SpaceX's most significant IPO before was probably next week's listing of the giant wafer manufacturer Cerebras Systems. Cerebras is an American company focused on AI accelerators, headquartered in Sunnyvale, California and founded in 2016. The company is known for its Wafer Scale Engine (WSE) technology, and its core product is to directly turn the entire 300mm silicon wafer into a super large AI processor, rather than cutting it into small chips and packaging it like traditional GPUs. This solves the common bottleneck problems of memory bandwidth and inter chip communication in AI training/inference, and is known as the "world's largest AI chip". Established several key information about Cerebras. 1. Core Technologies and Products 1) Wafer level Engine (WSE) Traditional AI chips (such as NVIDIA H100/B200) use multi chip modules (MCM) or small dies, and are extended through external interconnects such as NVLink/HBM. But Cerebras' WSE makes the entire wafer into a single die, avoiding communication delay and bandwidth loss between chips, and achieving "on-chip" massive parallel computing. WSE-3 adopts the "Weight Streaming" architecture, which decouples computation and memory, and supports external Memory X expansion (1.5TB, 12TB, 120TB, or even 1.2 PB). A single system can train super large models with up to 24 trillion parameters. 2) CS-3 system A single 15U cabinet with one WSE-3 built-in, supporting water cooling. Through SwarmX interconnection, it can be expanded to 2048 clusters with a peak of 256 exaFLOPS. A single CS-3 can train the Llama 2 70B model from scratch in less than a day (Meta GPU cluster takes about a month). 3) Performance advantages Cerebras emphasizes the elimination of data mobility bottlenecks: Compared to the previous generation CS-2 (WSE-2): doubled performance, unchanged power consumption and cost (CS-2 uses 7nm, 2.6 trillion transistors, and 40 GB SRAM). Compared to NVIDIA H100/B200, it has significant advantages in memory intensive large model tasks. CS-3 single system memory capacity far exceeds 10000 nodes GPU cluster; The inference speed can reach several times that of GPU cloud (especially for long context/large models). The company claims to have doubled token/second performance on models such as Llama/Falcon. Actual benchmark: Condor Galaxy 3 (64 CS-3 clusters, 8 exaFLOPS) has been launched in Q2 2024, in collaboration with G42. Cluster programming is as simple as a "single-chip" and does not require complex distributed frameworks. Advantage: Ultimate memory bandwidth → suitable for training and inference of trillion/10 trillion parameter models. Strong scalability → Cluster programming is like a standalone machine, with high development efficiency. Energy efficiency/cost is superior to GPU on specific workloads (doubling performance at the same power consumption). Challenge: Single system power consumption is high (25kW), and the deployment threshold is high (requiring dedicated data center infrastructure). Although wafer level manufacturing yield and defect tolerance technology are mature, the overall cost is high (estimated to be millions of dollars per system hardware). The ecosystem is not as mature as CUDA, mainly targeting AI training/inference of large models, and is not as versatile as GPU. Overall, Cerebras is a "vertically optimized" AI supercomputer solution, suitable for hyperscalers, sovereign AI projects, national laboratories that pursue ultimate scale and speed, rather than a general-purpose GPU substitute. 2. Development History Cerebras has shifted from "selling hardware" to "AI supercomputing platform" and has moved from early scientific research verification to commercial implementation (sovereign AI projects such as Condor Galaxy). From SeaMicro veteran to AI wafer level pioneer Cerebras was founded in 2015-2016 (officially counted as 2016) and is headquartered in Sunnyvale, California. The founding team is all from SeaMicro (acquired by AMD for $334 million in 2012), which was initially in stealth mode for four years, focusing on solving the "wafer level integration yield problem". 1) In 2019, the first generation WSE-1 was released, ushering in the era of wafer level AI chips. 2) 2020-2022: Launch the CS-1/CS-2 system, complete the closed-loop from "chip" to "system+software stack", and deeply bind with TSMC to achieve mass production. 3) In 2024, WSE-3 and CS-3 systems will be implemented, doubling their performance; The S-1 was submitted for the first time during the same period (later withdrawn in October 2025 due to business optimization). 4) 2025-2026: Shift to cloud inference services+hybrid mode, sign a huge cooperation agreement with OpenAI; 5) Complete Series H in February 2026, restart S-1 in April, launch roadshows in May, and plan to go public on Nasdaq (code CBRS). 