何时打开:你想知道 AI 行业 2025-2026 大趋势 / 哪些新品类正在涌现 / 头部 VC 头部公司的预测。本 wiki 把深思圈 84+ 篇趋势文按"AGI 进度 / 新品类 / 旧 SaaS 替代 / 基础设施 / 头部预测"五类组织。
一句话核心:红杉 2026:AGI 已到来 + a16z:Computer Use 18 个月超越人类 + AI 颠覆摩尔定律 2028 临界点 + 新品类批量诞生:Agent Economy / 沟通即服务 / AI 原生设计 / Agent 操作系统 + 旧 SaaS 替代浪潮:Salesforce / ERP / Intercom 转型。
0. 5 大趋势板块
| 板块 | 关键叙事 | 代表文章 |
|---|---|---|
| AGI 进度 | 红杉 2026:已到来 / a16z 摩尔定律颠覆 | 红杉年度分享 / a16z 圆桌 |
| 新品类崛起 | Agent Economy / 沟通即服务 / Agent OS | YC 对谈 / a16z Top50 / 2000 万押注新品类 |
| 旧 SaaS 替代 | Salesforce / ERP / Intercom 转型 | Peter Thiel 3500 万 / 红杉 2500 万 |
| AI 基础设施 | Memory / Agent Git / 保险 / 邮箱 | 4 大新基建专题 |
| 头部预测 | Sam Altman / Karpathy / Claude Code 之父 | 多篇独立预测帖 |
1. AGI 进度
1.1 红杉美国 2026 年度分享
"2026 红杉美国年度分享:AGI 已经到来,重新定义 AI 革命"
含义:主流顶级 VC 已经 align"AGI 到来"的判断,不再是边缘论调。
1.2 a16z 18 个月超越人类(Computer Use)
"a16z 最新预测:Computer Use 让 AI Agent 像人类一样工作,18 个月内效率将超越人类"
1.3 a16z 颠覆摩尔定律(2028 临界点)
"AI 颠覆摩尔定律,未来超线性发展趋势预测,2028 达到临界点"
1.4 AGI Paradigm 新阶段
"AGI 要来了?AI Paradigm 的新阶段和新趋势"
1.5 Aschenbrenner Situational Awareness(王凯派交叉)
详见 王凯 收藏 / 2024-2026 AI 前沿论文与报告精读 + Scaling Law / AGI 时间表 / Software 3.0(跨作者主题综合):
- AGI by 2027
- 万亿美元集群
- The Project(美国国家级 AI 投入)
2. 新品类崛起
2.1 Agent Economy(YC)
"YC 最新对谈:Agent economy 真的要来了!"
2.2 沟通即服务(2000 万押注新品类)
"2000 万美元押注一个新品类:投资人为什么相信'沟通即服务'的未来,下一个 10 亿美元赛道?"
2.3 Agent 操作系统(浏览器)
"Accel 领投 2200 万美元,这家让 YC 创始人都心动的公司,打造浏览器 AI agent 的'操作系统'"
2.4 AI 视频帝国(a16z)
"a16z 合伙人最新分析:AI 视频正在吞噬世界,打造未来数十亿美元的 IP 帝国"
2.5 AI 漫画出海
"AI+漫画的出海未来如何?专访苗奘,从顶尖设计集团到创业融资过亿,打造触手 AI 再出发"
2.6 AI Memory 系统
"2025 AI 记忆系统大横评:从插件到操作系统,谁在定义下一代 Agent Infra?"
2.7 AI + 社交全新探索
"AI+社交全新探索,西部世界未来成真,这款产品种子轮即获 Coatue 投资"
3. 旧 SaaS 替代浪潮
3.1 Salesforce 终结(Peter Thiel 3500 万)
详见 深思圈 / AI B2B 垂直行业(8 大赛道独角兽案例) §1.1:
"Peter Thiel 花 3500 万美金投了一个销售领域的 Claude Code,号称要终结 Salesforce 时代"
3.2 ERP 替代(红杉美国 2500 万)
"红杉美国 2500 万美元的押注:这家公司想用 AI-native 的 ERP,干掉传统财务软件"
3.3 Intercom 自我重构
"深度分析|AI 客服崛起下,老牌巨头 Intercom 如何完成一场 AI-First 的自我重构"
3.4 Salesforce 砍掉界面(a16z 合伙人金句)
"Salesforce 主动砍掉了界面,a16z 合伙人说:护城河从来不在那里"
3.5 红杉:不卖工具,直接重建律师事务所
"红杉美国又投法律赛道,但这家公司不再卖工具给律师,而是直接用 AI 重建律师事务所"
3.6 2024 = 企业级 AI 元年
"2024 会是企业级 AI 发展的元年,16 个 2B 领域正在形成的趋势与'非共识'"
4. AI 基础设施 4 大新基建
| 基础设施 | 代表公司 / 数据 | 角色 |
|---|---|---|
| AI Memory 系统 | 大横评(2025) | Agent Infra 关键缺口 |
| Agent Git(GitHub 前创始人) | a16z 1700 万 | Agent 版本控制 |
| Agent 保险 | 1500 万 | 为 Agent 出错买单 |
| Agent 邮箱注册 | YC 600 万 | Agent 身份基础设施 |
4.1 AI infra 再现 3000 万
"AI infra 赛道再现 3000 万美元大额融资,当数据处理遇上 AI,如何重新定义多模态数据的未来"
4.2 强化学习训练栈(Cursor 公开)
详见 深思圈 / AI Coding 公司案例(Cursor/Lovable/Bolt/Devin/Replit) §1:
- microVM 训练环境
- MXFP8 microscaling
- 自定义 Blackwell 内核
5. 头部预测合集
5.1 Sam Altman
"Sam Altman 最新透露:OpenAI 未来的计划"
5.2 Claude Code 之父
"Claude Code 之父最新预测:编程被解决后,下一个是什么?"
