Cerebras Systems Inc.
CBRS 4.5Price and target zones价格与目标区间
What the company does公司做什么
Cerebras Systems Inc. operates as an artificial intelligence infrastructure company. It designs and manufactures an AI compute platform that comprises of proprietary systems and software that is delivered in racks for deployment in data centers up to supercomputer scale.
The company's pioneering wafer-scale engine, a chip encompassing a silicon wafer, designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI, and other AI applications.
Its customers include hyperscalers, foundation model labs, AI-native and digital-native businesses, enterprises, and leaders of Sovereign AI initiatives.
The company has operations in the United States, Europe, the Middle East, Africa, and internationally.
Read the full description查看完整介绍
Cerebras Systems Inc. operates as an artificial intelligence infrastructure company. It designs and manufactures an AI compute platform that comprises of proprietary systems and software that is delivered in racks for deployment in data centers up to supercomputer scale. The company's pioneering wafer-scale engine, a chip encompassing a silicon wafer, designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI, and other AI applications. Its customers include hyperscalers, foundation model labs, AI-native and digital-native businesses, enterprises, and leaders of Sovereign AI initiatives. The company has operations in the United States, Europe, the Middle East, Africa, and internationally. Cerebras Systems Inc. was founded in 2015 and is headquartered in Sunnyvale, California.
Company facts公司资料
- Founded成立时间
- 2015
- Headquarters总部
- Sunnyvale, California
- Employees员工人数
- 708
- Industry行业
- Semiconductors
- Country国家
- United States
Investment thesis投资逻辑 from the watchlist来自关注列表
📋 ANALYZED 8/27 — WATCH, and far out of zone. Wafer-scale AI inference; IPO 5/14/26 at $185, −52.6% off the $386 day-one high. The speed claim is real (top of the Artificial Analysis throughput board; no HBM/CoWoS/3nm dependency in a cycle where all three are constrained). The rest of "Nvidia but cheaper" is not: a CS-4 rack holds 132GB SRAM vs ~13.4TB HBM in an NVL72, and
Red flags
- WSE-3 Turbo kept the same 5nm node and the same 44GB — the memory fix is CS-6 in 2028. 14.2% GAAP gross margin vs Nvidia's 74%, at 42x EV/core-sales vs Nvidia's 20x EV/sales — the product is cheaper, the stock is 2x more expensive. Mix is going 30/70 to cloud, i.e. a capital-intensive datacenter operator, and AMD's own flagship partnership assigns Cerebras decode only with Helios doing prefill — a component in someone else's rack
- $1.022B of equity issued to customers in six months (OpenAI/G42/AWS) amortizing as contra-revenue to Oct 2031 — see [[pattern-customer-warrants-are-contra-revenue-that-buys-the-backlog]]
- $25.4B RPO is only ~15% inside 24 months, includes zero-margin pass-through, and grew just $0.8B in H1 — [[pitfall-rpo-inflated-by-pass-through-and-termination-rights]]
- Material weaknesses in internal control, unremediated, covering revenue recognition and equity administration — the exact lines this thesis turns on
- Nvidia's Groq 3 LPX shipped Aug 2026 — the same architectural thesis, bought for $20B, inside the ecosystem customers already buy
- Top 3 customers 76% of revenue; OpenAI is ~32% and lent Cerebras $1.0B and holds warrants for 14% of the company
- CEO sold $47.3M and CTO $151.1M on 8/21. FV $70-145 (central $105). Entry $85-110, strong buy <$85. Trim $280 (dollar-form deliberately — vendor forwardPE 143.6 is built on FY2027 EPS $1.27 while epsCurrentYear is −$0.47, so any "NNx fwd" would render below spot and invert the call). ⏳ Do not buy ahead of the unlock ladder: ~68M shares release Sep 30–Oct 28, then ALL remaining at the Q3 print or ~11/09, into a 94M float with 14% short interest.