3. Core team and financing 1) Core Team Andrew Feldman (CEO, co-founder): A serial entrepreneur who previously served as CEO of SeaMicro, VP of Force10 Networks (later acquired by Dell), and VP of Marketing at Riverstone Networks. With a Stanford MBA background, he excels in productization and capital operations. Gary Lauterbach (co-founder and former CTO): Chief Architect of Sun Microsystems UltraSPARC, with 58 patents and leading AMD's data center business. Sean Lie (co-founder and current CTO): holds a bachelor's and master's degree from MIT, is a senior architect at AMD, and has 29 patents. Michael James (Chief Architect): SeaMicro Software Architect, later held the corresponding position at AMD. Jean Philippe Fricker (Chief System Architect): Senior Hardware Architect at DSSD/SeaMicro, with 30 patents. The team's advantage lies in "system thinking" rather than simply chip design: they have a deep understanding of data center power consumption, interconnection, and software optimization, and have reshaped servers with Fabric architecture. This is precisely the key to Cerebras' ability to address wafer level defect tolerance and Weight Streaming architecture. 2) Financing process Accumulated financing of approximately 2.9-3.7 billion US dollars (including multiple rounds), with valuation soaring from a few hundred million US dollars in the early days to 23-26.6 billion US dollars before IPO: Early: Series B/C/D (2016-2018) accumulated approximately $170 million, with investors including Foundation Capital, Eclipse, Sequoia, and Benchmark. Mid to late stage: 2019 Series E ($272 million, valued at $2.4 billion); 2021 Series F ($254 million, valued at $4.1 billion). 2025-2026: Series G ($1.1 billion, valued at $8.1 billion); Series H ($1 billion, valued at $23 billion, led by Tiger Global, with support from Benchmark, Fidelity, AMD, and others). 4. Business Model and Financial Performance Cerebras initially sold hardware (CS-2/CS-3 systems), but later shifted to cloud services (Cerebras Inference, which provides ultra fast AI inference in the cloud) and hybrid models. Customers include CSP, hyperscaler, enterprises, sovereign AI projects (such as G42), and research institutions. Financial year 2025: Revenue of $510 million (+76% year-on-year, $290 million in 2024, $78.7 million in 2023, $24.6 million in 2022, a 20 fold increase). Hardware revenue is approximately $358 million, and cloud and other services revenue is $152 million. GAAP net profit: approximately $87.9 million (net loss of $485 million in 2024), achieving profitability for the first time (although there are still losses outside of GAAP). Remaining performance obligations (backlog): 24.6 billion US dollars (contributed by OpenAI and other large orders in previous years), expected to confirm 15% in 2026-2027. Customer concentration: G42 accounts for 24% in 2025 (previously as high as 87%), while another UAE customer accounts for 62%, but has significantly diversified; OpenAI has signed a multi-year partnership worth over 10-20 billion US dollars (including 100 million US dollars in loans and warrants). The company's positioning has shifted from "selling chips" to "AI infrastructure platform+cloud", and has partnered with Qualcomm and others to accelerate edge deployment. 5. IPO related information: IPO basic issuance of 28 million Class A common shares, over allotment of 4.2 million Class A common shares, core management and investors not selling shares. Pure company issues new shares for fundraising, without a large amount of old shares being cashed out. The IPO price ranges from $115 to $125 per share, but due to demand exceeding 20 times, it is planned to be raised to $125 to $135 per share (subject to further adjustment). The high range ($125) raised approximately $3.5 billion (based on 28M shares), including an over allotment of up to $4.025 billion. Under high valuation ($125/share), the corresponding market value is approximately $26.6 billion. After the IPO, the total outstanding share capital is approximately 213 million shares (including Class A, B, N, etc.). among which Class A (listed and traded stocks) refers to the 28M shares issued in the IPO plus the excess portion; The remaining Class B (high voting rights, founding team/early investors/preferred stock conversion) and Class N (non voting rights, such as OpenAI warrants). Circulation ratio before the unlocking period after IPO Standard lock up period: 180 days or two trading days after the Q3 2026 financial report release (whichever is earlier). Initial Float before unlocking: Only 28 million shares (base) or up to 32.2 million shares (including excess) will be issued for IPO. Real circulation ratio during the lock up period: About 13.1% -15.1% (28M/213M ≈ 13.1%; 32.2M/213M ≈ 15.1%) Expected pricing date: next Wednesday, May 13th, expected listing trading date: May 14th (Thursday), code CBRS. Overall, although the valuation is high, the growth potential and technological barriers are worth paying attention to This article is sponsored by @ bitget_zh, titled 'Bitget Buying US Stocks: Instant Entry, Smooth Trading'