5.3 Karpathy(Software 3.0)
详见 Karpathy 体系(跨作者主题综合) §6 Agentic Decade:
- 不是 Agent Year,是 Agent Decade(2025-2035)
- 半自主 > 全自主
- 半自主产品 4 原则
5.4 a16z Top100 应用榜单
"A16Z 合伙人解读 2026 年 TOP 100 应用榜单背后的趋势"
5.5 AI 时代创业指南
"AI 时代的创业指南:四个全新趋势"
6. AI Agent 下一爆发场景预测
"除了代码和客服,AI agent 下一个爆发的场景会在哪?"(深思圈深度分析)
典型预测:
- 销售(已部分起飞,但还有空间)
- 法律(独角兽诞生中)
- 医疗 Copilot
- HR / 招聘(竞争已激烈)
- 金融基建
- 沟通即服务(2000 万押注新品类)
- AI Memory + Agent OS(基础设施层)
7. AI 投资三层框架(深思圈贯穿)
详见 王凯 AI 行业判断与大公司竞争格局观察 §三层框架 + newtype AI 商业、行业与投资观察:
| 层 | 内容 | 深思圈代表 |
|---|---|---|
| 1. 技术栈 | 大模型 / Agent 协议 / Infra | Karpathy / Cursor 技术内幕 / AI Memory |
| 2. 价值链 | 哪个环节有壁垒 | a16z 平台公司剧本(Google/微软/AWS/Palantir) |
| 3. 竞争格局 | 头部公司动作 | OpenAI vs Anthropic 估值翻倍战 |
8. RPA → AI Agent 3.0 演化
"2000 万融资风向标:从 RPA 到 AI agent,企业自动化进入 3.0 时代"
时间线:
- RPA 1.0:屏幕录制 / 规则
- AI 2.0:NLP + ML 辅助
- AI Agent 3.0:自主推理 + 工具调用 + 多步
9. UX / UI 设计趋势
"2024 年的 UX_UI 设计趋势,关于未来科技体验的可能性"
(与 审美 / 创作平民化(跨作者主题综合) 交叉)
10. AI 时代关键转折点(2024-2026 时间线)
| 时间 | 关键事件 |
|---|---|
| 2024-01 | 企业级 AI 元年 / 16 个 2B 趋势 |
| 2024-11 | Computer Use / Operator 发布 |
| 2025-Q1 | Karpathy Vibe Coding 推文 / Software 3.0 |
| 2025-06 | Karpathy YC AI School 演讲 / Agentic Decade |
| 2025-08 | Cursor Composer + 开源 AI 编程工具 3200 万 |
| 2025-11 | Cursor 首度揭秘训练即产品 |
| 2026-01 | a16z 开年观点供给侧跃迁 / 红杉 AGI 已到来 |
| 2026-05 | Anthropic 护城河 = 权限论 |
11. 跨作者交叉点
| 主题 | 深思圈本 wiki | 姊妹 wiki |
|---|---|---|
| AI 行业判断综合 | 一级市场 + 案例视角 | AI 行业判断与公司竞争(跨作者主题综合) / 王凯 AI 行业判断与大公司竞争格局观察 / newtype AI 商业、行业与投资观察 |
| AGI / Scaling Law | 红杉 2026 + a16z 18 个月 | Scaling Law / AGI 时间表 / Software 3.0(跨作者主题综合) / 王凯 收藏 / 2024-2026 AI 前沿论文与报告精读 |
| 机会预判 | a16z 供给驱动 + YC Agent Economy | 机会预判 / 商业预判 / 终局视角(跨作者主题综合) / 一堂机会预判方法论 |
| 新品类涌现 | 沟通即服务 / Agent OS | AI 时代 PM / Design / Eng 三人组协作(跨作者主题综合) |
| 基础设施 | 4 大新基建 | MCP 协议生态(跨作者主题综合) / RAG / 知识库工程(跨作者主题综合) |
12. 怎么用 —— 按身份的趋势路径
| 你是谁 | 主用 | 起点 |
|---|---|---|
| 判断 AGI 时间表 | 红杉 + a16z + Karpathy | 2026 已到来 / 2027 / 2028 临界点 |
| 押新品类 | Agent Economy / 沟通即服务 / Agent OS | 4 大新品类 |
| 替代旧 SaaS | Salesforce / ERP / Intercom 案例 | "终结 X 时代"模式 |
| 建 AI 基础设施 | Memory / Agent Git / 保险 / 邮箱 | 4 大新基建 |
| 看头部预测 | Sam / Karpathy / Claude Code 之父 | 5 个独立帖 |
| 押下一爆发场景 | "除了代码和客服" | 销售 / 法律 / 医疗 / HR / 金融 |
通用纪律:
- AGI 已到来(红杉 2026 共识)
- 18 个月超越人类(a16z Computer Use)
- 新品类批量诞生 —— 沟通即服务 / Agent OS / Agent Economy
- 旧 SaaS 替代浪潮(Salesforce / ERP / Intercom)
- 2024 = 企业级 AI 元年(深思圈定调)
- 下一爆发不只是代码客服 —— 销售/法律/医疗/HR/金融
13. 引用清单
- 深思圈作者全集入口 —— 全集入口
- 主要 source:
sources/learning/深思圈/含 "趋势" / "预测" / "未来" / "AGI" / "新品类" 等 ~84+ 篇 - AI 行业判断与公司竞争(跨作者主题综合) —— 综合 AI 行业判断
- Scaling Law / AGI 时间表 / Software 3.0(跨作者主题综合) —— Scaling Law / AGI
- 机会预判 / 商业预判 / 终局视角(跨作者主题综合) —— 机会预判方法论
- Karpathy 体系(跨作者主题综合) —— Karpathy Agentic Decade
- 深思圈 / 头部 AI 公司动作(OpenAI/Anthropic/Google/Apple/Perplexity) —— 头部公司动作
- 深思圈 / AI Agent 公司案例库 —— Agent 公司案例
- 深思圈 / AI B2B 垂直行业(8 大赛道独角兽案例) —— 旧 SaaS 替代赛道
来源与关联资料
- https://mp.weixin.qq.com/s?__biz=MzAxNjE1NTg2Ng==&mid=2247492593&idx=1&sn=fc40e30ef94ded4ed58c17b551a63fb2&chksm=9bfbb644ac8c3f526e0459b12301462eb3084e4f35537c714a512278ad09e82a391797a178a7
