📖 中文版在下方 — scroll down for the Chinese version.

Today's Google (Alphabet) stands at a crossroads. It's caught between the anxiety of its search ad empire being eroded and the trap of chasing OpenAI's playbook in the generative AI wave. This posture — inventing the future but being held hostage by the existing cash cow — looks eerily similar to Kodak, which invented the digital camera and still went bankrupt.

But Google's biggest mistake isn't falling behind on technology. It's blindly treating OpenAI as the only paradigm for AI.

Kodak's Lesson: Invented the Future, Died in the Past

In 1975, Kodak engineer Steve Sasson built the world's first digital camera. Management's first reaction wasn't excitement — it was fear: "This thing will kill our film business."

Kodak's film empire was enormously profitable: high margins, a massive printing ecosystem, global distribution channels. Management instinctively shelved the digital camera to protect the cash cow, and was ultimately destroyed by the very future they had invented.

This is what Clayton Christensen defined as the "Innovator's Dilemma" — successful companies trapped by their own success.

Google's "Kodak Symptoms"

Today's Google shows strikingly similar symptoms across multiple dimensions.

1. The Cash Cow Shackle

When users shift from "scanning a list of links" to "getting direct answers," Google's most profitable sponsored link placements lose their conversion value. Same logic as digital photos shutting down print shops.

2. