DeepSeek发布V4冲击闭源定价 硅谷异构计算与Token效率崛起

DeepSeek发布V4冲击闭源定价 硅谷异构计算与Token效率崛起 DeepSeek V4凭借极致的Token效率与低成本长上下文突破，为美国闭源大模型划定了结构性的商业“死亡线”，正倒逼全球AI产业链向非英伟达推理芯片与软硬件协同的异构计算加速演进。 本期节目深入探讨了 DeepSeek V4 凭借其混合注意力机制（CSA + HCA）、流形约束超连接（MHC）及 MUON 优化器等核心技术，在长上下文和推理成本上实现的极致工程突破。节目对比了中美大模型因资源多寡而形成的路线差异，指出中国厂商在“资源倒逼”下对 Token 效率的极致追求已为硅谷闭源巨头划下“商业死亡线”。同时，嘉宾们还分析了异构计算与非英伟达芯片适配的系统级挑战，并揭示了硅谷内部 Anthropic 凭高资本效率与专注度在企业端和二级市场逆袭 OpenAI 的内幕。 DeepSeek V4 通过混合注意力机制（CSA + HCA）、MHC 架构及 MUON 优化器等工程突破，在 1M 超长上下文下将计算成本降至三分之一、KV Cache 占用降至十分之一，推动行业竞争焦点全面转向 Token 效率。 (1:54) 中美大模型呈现技术路线分化：硅谷巨头依赖充裕算力追求极致智能与推理能力，而中国厂商在资源约束下优先攻克 Token 效率与低成本推理，为闭源巨头划定了商业“死亡线”并冲击其定价权。 (7:51) 非英伟达芯片在算子、通信和框架等环节面临系统级挑战，DeepSeek 通过推出“太龙”（TaiLong）语言及软硬件协同设计适配国产芯片，加速了数据中心在训练与推理端走向异构化分工的趋势。 (16:51) 企业实际业务中已有约 80% 的常规任务转向开源模型，动摇闭源商业模式基石；而在资本市场中，资本使用效率更克制的 Anthropic 在二级市场受追捧，反衬出 OpenAI 面临的基础设施过度开支质疑。 (59:59)

《硅谷101》对话技术与投资专家指出，DeepSeek V4凭借极致的Token效率与软硬件协同设计，打破了长上下文与低成本推理的门槛；这不仅倒逼硅谷重构大模型的定价逻辑与资本预期，也加速了AI算力向异构芯片分化。 架构创新推动Token效率与长上下文革命 (1:54) 华为云及相关技术专家肖志斌指出，DeepSeek V4的突破在于将CSA（压缩稀疏注意力）、HCA（重度压缩注意力）与滑动窗口相结合，辅以MHC（流形约束超连接）和MUON与ADAM混合优化器，克服了长上下文推理中KV Cache与计算成本呈平方级激增的问题。在100万（1M）超长上下文场景下，其PRO版本的计算成本降至原先的三分之一，KV Cache占用降至十分之一。 面对高频消耗Token的Agent时代，中国厂商在GPU供应与资本约束下激进推行软硬件协同设计，大幅拉低了API价格。DeepSeek在论文中对硬件厂商提出通信带宽与计算力的配比建议，警告过高带宽会挤占计算面积，展示了以架构与工程手艺压榨硬件极限的路径。 国产芯片适配工程与算力异构化演进 (16:51) 据肖志斌分析，DeepSeek V4适配华为昇腾、寒武纪等芯片，不仅涉及矩阵乘法，还需攻克融合MOE注意力、稀疏注意力与低精度算子等Kernel级挑战。DeepSeek为此开发了类Triton的国产领域特定语言“太龙”（TaiLong）。此外，MOE模型的调度通信、PD分离与连续批处理等Serving Runtime架构、训练容错机制以及全套编译调试生态均是必经难关。 业内预测，芯片市场未来将摆脱“单卡打天下”的局面走向异构化。超大规模紧耦合训练仍依赖英伟达等高带宽互联方案，但推理负载、内存敏感的长上下文处理及低延迟Agent流，将分流至国产芯片、TPU或专用LPU上，进而催生出自动匹配模型与硬件的AI决策调度层。 开源低成本路线重塑硅谷商业死亡线 (34:59) 硅谷投资人与创业者视角显示，DeepSeek等开源模型对美国闭源模型巨头划定了一条“死亡线”。企业实际落地中，约80%的常规任务已被转移至小参数开源模型组合，仅20%的极复杂任务保留给闭源大模型，直接侵蚀了闭源厂商的收入基石。每当开源生态推出对标方案，闭源公司的第二梯队模型定价权便遭遇腰斩。 业内观察到，OpenAI推出的GPT-5.5在长文本Pro版上每百万Token收费高达180美元，与DeepSeek的低价路线形成鲜明反差。投资人预期在未来12个月内，企业为边际智能提升支付超额溢价的意愿将大幅缩水，AI竞争正从片面追求Benchmark智能转向严苛的投资回报率（ROI）考量。 巨头分化与大模型IPO前夕的资本考验 (44:59) 在美股资本市场，Anthropic与OpenAI的IPO前景呈现明显分化。由于Anthropic坚持克制的算力开支，并在Claude Code企业编程与合规安全赛道实现突破，资本利用效率显著高于OpenAI，其老股在二级市场受追捧；而OpenAI因激进采购算力和多线作战，导致机构在上市前寻求套现，市场上频现千万美元级别的老股转让。 从大模型估值周期来看，OpenAI与Anthropic已进入严审营收、留存与烧钱率的“看指标”第二阶段；而DeepSeek仍处于看技术愿景与团队的“第一阶段”。正如OpenAI绑定微软、Anthropic绑定AWS，DeepSeek选择在此节点融入华为昇腾生态，旨在借力重构中国本土的软硬件AI闭环。 行业竞争的演变表明，迈向AGI的核心瓶颈不仅在于提升智能上限，更在于推理效率与工程普及度；然而，DeepSeek未来同样需要穿越从技术愿景走向严苛商业财务验证的周期考验。

[0:00] 没有效率，AGI 就只能是个 Demo；有了效率，AGI 才能成为真正的产品和基础设施。 [11:15] 单个任务花多少钱 [22:54] 肖志斌: 一张卡打天下 [22:54] 肖志斌: 芯片混用 [30:31] Token 预算管理器 [34:59] Jenny Xiao: 把模型做得既好又便宜 [34:59] 效率即智能 [34:59] Jenny Xiao: 死亡线 [34:59] Jenny Xiao: 顶在他们背后的一支枪 [44:59] 成熟的成年人 [44:59] Peter Steinberger: Claude Code 就是我的电脑 [49:59] 模型能力倡议 [59:59] 快速行动、打破常规 [64:59] 指数级增长的梦想 [64:59] 卖梦想 [64:59] 看指标

[0:00] Deep Seek V4来了那么最近模型之战又打起来了Deep Seek之外呢我们看到Kimi K2.6OpenAI的GPT 5.5还有谷歌新一代的TPU以及Android新融资消息几乎是同时出现了在如此多的声音当中今年的Deep Seek Moment对于AI市场来说意味着什么呢工程的这个完成度有非常大的坚细提高token efficiency的这个商量继续仪器绝成了合处的机制CSA加HCAMHC还有就是那个Muron的优化器token efficiency是达到AGI或者更强Agent System的这个必备之路或者是基础条件没有效率AGI就只能是个demo但是有了效率AGI才能成为真正的产品和基础设施短期看呢英伟达并不会被取代因为英伟达的优势并不仅仅是一个GPU我认为最大的重点是它是一个死亡的或者是一个死亡的发生的发生在美国的发生如果你是发生的发生而你会被创造成的开放的发生你的发生是一种货币的我认为这种课题是在这种课题里的发生发生的发生 Asked by an open source company, the value of your business is essentially zero. I think the lesson here is that efficiency is a part of intelligence. Open source models like DeepSeq is one of the biggest structural threats to closed source business model in the U.S. Antropic and OpenAI are going public this year. There is way more appetite for Antropic than there is for OpenAI. A lot of institutions are selling OpenAI. DeepSeek V4's specialization is not only the technology, but it is a support of 100 million tokens. The agenda is called Agenda Coding and the more-to-be-reviewed. And it has been reported that it is a way to use the smartphone, which is the technology and technology that is not only the most important part of the world. This video is a podcast with us with the two guests. We have talked about DeepSeek's development and innovation and innovation. The development of the smartphone and the development of the new smartphone. The development of the smartphone and the development of the openAI and the endobabic. and from the public side of the topic, and from the financial markets and the economic markets, and from the financial markets and the perspective of the public. First, I am with the guest of肖志斌. He is a computer-based software, he is a software-based software, and is a software-based software, and the CEO of the Huawei. Thank you,志斌. Thank you. Thank you. Let's talk about DeepSecretes' 8th. Do you think you have any other exciting points? Or are you all in the ideas? I think the 次DeepSick的发布对你来说你觉得有什么惊喜的点吗还是说都在意料之内我觉得大的方向是在意料之内但是工程的完成度有非常大的惊喜为什么说在意料之内因为整个AI行业包括DeepSick以前发布的很多的论文大家过去比的是模型参数训练的这种规模Benchmark现在越来越重要的是Token Efficiency然后Deep Seek从去年年初的时候开始移民金人到注重Token Efficiency到现在这个新的这个发布也是一直在这个提高Token Efficiency的这个上继续一记绝成了好的好的我们来说一下这个这次主要的三个核心技术好不好就分别是我看到有混合注意力机制分别是CSA加HCA然后还有MHC也就是一个Transformer的升级架构还有就是那个MUON的优化器能不能分别跟我们说一下这三个技术突破是如何帮助Dipsick v4实现更好的模型效率的因为Dipsick的论文它的那个论文里面其实已经着重强调的是这三个核心技术其实它还有其他很多其他的技术但是最重要的是这三个跟别的公司或者别的模型与众不同的首先是这个CSA跟HCA其实它解决的就是长夏文推理的这个Attention跟KVCash的成本问题大家都知道这个Attention机制每个Token要跟以前一个Sequence里面的Token所有的ContextKVCash做这种Coreation然后这个成本是平方的这个关系然后CSA跟HCA是一个混合的一个Attention的机制它其实解决的是说当前的Token并不是说跟所有过去的Token都相关CSA其实是Compressed Sparse Attention那这个它的这个思路其实是相当于是做一个精度的这个非常精确的搜索他把以前多个token的KVCash压缩成一个compressed KV Entry然后再通过Spot's Attention选择最相关的一部分的compressed KVCash来做Attend就是说做相关性的这种检索然后HCA是Heavily Compressed Attention你可以认为是作为一个全局的一个快速的 [4:59] attention的检索它相当于是把以前当前token之前的这些context做深度的压缩然后再做一个dense的attention的这种相关性的curation当然中间其实还有一个技术叫sliding window就是说在通过这种吸收或者是这种压缩的方式它的attention有可能是跟最近的这些token是有强相关性的这一部分不能丢掉所以CSA跟HCA加sliding window这三个技术保证其实是长上下纹的attention的压缩这样可以降低的其实重要的是降低推理的成本这是第一个技术那第二个技术是MHCMHC是manifold Constraint Hyperconnections如果训练过这种神经网络都知道这种层跟层之间需要做一个信息的流转所以它有个residual的这种link然后把层跟层之间的信息可以传输到不同的层保持这个训练的稳定性然后其实HC hyperconnection其实是前面一个paper提出来的然后Deepseq提出的是Manifold Constiint Hyperconnection其实是相当于是原来的Resident Link是单条的高速通路HCA是做了一个多条的高速通路在这个层跟层之间然后Manifold Constraint是在这个多层的这个Resident