- https://mp.weixin.qq.com/s?__biz=MzAxNjE1NTg2Ng==&mid=2247494188&idx=1&sn=394115e9006df186ed1c2944a1aa01b8&chksm=9bfbbf99ac8c368fb1821f0d52491195fcaaf55d2c6557c7c53c6621674a9840ece831b72246
- 深思圈作者全集入口
- 深思圈 / AI 一级市场 VC 视角(a16z + 红杉 + YC)
- 深思圈 / 头部 AI 公司动作(OpenAI/Anthropic/Google/Apple/Perplexity)
- 深思圈 / AI Agent 公司案例库
- AI 行业判断与公司竞争(跨作者主题综合)
- Scaling Law / AGI 时间表 / Software 3.0(跨作者主题综合)
- 机会预判 / 商业预判 / 终局视角(跨作者主题综合)
2026-10-09 增补：用共同经历延续游戏角色的陪伴关系
角色资料解决“它是谁”，共同经历才解释“我为什么想再见它”。 深思圈对 Yoroll / Yyo 的体验描述提供了一个产品假设：先在互动故事里让玩家认识角色，再把已经发生的故事与关系带入持续聊天。另一路径是在陪玩、小游戏或一起观看内容时积累新的共同经历。两者都比不断添加聊天话题更具体，但是否改善长期留存仍需要实验。
从剧情到陪伴，要传递可核对的状态
通关后角色如果忘记关键选择，甚至把另一条分支当成双方经历过的事，新增聊天入口反而破坏关系。实现时至少区分角色固定设定、玩家实际走过的剧情、双方后来发生的互动。关键事件应能追溯到来源，并能被用户纠正；重复开档、切换角色和删除记忆也需要明确规则。这些是依据文章提出的设计检查项，不是对 Yyo 内部实现的确认。
共同经历需要在合适的场景被再次提起。例如一起输过的一局游戏，可以成为下一次挑战的开场；但每次都机械复述旧事，或在不相关的对话里强行引用，会让“记得你”变成打扰。角色语气、价值取向和记忆连续性应与多模态表现一起验收，视频更流畅本身不能证明关系更可信。
先验证关系延续，再推算商业化
可以比较有剧情经历与直接进入聊天的用户，在相同观察窗口内的回访、主动互动、记忆纠错和付费表现；两组原有兴趣不同，应控制选择偏差。一次性游戏消费转订阅只是商业假设，还要扣除生成、实时语音视频和内容维护成本。高聊天量也可能来自模型反复理解失败，不能单独作为关系质量指标。
把体验报道与技术证明分开
文章提及 Hermes Agent 底座、视频模型速度、屏幕识别后的删除行为及多项市场数字。本轮未独立核验，不据此判定其隐私保障、技术领先或收入潜力。若设计屏幕陪伴，仍需明确用户授权、采集范围、保留期限与停止入口；角色“不谈论敏感页面”不能替代数据处理层面的验证。
这一路径与消费 AI 娱乐的交叉点是“游戏如何提供相识过程”。体验报道与融资统计的证据范围不同，不能互相替代。
来源：本次参考文章或访谈。
October 9, 2026 addition: Shared experiences can extend a game character's companionship
A character profile explains who the character is; shared experiences explain why someone wants to see them again. SenseAI's account of Yoroll and Yyo suggests a product hypothesis: let players get to know a character through an interactive story, then carry the events and relationship into ongoing conversation. Another route is to build new shared experiences through games, co-play, or watching content together. Both provide a more concrete starting point than continually adding conversation topics, but their effect on long-term retention still requires testing.
Carry verifiable state from the story into companionship
If a character forgets a crucial choice after the player finishes the game, or remembers events from a different branch as shared experiences, adding chat may damage the relationship. At minimum, distinguish the character's fixed background, the story the player actually experienced, and their subsequent interactions. Important memories should be traceable to their source and correctable by the user. Starting a new save, switching characters, and deleting memories also need explicit rules. These are design checks derived from the article, not confirmed details of Yyo's implementation.
A shared experience needs to resurface at an appropriate moment. A game lost together might open the next challenge. Mechanical repetition of old events, or inserting them into unrelated conversations, can instead turn recognition into an interruption. Character voice, values, and memory continuity should be tested alongside multimodal presentation. Smoother video alone does not establish a more credible relationship.
