下一轮科技权力，取决于谁能改变竞争的条件

从智能眼镜、低成本模型到 AI 生命科学：企业家的认知边界、资本纪律与国家创新能力

资料核验截至 2026 年 9 月 10 日；历史数据按其对应时期标注。

科技竞争最容易被误读的时刻，是新产品不断出现、旧企业仍然强大，而下一套商业规则尚未形成的时候。此时，人们习惯于比较模型排名、设备参数、投资规模和创始人的胆量，却较少追问：这些变化究竟降低了哪一种成本，又把稀缺性转移到了哪里？如果这一点没有想清楚，规模可能被误认为壁垒，支出可能被误认为远见，技术演示也可能被误认为已经到来的产业革命。

理解下一轮科技权力，需要更换分析单位。与其猜测哪一家企业将消灭另一家，不如研究谁能够改变客户完成任务、企业组织生产、科学家验证知识的条件。本文的核心判断是：长期优势不只来自拥有更强的技术，还来自把技术进步转化为可采用、可验证、可负担的结果，并在这一过程中形成可持续的价值分配机制。

这一视角既保留了企业家精神的重要性，也对它提出了更高要求。真正的战略认知，不是把未来描述得更壮观，而是看见既有业务之外的可能性，同时知道哪些障碍尚未消除、哪些证据足以支持下一笔投入，以及什么情况下应当改变方向。

这里所说的“科技权力”，不是市值或知名度的别称，而是影响任务入口、合作方接入条件、技术标准和收益分配的能力。技术进步只有转化为这些位置上的持续影响，才从一种产品优势上升为结构性优势；而这种影响，也可能受到替代技术、合作伙伴选择和制度约束的重新塑造。

战略的三个层次：提高效率、改变规则、建设能力

企业的战略选择，可以分为三个相互联系、但不能简单排列高低的层次。

在现有规则中提高效率： 降低成本、改善推荐、提高产品质量和服务效率。这类改进往往直接产生现金流，也是长期探索的资金来源。

改变竞争的条件： 通过新的交互入口、技术架构、分发方式或商业机制，让过去的优势不再具有同样的决定性。

建设他人能够继续创新的基础能力： 提供工具、数据、标准、实验平台和可信的制度安排，使更多参与者能够在其上创造新价值。

第三个层次并不天然优于前两个。基础设施可能建设过早，平台可能没有足够需求，公共研究也未必能够转化为某一家企业的利润。问题不在于业务名称是否宏大，而在于它解决的约束是否真实，以及组织是否具备跨越相邻环节的能力。

由此可以提出一组更有解释力的战略问题：当生成内容越来越容易时，可信的行动是否成为新的稀缺品？当模型调用更便宜时，任务验证是否比生成速度更重要？当科学预测更快时，实验与临床证据是否反而成为决定性约束？技术进步不会让所有瓶颈同时消失，它往往只是改变瓶颈的位置。

本文据此采用以下分析框架。它不是一套经过统计估计的评分模型，而是用来分辨“技术进步”与“战略优势”之间尚缺少哪些环节。

关键转换	容易发生的误判	真正需要建立的能力
从信息到行动	占据屏幕，就等于拥有任务入口	取得上下文、获得授权、可靠执行
从模型调用到合格工作	单次生成便宜，就等于生产成本低	工作流程整合、验证与失败处理
从预测到可用知识	预测数量巨大，就等于科学问题已经解决	新实验、外部验证、因果检验
从投入到产业能力	长期支出，就等于长期竞争力	阶段性学习、互补能力建设和资源再配置

现金流购买的是选择权，不是成功保证

Meta 提供了一个适合观察这种转换的样本。其 2025 年收入为 2,009.66 亿美元，经营活动现金流为 1,158 亿美元；同年 Reality Labs 收入为 22.07 亿美元，经营亏损为 191.93 亿美元。这些数据同时说明两件事：成熟业务能够为长期探索提供显著资源，而新业务的投入规模与已实现收入之间，仍存在很大的距离。

不能仅凭亏损就判断探索没有意义，也不能因为投入巨大就认定未来回报必然巨大。更合适的理解是，企业正在为未来的产品、技术能力和平台位置购买选择权。这种选择权的价值，取决于项目能否持续减少关键不确定性，而不是取决于企业能否持续承受支出。

尤其需要避免把所有投资混为一谈。Reality Labs 的分部经营亏损，不等于智能眼镜单一产品的亏损，更不等于集团全部 AI 基础设施支出；集团收入和现金流，也不能证明某一个前沿方向已经建立了独立的商业模型。上述数字应被用于描述不同业务之间的资源关系，而不是拼接成一个无法核验的“未来回报率”。

因此，评价长期布局，应同时看两张表：财务表记录投入、现金流和承受能力；学习表记录哪些技术、采用和商业假设已经得到验证。只有前者，容易把长期主义变成预算惯性；只有后者而没有资金约束，又容易把技术探索变成无法持续的实验。

智能眼镜的战略意义：争夺任务入口，而非宣布手机终结

判断智能眼镜，首先要区分不同技术阶段。Meta Ray-Ban Display 于 2025 年 9 月 30 日在美国开售，发布时售价为 799 美元，包含 Meta Neural Band；其腕带通过表面肌电信号识别手部动作，而非读取人的思想。（Meta 产品公告）Orion 则是另一类更完整的增强现实探索，Meta 官方仍将其描述为不向公众销售的原型，而不是已经确定规模化交付的消费产品。（Orion 官方介绍）

这些区别影响的不只是产品命名，而是战略判断。带摄像头和音频交互的眼镜、带有限显示能力的眼镜、能够呈现更完整空间界面的 AR 系统，需要解决的工程问题、适用任务和采用门槛并不相同。不能把其中一类产品取得的进展，直接当成另一类产品已经成熟的证据。

市场进展也应按正确口径衡量。EssilorLuxottica 披露，2025 财年售出的 AI 眼镜超过 700 万副；这一数字覆盖其相关 AI 眼镜业务，不是 Display 单一产品的销量，更不是每日活跃用户数。（EssilorLuxottica 年度业绩公告）它说明这类产品已经形成值得重视的销售规模，但持续佩戴、使用频率与任务留存，仍是判断入口价值需要另外观察的变量。

