THE SHORT VERSION
- In an interview, Sam Altman said it is quite terrible for people to keep a large share of the world's limited DRAM in personal computers used a small fraction of the time.在一次采访中，Altman 称让世界有限的 DRAM 留在个人电脑里“很糟”，因为这些机器大部分时间并不使用它。
- He argued that a supply-constrained market needs maximum utilization in the cloud, which is where OpenAI's closed-source models run.他主张在供给受限的市场里应最大化云端利用率，而 OpenAI 的闭源模型正运行在数据中心。
- Tom's Hardware read the remark as a defence of OpenAI's business model, and noted that the argument assumes data-centre buildouts continue at pace.Tom's Hardware 认为这番话是在为公司商业模式辩护，并指出该论证依赖数据中心建设持续高速推进。
What he actually said他到底说了什么
During an interview, OpenAI chief executive Sam Altman was asked about local inference: the idea that open-source models will soon run well enough on a personal computer that users will not need data-centre capacity for everything. The Verge reported that he called it "terrible" to have RAM in personal computers instead of data centres, and gave the reason in terms of global supply rather than product design.在一次采访中，OpenAI 首席执行官 Sam Altman 被问到“本地推理”——即开源模型很快会在个人电脑上运行得足够好，让用户不必事事依赖数据中心。The Verge 报道称，他把把内存留在个人电脑而非数据中心称为“很糟”，并给出的理由是全球供给，而不是产品设计。
His argument is about allocation. If the world's DRAM supply is limited, he said, it is quite terrible for everybody to hold a share of it in a machine that uses it five percent of the time, and the better use is maximum utilization in the cloud. That is a claim about efficiency, and it is also, as the interviewer's question implied, a claim about which architecture AI should be built on.他的论证关于分配。如果全球 DRAM 供给有限，他说，让每个人都把其中一部分放在一台只有 5% 时间在用它的机器里，是相当糟糕的；更好的用法是在云端实现最大化利用率。这是一个关于效率的主张，同时——正如提问所暗示的——也是关于 AI 该建立在哪种架构上的主张。
The utilization argument, stated fairly把利用率论证说完整
The logic is not exotic. DRAM is produced in fixed quantities over a given period, and demand from AI data centres has been strong enough to tighten supply across the industry. In that environment, every gigabyte installed in a desktop is a gigabyte that is idle most of the day, and idle capacity is expensive when the alternative is a server running inference around the clock for many customers at once.这套逻辑并不稀奇。DRAM 在给定周期内的产量是固定的，而 AI 数据中心的需求强到足以让全行业供给吃紧。在这种环境里，装进台式机的每一 GB，都是大部分时间闲置的一 GB；而当替代方案是服务器昼夜不停地为许多客户做推理时，闲置产能的代价很高。
By that measure the cloud wins, and Altman's phrasing is consistent with how capacity planners talk about a shortage. The weak point is the comparison itself. A consumer machine is not competing with a data-centre rack for the same workload; it is doing local inference for one person, and the question is whether that use is worth the memory it occupies. The quote answers the macro question and skips the micro one.按这个标准，云端胜出，Altman 的措辞也符合容量规划者谈论短缺的方式。弱点在比较本身。消费级机器并不与数据中心机架争夺同一类工作负载；它为一个人做本地推理，问题是这种用途是否值得它占用的内存。这段引语回答了宏观问题，跳过了微观问题。
It is also a business argument它同时也是商业论证
Tom's Hardware made the commercial reading explicit: OpenAI has some open-source models, but its primary business is closed-source models that run in data centres, which makes the remark a defence of the company's model. That does not make the allocation argument wrong, but it explains why the chief executive of a data-centre business would reach for it.Tom's Hardware把商业解读挑明了：OpenAI 确有一些开源模型，但主营业务是运行在数据中心的闭源模型，因此这番话是在为公司模式辩护。这并不使分配论证变成错的，但解释了为什么一家数据中心企业的首席执行官会伸手去用它。
The outlet also pointed to the subscription angle. If local models get good enough, users could pay less for a service like GPT-6 Luna because they can run something similar at home, which is exactly the pressure a company selling model access would want to keep at bay. The interview was wide-ranging, covering PHP vulnerabilities and sandboxing models during training, but this is the part with a commercial interest attached.该媒体还提到订阅这一层。如果本地模型足够好，用户就可能少为 GPT-6 Luna 这类服务付费，因为他们可以在家里运行相近的东西——这正是销售模型访问权的公司希望压住的压力。这次采访内容广泛，还谈到 PHP 漏洞与训练中给模型加沙箱，但带商业利益的是这一段。
