumulative private capital: Approximately $720 million-plus through multiple rounds. Pre-IPO valuation around $4 billion.
• Latest round (Series F, 2021): $250 million at approximately $4 billion valuation, with Alpha Wave Ventures, Abu Dhabi Growth Fund, and other investors.
• Earlier capital: Foundation Capital, Eclipse Ventures, Benchmark, Sequoia Capital, Coatue, Eclipse, Altimeter Capital, and other investors.
• IPO: Filed October 2024 on the Nasdaq; subsequently paused pending CFIUS review of the G42 customer relationship.

The G42 commercial-customer relationship has been characterized in industry coverage as both a commercial advantage and a regulatory complication for Cerebras' planned IPO. The Microsoft April 2024 investment in G42, with divestiture of certain Chinese technology relationships at the request of the US government, has been considered a contributor to the favorable CFIUS review trajectory through 2025 to 2026.

Industry position

Cerebras occupies a structurally distinctive position as one of the principal commercial alternatives to NVIDIA-based AI compute, with the Wafer-Scale Engine product line, the fastest commercial inference service for open-weights large language models, the Condor Galaxy AI supercomputer build-out with G42, and deployments at the US Department of Energy national labs. Industry coverage has consistently characterized Cerebras as one of the principal alternative-AI-compute companies globally, alongside Groq, SambaNova Systems, Tenstorrent, and other peers.

The 2022 to 2026 supply-chain constraints on NVIDIA H100 and H200 GPUs have continued to anchor commercial demand for alternative AI compute, although the scale-out of NVIDIA's H200, B100, and B200 production through 2025 to 2026 has shifted the supply-demand dynamic.

Competitive landscape

• NVIDIA Research. Direct competitor on the dominant AI compute platform. NVIDIA H100, H200, B100, and B200 GPUs are the principal alternative to Cerebras' wafer-scale architecture.
• Groq, SambaNova Systems, Tenstorrent, Mediatek. Alternative-AI-compute peers with different chip-architecture approaches.
• Together AI, Fireworks AI, Anyscale, Replicate. GPU-based inference services that compete with Cerebras Inference on commercial inference latency and pricing.
• Huawei Noah's Ark Lab Ascend. Chinese alternative-AI-compute peer with different supply-chain and customer-base dynamics.
• G42. Principal commercial customer for the Condor Galaxy AI supercomputer build-out and strategic-customer relationship.
• US Department of Energy national labs including Argonne National Lab, Lawrence Livermore National Lab. Scientific-computing customer base.

Outlook

• The IPO timeline through 2026 to 2027, contingent on CFIUS review and broader US-UAE technology policy.
• The continued WSE product iteration and the next-generation chip-design timeline.
• The Cerebras Inference commercial-service expansion and the continued open-weights model hosting trajectory.
• The continued Condor Galaxy AI supercomputer build-out with G42 and other deployment-region expansion.
• The competitive dynamic with NVIDIA on the dominant AI compute platform, particularly under the H200/B100/B200 production scale-out.
• Continued senior engineering recruitment and senior leadership stability through 2026 to 2027 commercial expansion.

Sources

• Cerebras official site. Company reference.
• Wafer-Scale Engine product page. WSE-3 chip reference.
• Cerebras Inference. Commercial inference service reference.
• Condor Galaxy AI supercomputer. G42 partnership reference.
• Cerebras IPO filing (S-1). October 2024 SEC filing.