Connection上面加了一个约束保证数值的稳定性这其实是提升了这个生成模型的训练的稳定性然后第三个技术其实是MUON它其实是用来提升这个AI模型的训练的收敛的速度和训练的稳定性然后这样可以用来训练更大和更大更复杂的模型所以这是个优化器所以总结来说呢三个方向那个CSA HCA是降低上下文的推理成本MHC是提升复杂生成架构的这种训练的稳定性然后MUON这种新的Optimizer是提升训练的收敛的效率和稳定性当然它也不是完全它里面所有的很多的模块并不是完全用MUON的Optimizer替代它还有原来的那个ADAM的Optimizer还会有一些模块还是用这个ADAM的Optimizer所以三者叠加Deep Seek才可以达到这个Token的Efficiency的一个新的高度模型效率这个Token Efficiency是目前几乎很多的模型厂商都在追求的事情特别是中国的模型厂商那你观察到就是中国的模型厂商和硅谷的模型厂商他们在技术路线上大家追求的有什么样的区别吗我觉得两者呢其实不是完全不同的路线但是它的优先级有一些先后顺序就是硅谷的这些头部的公司因为有更多的算力资本和资源所以他们更多的是追求更强的模型更好的智能比如说更好的 reasoning的能力多模态的能力agentic的这种能力还有安全性的能力和整个完整的一个产品生态这就是硅谷的头部的模型公司那中国的模型厂商呢像DeepSig,KIMI,MIMAX他们大部分情况是在更强的一个资源约束情况下做创新因为受到GPU的供应啊训练成本还有商业化的价格压力都很大所以会更激进的采取更多的hardware software code designmodel hardware code design和system hardware code design去追求模型的效率比如说像MOE的架构更激进的MOE的架构更激进的这种上下文压缩的这个技术还有低成本的推理还有非英伟达芯片的适配像API的这个价格也弄得也非常低所以我觉得是中国的模型厂商是被资源倒逼更早的进入了一个追求Token或者是Efficiency的这个创新然后硅谷的模型厂商呢在这个模型智能上面其实还有领先性但也会越来越重视Token Efficiency因为Agent的时代很不一样Chatbot是一次性的问答那Agent是一个常任务多能攻击的调用还有这种反思和规划Token的消耗是原来的10倍甚至100倍所以没有Token Efficiency下一步这个agent的这种system很难大规模的商业化所以硅谷的这些模型公司也会把这个token efficiency就是或者模型的efficiency模型跟硬件的匹配然后提到更高的这个优先级所以你是觉得DeepSeek目前还有Kimi的这些尝试其实是能够把这个 [9:59] 从主到这些硅谷的大模型厂商他们去进一步的去规划下一步他们怎么去做Token Efficiency的事吗是的特别是Deep Seek的论文因为对于我们做AI Infrastructure来说读Deep Seek论文读的就非常的亲切非常的带劲因为它里面有很多这种很小的各种各样的这种优化的策略然后这些优化的策略其实在学术界里面其实很多其实很多idea其实学术学里面有各种各样的尝试但是第一次这样用在一个这么大规模的模型上面而且是稳定的训练出来结果这是非常激动人心的对对你觉得他们能够把这些小技术然后拼接在一起然后能够出稳定的结果这件事情它的难点在哪里呢它是一个很engineering上面的一个难点吗它其实是首先是必须要有信心就是训练这种非常规的比如说 Sparse Attention因为我们以前也做过这种稀疏的这个模型压缩那其实首先是要有信心还有一个其实是真正的有压力就是说正是因为有这种计算资源的限制所以我必须倒逼我必须去做这样的尝试和创新我有一个观点就是其实大模型的架构其实可能不是特别的重要更重要的是数据然后所以模型架构要去匹配硬件的架构这样就可以再搭配适量的一样的数据那这样可以提高efficiency那Deepseek其实也是走的这个方向在Deepseek论文里面其实还提了一些这种对硬件厂商的这个建议比如说通信跟communication的这种配比Deep Seeker论文里面提到有一个当超过一定的这个bandwidth以后这个GPU跟GPU之间的这个带宽并不是越高越好因为这样的话其实是更多的这种芯片的面积是放在了这种通信而不是放在compute所以这里面有一个精确的这个配比一个是信心第二个是工程上的难度也非常大其实因为你不可能是每一个技术都要在一个全的模型上面去验证所以一定是在一点一点的验证一点一点的这种集体的智慧汇集到最后是训练一个大的把所有技术用在一起的这种模型好呀然后Deep6V4号称这个计算成本是之前版本的三分之一然后内存占用的是降到了十分之一你觉得就会给这个硅谷的领先模型厂商这几家OpenEye 和Drabic 和Gemini带来什么样的压力吗对 我觉得这里要准确一点就是说它的计算成本并不是说所有场景下都是三分之一或者KV Cash的size不是所有场景下都是十分之一它是在一兆这种大的这种上下文的场景下面就是100万的这种Token的上下文的场景下面它的PRO的版本可以计算成本是之前的三分之一然后KVCash的size降到十分之一DeepSick这个V4更多解决的是一个长下文推理场景下的巨大的效率的提升这也是下一个AI agent时代的这个核心的一个Botterneck或者就是Challenge但是即使是这样对硅谷的这种模型公司压力其实也是有一定的压力的因为首先是价格的压力因为这种Enterprise的客户最终关心的是每一个任务多少钱而不是模型参数多大所以如果一个同样的模型在上下文代码Coding Agent Workflow上面做到足够好但是它的价格非常低那么这种B原的模型的API就会面临重新的一个定价的压力然后第二就是长上下文的压力以前比如说Google的Gemline长上下文做得非常好那现在这个Deep Seek可以把长上下文跟更低的这种KVCash成本结合起来说明长上下文的这种Case可以更高质量但是可以低成本的使用所以对其他的这些模型公司也是一个很大的压力第三个是Agent的成本就是Deepseek这种长上下纹的这种推理成本的降低可以大大的降低Agentic这种Workload的这种Token的消货量所以不能说Deepseek全面超越Deepseek的整个模型的质量在有些方面是有优势有些方面跟硅谷的这些模型公司还是有一些差距的但是Deepseek的真正带来的是一个效率问题所有的这些模型厂商必须回答这个如何提高你的token的efficiency好那接下来token efficiency那会不会是所有模型都想去追求的东西但是它与这种追求更智能的AGI它会是两条技术路线吗你觉得 [14:59] 我觉得不是两条路线恰恰相反Token Efficiency是达到AGI或者更强Agent System的必备之路或者是基础条件所以未来更强的模型不仅仅是参数更大而且是需要更多的这种Test Time Compute更长的Raisoning更长的上下文更复杂的工具调用更长时间的这种任务的这种连续性如果每个Token都很贵模型就不能长时间思考也不能大规模的服务用户所以我觉得AGI其实是构建于这种Token Efficiency之上的所以真正的这种可用的AGI或Jetting System一定是Token Efficiency是非常高的所以有一句话叫做没有效率AGI就只能是个Demo但是有了效率AGI才能成为真正的产品和基础设施这个观点挺好的然后就是来讲讲芯片然后我看到这个路透书的报道说Deepseek V4已经是适配华为的升腾还有韩武器等国产芯片并且部分的V4 Flash训练也是使用了华为的芯片那这个信号对于英伟达来说意味着什么我觉得这个要区分这种报告和那个Deepseek的这种论文Deepseek的技术报告里面并没有太多的讲这种国产芯片的适配但是他提了很多对于硬件厂商的这种软硬件协同设计的一些需求或者是他们看到的一些insight但是怎么说呢短期看了英伟达并不会被取代因为英伟达的优势并不仅仅是一个GPU它是就像老黄说的英伟达的优势其实是Cuda的生态各种通信的库优化的算子NV-LinkInfinity Band的这种Networking还有Profiling Tool还有整个开发者的生态包括英伟达的成熟的供应链所以英伟达的这个优势还是非常的巨大的但是长期来说Deep Seek可以降低这种long contest的这种attention的这种对compute和kvcash的要求所以可以降低硬件能够跑这种大规模的推理模型的这个难度这会让很多这种非英伟达的芯片有机会可以承接部分的这种推理的workload尤其是对于这种固定模型固定场景的大规模部署的推理任务有非英伟达不仅仅是国产的也有可能是其他非英伟达包括AMD和Google的TPU包括HyperScaler自己做的芯片在推理场景上面可能会有更大的这种deployment这个有意思但是在pre-training上面还是得用GPU最好是吗对在training上面的话长期来看还是GPU是default在推理上面其实会有更多各种各样的芯片出现但是训练上面有没有任何的可能可以去替代掉英伟达的整个生态呢现在还没有任何的其实Google已经在做一些相关的事情像Anthelopic很多模型包括Google的Gemini都是在TPU的生态上面去训练出来的这其实是证明了这种如果你能够对整个软件生态包括硬件和系统都有掌控的话是可以在训练上面也能够替代部分替代英伟达的GPU的芯片的那他这个就是得特别的customized然后就针对着他的需求去训是吗对而且要有很多的资源的support有强大的信心还有对这种time to market的这种就是说有一定的浓冷度因为如果对time to market非常要求非常高的话还是肯定是在现成的GPU的这种英伟达的生态里面训练模型会更快我换个方式来问这个问题那比如说包围深藤还有含武器这类国产的芯片如果要承接Jipsick还有KIMI这样的模型的话它最大的难点是什么是不是可以比如说训练跟推理方面分别来跟我们讲一讲我觉得训练跟推理的话其实整体来说其实不是一说单颗芯片的一个算力而是整个系统软件站和整个集群的这种工程能力的一个一个适配吧因为这里面分成不同的层次首先是算子层面因为Deep Seeker论文里面讲到的这个算子不仅仅是有GEM这种矩正存法还有这种Fused MOE的AttentionSparse AttentionMOE的LoutingTop K然后All to AllKVCash管理低精度包括FP4的这种试词这些都需要非常强的这种Kernel Level的支持Deep Seeker也就是 [19:59] 一个这样泰浪这样的国产的就是国内开源的这种domain specific的编程语言来做这个kernel level的这个优化这是第一点kernel跟算子要做很多的这个工程的工作第二是针对通信就是MOE的这种模型对这种有各种的这种并行的这个调度而且还有很多对于通信的这种pattern的这种支持所以里面会需要大量的支持dispatch combine all to all这种interconnection的库芯片的率力即使很高如果通信communication做不好的话也会被拖垮第三个是serving的这种long time跑通模型并不等于高效服务上线服务也还有很多的当然这个大部分可以在基于开源的框架在做比如说VM SG-Long做这种contincent batchingPD分离有KVCache的管理等等这里面其实不同的芯片或者不同的系统要做很多的适配第三刚才讲到训练训练的这种稳定性其实是非常难的在训练的模型过程中因为有一个数值的精度的这个稳定性的问题所以大规模的训练训练的这种稳定性其实是非常难的在训练的模型过程中因为有一个数值的精度的稳定性的问题所以大规模的训练需要长时间的稳定运行所以对于芯片和软件系统的这种Fold Tolerance 容错 Checkpoint 数值一致性都有很高的要求推理的迁移相对来说要容易一些第五个就是开发者生态就是从这种Train in the Framework到整个像类似Chiton或者是TaiLong这样的这种framework还有像Nicco这样的通信库所以整个国产或者是非英伟达体系需要在compiler,debugger,profile还有其他的很多这种工具链上面的这种生态都需要足够的成熟所以非英伟达芯片我们不一定是说国产芯片要补的不是一个点而是一个整个AI-infer的soft stack但是有一点比较好因为现在AI agent或者是AI的能力非常强在kernel这个level其实在通过AI的能力其实是可以补齐的但是在更上层的这种system level的这种orchestration其实是需要深度的这种soft hardware code design需要很多的工作对有一个大概的时间吗就是如果大家非英伟达的这些厂商all in去做这件事情有一个时间线他可以在几年之内完成这样的一个事情其实现在像Google的TPU已经在很多场景下面其实是在推理上面是可以承接自己的这种就是说可以代替英伟达的GPU的所以我是觉得在推理上面会进展会比较快在训练上会相对会慢很多因为现在AI模型的这种竞争是非常的激烈就是你可能很难是说花大量的时间去做适配更应该的是大部分如果是实在是没有办法才会去做这种适配如果不是的话大部分情况还是应该用英伟达的这种GPU来做训练嗯嗯嗯嗯明白对然后刚才你也说到嘛就是那个4月份的时候我们看到谷歌其实也是发布了它最新一代的TPU然后呢它是把AI训练跟推理任务拆分成两款独立的芯片来就是应对LM的一个极致的需求那你觉得目前TPU给英伟达造成了多大的压力呢其实压力还是蛮大的因为首先就是至少是Google的Gemini包括Anthlopic的模型在Google的TPU集群上面训练包括推理这个已经证明是可行的所以证明如果hyperscala可以通过这种4stack的soft hardware code design完全是可以降低对英伟达的这种依赖的但是这种模式呢对于很多外面的这种芯片公司其实很难复制因为Google呢有自己的模型有自己的cloud有自己的数据中心有自己的compilerlong time和自己的芯片和系统所以它可以做一体的模型跟硬件的优化如果不做这种深层次的这种co-design的话其实就像老黄说的芯片公司同样的AI芯片公司如果它芯片价格卖就是不要钱他都很难比得过英伟达的这个就是很难竞争的过英伟达的这个芯片和系统所以这种模式对芯片公司其实是蛮难复制的但是Google这种拥有全套这种解决方案的公司其实就很有效所以TPU最大的意义就是说让AI-INFRA从 [24:59] 所有的workload的默认就是GPU变成了这些workload的也许用在推理上面用专用的ASIC更划算这件事情是证明了这件事情然后在对英伟达压力的另外一个压力就是Google先行后面Amazon