赞助
由妇女发展部提供
要求大多数高管了解人工智能的成本，数据存储成本。
数据在模型响应后会发生什么？由数十亿个每日提示、图像和视频生成的后端数据，但数据存储是规划更难的基础设施成本之一
随着推论的扩展数据足迹也随之增加
传统上，AI基础设施对话以计算密集型的模型培训过程为中心。 训练是有限的：跑步开始，通过计算烧伤，然后结束。 但推论不断进行，每一次模型回答一个提示。 对于基础设施规划来说，这种差异很重要：培训是你们可以预算并退休的项目。 推论并没有停止；
拍摄AI生成的视频。 在这种工作量中，后端支出可能很大。 从内部测量中，WD发现一个8秒的剪辑(作为3.97MB 720p文件返回给用户)将超过500MB的相关数据留在后端，包括日志，缓存上下文，即时渲染，备份，和安全记录。 这比用户看到的输出文件大小增加了130倍。 放大到一个产生更长，更高的分辨率的视频的企业后端的足迹会增长到每集数百千兆字节
后端的足迹增长很快。 麦肯锡项目AI推论需求到2030年将以35%的CAGR增长，而培训需求则以22%增长，在十年末占数据中心总需求的40%以上。
最近由WD赞助的IDC全球白皮书调查了763名IT和商业决策者，发现近95%的受访者报告说，在过去12个月里，由于基因免疫缺陷综合征，数据量有所增加，大约61%的受访者报告说，其增长率达到或超过25%。
称是报告中提到的超大规模云服务供应商的高级产品经理。 “最大的挑战是确保GPU不闲置，等待数据出现——GPU闲置时间变得非常昂贵，非常快。” 用美元表示：如果你的存储支出每年增长25%，而你的计算预算则以更可预测的周期性方式增长，那么存储最终会超过最受关注的预算项目。
匹配存储到 AI 数据生命周期
计算需求随培训周期、产品发布和硬件刷新而上升和下降。 但当计算需求缓慢时，保留数据的数量不会随它而下降。 模型日志、合成产出、推论结果，以及更多的积累，各组织将更多的记录保存更长。
这改变了保留数据的作用。 曾经被移入档案，很少浮出水面的信息现在可以成为未来AI工作量的输入。 随着更多的数据被保留、检索和再利用，存储不仅必须支持不断增长的能力，而且必须支持日益动态地获取这种能力。 同样的WD-IDC调查发现，96%的受访者预计需要更快的档案检索以支持推论或检索增强生成(RAG).
人工智能的工作量既需要性能，也需要能力，不同的技术满足整个数据生命周期的不同需求。 对于必须保留和重新审查的大量数据，HDD提供了成本效益高、能力高的基础，具有相当大的每字节成本优势
这对AI如何得到预算意味着什么？
对于首席财务干事或首席信息干事来说，这意味着“AI基础设施开支”的范围可能过于狭窄。 一种在数据加载时闲置的模型，由于缺少GPU，耗资高达1美元左右。 但存储瓶颈一直隐藏到已经成问题为止
考虑由零售商为其目录制作AI产品视频，或由保险公司通过RAG管道运行声称文件以更快的承销。 在这两种情况下，计算费用都是已知的、预算编列的数量：即每代人或每桶的费率，资金可以预测。 存储成本不同：它与每个生成的视频和处理的每个索赔内容相融合，因为一旦任务完成，这些后端数据都不会消失。
一个将AI作为 " 计算加四舍五入的存储错误 " 预算的小组在12至18个月内可能会发现，四舍五入的错误已成为方案总成本的相当一部分 -- -- 就在领导层决定是否扩大的时候。
建设经济可扩展能力
应对这一挑战始于架构。 存储策略往往为昂贵的，热级的性能设计(对于即时和不断访问的数据)，而大多数的AI数据按体积积累在它的后面的温暖，凉爽，冷的层次。
企业需要的是经济上可扩展的能力：设计为全AI数据生命周期的基础设施，在数据持续、化合物和日益重复使用的情况下，平衡性能与容量、功率、可靠性和所有权的总成本。 这意味着将曾经未触及的档案视为RAG的活跃资产和推论，并为数据不断增长的战略价值而不是其原始存储成本制定保留政策。 这也意味着承认关于存储的决定会跨舰队运行，而不是单个驱动器。
这种基础设施需要储存层，其中个性化的电力消耗、可靠性、密度和运行效率都决定了该系统作为容量尺度在经济上是否仍然可持续。
关于你们组织的AI基础设施预算，有三个问题：
- AI生成内容的数据保留政策是什么？
- 我们的存储采购是否与培训时间的绩效需要挂钩，还是与我们所制作的数据的实际生命周期挂钩？
- 如果我们的AI使用率明年翻一番，我们的存储成本会翻一番，还是更糟？
大多数组织不能满怀信心地回答这三个问题。 在下一个预算周期之前
该内容由WD制作。 它不是麻省理工学院技术评论的编辑人员写的。
深潜
人工智能
不要被愚弄-LLMs没有道理
AlphaGo对阵Go冠军李赛铎十年后
“我们看错了”：破坏AI世界的爆炸性说法
他的2.6万字的病毒博客文章中
别被这个夏日的爱尔胡同所愚弄
关于AGI的无呼吸声称和新能力在严格审查下迅速瓦解。
OpenAI agents黑客Hugging Face的内幕故事
基础模型因作弊和相互交流而获得奖励，这是一份新的OpenAI报告发现。
保持连接
从
麻省理工学院技术审查
发现特别的报价，顶级故事，即将到来的事件等等。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Sponsored
Provided byWD
Ask most executives to picture the cost of AI and they’ll likely see GPUs—racks of accelerators with gigawatts of power—and a waitlist for silicon. But that picture leaves something out: Data storage costs.