眼镜值得重视的地方，在于它可能降低某些任务的“意图到行动距离”。用户不必先掏出手机、打开应用、输入上下文，而可以在观察现实世界时提出问题，再由系统帮助完成下一步。这不是单纯把屏幕放到眼前，而是尝试把感知、理解、授权和行动连接起来。

但交互步骤减少，并不意味着总成本必然下降。佩戴舒适度、续航、视觉负担、误触、隐私顾虑和公共场合的接受程度，都可能抵消操作上的便利。Meta 对 Display 眼镜给出的混合使用续航约为 6 小时，这本身就提示，全天候使用仍需要具体的补能和使用安排，而不是依靠“无处不在”的想象。（Meta 产品说明）

更有意义的商业假设，不是眼镜将在某一年消灭手机，而是某些高频任务会逐步从手机迁移出来。导航、翻译、现场指导或信息提示，可以分别接受检验；它们不需要等待一种设备替代全部计算活动。新终端甚至可能增强既有服务的使用频率，而不是使这些服务失去意义。

从战略上看，这是一场围绕任务入口和平台议价权的竞争。如果某个系统掌握了用户的上下文、授权和下一步行动，它就可能获得新的分发位置。但这个位置只有在用户持续使用、服务方愿意接入、错误能够得到控制时，才具有经济价值。硬件原型购买的是尝试建立入口的机会，并不自动授予入口控制权。

DeepSeek 的启示：效率改变门槛，可靠性决定收益

算法效率确实可以改变竞争条件，但应当用具体技术事实说明，而不是把它写成“数学战胜硬件”的单线故事。DeepSeek 于 2026 年 9 月 10 日正式发布 V4.1-Flash，官方说明其具有原生多模态视觉理解能力。（DeepSeek 更新记录）其模型卡描述了 Causal Encoder-Decoder 架构、压缩稀疏注意力以及面向输入密集型代理任务的缓存优化；其中，FP4 全局 KV 缓存被报告为每 token 890 字节，约为 V4-Flash 的四分之一，这属于官方给出的特定架构指标。（DeepSeek 模型卡）

这类进展的意义，是改变达到某种能力水平所需的资源组合。更有效地处理上下文、缓存和计算，可以让原本成本过高的工作负载进入可用范围。但特定缓存指标的改善，不等于整套系统不再需要高性能内存，也不等于所有用户、所有任务都获得同一个比例的成本下降。

DeepSeek 的正式公告也将这一改进表述为：与上一代相比，KV 缓存所需 HBM 为四分之一，SSD 存储为八分之一，而不是取消了这些硬件需求。（DeepSeek 发布公告）这恰好说明，值得关注的是算法与内存层级的协同设计，而非把“智力”和“算力”设定为互斥选项。

截至 2026 年 9 月 10 日，官方价格页列出的 deepseek-flash 输出价格为每百万 token 非高峰时段 0.60 美元、高峰时段 1.20 美元，输入及缓存命中的收费另行计算。（DeepSeek 定价页）这个价格是采用决策的一部分，却不是最终的生产率指标。企业真正购买的，并不是一段廉价文本，而是能够被接受的工作结果。

因此，评价 AI 经济性的单位应当从“每 token 成本”推进到“每项合格任务的总成本”。这个口径至少包含模型调用、检索和工具使用、人工复核、重试，以及错误造成的损失。一个单次调用更便宜的模型，如果需要更多轮修正，未必更经济；一个价格较高的模型，如果显著减少人工介入，也未必更昂贵。

这不是否定低成本创新，恰恰是识别其真正价值的必要条件。当效率改进能够穿透完整工作流程，它会扩大 AI 的可用场景，降低进入门槛，并迫使竞争者重新说明自己的溢价来自哪里。相反，如果优势只停留在报价表，生产组织方式就未必发生根本变化。

效率提高也不必然降低全社会对算力的总需求。每次任务更便宜，可能促使更多任务被自动化，也可能让原有任务使用更长上下文和更多验证步骤；实际结果取决于新增需求与单位消耗如何共同变化。因此，算法创新与基础设施投资可以相互强化，而不只是彼此替代。

更深的战略变化在于：当基础模型更容易取得，差异化可能向工作流程、行业数据、权限管理、可靠性和客户关系迁移。模型能力仍然重要，但仅仅拥有模型，不等于拥有完成任务的整个系统。真正受到挑战的，是无法说明自身增值环节的商业模式，而不是所有既有企业。

开放不是立场宣言，而是边界与收益的设计

搜索、模型和科学数据看似分属不同领域，却共享一个问题：知识如何被发现、被使用，并与外部贡献形成反馈？一个平台既需要保护隐私、安全和服务质量，也需要让合法的信息流动与合作发生。战略上值得讨论的，不是抽象的“开放者”和“封闭者”，而是哪些边界提高了系统价值，哪些边界只是延缓了学习。

需要把至少四种开放分开：能否读取信息，能否取得模型或数据，能否修改和部署，以及能否把成果迁移到其他系统。它们对应不同的技术接口和权利安排。提供下载，不等于提供全部训练信息；允许研究，不等于允许任意商业用途；提供接口，也不等于保证长期可迁移。

具体许可证比“开源”标签更能说明边界。例如，2023 年的 Llama 2 社区许可证要求，在版本发布时达到其条款规定的超过 7 亿月活用户门槛的使用者，另行向 Meta 申请许可，是否授权由 Meta 自行决定。（Llama 2 许可证）而 Meta 于 2026 年 9 月 2 日介绍 Muse Spark 1.3 时，列出了 API 等使用渠道，却把开放权重发布列为后续路线图事项。（Meta 研究公告）两者不能被合并为一个永远不变的开放身份；不同模型、不同阶段的权利安排，需要分别判断。

开放的商业意义，同样不只是免费。企业可以降低某个基础层的价格，让更多参与者进入，从而扩大互补产品的需求；也可以开放工具，却在托管服务、应用关系或其他环节获得回报。这样的安排既可能扩大竞争，也可能形成新的依赖，必须具体分析价值在哪里产生、由谁承担成本、又由谁取得收益。