MicrosoftMeta包括OpenAITikTok这些甚至于其他一些大模型厂商都会认真的评估自研芯片的可能性特别是在推理上面对好呀对你刚才说到谷歌有这样的一个全战的能力吗那如果我们对比一下国产的话华为也有这样的能力对不对然后他现在如果跟Deepseek如果是一直从一开始就适配的话感觉也是能够跑出来的对这里其实有一个这种产能的问题了就是说这个supply chain能不能供应出同样性能的这个芯片就设计能力其实大家会越来越接近但是还有一些physical的这种limitation就是物理的现实的这种限制对理解理解对然后刚才我们其实说到TPU它其实是分成了两款独立的芯片你觉得就是未来的芯片会不会不再是一张卡打天下的这种感觉而是比如说训练芯片推理芯片然后还有什么常上下的芯片agent workflow这些芯片分化的一个情况对我觉得这是一个必然的趋势我觉得就是包括训练推理常上下文agentic workload因为这些这种workload对于因为计算资源计算存储和通信的需求完全都不稀样比如说训练你需要这种大规模的集群非常close coupledGPU跟GPU之间要做大规模的同步要有高带宽的这种互联和有高可靠性这是训练推理的话又需要你要低成本低延迟和高吞吐量如果长上下稳的话你需要更大的kv cache所以Memory bandwidthMemory hierarchy要精准的设计然后对于这种MOE的模型你需要这种支持Louting All-to-All的communication对于这种通信的要求又比较高对Agentic workload它是一种小的batch长的这种horizon或者是长时间的这种频繁的工具调用动态的调度还有低延迟的响应所以所有这些workload其实它都有不同的对于计算存储和通信的要求这些需求不可能有一种芯片完美的覆盖其实是很难的做芯片设计的所以后面可能会针对不同的场景包括像Google的这种TPU V8它有圈林的版本也有推理的版本然后包括硅谷最近也有很多公司在做这种KVCash Offload的这种芯片针对Memory做扩展然后包括这种因为大家收购了Glock的LPU的这种架构的芯片就是针对Agentic的低延时的Workload的这种Accelerator所以后面会整个数据中心里面的这种特别是推理会变得越来越易购那这也是我们我现在这个公司在解决的一个问题就是硬件越来越分化模型越来越diverse然后我们怎么样去把这些模型匹配到更合适的芯片上面最后把所有的这种软件可以在一个系统上面做自动的调度和自动的优化这里其实是有很多可以做的事情对这个挺有意思的那以后我们会不会看到这个芯片混折用的情况比如说训练在英文的GPU上然后推理又在比如说这个国产芯片或TPU上然后这个agent的这个workload又在其他的芯片上会出现这样的情况吗那对于整个芯片的computing industry来说会意味着什么样的一个竞争格局的一个改变吗对我觉得我的判断是会而且这可能是未来的一个常态就是会有各种各样的芯片比如说训练是英伟达的芯片但是部署是在国产芯片上面或者是非英伟达的芯片上面这是有可能的这个以前在做这件事情的时候实际上是有一些数值进度或者是一些移植性的问题可移植性的问题那如果后面训练跟推理的兼容在数值上的兼容会更好的情况下面训练跟推理的这种infrastructure会越来越异构所以对于英伟达来说对英伟达倒不是一个短期的灾难但会改变它的一个市场结构就是说以前英伟达是所有的AI的workload是一个默论的选择那后面的话就是英伟达变成了这种异构AI任务里面对于high throughput比如说GPU架构变成high throughput的这种 [29:59] 这种prefirm这种workload的最优先级的这个硬件或者系统当然英伟达也不会只局限于他的GPU的accelerator这也是为什么他要把LPU也在跟他的GPU做融合同时要把CPU跟他的GPU跟LPU做融合所以英伟达本身已经意识到这个异构的趋势所以这也是为什么老黄一直说因为大家做的是xerate computing他并不是说做GPU的他是只要能对未来的这种workload做xerating的这种架构都属于英伟达的默认的这个架构的这个演进方向所以这个趋势对于我们做这种异构计算的这种调度和优化的公司其实是非常有利的因为异构系统最大的问题是这个劫测的复杂度爆炸因为有这样多的不同的workload然后有这样不同的模型不同的芯片不同的并行策略不同的这种kv cache memory管理不同的serving runtime所以这个组合空间非常非常的大如果只是靠能或者是能工或者是像以前一样就是说针对不同的芯片每次都需要花能力来做这种适配那肯定是不够的所以最终的话我们是需要一个AI的一个Decision的一个Layer它可以做自动的Profiling自动的仿真自动的搜索告诉客户给定一个模型给定一个硬件你怎么样去做最优的这种匹配降低token costopen cloud之后呢就是全球AI公司是出现了一个模型调用token量的一个飙升那对于token efficiency来说那未来还可以做哪些优化工作以及你觉得deep seek的这个路线对于市场来说有没有什么样的启发性呢对啊这个deep seek这个整个路线其实是对于市场意味着所有的公司都要想着以Token Efficiency为重要的这个优化的目标未来的Token Efficiency目前我觉得会有很多个优化方向就是其实Deep Seek这个技术报告里面已经给出了很多的这个方向其实这个方向还会持续往前走首先是模型架构的这种优化模型架构叫Model Hardware Code design然后第二个就是memory system的这种优化就降低怎么样去降低KV cache的成本包括Deep Seek这个技术论文里面也讲了各种各样的这种sparse attention的这种优化方法第三个是在run time调度的层次在模型的这个调度层次就是模型不同的batch size它的contest length不同的客户对于latency的要求的情况下面怎么样去做最优的这种run time的调度还有怎么样自动决定自动决定pt分离pt分离以后这种变形度和这种memory怎么样去做location这里面有很多可以做优化的地方第四就是基于通信就是以后的这个推理也会变成这种distribute inference其实现在已经是了所以针对这种interconnection的这种优化第五的话其实是agentic workflow的这种优化其实我们以前大量的时间是花在这种芯片系统或者是这种serving long time级的这种优化但是agentic这一层次包括我们公司前一阵子也做了openClaw的这种AutoTuner工具就我们发现在Agentic Workflow这个城市有太多的优化空间有很多的LowHunt Influte因为在这个城市对Token的浪费其实是非常非常大的它这个浪费不是一个模型的问题而是这个Workflow的问题比如说你重复调用工具重复读你的上下文重复读文件没有这种memory的这种hierarchy就是冷冷冷冷冷冷冷冷冷冷冷冷的或者温的数据的这种啊调度也没有这种啊reflection for reasoning的这种iteration的控制所以啊每个未来的agent的需要有一个这种token budget managermanage啊这种token的使用量而且这个里面优化的空间是非常大的最后一个其实是针对不同的硬件和系统要做深度的定制和优化所以Deepseek这个其实告诉大家我们就是说模型的竞争不仅仅是参数量的竞争更多的是一个token efficiency的竞争这个总结的非常好好的那以上就是我全部的问题了而在技术角度之外我们也想从整个硅谷的视角来看看Deepseek带来的影响和启示那第二位嘉宾和我们对谈的是 [34:59] The former former OpenAI researcher, as a former former entrepreneur, is a former partner of the and also DeepSeq R1 was definitely the defining moment for Silicon Valley. Like early last year, a lot of people in Silicon Valley said we had a DeepSeq moment because at that time, very few Silicon Valley people believe that China has the capability to train a frontier model. I would say the reaction to DeepSeq V4 is still happening in real time. But I think the scale of the reaction is definitely less intense compared to the last DeepSeq release. And I think even in China, the reaction is a lot less intense compared to the previous release. However, I think Silicon Valley views DeepSeek V4 as a really great engineering breakthrough, especially the efficiency gains. I think the Chinese models are significantly better than the American Foundation Model Labs at building something that not only good but also really cheap And I think people really respect that At the same time a lot of people have competitive concern towards deep seek and Chinese models more generally, especially the competitive pressure, you know, against the Foundation Model Labs. I know that like, you know, in the US, there's always this sense of, oh, yeah, it's impressive engineering, but it's also not transformational yet. And then a lot of people are very, you know, defensive about the DeepSeq, saying that numbers aren't real, or the Chinese models tend to exaggerate their performance on benchmarks. I think with the DeepSeq V4 release, I'm seeing a lot less of that and more of the congratulatory attitude towards the Chinese lab. So I think that's a positive sign. Great. And what is the part that should worry U.S. model companies the most? Is it benchmark performance, cost efficiency, long text capability, open weights, or hardware independence? Yeah, I think that's a really interesting question. I definitely think benchmark performance is not the most concerning thing for American companies, just given that benchmark performance are like you know people always joke about benchmark performances and everyone trying to inflate them It more like a marketing function than anything else And also in terms of efficiency I think it is really important But at the same time if OpenAI Anthropic really doubles down on it it's not something that they're not able to catch up on. I think the biggest risk of DeepSeek is that it creates a death zone or a kill line for foundation model companies in the U.S. We actually wrote about this earlier last year. the idea that open source models create a kill line for all foundation model companies. And the idea there is that the AI business model is very binary. If you're a foundation model company and you get surpassed by open source company, the value of your business is essentially zero. Just an exercise, right? If Anthropic is not the