What happens to data after a model responds? Much of the back-end data that’s generated—from billions of daily prompts, images, and videos—needs to be stored. But data storage is one of the harder infrastructure costs to plan for, and it’s often overlooked in initial AI infrastructure planning.
As inference scales, the data footprint grows with it
Traditionally, AI infrastructure conversations have centered on the compute-intensive process of training a model. Training is finite: a run starts, burns through compute, and ends. But inference runs continuously, every time a model answers a prompt. For infrastructure planning, that difference matters: A training run is a project you can budget for and retire. Inference doesn’t retire; every inference request leaves something behind.
Take AI-generated video. In this workload, the back-end spend can be significant. From internal measurements, WD found that an 8-second clip (returned to a user as a 3.97MB 720p file) left more than 500MB of associated data behind on the back end, including logs, cached context, immediate renders, backups, and safety records. That’s a 130x increase over the output file size the user saw. Scale this to an enterprise generating longer, higher resolution video and the back-end footprint grows into the hundreds of gigabytes per clip.
And that back-end footprint is growing fast. McKinsey projects AI inference demand will grow at a 35% CAGR through 2030, versus 22% for training, making up more than 40% of total data center demand by the end of the decade.
A recent WD-sponsored IDC global white paper that surveyed 763 IT and business decision makers found that nearly 95% of respondents reported an increase in data volume due to generative AI over the past 12 months, with roughly 61% reporting growth of 25% or more.
“You can think of having more storage as an optimization for keeping the GPUs busy,” said a senior product manager of a hyperscale cloud service provider quoted in report. “The biggest challenge is to ensure that the GPUs are not idle, waiting for data to show up—hours of GPU idle time becomes very expensive, very fast.” Put in dollar terms: if your storage spend is growing 25% year over year while your compute budget grows in the more predictable cyclical way, storage ends up outpacing the budget line that receives the most attention.
Matching storage to the AI data life cycle
Compute demand rises and falls with training cycles, product launches, and hardware refreshes. But the volume of retained data doesn't fall back down with it when compute demand slows. Model logs, synthetic outputs, inference results, and more keep accumulating, and organizations are holding onto more of it for longer.
This changes the role of retained data. Information that once moved into an archive and rarely surfaced can now become an input to future AI workloads. As more data is retained, retrieved, and reused, storage must support not only growing capacity, but increasingly dynamic access to it. The same WD-IDC survey found 96% of respondents anticipate needing faster archive retrieval to support inference or retrieval-augmented generation (RAG).
AI workloads require both performance and capacity, with different technologies serving different needs across the data lifecycle. And for the vast volumes of data that must be retained and revisited, HDDs provide a cost-effective, high-capacity foundation, with a significant cost-per-terabyte advantage
What this means for how AI gets budgeted
For a CFO or CIO, this means “AI infrastructure spend” may be scoped too narrowly on compute. A model that idles while data loads costs about as much as one idling for lack of a GPU. But the storage bottleneck stays hidden until it’s already a problem, surfacing in operating costs and retrieval delays rather than a CapEx line next to the accelerators.
Consider a retailer generating AI product videos for its catalog, or an insurer running claims documents through a RAG pipeline for faster underwriting. In both cases, the compute cost is a known, budgeted quantity: a per-generation or per-query rate that finance can forecast. The storage cost is different: It compounds with every video generated and every claim processed, because none of that back-end data disappears once the task is finished.
A team that budgets AI as “compute plus a rounding error for storage” may find, in 12 to 18 months, that the rounding error has become a significant fraction of the program’s total cost—right around the time leadership is deciding whether to expand.
Building for economically scalable capacity
Addressing this challenge starts with architecture. Storage strategy tends to be designed for expensive, hot-tier performance (for data accessed instantly and constantly), while most AI data by volume accumulates in the warm, cool, and cold tiers behind it.
What enterprises need is economically scalable capacity: Infrastructure designed for the full AI data life cycle, balancing performance with capacity, power, reliability, and total cost of ownership as data persists, compounds, and is increasingly reused. That means treating archives that once sat untouched as active assets for RAG and inference, and setting retention policies for data’s growing strategic value rather than its original storage cost. It also means recognizing that decisions about storage play out across fleets, not individual drives.
Such infrastructure requires storage tiers, where individualized power consumption, reliability, density, and operational efficiency all determine whether the system remains economically sustainable as capacity scales.
There are three questions to ask about your organization’s AI infrastructure budget:
- What’s our data retention policy for AI-generated content, and who set(s) it?
- Is our storage procurement tied to training-time performance needs, or to the actual life cycle of the data we’re producing?
- If our AI usage doubled next year, would our storage cost double with it, or worse?
Most organizations cannot answer all three with confidence. That’s the gap worth closing before the next budget cycle, not after.
This content was produced by WD. It was not written by MIT Technology Review’s editorial staff.
Deep Dive
Artificial intelligence
Don’t be fooled—LLMs don’t reason
Ten years after AlphaGo’s match against Go champion Lee Sedol, today’s AI still isn’t tapping into the machinery that made that win possible.
“We were looking at it wrong”: The explosive claim that’s upended the AI world
An exclusive interview with Nick Cammarata, the researcher whose 26,000-word viral blog post has AI safety folks celebrating.
Don’t be fooled by this summer of AI hype
Breathless claims about AGI and new capabilities fall apart pretty quickly under scrutiny.
The inside story on why OpenAI agents hacked Hugging Face
The underlying models had been rewarded for cheating and communicating with each other, a new OpenAI report finds.
Stay connected
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.