因此，围墙是否合理，不应只从平台能否保住流量判断。还应追问：创作者是否获得更多发现机会，开发者是否能够形成稳定业务，用户是否拥有有效选择，以及外部知识能否改善系统表现。如果封闭减少了滥用和质量风险，它可能创造价值；如果封闭主要阻止替代和互操作，它也可能消耗长期适应能力。

这一点在 AI 时代尤其重要。模型本身的开放程度，与用户能否真正控制数据、工作流程和迁移路径，是相关却不相同的问题。更成熟的战略，是设计可以长期被合作伙伴接受的依赖关系，而不是把短期锁定能力误认为永久竞争力。

对搜索与内容平台，还可以提出一个更直接的检验：当某种边界减少外部知识进入、优质内容被发现或创作者触达外部受众的机会时，平台应当证明，新增的安全与质量收益足以补偿损失的发现与学习能力。这既不是要求取消所有边界，也不是否定内部生态，而是把边界本身纳入长期价值核算。

AI 生命科学：突破的是发现方式，不是取消验证

生命科学将上述问题推进到了更高的证据门槛。Biohub 于 2026 年 4 月 29 日宣布为期五年、总额 5 亿美元的 Virtual Biology Initiative，其中 4 亿美元用于大规模数据生成及下一代测量、成像和工程技术，1 亿美元用于支持外部研究和全球协同数据生成；NVIDIA 被列为技术合作伙伴。（Biohub 计划公告）这一资金配置值得关注，因为它没有把进步仅仅理解为购买更大模型，而是同时投入产生新知识所需的测量条件。

讨论这一方向，必须保留组织边界。上述计划的公告主体是 Biohub，CZI 与 NVIDIA 此前公布的合作围绕虚拟细胞模型及相关科研工具展开；这些科研项目不能被改写为 Meta 已经建立的生命科学营收业务。（Biohub 公告、CZI 合作说明）共同的发起人背景，不足以证明不同机构的资产、目标、知识产权和商业收益属于同一条链条。

这种区分反而有助于理解企业家认知的更大作用。突破既有业务边界，不一定意味着让同一家上市公司无限扩张；它也可以意味着用不同组织承载不同目标，让基础研究、公共工具与商业产品拥有适合各自使命的评价标准。科研基础设施的成功，不必以资助者垄断下游市场为前提。

AI 在这里最值得期待的贡献，是改善提出假设、选择实验和解释数据的过程。虚拟细胞的战略价值，不是让计算替代所有生物现实，而是帮助研究者决定下一次值得测量什么、干预什么，以及哪些预测最需要被推翻。如果系统能够把预测、实验、反馈和修正连接起来，研究活动才可能形成更有效的学习循环。

与此同时，描述细胞状态不等于预测干预结果，预测干预结果也不等于证明人体治疗有效。一个模型能够拟合既有观察，仍然需要面对新的细胞类型、疾病阶段、实验条件和人群差异。数据规模提供了研究机会，却不会自动提供因果解释。

已经建立的数据基础相当可观。CZ CELLxGENE Discover 的同行评议平台论文报告，截至 2024 年 10 月 1 日，平台包含 9,360 万个独特细胞记录、超过 1,550 个数据集；这是对应日期的数据规模，而不是当前所有虚拟细胞模型的训练规模。（CELLxGENE 平台论文）另一方面，一项 2026 年 4 月的虚拟细胞基准预印本发现，在若干未见过的扰动条件下，模型难以可靠恢复细粒度、扰动特异性的效应；该研究尚属预印本，其结论应限定于所评估的数据、模型与任务。（虚拟细胞基准研究）这两类证据并不矛盾：数据基础正在扩大，跨越分布差异的有效预测仍需要检验。

这并不是某一个机构独有的机会。Google DeepMind 在 2026 年 9 月公布的 AlphaGenome Atlas，为约 90 亿种单核苷酸变异提供预测，同时明确说明其尚未经过临床用途验证。（AlphaGenome Atlas 官方说明）这个例子同时说明，成熟科技机构可以参与基础能力创新，而极大规模的计算成果仍然需要遵守临床证据边界。

Human Cell Atlas 也不是任何一位企业家的个人数据库，而是一个跨机构、跨国家的科学共同体；其全球社区页面列出了分布于许多国家和研究机构的参与者。（Human Cell Atlas 全球社区）因此，更准确的竞争图景，是多个科研组织、技术平台与专业机构共同建设能力，而不是某一家公司的算法接管全部生命科学。

这里最重要的稀缺性迁移，是从“能否生成更多预测”转向“能否更快、更可信地排除错误”。如果预测能力增长快于实验能力，测量质量、因果验证和数据标准会变得更加关键。拥有模型的一方未必自然拥有这些能力；连接模型、实验与专业责任的组织，才可能取得更深的优势。

巨大需求不等于可独占的收入

健康改善具有重大价值，但不能据此推导出无限支付能力或没有商业化障碍。需要治疗的人、决定采用的人、支付费用的人、批准产品的人，可能并不是同一个主体。技术即使有效，也仍需证明它对特定人群、特定用途和特定支付条件是否值得采用。

因此，从科学突破到商业收益，至少要回答三个不同问题：是否创造了真实效果，是否能够以可承受方式交付，以及提供者能够保留多少收益。这三者不能用一个庞大的市场规模数字代替。公共数据、基础工具、诊断、药物和医疗服务，也不应被视为具有相同商业模式的产品。

以 Avantect 胰腺癌检测为例，其官方说明明确区分了实验室检测与监管批准：检测由符合相应实验室资质的机构开展，但尚未获 FDA 批准或许可，并且不适用于平均风险人群的普遍筛查，也不构成确诊。（Avantect 官方问答）这里的启示不是评价某项检测优劣，而是说明“已经提供服务”“具备检测性能”“获准某种用途”与“证明改善健康结局”，不是同一种证据。

历史临床开发数据进一步显示了这条链条的难度。BIO、QLS Advisors 与 Informa 的报告分析了 2011 至 2020 年间 9,704 个开发项目的 12,728 次阶段转换，估计从一期临床走到 FDA 批准的总体概率为 7.9%；这一结果由各阶段推进概率相乘得出，包含科学、商业与项目终止等因素。（临床开发成功率报告）它不是对当代 AI 药物项目的成功率预测，也不能证明 AI 无法改善研发；它说明的是，即使前端发现更快，也不能把后续验证当作已经完成。