best performing model in coding, who's going to use cloud code, right? Where does Anthropic's revenue go down to? to and what is that you know what should anthropics valuation go down to so it's a survival question for foundation model companies in the u.s like deep seek is like holding a gun behind their back right if these companies don't run fast enough deep seek is going to catch up to them and and you know like just really really destroy their business right so do you see deep seek as a short competitive pressure or a sign of a more structural shift in the in the market So what is the biggest lesson that Silicon Valley should take from DeepSeek I definitely see it more as a structural shift and it'll leave Silicon Valley soul searching for a while. And I think what DeepSeek shows is that efficiency is a form of intelligence. Right. For the longest time, Silicon Valley companies have obsessed over smarter models. And, you know, these labs have been under investing in algorithmic efficiency for the longest time just because their business model actually encourages them to charge users more for models instead of like, you know, really driving algorithmic efficiency. I think it's really interesting because DeepSeq v4 and OpenAI's GPT 5.5 actually came out on the same day. And then v4 is so cheap compared to all the other models that have a similar level of intelligence. And GPT 5.5 is actually twice as expensive as GPT 5.4. So you can clearly see that, you know, DeepSeq is taking a very different approach compared to the big labs in the U.S. And I think the lesson here is that efficiency is a part of intelligence. I actually think because of like newer frameworks and harnesses such as OpenClaw, their capabilities of mid-tier models, such as DeepSeq, right? [39:59] such as like, you know, the other open source models in China that are not they're not the top three models yet. But these models capabilities are significantly better released as a result of model harnesses. And I definitely had that feeling that, oh, because of, you know, OpenClaw, FrameworkSide OpenClaw, I have a use case or I have like a better ways of making use of less capable models that are a lot cheaper. So I think, you know, I think like people are increasingly realizing that efficiency is a really big component in usability of AI. Right. As you mentioned that DeepSeq and other Chinese open source models like Kimi are now showing that open source models can catch up very, very quickly. Right. They're not the best, but good enough and much, much cheaper. So how does that change the competitive landscape for companies like OpenAI, Anthropic, as well as Google Gemini? What does that do to the pricing power of the frontier model companies? Yeah. One thing that we actually we have data on this internally at our fund. We look at the data of the frontier models released by OpenAI and Anthropik and Google Gemini And then we also look at the pricing power of second tier models that are released by these companies And one trend that we're seeing over and over and over again is that whenever you have an open source model, the pricing power of the second tier models gets slashed. There's still significant pricing power. And when it comes to state-of-the-art models, whenever an open source equivalent comes around, Like OpenAI simply cannot charge that much for, you know, its earlier, earlier models. The same goes for Anthropic. Now, it's really interesting because in Silicon Valley, for the longest time, people believe that a lot of companies are willing to pay for the extra intelligence. Right. So even if like GPT 5.5 or Opus 4.6, 4.7 are really, really expensive, there's still a lot of roles where you require the best intelligence and the more expensive models to do. and a lot of American companies are willing to pay that price. However, I think as open source models like DeepSea get better, I think in the next 12 months or so, there's going to be an inflection point where the marginal value of additional intelligence is not there anymore, right? If DeepSea v5 or v6 can achieve the performance of an Opus 4 then maybe there very little utility in having a smarter model especially if I doing just you know like normal enterprise tasks Right. Also, I think DeepSeq represents a path which model companies seek extreme token efficiency, cost efficiency, and computing optimization, right? So is this a different path from Silicon Valley model companies who are now raising hundreds of billions of dollars to change the HGI, which is the right path in your opinion? Yeah, I don't think there's a dichotomy here. I think in every industrial revolution, it's not just about how good the technology is, it's also about how cheap it is, right? The types of technologies that are able to drive change are the ones that are cheap enough for people to use them. Sometimes when I'm using Opus 4.6 for coding tasks or in my open claw, I always ask myself, is it really worthwhile paying $200 a day for something that I can, you know, just use a human and maybe do it even more cheaply, right? So I think in order to achieve AGI, especially in order to achieve AGI that's useful for like the masses, we definitely need a model that intelligent enough and also cheap enough And I think the question to whether you know like the DeepSeek approach or the open AI and the throttling approaches right really depends on whether scaling continues to yield intelligence gates. And there's some debate around that, especially if you look at the cost, right? And if scaling returns diminish proportional to the cost, I think DeepSeek's approach will seem really smart in hindsight in the future. Great, great. And let's talk about some of the top model companies here in the Silicon Valley. First of all, I want to ask about Anthropic. So Anthropic has recently received a lot of attention for its growth, enterprise traction, valuation momentum as well. What is driving Anthropic's rise in Silicon Valley right now? I think it's basically three things. I think cloud code definitely is a defining moment for Anthropik, even though Anthropik's previous models were already good enough in coding, but I do think