“现在几乎肯定存在一些泡沫。” 城堡创始人兼CEO肯·格里芬10月20日接受彭博专访时说，任何对经济的彻底改造，通常都和泡沫缠在一起。 当年给互联网公司估值时，有些商业模式相当离谱，但他只恨自己当年没留着亚马逊的股票。 泡沫过后留下的东西常常改变了世界，而在泡沫破了以后还活着的公司，会当几十年的市场领袖。 过去两年大家做的是主题投资，策略是“我全都买”：买云巨头、买芯片公司、买五年前听都没听过的韩国、台湾、日本公司。 但现在市场开始做研究，大家变得更挑剔，开始往后看两三年，格里芬也抛出了他真正担心的三件事。 第一个担心，是GPU的供给如果超过需求，租价会自由落体。 他说GPU像航空公司，多跑一点的成本很低。 一架没坐满的飞机，多出来的座位卖20美元也是赚的。 航空业之所以是噩梦，就是因为一旦运力过剩，价格就自由落体。 2000年电信大建设就是这样，光纤铺得比需求多，电信价格一路往下掉。 如果GPU的供给超过需求，租出去能收的价钱真的会自由落体，这会推翻今天大家的投资假设。 第二个担心，是GPU到底能用几年。 现在有人用五六年前的GPU还能卖出非常高的价钱，但这大概撑不到再过五六年。 买了GPU、把算力租出去的人，在芯片用到第七、第八年以后还能从上面收回多少钱，已经不能再按12个月前那样去算。 第三个担心，是买单的企业正在压账单。 整个美国企业界都在看着Anthropic和OpenAI的账单，盘算这笔钱花下去到底能不能帮公司干出活来。 最成熟的美国公司正在分清，哪些业务要付高价用最先进的模型，哪些业务改用开源模型、只花零头。 《华尔街日报》提到，正在准备上市的Anthropic在招股书里写了未来7500亿美元的开支。 一边是模型公司加码，一边是企业压成本，格里芬承认这确实矛盾。 关键看Anthropic能不能让自己的模型变成非用不可，比如帮药企找到癌症的治法，让药企愿意付任何价钱，而我们不知道答案。

IMO, today was another DeepSeek moment Every 2nd/3rd tier inference provider now has a model and a recipe that can generate 90%+ margins on their compute DeepSeek is also promising future model support and optimizations This creates a huge incentive for inference providers to serve, as well as push their customers to use more DeepSeek models and utilize the DeepInfer inference framework DeepSeek gets proliferation and entrenchment, while the inference providers get to make 90% margins

穆尼尔等闲不会去，要去就应该有一定成果

Deep|LLM: Enterprise AI Adoption Survey (Vol. 6) - In Our Sample: AI Spending Growth Expected to Slow; Routine Work Shifts to Lower-Cost Models AI spending growth is expected to slow at most companies in this sample. The fintech group’s spending growth is expected to slow after a rapid initial AI rollout. At the pharmaceutical group, better model selection and procurement terms are expected to curb spending growth. Spending growth at the pharmacy and health insurance group is expected to remain broadly steady. Claude-based workflows can replace specialist software for feature engineering. The subprime auto lender’s in-house workflow lets its team direct the search for variables, require explanations of their business meaning and account for compliance constraints. With that workflow meeting its needs, the lender plans to cancel or not renew a specialist tool costing ~$80K-100K a year. Matching models to tasks can cut AI spending even as token use rises. After training, some teams at the fintech group roughly halved AI spending while using more tokens. The pharmacy and health insurance group’s expert estimates that better routing and model choice could reduce API costs by 20%-30% for the same workload; its automatic router is still under evaluation. Some companies tie additional AI funding to expected cost savings, revenue gains or less manual work. The brewing group asks business owners to explain expected cost savings or additional revenue and whether they will commit to changing workflows and reducing headcount after full implementation. At the pharmacy and health insurance group, proposals to scale a use case must explain how it will reduce manual processing, handling time or rework. Most companies in this sample have not adopted Fable for routine production; one interviewed company reports using Astra for cross-system commercial analysis. Cost, a lack of recurring tasks that justify it, and data-retention terms constrain broader Fable adoption. The mega brewing group uses Astra to combine market strategy, brand plans and financial information with customer-pricing information. Self-hosted open-weight models already handle about half of API calls at one interviewed company. The pharmaceutical group has incorporated model downloading, scanning, verification and tuning into its internal process. Its expert roughly estimates model-use costs to be ~90% lower than with external APIs. These workloads avoid external per-token charges but still incur internal deployment and operating costs. Detailed Report fundaai.substack.com/p/deepl…