Test relationship continuity before projecting revenue
Compare users with prior story experience against users who enter chat directly, using the same observation window for return visits, user-initiated interactions, memory corrections, and payments. Account for selection bias: the groups may already have different levels of interest. Moving from a one-time game purchase to a subscription is a business hypothesis. Generation, live audio and video, and content maintenance costs must also be considered. High message volume may reflect repeated misunderstandings by the model, so it cannot independently measure relationship quality.
Separate an experience report from technical proof
The article mentions a Hermes Agent foundation, video-model speed, deletion after screen recognition, and several market figures. These have not been independently verified in this review and do not establish privacy guarantees, technical superiority, or revenue potential. Screen-based companionship still needs clear user authorization, capture scope, retention periods, and a way to stop. A character avoiding discussion of a sensitive screen is not a substitute for verifying how the underlying data is handled.
The connection to consumer AI entertainment is how games introduce people to characters. An experience report should retain its own scope rather than being treated as equivalent to financing statistics.
Source: Article or interview used for this addition.
2026-10-17 增补：把算力需求拆到买方与任务，再判断投入
来源为深思圈对黄仁勋观点的中文转述，尚未独立核对原节目。英伟达是供给方，其增长判断有商业立场；本节只提炼调查方法，不采用文中投资规模与增长预测作为已验证结论。
同一笔基础设施叙事里，可能装着不同需求
黄仁勋把需求讨论扩展到存量计算迁移、生成式AI和Agent任务。用于产品判断时，可以把这三个方向拆开调查，避免把整个行业的投入都归到少数模型公司的收入上，也避免因为有多个需求方向就断定投入一定能收回。
| 需求方向 | 先找谁与什么任务 | 需要补的证据 |
|---|---|---|
| 存量计算迁移 | 原有计算任务的预算负责人；哪些工作考虑改用新的计算方式 | 迁移成本、兼容性，以及在同一任务下的总成本和结果 |
| 生成式AI | 为生成、推理能力付费的组织或用户 | 真实使用量、付款与续用，而非仅有注册或演示 |
| Agent任务 | 希望持续委托完整流程的业务负责人 | 成功任务量、重试、人工接手和可承担的价格 |
这张表是本库的调查工具，不是已证实的市场分布。记录“谁付款、从哪个预算出钱、解决哪项任务”后，再检查预算是否新增或从旧支出迁移。同一客户可能横跨三类，不能直接相加成市场规模。
单次便宜与总账变小，要分别验证
单次推理成本下降后，用户可能委托更多任务，也可能增加每个任务的计算量。总支出是否下降，需要同时看任务数量、每项任务消耗、重试与人工处理。对于自己的项目，可先固定一组任务比较成本，再单独观察使用量变化；不要把供应商的长期扩张预测直接当作自己的收入计划。
何时用。 适合讨论算力采购、Agent定价或行业机会时澄清需求来源。若没有付款方、真实任务和持续使用证据，先把它记为待验证假设。土地、电力和建设等交付约束也应单列：需求存在不代表供给能够按预期时间落地。相关投资判断见创业与投资，本节不形成买卖结论。
2026-10-17 Addition: Identify buyers and tasks behind compute demand before committing resources
The source is Shensiquan’s Chinese account of Jensen Huang’s views; the original program has not been independently checked. NVIDIA is a supplier with a commercial interest in growth. This section extracts a research method and does not treat the source’s investment totals or growth forecasts as verified conclusions.
One infrastructure narrative can contain several kinds of demand
Huang extends the demand discussion to migration of existing computing workloads, generative AI and agent tasks. For product decisions, investigate these separately. Avoid attributing all industry investment to the revenue of a few model companies; equally, several potential demand categories do not prove that investment will pay back.
| Demand category | Buyer and task to identify first | Evidence still needed |
|---|---|---|
| Migration of existing computing workloads | The current budget owner and the workloads being considered for a different computing approach | Migration cost, compatibility, total cost and outcomes for the same task |
| Generative AI | Organizations or users paying for generation or inference capabilities | Actual usage, payments and continued use, not registrations or demonstrations alone |
| Agent tasks | Business owners seeking to delegate complete workflows on an ongoing basis | Successfully completed tasks, retries, human takeovers and an affordable price |
This table is a research tool proposed by this knowledge base, not an established market breakdown. Record who pays, which budget pays and which task is served, then establish whether the budget is additional or shifted from existing spending. A single customer may span all three categories, so do not simply add them together to calculate market size.
Lower unit cost and lower total spending require separate verification
When inference becomes cheaper, users may delegate more tasks or use more computation per task. Whether total spending falls depends on task volume, consumption per task, retries and human handling. For your own project, first compare costs using a fixed set of tasks, then separately observe changes in usage. Do not convert a supplier’s long-term expansion forecast directly into your revenue plan.
When to use it. Use this approach to clarify the source of demand when discussing compute procurement, agent pricing or industry opportunities. Without an identifiable payer, real tasks and continued-use evidence, keep the claim as a hypothesis. Track delivery constraints such as land, power and construction separately: demand does not mean supply can arrive on schedule. See Startups and investing for related investment discussions; this section makes no buy or sell recommendation.