it's first party product, cloud code is what's driving most of the revenue for Anthropik. I think the second thing for Anthropik is enterprise trust. I have a lot of enterprises, [44:59] within our network. And a lot of them say that like they just trust Anthropics models more because they have, you know, like a commitment towards safety. And also, you know, the whole, you know, Anthropics versus the Pentagon thing, a lot of enterprises, they bought into Anthropics, like messaging there. And I think the third part is that Anthropics vis-a-vis OpenAI, Anthropics just seemed like the adult in the room, whereas like OpenAI was going through so much chaos personnel-wise. That's why a lot of enterprises decided to go with Anthropics. And even like within a small fund like us, we're using Anthropic instead of ChatGPT. And the reason is that we think Anthropic's models genuinely perform a lot better compared to ChatGPT in work-related tasks. And initially when Anthropic was releasing their first models back in like 2021, 2022, and even like till, I say like 2023, Anthropic was really struggling to find product market fit. And the reason is that in the very early days of LLMs, you know, like the main product was the chatbot. And Anthropic doesn't really have the consumer DNA or the ability to ship really quickly, given their security culture. Because they didn have that they were always you know like lagging behind open AI And now in the world of enterprise AI adoption as well as AI coding I think Anthropic is finding their product market fit and overtaking OpenAI as like the number one company in AI in the Silicon Valley. Yeah, as you mentioned, I think Anthropic's valuation in the secondary market surpassed OpenAI, reaching $1 trillion, right? Why is that? Like what happened for the tipping point? Yeah, it's really interesting because last summer we actually did a 10,000-word deep dive on OpenAI. And one of our biggest takeaways from that deep dive is that OpenAI is trying to build an everything platform. It's trying to become everything for everyone. Now, I think it's that strategy that actually made OpenAI lose its technical lead vis-a-vis Anthropic. OpenAI had a hardware division. They were trying to build their own chips. They were trying to build a shopping app. Like, the core chat GPT experience was just not there. And that's why, like, you know, since late last year, OpenAI has been doubling down on the core experience and just like cutting a lot of like side projects, including Sora. Whereas Anthropic has always been super, super focused. It's always been focused on safety. It's always been focused on like, you know, enterprise related tasks. And I think the focus on coding is also really smart because I think coding really is the most important step towards AGI Like whoever owns probably will be the dominant player in AGI as well Yeah I totally get it I mean, coding is important. I get it. But does winning the coding market mean winning the AGI race? I mean, you still have the chatbot, you have the text to video, you have the multi-modality, right? Why is coding so important? I remember reading this article from Peter Steinberger, the founder of OpenClaw. And the article title was something like Claude Code is my computer. Right. The idea is that he's basically using Claude Code to control his computer before he started OpenClaw and before you had a consumer interface to access the model's capabilities. So I guess most tasks that you run on a computer is a coding task, whether it is video editing, whether it is website design. Everything you do on a computer fundamentally is a coding task. And that's why coding is such an important landscape for AI companies to compete in. And also, unlike other use cases, coding is very measurable. It's very easy to see if it's succeeding or failing. and also you have an insane amount of data And plus engineers are usually the first to adopt a new technology And that why with all of those combined coding is the perfect first task for AI models to tackle on its way to AGI Right. So I think Anthropic has not released any video models or image models as of now. Right. That's correct. Yeah. Do you think they will totally give up that multi-modality race or they will somehow one day come up with the multi-modality models or achieve that simply by coding? I think on the multi-modality side, I think Anthropic is less likely for Anthropic to go after video just because I think like ByteDance is, you know, for sure going to be the lead player in video, right? If you look at like the amount of video data, like ByteDance has way more data than anyone else. Even if you look at Google, like Google has way more video data than OpenAI or Anthropic. Like I remember this was early 2024 when Sora just came out and people were saying that like, oh my God, OpenAI just came out with Sora. It's so amazing. It was so mind blowing at the time. And my initial reaction towards it is that this feels like a side project at OpenAI, especially given that OpenAI doesn't really have the amount of information. [49:59] data that it needs to, you know, create a really big and useful AI video model. And, you know, two years later, OpenAI basically cut that project. And I think it's not, I don't really see a reason why anthropo-people want to do that on its own, like why anthropo-people would want to go after the multimodal models on its own, unless there's a particular strategic reason for it to do so. I know for a fact there are some explorations and like other appeals, like world models, as well as like robotics but i just think that like video generation is so far from anthropics core competency that like maybe they don't really want to distract themselves by doing that i have friends who are early employees at xai and one of their complaints that they you know told me repeatedly is that elon changes his mind every other day he sees a new model come out by from open ai or from you know google and they're like why don't we have the same right that's why they went after coding they went after image generation and video generation they have like a gazillion different things and i think that kind of