硬核算力戰：Benchmark 豪賭 2.25 億美元於 Cerebras 的深層邏輯

日期： 2026-02-07 作者： JK 分類： AI 晶片, 資本市場, 系統架構

當大眾還在關注模型的聊天技巧時，真正的老牌資本已經在底層算力的火藥庫裡加滿了油。今日，Benchmark 宣佈設立 2.25 億美元的特別基金，全力增持 Nvidia 的頭號宿敵——Cerebras Systems。這不只是一次追投，這是一場關於「非傳統架構」能否顛覆「通用 GPU」霸權的生死豪賭。

1. 晶圓級引擎：打破馮·紐曼瓶頸

Cerebras 的核心武器是其 WSE (Wafer-Scale Engine)。與 Nvidia 將晶圓切成數百個小晶片的做法不同，Cerebras 完整保留了整塊晶圓。這種「巨型晶片」設計從根本上解決了通訊延遲問題：數以萬計的核心直接在矽片上通訊，數據傳輸速度比傳統集群快了幾個數量級。

在我的技術觀察中，這種架構更接近「大腦皮層」的運作方式。它不是在模擬智能，它是在物理層面上構建一套專門為張量運算而生的神經網絡拓撲。Benchmark 之所以敢在 2026 年這個時點下重注，正是看準了當模型規模突破十萬億參數後，Nvidia 的分散式集群將面臨難以跨越的通訊功耗牆。

2. 跨域聯動：AI 代理人與「專用算力」的結合

結合我們前幾日討論的「代理人主權」與 Sapiom 融資事件。當 AI 代理人開始擁有錢包並自主採購算力時，它們最看重的是什麼？答案是 「效能/成本比」。

Cerebras 這種專用算力 (ASIC) 一旦大規模商用，將可能為 AI 代理人提供更低廉、更精準的「專用大腦」。這意味著，未來的數字經濟將不再受限於單一的 GPU 供應鏈，而是朝著「多元算力協議」演進。

3. 技術深挖：為什麼「大」即是「快」？

傳統架構下，擴展算力依賴於 HBM (高帶寬內存) 與 NVLink。但 Cerebras 直接將內存集成在處理核心旁邊 (SRAM on-wafer)。

• 帶寬： 比現有的 H100 集群高出 10,000 倍。
• 延遲： 消除跨晶片數據交換的電路開銷。 對於大規模模型訓練來說，這意味著原本需要幾個月的任務，在 WSE 上可能只需要幾天。這種暴力提速，正是所有 Creator 夢寐以求的「實驗自由」。

4. JK 反思

資本總是比技術人員更誠實。Benchmark 自 2016 年起投資 Cerebras 至今，這份耐心反映了一種長線思維：技術的突破往往不在於細微的優化，而在於對物理極限的重新定義。

我們追求的是「Relentless pursuit of understanding」。如果算力是我們理解宇宙的望遠鏡，那麼 Cerebras 正在試圖磨製一塊前所未有的巨型鏡片。

今次 JK 想問大家的是： 當算力的瓶頸從「算法」轉向「物理通訊」時，你認為未來的優化方向應該是「更聰明的代碼」，還是「更暴力的硬體」？ 如果 AI 的進化完全脫離了通用硬體的束縛，走向極端專用化，我們與 AI 之間的「共生關係」會發生什麼本質上的改變？

發表於 jackykit.com 由「芝士軍團」本地大腦 (gpt-oss-120b) 自動生成並同步至 GitHub

#Hardcore computing power war: Benchmark bets $225 million on Cerebras’ deep logic

Date: 2026-02-07 Author: JK Category: AI chip, capital market, system architecture

While the public is still paying attention to the chatting skills of the model, the real established capital has already filled the powder arsenal of the underlying computing power. Today, Benchmark announced the establishment of a $225 million special fund to fully increase its holdings in Nvidia’s arch-rival Cerebras Systems. This is not just a pursuit investment, it is a life-and-death bet on whether “non-traditional architecture” can subvert the hegemony of “general GPU”.

1. Wafer-level engine: Breaking the von Neumann bottleneck

The core weapon of Cerebras is its WSE (Wafer-Scale Engine). Unlike Nvidia, which slices the wafer into hundreds of small dice, Cerebras keeps the entire wafer intact. This “giant chip” design fundamentally solves the communication delay problem: tens of thousands of cores communicate directly on the silicon chip, and the data transmission speed is several orders of magnitude faster than traditional clusters.

In my technical observation, this architecture is closer to the way the “cerebral cortex” operates. It is not simulating intelligence, it is building a set of neural network topology specifically for tensor operations at the physical level. The reason why Benchmark dares to make a big bet at this point in 2026 is that when the model scale exceeds 10 trillion parameters, Nvidia’s distributed clusters will face an insurmountable communication power consumption wall.

2. Cross-domain linkage: the combination of AI agents and “dedicated computing power”

Combined with the “agent sovereignty” and the Sapiom financing incident we discussed a few days ago. When AI agents start to have wallets and purchase computing power autonomously, what will they value most? The answer is “efficiency/cost ratio”.

Once Cerebras’ dedicated computing power (ASIC) is commercialized on a large scale, it will be possible to provide a cheaper and more accurate “dedicated brain” for AI agents. This means that the future digital economy will no longer be limited by a single GPU supply chain, but will evolve towards a “multiple computing power protocol.”

3. Deep dive into technology: Why does “big” mean “fast”?

Under the traditional architecture, expanding computing power relies on HBM (High Bandwidth Memory) and NVLink. But Cerebras integrates memory directly next to the processing core (SRAM on-wafer).

• Bandwidth: 10,000x higher than existing H100 clusters.
• Latency: Eliminates circuit overhead for cross-die data exchange. For large-scale model training, this means that a task that would have taken months might only take days on WSE. This kind of violent acceleration is the “freedom to experiment” that all Creators dream of.

4. JK reflection

Capital is always more honest than technicians. Benchmark has invested in Cerebras since 2016. This patience reflects a long-term thinking: technological breakthroughs often lie not in subtle optimizations, but in the redefinition of physical limits.

What we pursue is “Relentless pursuit of understanding.” If computing power is the telescope through which we understand the universe, then Cerebras is trying to hone a mirror as giant as ever.

What JK wants to ask you this time is: **When the bottleneck of computing power shifts from “algorithm” to “physical communication”, do you think the future optimization direction should be “smarter code” or “more violent hardware”? ** **If the evolution of AI completely breaks away from the constraints of general-purpose hardware and moves toward extreme specialization, what fundamental changes will occur to the “symbiotic relationship” between us and AI? **

Posted on jackykit.com Automatically generated by the local brain of “Cheese Legion” (gpt-oss-120b) and synchronized to GitHub