硅谷AI创业公司正经历从高速扩张向成熟阶段的重大转变。根据Cognition公司资深AI工程师马培元的观察，这一转变体现在招聘、组织架构、成本控制等多个维度的系统性调整[1]。
在招聘标准上，AI原生能力开始成为更重要的选拔要素[1]。马培元在Cognition的面试经历体现了这一点——他遇到了一位19岁的面试官，反映出公司不拘一格的选才思路[1]。组织架构方面，人力经理职位正在消亡[1]。以Cognition为例，其60多人的工程团队中仅有一名真正的人力管理者，其他管理者都需要直接从事代码编写工作[1]。这种扁平化趋势已在更大的企业中得到印证，腾讯和字节都先后宣布取消中层[1]。
成本优化成为硅谷AI公司的当下核心关切。曾经不计代价消耗Token的热潮已然褪去，效率提升成为新的评判标准[1]。Cognition在6月4日推出的"AI生产力保证"计划明确承诺，效率提升若未达预期，最高将赔偿1000万美金[1]。更为明显的例子是，Uber 2026年全年的AI专项预算在前四个月基本烧完，迫使大公司开始重新严控Token用量[1]。在技术方向上，单一模型的崇拜也在瓦解，融合调度策略获得重视——Cognition内部推出的融合调度功能(fusion mode)整合廉价模型与顶尖模型，实现成本与性能的平衡[1]。Cognition本身的增长速度也印证了这些变化的市场认可度，该公司在8个月内估值从102亿美元升至260亿美元，融资超过10亿美元[1]。马培元用一句话总结了这个时代的特点："在硅谷AI公司内部，大于60%的判断可能一个月之后就反着来了"[1]。
Silicon Valley's artificial intelligence startups are transitioning from explosive growth to a more mature operational phase, marked by significant shifts in organizational structure, hiring practices, and technology priorities. Through interviews with senior AI engineers at Cognition, industry observers have identified emerging trends that signal fundamental changes in how these companies operate and invest [1].
Cognition exemplifies this transformation, having grown from a $10.2 billion valuation to $26 billion in just eight months while raising over $1 billion [1]. The company's approach to talent reflects the new reality: rather than prioritizing traditional credentials, Cognition's engineering team of over 60 people includes a 19-year-old interviewer, demonstrating that native AI skills now outweigh years of experience [1]. More strikingly, the company maintains only one dedicated people manager, with all other managers required to write code [1]. This flat organizational structure mirrors a broader industry trend, as tech giants Tencent and ByteDance have both announced the elimination of middle management layers [1].
Cost efficiency has become a central preoccupation as initial euphoria around token usage fades. Uber's annual AI budget for 2026 was largely exhausted within the first four months, prompting larger companies to reimpose restrictions on token consumption [1]. Rather than chasing raw model performance, companies are now adopting hybrid scheduling strategies that combine cheaper models with premium ones, exemplified by Cognition's internally developed "fusion mode" feature [1]. On June 4th, Cognition introduced a "productivity guarantee," offering compensation of up to $10 million if efficiency gains fall short of promised levels [1]. As one observer noted, "In Silicon Valley AI companies, judgments that seem correct today may reverse within a month" [1], underscoring the industry's rapid evolution and the diminishing credibility of early certainties about optimal approaches.