What the argument assumes这套论证假设了什么
Two assumptions carry the claim. The first is that data-centre buildouts keep going at their current pace, so the demand for DRAM stays high and supply stays tight. The second is that RAM used continuously in an AI data centre is worth more than RAM used for spreadsheets, browsing, games and the occasional local model. Both may be true; neither is self-evident.有两个假设支撑这个主张。第一是数据中心建设按当前速度继续，因此 DRAM 需求维持高位、供给维持紧张。第二是 AI 数据中心里昼夜运转的内存，比用于表格、浏览、游戏与偶尔本地模型的内存更有价值。两者都可能成立，但都不是不言自明。
The second assumption is the more contestable, because it compares very different outputs. A cloud GPU running tokens for paying customers has a measurable revenue figure attached. A desktop with 32 gigabytes of memory has a person using it, and the value of that is not captured by utilization percentages. An argument about scarce memory that treats idle consumer RAM as waste has already decided what counts as productive use.第二个假设更可争议，因为它比较的是截然不同的产出。为付费客户跑 token 的云端 GPU 有可度量的收入。而一台装着 32 GB 内存的台式机背后是一个人在使用它，这种价值无法用利用率百分比捕捉。一个把闲置消费级内存视为浪费的稀缺性论证，已经预先决定了什么算“生产性使用”。
The case for keeping memory local把内存留在本地的理由
The reason the question keeps coming up is that local inference has real advantages that have nothing to do with memory economics: it works offline, it keeps data on the machine, and it does not require a subscription or an account. This site has covered Mistral opening the weights of its flagship, one of several moves that make running a capable model on your own hardware more plausible than it was a year ago.这个问题不断被提起，是因为本地推理有与内存经济学无关的真实好处：可离线工作、数据留在本机、不需要订阅或账号。本站报道过Mistral 开放其旗舰模型权重，这是让“在自有硬件上运行一个够强的模型”比一年前更可行的若干动作之一。
The counter-pressure is hardware. Memory prices have risen as data-centre demand competes with consumer supply, and this site reported on OpenAI's appetite for compute and chips as part of the same squeeze. If DRAM stays expensive and scarce, the local option becomes a premium choice rather than a default one, and Altman's allocation argument gets stronger for reasons that have nothing to do with his company's business model.反向压力来自硬件。随着数据中心需求与消费级供给竞争，内存价格已经上涨；本站也报道过OpenAI 对算力与芯片的需求是同一轮挤压的一部分。如果 DRAM 持续昂贵且稀缺，本地选项就会变成高端选择而非默认选项，Altman 的分配论证也会因为与其公司商业模式无关的原因而变得更强。
What is established, and what is not哪些已经确定，哪些还没有
Established: Altman said in an interview that keeping the world's limited DRAM in personal computers is quite terrible and that supply-constrained markets need maximum cloud utilization; he was answering a question about local inference with open-source models; OpenAI's main business runs closed-source models in data centres. Not established: whether local inference will stay a niche, whether the sentiment of the remark reflects OpenAI's roadmap, or how much consumer memory prices will change the calculation.已经确定的：Altman 在采访中称，把世界有限的 DRAM 留在个人电脑里很糟，供给受限市场应最大化云端利用率；他回答的是关于开源模型本地推理的问题；OpenAI 的主要业务是运行在数据中心的闭源模型。尚未确定的：本地推理是否会一直是小众、这番话是否反映 OpenAI 的路线图，以及消费级内存价格会在多大程度上改变这道算术。
For readers, the useful way to read this is as a stated preference from a company with a position in the outcome. The technical argument about utilization is worth taking seriously, and it is not the whole argument. Whether the world's memory should sit in a rack or on a desk is a question about who gets to run models and on what terms, and that is a policy choice as much as an engineering one. This site's coverage of Meta's AI infrastructure financing shows how much capital is being committed to one answer.对读者而言，有用的读法是把这当作一家在结果中有利益的公司所表达的偏好。关于利用率的技術论证值得认真对待，但它不是全部论证。世界的内存该放在机架上还是桌上，是一个关于谁能运行模型、以什么条件运行的问题，而这既是工程选择，也是政策选择。本站关于Meta 的 AI 基础设施融资的报道，说明了有多少资本正押在这一个答案上。
Sources & reporting limits来源与核查范围
Follow Industry to find related stories on your next visit.关注产业动态，下次访问时阅读相关文章。