📋 8/27 已分析 —— 观察，且远低于买入区间。 ** 晶圆级 AI 推理；2026-05-14 以 $185 上市，较 $386 的上市首日高点下跌 52.6%。 速度优势是真实的（在 Artificial Analysis 吞吐量榜单上位居第一；在 HBM、CoWoS、3nm 三者全部供应紧张的周期中，它对这三样都无依赖）。 但「更便宜的 Nvidia」这个说法的其余部分不成立： 一个 CS-4 机架只有 132GB SRAM，而 NVL72 有约 13.4TB HBM，并且 🚩 WSE-3 Turbo 仍沿用同样的 5nm 制程和同样的 44GB —— 内存问题要到 2028 年的 CS-6 才能解决。 GAAP 毛利率 14.2%，而 Nvidia 为 74%；估值为 42x EV/核心营收，而 Nvidia 为 20x EV/营收 —— 产品更便宜，但股票贵了一倍。 业务结构正在向 30/70 的云倾斜，也就是一家资本密集型的数据中心运营商；而且 AMD 自己的旗舰合作方案只把解码（decode）交给 Cerebras，预填充（prefill）由 Helios 负责 —— 它是别人机架里的一个部件。 🚩 六个月内向客户发行了 $1.022B 的股权（OpenAI/G42/AWS），作为冲减收入项摊销至 2031 年 10 月 —— 见 [[pattern-customer-warrants-are-contra-revenue-that-buys-the-backlog]]。 🚩 $25.4B 的剩余履约义务中只有约 15% 在 24 个月内确认，且包含零毛利的转付收入，上半年仅增长 $0.8B —— 见 [[pitfall-rpo-inflated-by-pass-through-and-termination-rights]]。 🚩 内部控制存在重大缺陷且尚未整改，涉及收入确认与股权管理 —— 正是这个投资逻辑所依赖的科目。 🚩 Nvidia 的 Groq 3 LPX 已于 2026 年 8 月出货 —— 同样的架构思路，以 $20B 买下，并且放进了客户本来就在采购的生态里。 🚩 前三大客户占收入 76%；OpenAI 约占 32%，同时借给 Cerebras $1.0B，并且持有相当于公司 14% 的认股权证。 🚩 CEO 在 8/21 卖出 $47.3M，CTO 卖出 $151.1M。 ** 合理价值 $70-145（中枢 $105）。 买入区间 $85-110，低于 $85 为强力买入。 减持 $280（刻意使用美元形式 —— 供应商的 forwardPE 143.6 建立在 FY2027 每股收益 $1.27 之上，而 epsCurrentYear 为 −$0.47，因此任何「NNx fwd」写法都会渲染出低于现价的水平，把结论完全颠倒）。 ⏳ **不要在解禁阶梯之前买入：约 68M 股在 9/30 至 10/28 之间解禁，其余全部股份在第三季度财报或约 11/09 解禁，而流通股仅 94M，做空比例达 14%。 **
Historical results历史业绩 past fiscal years — these values do not change过去财年——这些数值不会变动
Profit margins利润率 the share of revenue that stays as profit营收中留存为利润的比例
of every $1 of revenue每 $1 营收中的占比
- Revenue growth营收增长
- +74%
Competitive advantages竞争优势 from the thesis and the reports来自投资逻辑和报告
- Moat The quantitative base is not a moat yet ROIC is not meaningfully computable — operating income is negative and invested capital was negative until the IPO recapitalized the balance sheet.
- 🚩 The moat is currently bought , not earned In six months Cerebras issued $1.022B of equity to its customers — $822.9M to OpenAI, $149.4M to G42, $50.0M to AWS — against H1 revenue of $373.5M.
- A moat means you can raise price; Cerebras is paying customers to take the product.
- For a developer consuming an inference API, switching cost is roughly a base-URL change.
- You are right that CUDA lock-in largely does not apply here — and bear theses leaning on porting friction attack the wrong target.