distraction is part of the reason why xai is really struggling to keep up with the frontier model race today right so maybe the the key for anthropic success is that they are so so focused They just focus on coding while OpenAI wants to do everything And although I think they canceled Sora app, but they still maintain the video team, right? But I heard they actually combine the team with the robotics team. So probably they will use the video model for the world model to train the robotics. Yeah, 100%. I think like it makes total sense for them to do that, especially if the goal is AGI. I don't think like video as its own team, especially for like design or, you know, like creative work. I don't think that's the way that you do it, especially if you want to have like a main goal. And for OpenAI, the goal is to build AGI. And I think that's the same for Anthropic. Right. And OpenAI actually released a GPT 5.5 on the same day when DSTIC released before. I heard a very positive feedback from the developer community. What does it mean for the battle between OpenAI and Anthropic? So does that make OpenAI feel better? Yeah I think between the top three labs in the US the OpenAI Anthropic and Google the gap between these three companies has never been huge right People always I think in Silicon Valley people have FOMO and people always think that like there's a clear winner in AI. I remember like a year ago, it was all about open AI. It was all about like how amazing open AI's models are and how far behind everyone else is and how open AI is going to become the next like, you know, like the next Google, right? And then the narrative like six months ago or like a little over six months ago was that Gemini is a clear winner. And then, you know, the last like three months or so, or maybe a little longer than that, the narrative has been like, oh, Anthropic is the biggest winner. Right. So like the, you know, number one keeps flipping between these model companies. And therefore, like, you know, the conclusion might be that there's no clear single winner who's able to maintain a lead for more than a year. and I think the gap between these model companies are actually relatively small like if one company just you know loses a step right it's gonna like you know it's gonna fall behind if another company just like gets their shit together just really like pulls ahead they're able to do that and I think with GPT 5 even I thought it was a really powerful model in cyber security and then for 5 I think this is again OpenAI trying to be back on the offense in enterprise after being distracted for a really long period And I have heard a lot of really great things about GPT 5 However, a lot of people are complaining about the cost as well, which is why, you know, models like DeepSeek still play a role in the developer community. Yeah, I think 5.5 is super expensive, right? It's really expensive. especially the pro um i remember for the long text one was like 180 per million tokens if i remember correctly that's i think so yeah yeah right um so as you mentioned 5.5 has a very strong emphasis on real world capabilities such as coding research data analytics document creation and tool use what does that tell you about where the u.s sota models are heading to yeah no i think that's a I forgot to mention this in an earlier question about OpenAI versus Anthropics valuations. And I think one of the reasons why Anthropics is valued a lot higher compared to OpenAI today because OpenAI is actually trading below its last round valuations and then Anthropics is trading way above its last round valuations. And part of the reason is that Anthropics has a very clean enterprise focus. [54:59] Its revenue is very heavily concentrated in enterprise. And for U.S. investors, enterprise revenue is seen as stickier and seen as higher quality. And also, like, it expands the TAM for the company as well. U.S. investors historically have preferred enterprise revenue over consumer revenue, especially given that OpenAI's revenue predominantly comes from the ChatGPT subscription. Now, I think OpenAI is definitely being on the offensive in terms of going after enterprises. And I've heard a lot of companies receiving pitches from the OpenAI. I think OpenAI Frontier, that's their enterprise product. So I think ahead of their IPOs, OpenAI and Anthropik definitely want to position their revenue as predominantly enterprise and then only secondarily consumer. And especially for a company like OpenAI, that's what they want to do. And in order to do that, they need to make their models perform well on a real world task instead of just benchmarks. Historically, OpenAI has been the company that focused a lot on benchmark performance, whereas Anthropoc has been a lot more chill about benchmark performance, but they care a lot more about real world performance. And I think it positive for everyone that these foundation model companies are giving up benchmarks or giving up their sole focus or their obsession with benchmarks and just focusing a lot more on real performance and tasks And what happened to XAI and Meta? Can they still catch up? Yeah, I really think that foundation models, the foundation model business is a very tough business, right? As I mentioned, there was a huge kill line that's created by DeepSeek. And as open source models are getting better and better, it's going to be harder and harder for closed source models to compete. I think for Meta specifically, in the last month or so, they released a new model called MuseSpark. And a lot of people are saying, oh, this is a great sign that Meta is making a comeback. And they're actually doing something with the new super intelligence team. But after digging into the model, I don't think it is commercially differentiated vis-a-vis the other open source models out there. I don't even think it's like differentiated vis-a-vis like, yeah, like, I don't know, second tier models from the from OpenAI Anthropic. And that why I think it going to be pretty hard for them to really pull ahead However there is a very controversial recent initiative that Meta is pursuing which is the model capability initiative where they are