2026-10-22 增补：长期算力规划先写不可让渡的目标
当算力交付要提前数年规划，今天的资源分配与未来的位置并不处在同一个时间尺度上。深思圈对 Sundar Pichai 的转述提供了一个战略案例：先规定保持 AI 前沿所需的研究投入，再在这个底线之上权衡云客户的合同需求和短期收入。它描述的是 Google 的取舍，不是其他公司都应该优先研究的通用答案。
用于自己的项目时，可先写清必须保住的目标、所需最低资源和重新审视条件，再比较新增订单、研究任务和产品使用的机会成本。需求信号、合同承诺、内部实际使用和模型发展预期应分别记录，不能把其中一种当成全部需求。交付周期、组件瓶颈与外部冲击也应单列；长期方向成立，仍可能买错规模或时间。
这与本页“需求拆到付款方与任务”的方法互补：前者回答需求从哪里来，这里回答资源紧缺时按什么约束分配。所有预算门槛和重审时间应由团队依据自身目标确定，原文未提供可直接照抄的最优比例。模型榜单、token 总量、上市时间和行业规模预测不作为本节的验证结论，本节也不形成投资买卖建议。
来源：深思圈对 Sundar Pichai 访谈的中文转述。
October 22, 2026: Protect Strategic Resource Commitments in Long-Term Compute Planning
When computing capacity must be planned years ahead, current allocation and future competitive position operate on different time horizons. Shensi’s account of Sundar Pichai offers a strategic example: establish the research capacity needed to remain at the AI frontier, then weigh cloud contracts and short-term revenue within that constraint. It describes Google’s tradeoff, not a universal rule that every company should prioritize research.
For an individual project, first specify the goal that must be protected, its minimum resource needs and the conditions for reconsideration. Then compare the opportunity costs of new orders, research and product usage. Demand signals, contractual commitments, actual internal use and expectations about model development should be recorded separately. None represents all demand. Lead times, component bottlenecks and external shocks also deserve separate treatment: a sound long-term direction can still produce the wrong purchase size or timing.
This complements the existing method of tracing demand to a payer and task. That method asks where demand comes from; this one asks which constraints govern allocation when resources are scarce. Teams must set their own budget thresholds and review dates. The account supplies no optimal allocation ratio to copy. Model rankings, token totals, listing timelines and market-size forecasts are not treated as validated conclusions, and this section provides no investment recommendation.
Source: Shensi’s Chinese account of an interview with Sundar Pichai.
2026-10-23 增补：资本供给与实际使用需求分开看
判断 AI 基建能否持续交付，除了芯片、电力、材料，还需要检查资金条件。Semafor 对星展集团 CEO Tan Su Shan 的访谈，把资本也放进供给约束：贷款方会要求与风险相匹配的回报，长期投入未必能按预期时间转成现金。她谈到大型云厂商从现金创造者转向大额现金使用者，以及集中流入某类增长资产后的回报压力。这是受访者的风险判断，不证明资本短缺已在所有市场发生；“接下来两年”也是她的预测，不当作行业时钟。
本库应用：为一个基建项目分别列出真实使用者和付款方、建设与运行需要何时付款、收入和现金何时可能到账，以及资金期限和持续供给的条件。需求强、客户愿意试用、融资可获得，分别是不同证据，不能相互替代。项目收入增长也不等于已经有足够现金支撑后续投入。
本页已有“需求拆到付款方和任务”及“长期研究资源底线”。这里新增的是资金时间与交付时间的对应检查：资源预算获批之后，仍需确认钱能按期到位、使用回报的证据是否更新。它用于解释方案依赖和风险，不给出投资买卖、融资工具选择或统一回报门槛。论坛转发帖只作为发现路径，判断归属保留给 Tan Su Shan 的原始访谈，不归为王凯已核验结论。
来源：https://www.semafor.com/article/10/16/2026/the-ai-bet-is-running-straight-into-a-capital-crunch
English edition
Shensi: AI Industry Trends, Emerging Categories and Forecasts
Historical source note: This edition translates the complete published trend synthesis. Headlines, funding figures, dates and predictions are retained as historical source claims, not independently reconfirmed announcements or proven forecasts. Expressions such as “AGI has arrived” and “replace Salesforce” describe the cited narrative, not an established conclusion of this translation.
When to read: when looking for the broad AI trends of 2025–2026, emerging categories, or predictions from major investors and companies. The published synthesis organizes 84+ Shensi trend articles into five groups: AGI progress, new categories, replacement of legacy SaaS, infrastructure, and leading figures' forecasts.
Its central narrative: Sequoia's 2026 claim that AGI had arrived; a16z's forecast that computer-use agents would surpass human efficiency in 18 months and that AI would reshape Moore's law around a 2028 inflection point; categories such as the agent economy, communication as a service, AI-native design and agent operating systems; and the transformation of Salesforce, ERP and Intercom.
0. Five Main Trend Areas
| Area | Key narrative | Representative material |
|---|---|---|
| AGI progress | Sequoia's 2026 “already here” framing; a16z on Moore's law | Sequoia annual presentation; a16z roundtable |
| Emerging categories | Agent economy, communication as a service, agent OS | YC discussion; a16z Top 50; a $20 million category bet |
| Legacy SaaS replacement | Transformation of Salesforce, ERP and Intercom | Peter Thiel's $35 million bet; Sequoia's $25 million bet |
| AI infrastructure | Memory, agent Git, insurance and email | Four new infrastructure categories |
| Leading forecasts | Sam Altman, Karpathy and Claude Code's creator | Separate forecast articles |
1. AGI Progress
1.1 Sequoia US: The 2026 Annual Presentation
The cited headline says: “Sequoia US's 2026 annual presentation: AGI has already arrived; redefining the AI revolution.” The synthesis interprets this as a sign that an “AGI has arrived” judgment had entered mainstream leading-VC discussion rather than remaining marginal. That interpretation does not settle the definition or attainment of AGI.