这一逻辑也适用于消费科技与跨境平台。获得大量用户，不等于形成稳定利润；降低商品价格，不等于全链条成本已经下降；取得技术入口，也不等于能够自由收取任何费用。企业可以创造很大的社会价值，却只获得其中一部分商业收益，也可能在价值分配失衡时失去合作伙伴和市场准入。

因此，最高层次的战略判断，不是寻找一个“需求无限”的市场，而是理解价值如何穿过技术、组织、规则和支付体系。市场的边界并非完全固定，但扩大边界需要补齐真实约束，而不是把愿望换算成收入。

全球化不是同一场战争：分发、供应链与制度准入

把 TikTok、Temu 与 DeepSeek 放在一起分析，有价值的原因是它们可以代表不同环节的竞争，而不是它们构成了一套相同的颠覆模式。对内容平台，可以追问发现与分发效率如何变化；对电商平台，可以追问供给组织、价格与交付如何变化；对模型供应商，则要追问能力、使用成本与部署条件如何变化。

TikTok 在 2026 年 8 月 31 日介绍其第七份《数字服务法》透明度报告时，称其在欧盟每月触达 1.89 亿人，该报告覆盖 2026 年上半年。（TikTok 欧洲透明度公告）这一自报触达口径体现了平台覆盖面，但不能直接换算为注意力份额、用户忠诚度或盈利能力。对这类平台，战略价值还取决于能否持续连接创作者、受众与商业需求。

Temu 则说明，全球规模需要与全球经营责任同时建设。欧盟委员会于 2026 年 5 月 28 日宣布，对其未充分识别、分析和评估平台非法商品系统性风险的行为处以 2 亿欧元罚款。（欧盟委员会决定公告）这一具体决定不等于否定一种商业模式，却表明商品风险识别和平台治理，是规模化经营的一部分，不能被放在增长之后无限延期。

由此也能看清低价竞争的战略检验标准：更低的价格是否来自更有效的供给组织，还是依赖补贴或把成本转移到质量、交付和风险处理环节？这个问题应当用可持续的全链条经济性回答，而不能仅凭一次促销价格、短期增长或对某位创始人的评价回答。

这对智能眼镜同样重要。即使计算入口迁移，内容、购物、支付和履约需求也未必消失；既有平台可能迁移、合作，甚至成为新终端不可缺少的服务方。因此，新入口的建设者与现有平台之间，既可能发生替代，也可能形成互补。科技权力的重新分配，往往表现为合作条件与议价位置变化，而不是所有旧业务同时归零。

耐心资本的核心，是让学习快于不可逆承诺

长期投入与财务纪律并不对立。真正需要协调的是不同决定的时间尺度：某些能力需要提前建设，某些高额承诺则应当等待更多信息。这要求企业区分“必须现在获得的条件”与“可以以后再决定的资产”。

Meta 在 2026 年第二季度电话会议中给出的全年资本开支指引为 1,300 亿至 1,450 亿美元，口径包含融资租赁本金支付，属于指引而非已经实现的支出；相关披露见会议 29 分 18 秒。同一会议讨论更远期容量规划时，财务负责人说明了保留土地和电力条件、同时延期决定芯片等大额采购的思路，见 1 小时 2 分 35 秒。这一表述体现的是在长期建设中保留调整空间，而不是对所有未来需求一次性作出确定判断。

这种纪律可以推广为一个管理原则：对探索方向保持足够耐心，对具体实现路径保持足够弹性。战略承诺不应等于技术路线不可更改，项目终止也不应自动被视为领导者失去远见。能够在新证据出现时重新配置资源，本身就是长期能力的一部分。

创始人的控制力可能扩大组织承受短期压力的空间，但同样可能延长错误判断的寿命。职业经理人可能过度重视近期表现，也可能通过更严格的验证机制保护长期投入。这些都是治理机制需要处理的风险，不能通过个人出身、性格标签或英雄叙事得到解释。

对投资项目，更有效的考核不是笼统追问“今年赚了多少钱”，也不是豁免一切短期检验，而是要求不同阶段提供不同证据：原型阶段消除了什么技术障碍，试点阶段证明了哪些重复需求，扩张阶段出现了怎样的单位经济，平台阶段又是否形成了独立于补贴的生态参与。不同阶段可以有不同容忍度，但不能没有证据标准。

从企业竞争到国家能力：决定差距的是转化体系

企业家的认知能够影响一家组织愿意进入哪些领域，但国家层面的科技能力，不能被简化为是否出现了某位特别大胆的创始人。一个更完整的分析单位，是把想法变成可验证技术、把技术变成可靠产品、再把产品扩散到经济活动中的转化体系。

这一体系至少需要连接四类能力：提出和检验新知识的科研能力，将结果变成可制造、可维护系统的工程能力，承受不同时间尺度风险的资本与组织能力，以及提供授权、责任、采购和市场准入条件的制度能力。任何一个环节长期脱节，都会使前端的突破难以产生后端的广泛收益。

这也意味着，科技战略不能只统计模型数量、论文数量或算力规模。更值得追踪的问题是：普通企业是否有能力使用这些工具，独立团队是否能够取得合规数据和验证资源，公共研究成果是否容易复用，以及真正有效的产品能否跨越采购和部署障碍。这些指标衡量的是社会吸收技术的能力，而不是单一机构展示技术的能力。

自主能力与开放合作也不必被理解为非此即彼。更有韧性的安排，是在关键环节保留选择权，同时降低合法合作、知识复用和跨机构验证的成本。完全依赖外部系统可能缺少议价能力，但把所有外部知识都视为风险，同样可能降低本土系统的学习速度。

从这个意义上说，基础设施的战略价值有两个不同面向：对企业而言，它可能形成商业控制点；对社会而言，它也可能降低更多人的创新门槛。优秀的制度设计，应当分辨这两种价值何时相互促进、何时发生冲突，而不是把任何一种平台主导地位自动等同于国家利益。