rolling out internal tracking on their own employees. And it's very, very controversial because Meta employees are like, yeah, I don't want to become trained data for Meta's AI model. But I actually think this is ironically one of the only ways for meta to have a data mode when it comes to coding or agentic tasks and interestingly over the last like two or three months i've heard a few pitches from startups on um you know that want to do exactly the same thing which shows that foundation model labs want this kind of telemetric you know like a computer used data for model training so i think maybe this is the right approach for meta like maybe this is going to help them create an edge but i still think it's like very, very competitive. And I think they're already behind and I'm not sure what their internal strategy is. And then for XAI, I think it was very chaotic like over the last year or so, like given all the Elon's temperament changes, like him being very distracted, team being very distracted with like different product decisions Very interestingly I actually invested in an early XAI employee who got fired by Elon because you know he was fighting back on some of his boss requests But, you know, I think I think there's more to be desired in terms of their organizational focus as well. I think it's still the best tool in terms of like query real time data on Twitter. But beyond that, I really don't use Grok very much for any other use case. so do you think in the u.s other open source models would have a chance to compete with the chinese open models how about reflection ai you know i i have to be harsh here i actually don't think they have a chance including for you know meta meta's lava models like i just like you know in terms of the open source models that i use almost 100 of them are like chinese i sometimes will use lava models but like the vast majority of them are chinese models i think the chinese developers are very smart to open source all their models, including their model architecture. Not only does this encourage open source development around the world, but also because they are not the top tier models out there, they're not the frontier models. I think in that case, right, it makes sense for them to just open source everything, have people adopt it, and just have the community give them feedback on their model performance and just iterate really, really quickly. [59:59] So I think it's going to be very hard for a U.S. company to outcompete the Chinese companies, especially given that there's so much development, the cadence of new model iteration is just incredible. But do you think the Chinese open source models are going to chip away GPT or clause revenues? I think, you know, if you ask these companies, the answer is definitely no. Right. But if you just think rationally for yourself, the answer is definitely yes, because if you look at the mix of models that people are using, if you ask them what models you're using for your products, I would say like 80 percent of their tasks are run on a combination of smaller open source models and only 20 percent of the most intellectual or sophisticated tasks are run on closed source models. Right. So I think definitely open source models are chipping away at the foundation of models revenue. I wouldn't go as far as to say, like, you know, OpenAI or Anthropik is going to be dead because of DeepSeek. But I do think I do think like open source models like DeepSeek is one of the biggest structural threats to the closed source business model in the US. Cool And I think Anthropik and OpenAI are going public this year How is the sentiment in the capital market in your opinion You feel like the money is waiting for them or people think their valuation has been too high How do you think the IPOs will be? Yeah, I think just going back to my earlier comment about OpenAI and Anthropik's business, the revenue like a breakdown I think there is way more appetite for anthropic than this is there there is for open AI and I think that's the consensus right now when I talk to public investors right everyone wants to get access to anthropic and it's interesting because a lot of people that are buying anthropic secondaries today are actually hedge funds and public market investors because they think the stock will pop on the public market, given that the revenue is growing really aggressively. And I think the sentiment is very different for OpenAI. A lot of institutions are selling OpenAI. Even within my network, I get asked whether I'm buying OpenAI on a daily basis. Someone's like, oh, we have a 10 million chunk of OpenAI. Is there a fund buying or do you know anyone who buying OpenAI So it just shows that a lot of institutions are trying to sell OpenAI ahead of the IPO And I think OpenAI definitely definitely definitely need to fix its spending problem We actually have an internal chart on OpenAI versus Anthropics efficiency and spending And you can see that for the same amount of revenue that Anthropics is generating, it's significantly more capital efficient compared with OpenAI. Like OpenAI has just, you know, procured way too much of compute and, you know, just spending way too much, especially on infrastructure before its revenue has grown to that level. That's interesting. Like why has OpenAI spent much more than Anthropic on the computing? I think a big part of it is the philosophy, like Anthropic throughout the company's history. they've always exercised restraint. Since day one, they were very careful about not overly committing themselves to GPU and infrastructure spending. They don't really want to be in a situation where if the revenue doesn't catch up, they are beholden to a bunch of GPU purchases. They don't really want to be in that situation, whereas OpenAI has always been a move fast break things they just expand really quickly kind of company and they been a lot more aggressive from day one I actually this is a guess This is just my personal guess I actually think ahead of the IPO there might be like investor pressure on OpenAI to like change a CEO because like Sam is not someone who's very like, you know, stingy. Like he doesn't save money. Let's