Google已成为 AI 领域中算力最充裕的公司之一，因为它结合了庞大的云端业务、自研晶片，以及与 Anthropic PBC、Meta Platforms Inc. 等公司的晶片共享协议；但这种优势也让它自己的资源变得稀缺。现任与前任员工表示，Google 研究人员越来越需要与付费客户和内部团队竞争，尤其是在公司搜寻与云端部门需要使用其张量处理单元，也就是 TPUs 时。Andrew Dai，曾任职于 Google 的 AI 实验室，说他是在意识到自己无法取得足够的运算能力、无法在公司内部推进视觉推理工作后离开的；他甚至在用一张棋盘游戏图片测试 Gemini、Google 的旗舰模型时，找到了它的盲点。

这篇文章把算力取得描绘为塑造 Google DeepMind 内部研究选择、升迁与结盟的一大力量。主管们表示，算力是透过一套严谨流程来分配的，需在客户需求与长期研究之间取得平衡，而 Sundar Pichai 也表示 DeepMind 必须获得建构前沿模型所需的资源。Alphabet 另表示，Google Cloud 的积压订单量较前一季几乎翻倍，超过 4600 亿美元，而 Pichai 说 Google 在「短期内受到算力限制」。前研究人员描述，ChatGPT 在 2022 年亮相后，Google 转向大型语言模型，之后又转向以程式码为重点的模型；而被视为较不具即时效益的专案，则面临更紧的预算与延迟。

前员工表示，算力稀缺甚至可能迫使研究人员创办新创公司，因为那样他们可以从多个来源搜集晶片，并避开内部官僚程序。Ioannis Antonoglou 说，Google 在 AlphaGo 上有充足算力，但对于后训练与强化学习的优先事项却不够；Anna Goldie 则说，Google 曾提供她更多算力，试图留住她不要离开，而她的创业公司之后募集到 335 million 美元。Google 在 2023 年将 DeepMind 与 Brain 合并，结束了那种较为自上而下的实验室与另一个研究人员使用波动不定的内部晶片额度的分裂状态；然而到了 2024 年，一次大型训练运行据报仍让部分研究暂停了约四分之一。更广泛的意涵是，AI 竞赛如今不仅取决于谁拥有最多算力，也取决于谁能最有效地使用它，以及稀缺资源是否会迫使 Google 优先做与 Gemini 相关的短期工作，而不是更具实验性的想法。

Google has become one of the most compute-rich companies in AI because it combines a large cloud business, in-house chips, and chip-sharing deals with firms such as Anthropic PBC and Meta Platforms Inc., yet that advantage has also made its own resources scarce. Current and former employees say Google researchers increasingly have to compete with paying customers and internal teams, especially when the company’s search and cloud units need its tensor-processing units, or TPUs. Andrew Dai, formerly in Google’s AI lab, said he left after realizing he could not get enough computing power to pursue visual reasoning work inside the company, even after finding a blind spot in Gemini, Google’s flagship model, while testing it on a board game image.

The article depicts compute access as a major force shaping research choices, promotions, and alliances inside Google DeepMind. Leaders say compute is allocated through a rigorous process that balances customer needs with long-term research, and Sundar Pichai has said DeepMind must get the resources needed to build frontier models. Alphabet also said Google Cloud’s backlog nearly doubled from the prior quarter to over $460 billion, and Pichai said Google is “compute constrained in the near term.” Former researchers described a shift after ChatGPT’s 2022 debut pushed Google toward large language models and then code-focused models, while projects seen as less immediately useful faced tighter budgets and delays.

Former employees said compute scarcity can even push researchers to found startups, where they can gather chips from multiple sources and avoid internal bureaucracy. Ioannis Antonoglou said Google had ample compute on AlphaGo but not enough for post-training and reinforcement learning priorities, while Anna Goldie said Google offered her more compute to try to keep her from leaving and that her startup later raised $335 million. Google’s 2023 merger of DeepMind and Brain ended a split between a more top-down lab and one where researchers used fluctuating internal chip credits; still, in 2024 a large training run reportedly paused some research for about a quarter. The broader implication is that the AI race is now defined not only by who has the most compute, but by who can use it best, and by whether scarce resources force Google to favor short-term, Gemini-related work over more experimental ideas.