1.2 a16z: Surpassing Humans in 18 Months Through Computer Use
The preserved headline predicts that computer use lets AI agents work like humans and surpass human efficiency within 18 months. This remains a forecast attributed to that article.
1.3 a16z: Reshaping Moore's Law, With a 2028 Inflection Point
The source headline predicts superlinear AI development and a critical point in 2028, framed as AI overturning Moore's law.
1.4 A New Stage in the AI Paradigm
A related headline asks whether AGI is approaching and describes new stages and trends in the AI paradigm.
1.5 Aschenbrenner's Situational Awareness: Overlap With Wang Kai's Reading
The related notes on frontier papers and scaling laws cover AGI by 2027, trillion-dollar clusters, and “The Project,” meaning US national-scale AI investment. These are the scenario and forecast topics in those sources.
2. Emerging Categories
2.1 The Agent Economy: YC
A cited YC discussion is framed by the headline “The agent economy is really coming.”
2.2 Communication as a Service: A $20 Million Category Bet
The article asks why investors believe in communication as a service, describing a $20 million bet and asking whether it could become the next billion-dollar market.
2.3 An Operating System for Browser Agents
A cited company reportedly raised $22 million led by Accel, attracted YC founders' interest, and set out to build an operating system for browser AI agents.
2.4 AI Video Empires: a16z
An a16z partner's analysis is presented as AI video consuming the world and enabling future IP empires worth billions of dollars.
2.5 AI Comics for Global Markets
An interview with Miao Zang traces a move from a leading design group to entrepreneurship, financing of more than RMB100 million, and a new venture with Chushou AI, asking what international expansion in AI comics could look like.
2.6 AI Memory Systems
A 2025 comparison of AI memory systems examines the shift from plug-ins toward operating systems and asks who will define the next generation of agent infrastructure.
2.7 New AI-Social Products
A source describes an AI-social experiment, evokes a Westworld-like future, and reports Coatue participation in a seed round.
3. The Wave of Legacy SaaS Replacement
3.1 A Salesforce Challenger: Peter Thiel's $35 Million Bet
The related vertical-B2B note covers a company described as a Claude Code for sales, reportedly backed with $35 million by Peter Thiel and claiming to end the Salesforce era. The replacement claim belongs to the source's positioning.
3.2 ERP Replacement: Sequoia US's $25 Million Bet
A cited $25 million investment is framed around AI-native ERP seeking to displace traditional financial software.
3.3 Intercom Rebuilds Itself
A detailed article examines how the established company Intercom undertook an AI-first transformation as AI customer support rose.
3.4 Salesforce Removes Its Interface
The headline reports Salesforce removing an interface and quotes an a16z partner's view that the moat had never been in that interface.
3.5 Rebuilding a Law Firm Instead of Selling Lawyers Tools
A Sequoia-backed legal company is described as using AI to rebuild a law firm itself, rather than selling another tool to lawyers.
3.6 2024 as the Opening Year for Enterprise AI
A 2024 source presents 16 emerging B2B trends and non-consensus views, calling that year the beginning of enterprise AI's development.
4. Four New AI Infrastructure Categories
| Infrastructure | Representative company or reported figure | Role |
|---|---|---|
| AI memory systems | A broad 2025 comparison | A key gap in agent infrastructure |
| Agent Git | A former GitHub founder; $17 million from a16z | Agent version control |
| Agent insurance | $15 million | Paying for agent mistakes |
| Agent email registration | YC; $6 million | Agent identity infrastructure |
4.1 Another $30 Million AI Infrastructure Round
A source reports $30 million in financing and explores how combining data processing with AI might redefine multimodal data infrastructure.
4.2 The Reinforcement-Learning Training Stack Revealed by Cursor
The related coding-company note covers microVM training environments, MXFP8 microscaling, and custom Blackwell kernels.
5. Forecasts From Leading Figures
5.1 Sam Altman
A cited article covers Sam Altman's statements about OpenAI's future plans.
5.2 Claude Code's Creator
A separate forecast asks: after coding is solved, what comes next?
5.3 Karpathy: Software 3.0
The related Karpathy note's Agentic Decade section emphasizes an agent decade rather than an agent year, 2025–2035, partial rather than full autonomy, and four principles for partially autonomous products.
5.4 a16z's Top 100 Apps
An a16z partner interprets trends behind the 2026 Top 100 application list.
5.5 Entrepreneurship in the AI Era
A related startup guide identifies four new trends.
6. Predictions for the Next Agent Breakout
A Shensi analysis asks which agent use cases might expand beyond coding and customer support. Its candidates are:
- Sales: partly established, with room to grow.
- Law: companies reaching unicorn status.
- Medical copilots.
- HR and recruiting: already highly competitive.
- Financial infrastructure.
- Communication as a service, linked to the reported $20 million category bet.
- AI memory and agent operating systems, at the infrastructure layer.
These are candidate opportunities from the source, not validated market outcomes.
7. A Three-Layer Framework for AI Investing
The synthesis connects to Wang Kai's industry-judgment framework and Newtype's business-trend notes.
| Layer | What to examine | Shensi examples |
|---|---|---|
| 1. Technology stack | Models, agent protocols and infrastructure | Karpathy; Cursor's technical details; AI memory |
| 2. Value chain | Which part contains a defensible barrier | The a16z platform-company playbook: Google, Microsoft, AWS and Palantir |
| 3. Competitive landscape | Leading companies' actions | OpenAI versus Anthropic and the race in rising valuations |
8. From RPA to AI Agent Automation 3.0
A headline uses $20 million of financing as a signal of enterprise automation entering a third era:
- RPA 1.0: screen recording and rules.
