概要
To control token costs and protect data privacy, there is a growing demand for private inference
• Top researchers from New York University, Stanford University, Dartmouth College, and University of Hawaii at Manoa are using B3IQ to support their AI work, including cancer research and specialized AI model training.
• A survey of 1,800 IT leaders by Broadcom in 2026 found that 56% of enterprises have already implemented or plan to run production-grade inference on private cloud infrastructure, while the usage of similar workloads on public clouds has dropped from 56% to 41% within a year.
• B3IQ currently operates a 27,000-square-foot facility in Oregon and plans to rapidly expand its inventory of US-assembled NVIDIA GPU systems to meet the growing demand.
Privacy is becoming an increasingly prominent issue in the field of AI. Every prompt sent to centralized AI service providers is like a deposit in someone else's bank vault: research data, business logic, and creativity flow into systems beyond the sender's control, with terms not set by themselves. The most straightforward solution is to run AI models on one's own hardware, but in the past, this method was out of reach for the vast majority: high-end GPU systems cost tens of thousands of dollars, are in short supply, and require space and professional operations.
Today, B3 Labs has released B3IQ to fill this gap: B3IQ is a groundbreaking AI infrastructure service designed to give universities, enterprises, and professional users greater control over the hardware, models, and data on which their AI workloads depend.
Prior to this, institutions requiring AI power had only two imperfect options: renting from cloud vendors, which means prices fluctuate, no ownership, and limited supply in times of GPU shortage; or purchasing an entire machine, which entails upfront costs, plus power, cooling, and maintenance burdens. B3IQ offers a third hybrid model that combines the benefits of both: economically you "own" it: assets and profits belong to you; operationally it is "hosted": the data center and maintenance are handled by the platform.
B3IQ 用戶可通過分期付款購置專屬 NVIDIA GPU 系統:由 B3 Labs 投資的美國 AI 系統製造商 Andromeda 在美國製造,託管於俄勒岡州。通過 B3IQ dashboard,機主可將閒置算力與算力需求撮合變現,收益可用於抵扣購機餘款,或留作收入。付清全款後,機主可選擇繼續由 B3IQ 託管,亦可安排硬件實物交付。
B3IQ 的早期用戶包括紐約大學、達特茅斯學院、夏威夷大學馬諾阿分校和史丹佛大學的教員、AI 研究人員與學生團隊。對這群人來說,GPU 短缺帶來的預算壓力格外真實:
「B3IQ 的『所有者掌控』模式,是介於租用與購買之間的一條可行路徑,這正是我們決定與 B3IQ 合作的原因。科研經費是固定且提前撥付的,而雲端計算成本是浮動的,可能在項目中途悄悄吃掉一整條預算線。把算力變成一項可預知、可編入預算的成本,規劃會更容易,向課題負責人或經費管理部門交代也更容易。」夏威夷大學 AI 研究員 Pavel Bushuyeu 表示,「擁有自己的算力,還能讓我們免受『算力荒』的影響。GPU 供應緊張時,中心化服務商會限量分配,學術用戶往往排在付費企業客戶之後。有了自己的節點,需要跑任務時就無需排隊爭搶。而當系統閒置時,將其閒置算力收益的一部分用於抵扣購置成本,也能攤薄這筆投入。」
對於無法把敏感數據交給 Anthropic、OpenAI 等第三方模型服務商的研究者和機構而言,這更是一個規模可觀的市場。夏威夷大學的 Pavel Bushuyeu 等 B3IQ 試點用戶,正在平台上運行專有模型進行癌症研究與機器人訓練,這些敏感數據根本不能外傳。此外,AI 服務商還會直接屏蔽特定關鍵詞與主題,可能讓整塊研究領域無從開展:紐約大學全球事務中心 Yorke E. Rhodes III 教授的 Ethical Tech CoLab 實驗室裡,碩士生們基於戰區撤離數據構建框架與模擬、推演國家級外交談判,這類工作會觸發商業模型的內容過濾,只能運行在團隊自有的基礎設施上。
B3 Labs 認為,這一模式現在能成立,靠的是開源模型的快速進步:如今可下載、可自行部署的模型已經足夠強,機構不再必須調用 OpenAI 等中心化廠商的介面,前提只剩一個:你得有自己的機器來跑。而這正是 B3IQ 補上的一環。
「機構希望更好地掌控 AI 在哪裡運行、數據如何被處理,以及為算力付出多少成本。」B3 Labs 首席技術官 Sean Geng 表示,「B3IQ 把這些決策收進同一套系統:用專屬硬件運行私有工作負載,再通過一個可自主選擇加入的網路,讓閒置 GPU 算力產生價值。」
無論你想購置 GPU、出租閒置算力,還是直接租用網路裡的算力,都可以在 b3iq.org 瞭解詳情。
B3 Labs 為企業級 AI 構建軟體與硬體。公司成立於 2024 年,創始團隊來自 Coinbase,已獲得 Pantera Capital、Coinbase Ventures 等機構逾 2,100 萬美元投資。B3 Labs 運營兩條產品線:B3OS 是企業 AI 執行引擎,讓 AI 智能體在企業系統內穩定、可控地執行任務;B3IQ 是客戶完全擁有、部署在美國本土的 GPU 基礎設施。更多資訊請訪問 B3OS.org 與 B3IQ.org。
歡迎加入律動 BlockBeats 官方社群:
Telegram 訂閱群:https://t.me/theblockbeats
Telegram 交流群:https://t.me/BlockBeats_App
Twitter 官方帳號:https://twitter.com/BlockBeatsAsia