just put it that way. And I think a lot of investors are not happy about his spending habits when it comes to like infrastructure investments. One more DeepSeq question. So does DeepSeq actually prove that scaling along is not enough while, you know, for OpenAI, you have to think about the efficiency, you have to think about the compute, right? With the pressure from the investors, from the IPO. Yeah, I would even reframe it as like DeepSeq has, I think the main thing that DeepSeq has shown is that efficiency is really important. And I think that's like a lesson that Silicon Valley keeps forgetting. And I think the reason why DeepSeek is able to get to such efficiency is because it's in an environment where it doesn't have the infrastructure or the chips that Silicon Valley companies have. Right. Like sometimes I even joke with people that Silicon Valley companies have too much cash and because they have so much cash, they don't have an incentive. [64:59] They're not put in an environment where they have to think about efficiency. And that's why, you know, like DeepSeek has a structural advantage there in showing that, like, hey, we can do things much more efficiently. And I think that's the biggest advantage that DeepSeek has vis-a-vis the foundation model companies in the U.S. But at the same time, right, I wouldn't say that, like, because of DeepSeek, because of the architectural improvements, it shows that the scaling law is no longer applying to AI. I actually think from an investor's perspective, investors almost don't really care about whether the scaling wall is working or not. I think the real question is at what cost, right? What is the marginal benefit of like an additional billion dollar, 10 billion dollar investment? Like how much better does it make the model? Like how much how much additional TAM or how much additional like, you know, like revenue can you generate with that model performance? I think like a few years ago is a very different mindset like among us investors and the mindset a few years ago is like hey this thing's still growing we're still on an exponential curve so we going to put more capital into it but like I think in today day and age the question is very different like the question is like even if we do put more capital into it what the roi and um i think unfortunately right the foundation model companies are forced to ipo because they really don't have that many other pools of capital to tap into they have to tap into the capital markets to continue to fuel um their capital spend right especially like a company for open ai you know if you look at the last few rounds of fundraising it's almost out of desperation They were going after the Saudis. They're going after private equity funds. Like they're trying to get every little piece of capital, every bit of, you know, every drop of capital they can. So I really I really want to reframe it as an ROI question. And I think DeepSeek shows that maybe there's a different way of doing things. Maybe the ROI for scaling the models or, you know, for scaling infrastructure really isn't there anymore. Right, and interestingly, DeepSeek is raising money as well. So of course, DeepSeek is raising in the Chinese market. In terms of valuation, how do you think the open source model companies will be valued? Will they be valued differently from the OpenAI and also Anthropix those kind of closed models Yeah I think AI foundation model companies go through two phases in their valuation Like in phase one, you're buying the story, you're buying the team, you're buying the methodology or the tech. Right. And I think DeepSea is absolutely in that phase. And this is the phase that OpenAI and Anthropics were in back in 2022 and 2023, when people were like really buying the dream they think that this company has a unique approach it's going to do something special whereas like opening an anthropic are in phase two of foundation model valuations which is the phase where investors start asking hey what's your revenue right what's your retention what's your burn like you know they start to ask about financial metrics i think it's still going to be a few years before deep seek enters that phase which is why i think the valuation today almost doesn't really matter it's just like a number on on paper, right? Like it's more about like how much capital does it take for this company to go to the next stage? That matters a lot more than like, how much is this company actually worth? But I think it's really smart for DeepSeek to raise capital at this particular moment because I think it's a really good opportunity for them to build an ecosystem right I sure you and the previous guest talked about it as well like DeepSeek is very deeply integrated in the Huawei Ascend ecosystem just the same way that OpenAI is integrated with Microsoft and Anthropix is integrated with AWS. That kind of integration gives DeepSeek a once-in-a-lifetime opportunity to build an ecosystem around its models, around its infrastructure, and just really rebuild the Chinese AI ecosystem around domestically produced chips. And I think that's why it's so smart for DeepSeek to do a fundraise at this moment. Right, right. But they have to go through this too as well. Yes, yes. And they have to show their product. Yes, absolutely. Right. I think those are all my questions. Is there anything else you want to add? I think that's about it. Yeah, alright. Thank you so much, Jenny. From the conversation between two guests, we can clearly feel that AI-like development of the development of the AI-like technology has become a change. DeepSig's meaning is not just that it has a strong design, it is that it lets you see that the design of the blockchain is from a benchmark to a system of the blockchain. The design of the blockchain, the token efficiency, the machine is correct, and the machine is correct. [69:59] 商业化 开源生态正在变成同一场战争的不同战场在硅谷过去最强的路径是更多GPU更大模型更强的必源产品但是Deepseek和Kimi这样的开源模型代表的另一条路径是更高效率 更低成本 更开放生态以及对本土算力战的适配这也许不是两种路线模型与算力效率这个关键技术突破点可能是人类到达AGI的必经之路而随着AI的发展步入到Agentic时代未来的模型赢家可能不止取决于谁的模型最聪明而是取决于谁能够用最低的成本最稳定的把智能交付给最多的开发者和企业那好以上就是这期视频的全部内容了我是陈倩你们的留言点赞和转发是支持我们硅谷101做好深度科技和商业内容的最佳动力那我们就下期视频再见了Bye