不预测唯一终局，而是观察决定性变量

下一轮竞争不必只有一种结果。智能眼镜可能成为重要入口，也可能长期与手机分工；开放模型可能压低基础能力价格，也可能让拥有数据、分发和工程系统的企业获得更多收益；AI 科学工具可能广泛扩散，而临床与商业价值则由多个专业机构共同实现。

判断这些情景，应当持续观察以下变量，而不是等待某一次发布会宣布胜负：

入口迁移是否真实发生： 新设备带来的是短期试用，还是持续使用和独有的高频任务？

成本优势是否穿透工作流程： 便宜的是单次生成，还是包含复核和失败处理的合格结果？

科研进展是否跨越验证边界： 改善的是已有数据上的预测，还是新的实验、外部样本和实际临床结果？

生态是否能够独立成长： 合作方是依赖补贴暂时参与，还是拥有可持续收益与合理的迁移权？

资本配置是否能够纠错： 新证据是在改变预算与路线，还是仅仅改变项目的宣传方式？

这些问题最终指向同一件事：战略不是对未来的修辞占有，而是对未来选择空间的持续建设。能够创造选择、验证选择，并在必要时放弃选择的组织，比只能不断强化原有叙事的组织，更有机会穿越技术周期。

真正值得重视的企业家认知，是同时理解技术能够做什么、组织能够承担什么、市场愿意接受什么，以及社会需要什么样的责任安排。下一轮科技权力未必属于最早宣布终局的人，而更可能属于那些能够改变竞争条件、又能把改变转化为广泛而可靠价值的人。

AI interfaces, efficient models, and computational biology reveal a larger strategic question: who can turn technical progress into reliable, widely adopted value?

Evidence checked through September 10, 2026. Historical figures retain their original reporting periods.

Technology competition is hardest to interpret when new products are arriving quickly, established businesses remain powerful, and the rules of the next market have not yet settled. Model rankings, investment announcements, hardware specifications, and founder personalities offer convenient narratives. They do not necessarily explain where durable advantage will emerge. For that, a different question is needed: which constraint is being relaxed, and where does scarcity move next?

The relevant unit of analysis is not simply the company or the product. It is the set of conditions under which people complete tasks, businesses organize production, and researchers establish knowledge. Lasting advantage requires more than technical capability. It requires translating that capability into outcomes that can be adopted, verified, afforded, and sustained within a viable system of value creation and distribution.

This perspective does not diminish the importance of entrepreneurial ambition. It makes the test of ambition more demanding. Strategic foresight means recognizing possibilities beyond an existing business while remaining precise about unresolved constraints, the evidence needed for further investment, and the circumstances that should trigger a change of direction.

“Technological power” here means the ability to influence task entry points, the terms on which partners participate, technical standards, and the distribution of returns. It is not a synonym for market capitalization or visibility. A technical advance becomes a structural advantage when it produces durable influence over these arrangements, and that influence remains subject to alternatives, partner choices, and institutional constraints.

Three levels of strategy

Strategic choices can be understood at three connected levels. They are not a simple hierarchy in which the most expansive undertaking is always the best.

Improve performance within existing rules: Lower costs, improve recommendations, raise quality, and serve customers more effectively. These activities often provide the cash that makes longer-term exploration possible.

Change the conditions of competition: Introduce an interface, architecture, distribution mechanism, or business arrangement that changes the significance of established advantages.

Build capabilities on which others can innovate: Supply tools, data, standards, experimental platforms, and dependable institutional arrangements that enable further value creation.

The third level is not inherently superior. Infrastructure can be built too early, a platform can lack sufficient demand, and valuable public research need not produce profits for its sponsor. What matters is whether the constraint being addressed is real and whether the organization can connect its contribution to the complementary capabilities needed downstream.

This leads to a more useful set of questions. If generating content becomes easier, does trustworthy action become more valuable? If inference becomes cheaper, does verification become a larger part of the cost of useful work? If scientific prediction accelerates, do experiments and clinical evidence become more important bottlenecks? Technological progress rarely removes every constraint simultaneously. More often, it changes which constraint matters most.

The following framework organizes those questions. It is an analytical device, not an empirically estimated scoring model.

Critical conversion	Common misreading	Capability that must be built
Information into action	Owning a screen means owning the task	Context, authorization, dependable execution
Model calls into accepted work	Cheap generation means cheap production	Workflow integration, verification, failure handling
Predictions into usable knowledge	More predictions mean a solved scientific problem	New experiments, external validation, causal testing
Expenditure into industrial capability	Sustained spending means sustained advantage	Staged learning, complementary capabilities, reallocation

Cash flow buys options, not guaranteed outcomes

Meta illustrates the relationship between an established business and a portfolio of uncertain opportunities. In 2025, it generated revenue of $200.966 billion and operating cash flow of $115.8 billion; Reality Labs reported revenue of $2.207 billion and an operating loss of $19.193 billion. The figures establish two facts at once: the existing business provides substantial resources for exploration, and the distance between investment and realized revenue in the newer segment remains considerable.

A loss does not, by itself, establish that an exploration is misguided. Nor does a large investment establish that a correspondingly large payoff is inevitable. A better interpretation is that the company is purchasing options on future products, capabilities, and platform positions. Those options become more valuable when the organization reduces decisive uncertainties, not merely when it demonstrates an ability to keep spending.

The accounting boundaries matter. Reality Labs' segment operating loss is not the loss of a single glasses product, nor is it Meta's entire AI infrastructure expenditure. Consolidated cash flow also does not demonstrate that any particular frontier project has acquired a self-sustaining business model. These figures describe how resources can flow across activities; they should not be assembled into a fictional return on an undifferentiated “future.”

Long-horizon investment therefore needs two scorecards. A financial scorecard tracks expenditure, cash generation, and resilience. A learning scorecard records which technical, adoption, and commercial assumptions have actually been tested. The first without the second risks turning patience into budgetary inertia. The second without the first risks producing experiments that cannot be sustained.

