The next AI winner may not be the company with the biggest model. It could be the company that figures out how to get more useful AI for every dollar it spends.Nvidia Corp.(NASDAQ:NVDA) is helping shape that shift by powering more efficient, task-specific AI systems that aim to squeeze more performance out of every GPU cycle.
下一個AI贏家可能並非擁有最大模型的公司，而可能是找到如何讓每投入一美元都能獲得更有用AI的公司。輝達公司（納斯達克：NVDA）正通過為更高效的、面向特定任務的AI系統提供算力支援，推動這一轉變，力求從每個GPU計算周期中搾取更多性能。
DigitalOcean Holdings Inc.(NYSE:DOCN) CEOPaddy Srinivasantold Benzinga in an exclusive email interview that AI builders are increasingly mixing different models for different jobs rather than relying exclusively on expensive frontier systems from companies such as OpenAI and Anthropic.
DigitalOcean Holdings Inc.（紐約證券交易所：DOCN）首席執行長Paddy Srinivasan在與Benzinga進行的一次獨家電子郵件採訪中表示，AI開發者越來越多地針對不同任務混合使用不同的模型，而非完全依賴OpenAI和Anthropic等公司提供的昂貴前沿系統。
"We believe in: right model, right cost, for every task," Srinivasan said.
Srinivasan表示：「我們相信：為每項任務選擇合適的模型、合適的成本。」
He pointed to healthcare AI companyHippocratic AIas an example, saying AI builders like Hippocratic "get better intelligence per dollar" as they optimize across models.
他以醫療健康AI公司Hippocratic AI為例，稱像Hippocratic這樣的AI開發者通過在多個模型間進行優化，「實現了每美元更高的智能產出」。
Hippocratic's connection to Nvidia makes that strategy particularly interesting. Nvidia says Hippocratic's Polaris architecture runs more than 25 task-specific AI models on Nvidia H200 GPUs, while its TensorRT-LLM software makes those models faster, smaller and more efficient, lowering costs and allowing more conversations to run on the same hardware.
Hippocratic與輝達的關聯使得這一策略尤為引人關注。輝達表示，Hippocratic的Polaris架構可在Nvidia H200 GPUs上運行超過25個面向特定任務的AI模型，而其TensorRT-LLM軟件則使這些模型更快、更小、更高效，從而降低成本，並允許在同一硬體上運行更多對話。
The AI Model Doesn't Have to Be the Most Expensive
AI模型未必非得是最昂貴的
Srinivasan said frontier models are typically needed for only about 25% of the job, mainly the hardest reasoning or specialized use cases. The remaining 75% can often be handled by open-weight models, which can offer lower-cost alternatives for less demanding tasks.
Srinivasan表示，前沿模型通常僅在約25%的任務中是必需的，主要是最難的推理或專業應用場景。其餘75%的任務往往可由開放權重模型處理，這類模型能為要求較低的任務提供成本更低的替代方案。
Read Also:EXCLUSIVE: The 75% of AI Workloads That May Not Need OpenAI or Anthropic
延伸閱讀：獨家：75%的人工智慧工作負載可能並不需要OpenAI或Anthropic
That creates a different optimization problem for AI companies.
這為AI公司帶來了另一種優化問題。
Instead of asking which model is the smartest, they can ask which model is smart enough for a particular task at the right price.
它們不再問那個模型最聰明，而是可以問：對於特定任務而言，那個模型在合適的價格下足夠聰明？
"Most AI Native companies today are already multi-model," Srinivasan said. "They all have a mixture of models and route specific prompts to the right model."
Srinivasan表示：「如今大多數原生AI公司已經是多模型架構。它們都混合使用多種模型，並將特定提示路由到合適的模型。」
DigitalOcean's Inference Engine is designed to route workloads based on factors including performance, latency, cost and customer preference.
DigitalOcean的推理引擎旨在根據性能、延遲、成本和客戶偏好等因素對工作負載進行路由分配。
Hippocratic Shows Why Nvidia's Hardware Matters
Hippocratic展示了為何輝達的硬體至關重要
Hippocratic is a useful example because its healthcare AI requires real-time responses while handling safety-sensitive conversations.
Hippocratic 是一個有用的例子，因為其醫療健康領域的人工智慧需要在處理安全敏感對話的同時提供實時響應。
The company's Polaris system runs on Nvidia H200 GPUs and uses more than a trillion parameters across its model constellation. The goal isn't simply to use the most powerful hardware or model available. It is to make the entire system more efficient so Hippocratic can handle more interactions without proportionally increasing its computing costs.
該公司的 Polaris 系統運行在輝達 H200 GPU 上，其模型叢集包含超過一兆個參數。目標不僅僅是使用最強大的硬體或模型，而是讓整個系統更加高效，從而使 Hippocratic 能夠處理更多互動，而無需同比例增加計算成本。
That is the strategy behind Srinivasan's "intelligence per dollar" argument.
這正是Srinivasan提出的「每美元智能」（intelligence per dollar）論點背後的戰略。
'Intelligence Per Dollar' Could Become the New AI Metric
「每美元智能」可能成為人工智慧的新衡量指標
Nvidia itself has increasingly emphasized the economics of AI, often focusing on concepts like performance per dollar and the cost efficiency of AI compute to describe the value businesses can get from their computing investments.
輝達自身也越來越多地強調人工智慧的經濟性，通常聚焦於「每美元性能」和人工智慧計算的成本效益等概念，以描述企業從其計算投資中所能獲得的價值。
That could change how investors view the AI race.
這可能會改變投資者看待人工智慧競賽的方式。
The industry's first phase was dominated by model size, training costs and the race to build increasingly powerful systems. As AI moves into everyday business applications, however, the economics of actually running those models become harder to ignore.
行業第一階段由模型規模、訓練成本以及構建日益強大系統的競賽主導。然而，隨著人工智慧進入日常商業應用，實際運行這些模型的經濟性變得越來越難以忽視。
If companies can use a mix of models and optimize the infrastructure underneath them, the winners may not necessarily be the companies with the biggest AI models.
如果企業能夠結合使用多種模型，並優化其底層基礎設施，最終的贏家未必是擁有最大人工智慧模型的公司。
They could be the companies that figure out how to get the most useful intelligence for every dollar of compute.
它們可能是那些能夠為每一美元的計算支出獲取最有用智能的公司。 (Benzinga)