- AI 2.0: NLP and machine-learning assistance.
- AI Agent 3.0: autonomous reasoning, tool use and multiple steps.
9. UX and UI Design Trends
A 2024 UX/UI design-trends article explores possibilities for future technological experiences. It overlaps with the related note on democratizing aesthetic creation.
10. Turning Points in the Published 2024–2026 Timeline
| Date in the historical synthesis | Event or framing |
|---|---|
| January 2024 | Enterprise AI's opening year; 16 B2B trends |
| November 2024 | Computer Use / Operator release framing |
| Q1 2025 | Karpathy's vibe-coding post; Software 3.0 |
| June 2025 | Karpathy's YC AI School talk; the Agentic Decade |
| August 2025 | Cursor Composer; a reported $32 million open-source AI coding-tool round |
| November 2025 | Cursor's first account of training as product |
| January 2026 | a16z on a supply-side leap; Sequoia's “AGI has arrived” claim |
| May 2026 | A permissions-based account of Anthropic's moat |
Dates and grouping above follow the published synthesis; they are not a newly checked product-release chronology.
11. Cross-Author Connections
| Topic | This Shensi synthesis | Related notes |
|---|---|---|
| Overall AI industry judgment | Primary-market and case perspectives | General AI industry judgment; Wang Kai industry judgment; Newtype AI business trends |
| AGI and scaling laws | Sequoia 2026; a16z's 18-month forecast | Scaling laws and AGI; Wang Kai's frontier-paper notes |
| Anticipating opportunities | a16z's supply-led view; YC's agent economy | Opportunity prediction; Yitang's opportunity framework |
| Emerging categories | Communication as a service; agent OS | AI product, design and engineering collaboration |
| Infrastructure | The four new infrastructure categories | MCP ecosystem; RAG and knowledge engineering |
12. How to Use This Note
| Reader's question | Main material | Starting point |
|---|---|---|
| Assess an AGI timeline | Sequoia, a16z and Karpathy | “Already here” in 2026; 2027; the 2028 inflection point |
| Explore a new category | Agent economy, communication as a service, agent OS | The four emerging-category framing |
| Challenge legacy SaaS | Salesforce, ERP and Intercom cases | The “end the era of X” pattern |
| Build AI infrastructure | Memory, agent Git, insurance, email | Four infrastructure categories |
| Read leading forecasts | Altman, Karpathy, Claude Code's creator | Five separate forecast articles |
| Explore the next breakout | The “beyond code and support” analysis | Sales, law, medicine, HR and finance |
The original note summarizes six recurring propositions: Sequoia's 2026 AGI claim; a16z's 18-month computer-use forecast; multiple emerging categories; replacement pressure on legacy SaaS; 2024 as the opening year for enterprise AI; and opportunities beyond coding and support. Treat these as a checklist of claims to investigate, retaining their attribution and dates.
13. References and Related Reading
- Shensi author overview — the collection entry point.
- The underlying Shensi collection contains approximately 84+ articles tagged around trends, predictions, the future, AGI and new categories. This English page does not expose private source-file paths.
- AI industry judgment — the combined framework.
- Scaling laws and AGI.
- Opportunity prediction — methods for anticipating opportunities.
- Karpathy's system — the Agentic Decade.
- Shensi: Leading Companies — major-company developments.
- Shensi: Agent Companies — the agent-company case library.
- Shensi: AI B2B Verticals — categories challenging legacy SaaS.
October 9, 2026 addition: Shared experiences can extend a game character's companionship
A character profile explains who the character is; shared experiences explain why someone wants to see them again. SenseAI's account of Yoroll and Yyo suggests a product hypothesis: let players get to know a character through an interactive story, then carry the events and relationship into ongoing conversation. Another route is to build new shared experiences through games, co-play, or watching content together. Both provide a more concrete starting point than continually adding conversation topics, but their effect on long-term retention still requires testing.
Carry verifiable state from the story into companionship
If a character forgets a crucial choice after the player finishes the game, or remembers events from a different branch as shared experiences, adding chat may damage the relationship. At minimum, distinguish the character's fixed background, the story the player actually experienced, and their subsequent interactions. Important memories should be traceable to their source and correctable by the user. Starting a new save, switching characters, and deleting memories also need explicit rules. These are design checks derived from the article, not confirmed details of Yyo's implementation.
A shared experience needs to resurface at an appropriate moment. A game lost together might open the next challenge. Mechanical repetition of old events, or inserting them into unrelated conversations, can instead turn recognition into an interruption. Character voice, values, and memory continuity should be tested alongside multimodal presentation. Smoother video alone does not establish a more credible relationship.
Test relationship continuity before projecting revenue
Compare users with prior story experience against users who enter chat directly, using the same observation window for return visits, user-initiated interactions, memory corrections, and payments. Account for selection bias: the groups may already have different levels of interest. Moving from a one-time game purchase to a subscription is a business hypothesis. Generation, live audio and video, and content maintenance costs must also be considered. High message volume may reflect repeated misunderstandings by the model, so it cannot independently measure relationship quality.
Separate an experience report from technical proof
The article mentions a Hermes Agent foundation, video-model speed, deletion after screen recognition, and several market figures. These have not been independently verified in this review and do not establish privacy guarantees, technical superiority, or revenue potential. Screen-based companionship still needs clear user authorization, capture scope, retention periods, and a way to stop. A character avoiding discussion of a sensitive screen is not a substitute for verifying how the underlying data is handled.