Smart glasses compete for task entry points, not a predetermined post-phone world

The first requirement in assessing glasses is to distinguish stages of development. Meta Ray-Ban Display went on sale in the United States on September 30, 2025, at a launch price of $799 including the Meta Neural Band; the wristband uses surface electromyography to interpret hand movements, rather than reading thoughts. (Meta product announcement) Orion represents a different, more expansive augmented-reality effort: Meta's official page describes it as a prototype unavailable to the public, not a consumer product with an established schedule for scaled delivery. (Meta's Orion overview)

The distinctions are strategically consequential. Camera-and-audio glasses, glasses with a limited display, and more complete spatial AR systems address different tasks and face different engineering and adoption constraints. Progress in one category cannot simply be counted as evidence that another has reached maturity.

Commercial progress also needs the right denominator. EssilorLuxottica reported more than seven million AI-glasses units sold in fiscal 2025; this covers its relevant AI-glasses business, not Display alone, and is not a count of daily active users. (EssilorLuxottica annual results) It establishes a meaningful sales base. Persistent wear, usage frequency, and retention around specific tasks remain separate variables in determining the value of an interface.

The opportunity is to shorten the distance between intention and action. Instead of retrieving a phone, opening an application, and describing the relevant context, a user might ask a question while looking at the world and receive assistance with the next step. The strategic proposition is not just a screen closer to the eyes. It is a tighter connection among perception, understanding, authorization, and action.

Fewer interaction steps, however, do not guarantee a lower total cost of use. Comfort, charging, visual fatigue, mistaken inputs, privacy concerns, and social acceptance can offset convenience. Meta's stated mixed-use battery life of roughly six hours for Display glasses is one concrete reminder that an all-day interface still requires a workable pattern of charging and use. (Meta product specifications)

The more useful commercial hypothesis is therefore that particular high-frequency tasks will migrate away from the phone. Navigation, translation, field assistance, or contextual prompts can be tested separately. They do not require a single device to replace all computing activity. A successful new interface might even increase demand for existing services rather than make those services obsolete.

The strategic contest concerns task entry points and bargaining power. A system that holds relevant context, obtains authorization, and coordinates the next action may acquire a new distribution position. But that position becomes valuable only if people continue to use it, service providers participate, and failures remain manageable. A prototype creates an opportunity to establish an interface; it does not confer ownership of the resulting market.

DeepSeek: efficiency changes access, reliability determines value

Algorithmic efficiency can alter the conditions of competition, but the argument should rest on specific advances rather than a story in which mathematics makes hardware irrelevant. DeepSeek officially released V4.1-Flash on September 10, 2026, describing native multimodal visual understanding. (DeepSeek release record) Its model card describes a Causal Encoder-Decoder architecture, compressed sparse attention, and cache optimizations for input-heavy agentic workloads; it reports an FP4 global KV cache of 890 bytes per token, roughly one-quarter that of V4-Flash, a specific architectural measure reported by the developer. (DeepSeek model card)

The significance is a change in the resource mix required to deliver a capability. Better treatment of context, memory, and computation can bring previously expensive workloads within reach. An improvement in a particular cache measure does not establish that the complete system no longer needs high-performance memory, or that every customer and workload receives the same cost reduction.

DeepSeek's announcement describes the KV cache as requiring one-quarter of the previous generation's HBM and one-eighth of its SSD storage, not eliminating either requirement. (DeepSeek release announcement) The relevant achievement is coordination between architecture and memory hierarchy. It is not a choice between intelligence and infrastructure as mutually exclusive sources of progress.

As of September 10, 2026, the official pricing page listed deepseek-flash output at $0.60 per million tokens off-peak and $1.20 during peak periods, with separate charges for input and cache usage. (DeepSeek pricing) These prices matter, but they are inputs into an economic decision, not a complete measure of productivity. A business is buying an acceptable result, not merely an inexpensive sequence of tokens.

The relevant denominator should consequently move from cost per token to total cost per accepted task. That includes model calls, retrieval and tool use, human review, retries, and the consequences of errors. A cheaper invocation can be more expensive if it requires repeated correction. A more expensive model can be economical if it materially reduces human intervention.

This is not a qualification that weakens the case for efficiency. It is the test that identifies its genuine value. When an efficiency gain survives the entire workflow, it expands feasible uses, lowers barriers to entry, and forces competitors to explain the basis of their premiums. When the advantage exists only on a pricing page, the organization of production may change much less.

Lower unit costs also need not reduce aggregate demand for computing. They may bring new tasks into use or encourage longer context, more extensive analysis, and additional verification. The outcome depends on both consumption per task and growth in the number and intensity of tasks. Algorithmic progress and infrastructure investment can therefore reinforce one another rather than act only as substitutes.

The deeper strategic implication is that differentiation may migrate as foundational models become easier to obtain. Workflows, industry data, authorization, reliability, and customer relationships can become more important complements. Models remain consequential, but possessing a model is not the same as possessing the complete system that performs useful work. What comes under pressure is an unexplained premium, not every incumbent indiscriminately.

Openness is a design of boundaries and returns

Search, AI models, and scientific data share a strategic problem: how can information be discovered and used, and how can external contributions improve the system? A platform has reasons to protect privacy, security, and quality. It also needs legitimate information flows and collaboration. The useful distinction is not between morally “open” and “closed” organizations, but between boundaries that increase system value and boundaries that unnecessarily restrict learning.

At least four forms of openness should be separated: the ability to read information, obtain a model or dataset, modify and deploy it, and move resulting work to another system. These involve different technical interfaces and rights. A download does not necessarily include training data. Research access does not necessarily authorize every commercial use. An API does not necessarily guarantee durable portability.

Specific licenses make these boundaries clearer than a general “open” label. The 2023 Llama 2 Community License, for example, required licensees meeting its release-time threshold of more than 700 million monthly active users to seek a separate license, which Meta could grant at its sole discretion. (Llama 2 license) Meta's September 2, 2026 announcement of Muse Spark 1.3 listed API access while placing an open-weights release on the future roadmap. (Meta research announcement) Neither example establishes an immutable identity: rights and access need to be assessed by model and stage.

The commercial purpose of openness is also more complicated than making a product free. An organization may lower the price of a foundational layer to attract participation and expand demand for complements. It may distribute tools while earning revenue from hosting, applications, or other relationships. Such arrangements can enlarge competition while simultaneously creating new dependencies. The analysis must identify where value is created, who bears the cost, and who can retain the returns.

The merits of a platform boundary should therefore not be assessed solely by whether it preserves traffic. Does it improve discovery for creators? Can developers build durable businesses? Do users retain meaningful alternatives? Can external knowledge improve system performance? Restrictions that reduce abuse and quality risks may create value. Restrictions that mainly prevent substitution may consume long-term adaptability.

In the AI era, the openness of the model and the user's control over data, workflows, and migration are related but distinct questions. A durable strategy designs dependencies that partners can continue to accept. It does not confuse a temporary ability to prevent exit with a permanent capacity to create value.

For search and content platforms, the test can be even more direct. When a boundary reduces access to external knowledge, discovery of useful content, or creators' ability to reach outside audiences, the platform should establish that its security and quality benefits justify the loss of discovery and learning. This is not an argument against every boundary. It is an argument for including the boundary itself in the calculation of long-term value.

AI biology changes discovery; it does not abolish validation

Life sciences bring these issues into a domain with more demanding evidence requirements. On April 29, 2026, Biohub announced a five-year, $500 million Virtual Biology Initiative: $400 million for large-scale data generation and next-generation measurement, imaging, and engineering, and $100 million for external research and coordinated global data generation, with NVIDIA named as a technology partner. (Biohub announcement) The allocation is strategically revealing because it treats improved measurement, not simply larger models, as part of the foundation for progress.

Institutional boundaries must remain visible. Biohub announced that initiative, while CZI's earlier NVIDIA partnership concerned virtual-cell models and associated scientific tools; neither announcement establishes a Meta life-sciences revenue business. (Biohub; CZI partnership announcement) Shared founder connections do not demonstrate a single chain of ownership, intellectual property, commercial objectives, or profit capture.

This distinction permits a more serious interpretation of entrepreneurial ambition. Moving beyond an existing business need not mean expanding the same public corporation into every promising field. It can mean establishing different institutions for different purposes, with standards appropriate to basic science, shared tools, and commercial products. Research infrastructure can succeed without its sponsor owning the downstream market.

AI's most promising contribution here is to improve the sequence of hypothesis formation, experimental selection, and interpretation. The strategic value of a virtual cell is not that computation replaces biological reality. It is that a model might help researchers identify what to measure next, which intervention to test, and which prediction most needs to be challenged. Prediction becomes more useful when it is connected to experiments, feedback, and revision.

Describing a cellular state is not the same as predicting an intervention, and predicting an intervention is not the same as demonstrating an effective human treatment. A model that fits existing observations still faces unfamiliar cell types, disease stages, experimental conditions, and populations. More data creates opportunities for investigation; it does not automatically establish causality.

The data foundation is already substantial. The peer-reviewed CZ CELLxGENE Discover platform paper reported 93.6 million unique cells across more than 1,550 datasets as of October 1, 2024; these are dated platform figures, not the training corpus of every contemporary virtual-cell model. (CELLxGENE platform paper) Separately, an April 2026 virtual-cell benchmark preprint found that models struggled to recover fine-grained, perturbation-specific effects under several unseen-intervention conditions; its findings remain those of a preprint and are bounded by the evaluated models, data, and tasks. (Virtual-cell benchmark) These observations are compatible: a growing data foundation does not remove the need to test generalization.

Nor is this opportunity confined to a single institution. In September 2026, Google DeepMind introduced AlphaGenome Atlas, providing predictions for approximately nine billion single-nucleotide variants while explicitly stating that it had not been validated for clinical use. (AlphaGenome Atlas announcement) The example challenges both the assumption that established technology organizations only defend existing businesses and the assumption that computational scale dissolves clinical evidence requirements.

The Human Cell Atlas likewise represents an international scientific community spanning many countries and institutions, not a database belonging to a single technology founder. (Human Cell Atlas global community) The more accurate picture is one of scientific organizations, technology providers, and specialized institutions building complementary capabilities, rather than one company's algorithm taking control of medicine.

The crucial shift in scarcity may be from generating more predictions to eliminating errors more credibly. If prediction improves faster than experimentation, measurement quality, causal testing, and interoperable data standards become increasingly consequential. A model developer does not automatically possess those capabilities. Organizations that connect modeling, experiments, and professional responsibility may occupy a more defensible position.

Large human needs do not imply unlimited commercial capture

Better health can create enormous value without implying unlimited purchasing power or frictionless commercialization. The patient, the clinician, the payer, and the authority deciding whether a product may be used can be different actors. Even an effective technology must establish why it should be adopted for a defined population, use, and payment arrangement.

Three questions therefore separate a scientific advance from a sustainable business: does it produce a meaningful effect, can that effect be delivered affordably, and how much of the resulting value can the provider retain? A large market estimate cannot answer all three. Public datasets, research tools, diagnostics, medicines, and healthcare services also have different economics and should not be collapsed into a single business model.

The Avantect pancreatic-cancer test illustrates the importance of evidence categories. Its official FAQ states that the test is performed in an appropriately accredited laboratory but has not been cleared or approved by the FDA, is not intended for general screening of average-risk individuals, and does not establish a diagnosis. (Avantect FAQ) The analytical point is not to judge the test. It is that commercial availability, analytical performance, authorization for a particular use, and demonstrated improvement in health outcomes are different propositions.

Historical clinical-development data underscore the difficulty of the chain. A report by BIO, QLS Advisors, and Informa examined 12,728 phase transitions across 9,704 development programs during 2011–2020 and estimated an overall Phase I-to-FDA-approval likelihood of 7.9%, calculated by multiplying stage-transition probabilities and incorporating scientific, commercial, and program-discontinuation factors. (Clinical development success-rate report) This is not a forecast for today's AI-enabled drug programs and does not establish that AI cannot improve development. It demonstrates why accelerating discovery should not be treated as completing subsequent validation.

The same logic extends beyond medicine. A large user base does not establish durable profitability. A lower retail price does not establish lower end-to-end costs. Ownership of an interface does not confer an unrestricted ability to impose fees. An organization can create substantial social value while retaining only part of the commercial surplus, or lose partners and market access when its distribution of value becomes unsustainable.

The strategic task is not to find a market whose needs can be called infinite. It is to understand how value passes through technology, organizations, rules, and payment systems. Market boundaries can expand, but expansion requires resolving constraints, not converting aspiration directly into projected revenue.