The connection to consumer AI entertainment is how games introduce people to characters. An experience report should retain its own scope rather than being treated as equivalent to financing statistics.
Source: Article or interview used for this addition.
2026-10-17 Addition: Identify buyers and tasks behind compute demand before committing resources
The source is Shensiquan’s Chinese account of Jensen Huang’s views; the original program has not been independently checked. NVIDIA is a supplier with a commercial interest in growth. This section extracts a research method and does not treat the source’s investment totals or growth forecasts as verified conclusions.
One infrastructure narrative can contain several kinds of demand
Huang extends the demand discussion to migration of existing computing workloads, generative AI and agent tasks. For product decisions, investigate these separately. Avoid attributing all industry investment to the revenue of a few model companies; equally, several potential demand categories do not prove that investment will pay back.
| Demand category | Buyer and task to identify first | Evidence still needed |
|---|---|---|
| Migration of existing computing workloads | The current budget owner and the workloads being considered for a different computing approach | Migration cost, compatibility, total cost and outcomes for the same task |
| Generative AI | Organizations or users paying for generation or inference capabilities | Actual usage, payments and continued use, not registrations or demonstrations alone |
| Agent tasks | Business owners seeking to delegate complete workflows on an ongoing basis | Successfully completed tasks, retries, human takeovers and an affordable price |
This table is a research tool proposed by this knowledge base, not an established market breakdown. Record who pays, which budget pays and which task is served, then establish whether the budget is additional or shifted from existing spending. A single customer may span all three categories, so do not simply add them together to calculate market size.
Lower unit cost and lower total spending require separate verification
When inference becomes cheaper, users may delegate more tasks or use more computation per task. Whether total spending falls depends on task volume, consumption per task, retries and human handling. For your own project, first compare costs using a fixed set of tasks, then separately observe changes in usage. Do not convert a supplier’s long-term expansion forecast directly into your revenue plan.
When to use it. Use this approach to clarify the source of demand when discussing compute procurement, agent pricing or industry opportunities. Without an identifiable payer, real tasks and continued-use evidence, keep the claim as a hypothesis. Track delivery constraints such as land, power and construction separately: demand does not mean supply can arrive on schedule. See Startups and investing for related investment discussions; this section makes no buy or sell recommendation.
October 22, 2026: Protect Strategic Resource Commitments in Long-Term Compute Planning
When computing capacity must be planned years ahead, current allocation and future competitive position operate on different time horizons. Shensi’s account of Sundar Pichai offers a strategic example: establish the research capacity needed to remain at the AI frontier, then weigh cloud contracts and short-term revenue within that constraint. It describes Google’s tradeoff, not a universal rule that every company should prioritize research.
For an individual project, first specify the goal that must be protected, its minimum resource needs and the conditions for reconsideration. Then compare the opportunity costs of new orders, research and product usage. Demand signals, contractual commitments, actual internal use and expectations about model development should be recorded separately. None represents all demand. Lead times, component bottlenecks and external shocks also deserve separate treatment: a sound long-term direction can still produce the wrong purchase size or timing.
This complements the existing method of tracing demand to a payer and task. That method asks where demand comes from; this one asks which constraints govern allocation when resources are scarce. Teams must set their own budget thresholds and review dates. The account supplies no optimal allocation ratio to copy. Model rankings, token totals, listing timelines and market-size forecasts are not treated as validated conclusions, and this section provides no investment recommendation.
Source: Shensi’s Chinese account of an interview with Sundar Pichai.
Source Links and Related Public Notes
- Reference 1
- Reference 2
- Reference 3
- Reference 4
- Reference 5
- Reference 6
- Reference 7
- Reference 8
- Reference 9
- Reference 10
- Reference 11
- Reference 12
- Reference 13
- Reference 14
- Reference 15
- Reference 16
- Reference 17
- Reference 18
- Reference 19
- Reference 20
- Reference 21
- Reference 22
- Reference 23
- Reference 24
October 23, 2026 Addition: Separate Funding Availability From Demand and Delivery
To assess whether an AI infrastructure project can keep delivering, check funding conditions alongside chips, power and materials. In Semafor’s interview, DBS Group CEO Tan Su Shan treats capital as another supply constraint: lenders seek risk-adjusted returns, while long-term spending may not turn into cash on the expected schedule. She discusses hyperscalers moving from cash generators to heavy cash users and return pressure after money concentrates in a category of growth assets. These are her risk assessments, not proof of a capital shortage across every market. Her reference to the next two years is a forecast rather than an industry clock.
Application in this knowledge base: separately record the real user and payer, when construction and operating costs become payable, when revenue and cash may arrive, and the term and continuity of funding. Strong demand, willingness to try and access to financing are distinct forms of evidence. Revenue growth does not itself establish sufficient cash for subsequent investment.
This article already separates demand by payer and task and discusses resource floors for long-term research. The additional check connects funding time to delivery time: an approved resource budget still requires money to arrive when needed and fresh evidence about returns from use. This explains dependencies and risks without giving trading instructions, selecting financing instruments or imposing a universal return threshold. The forum post is a discovery path; the judgments remain attributed to Tan Su Shan’s original interview, rather than presented as independently verified conclusions by Wang Kai.
Source: https://www.semafor.com/article/10/16/2026/the-ai-bet-is-running-straight-into-a-capital-crunch