Globalization is not a single contest

TikTok, Temu, and DeepSeek are useful to consider together because they illuminate different parts of competition, not because they constitute one uniform model of disruption. A content platform raises questions about discovery and distribution. A commerce platform raises questions about supply coordination, price, and delivery. A model provider raises questions about capability, usage costs, and deployment conditions.

On August 31, 2026, TikTok said it reached 189 million people in the European Union each month as it introduced its seventh Digital Services Act transparency report, covering the first half of 2026. (TikTok European transparency announcement) That self-reported reach measure establishes breadth, not attention share, loyalty, or profitability. The platform's strategic value also depends on sustaining useful relationships among creators, audiences, and commercial demand.

Temu illustrates why operating responsibilities have to develop alongside international scale. On May 28, 2026, the European Commission announced a €200 million fine for failures to diligently identify, analyze, and assess systemic risks involving illegal products on its platform. (European Commission decision announcement) The specific decision is not a verdict on an entire business model. It demonstrates that product-risk assessment and platform governance belong within the operating system of a scaled business, rather than being deferred indefinitely until after growth.

The corresponding test of a low-price strategy is whether the price reflects more efficient supply coordination or depends on subsidies and costs displaced into quality, delivery, and risk management. That question requires sustainable end-to-end economics. A promotional price, a period of rapid growth, or a founder's reputation cannot answer it.

The implication also reaches back to smart glasses. A migration of computing interfaces need not eliminate demand for content, shopping, payment, or fulfillment. Existing platforms may adapt, collaborate, or become indispensable services on a new device. Substitution and complementarity can coexist. The redistribution of technological power may appear as a change in negotiating positions and participation terms, not the simultaneous disappearance of every older business.

Patient capital should make learning faster than irreversible commitment

Long-term investment and financial discipline are not opposites. The practical challenge is to match different decisions to different time horizons. Some capabilities must be secured early, while other commitments become more valuable when postponed until additional information is available.

In its second-quarter 2026 earnings call, Meta guided to full-year capital expenditure of $130 billion–$145 billion, including principal payments on finance leases; this was guidance, not realized spending, disclosed at 29m 18s. Discussing more distant capacity needs, its finance chief described securing land and power while deferring chip and other major purchasing decisions, at 1h 2m 35s. The distinction is between preparing the conditions for future capacity and prematurely fixing every expensive element of its implementation.

That distinction suggests a broader management principle: remain patient about a promising problem while staying flexible about the proposed solution. Commitment to a strategic direction should not make a particular technical route untouchable. Closing a project after learning something important is not necessarily a retreat from long-term thinking. Reallocation is part of the capability that makes long-term thinking useful.

Concentrated founder control can create room to withstand short-term pressure, but it can also allow a mistaken belief to persist. Professional management can overemphasize near-term results, but it can also protect research through clearer accountability and evidence standards. These are governance possibilities to be managed, not conclusions that follow from a person's upbringing or a heroic account of their character.

The right evidence changes with the stage of investment. A prototype should remove an important technical uncertainty. A pilot should establish repeatable demand. Expansion should reveal credible unit economics. A platform should demonstrate that complementors can participate sustainably without indefinite subsidy. Patience can vary across these stages; the requirement to learn should not disappear.

National advantage depends on conversion capacity

Entrepreneurial cognition can influence which problems an organization is willing to attempt. National technological capability, however, cannot be reduced to the presence of an unusually ambitious founder. A more useful unit of analysis is the system that converts ideas into validated technologies, technologies into dependable products, and products into broadly distributed economic capabilities.

That system connects at least four kinds of competence: scientific inquiry and testing; engineering, manufacturing, and maintenance; capital and organizations suited to different risk horizons; and institutions that establish permission, responsibility, procurement, and market access. A persistent break in any of these connections can prevent an upstream breakthrough from producing substantial downstream benefits.

This changes what a serious technology strategy should measure. Counting models, papers, or installed computing capacity is not enough. Can ordinary businesses use the resulting tools? Can independent teams obtain lawful data and appropriate testing resources? Can public research be reused? Can an effective product pass through procurement and deployment without losing its economic advantage? These questions measure an economy's ability to absorb technology, rather than a single institution's ability to demonstrate it.

Strategic autonomy and openness need not be mutually exclusive. A resilient arrangement can preserve alternatives in critical functions while reducing the cost of legitimate collaboration, knowledge reuse, and cross-institutional validation. Dependence on an external system may weaken bargaining power. Treating every external contribution as a threat may weaken the capacity to learn.

Infrastructure therefore has two distinct strategic meanings. For a company, it can establish a commercial control point. For society, it can lower the cost of innovation for many participants. Sound institutional design must recognize when these interests reinforce one another and when they conflict. The dominance of a particular platform should not automatically be treated as the equivalent of national technological strength.

Watch the variables, not a predetermined ending

There is no requirement that this transition produce a single outcome. Glasses may become a major interface or remain a valuable complement to phones. Open models may compress the price of basic intelligence while increasing returns to distribution, proprietary workflows, or engineering competence. Scientific tools may diffuse widely while clinical and commercial value is realized by several specialized institutions.

The most informative indicators are therefore conditional rather than theatrical:

Is an interface shift actually occurring? Distinguish initial trials from durable use and high-frequency tasks that the new device performs distinctly better.

Does the cost advantage survive the workflow? Measure accepted outcomes after review, retries, and failure handling, not only the price of generation.

Does scientific progress cross a validation boundary? Separate better predictions on familiar data from new experiments, external validation, and clinical results.

Can the ecosystem sustain itself? Ask whether participants have durable economics and reasonable portability, rather than temporary incentives to join.

Can capital allocation correct itself? New evidence should change budgets and technical choices, not merely the language used to describe a project.

These questions define strategy as the construction of future options rather than rhetorical possession of the future. An organization that can create, test, and sometimes abandon options has a stronger basis for adaptation than one that can only intensify its original narrative.

The most consequential entrepreneurial judgment combines an understanding of what technology can do, what an organization can support, what users will adopt, and what responsibilities society will require. The next technology advantage is unlikely to belong simply to those who announce the ending first. It is more likely to accrue to those who change the conditions of competition and turn that change into dependable value at scale.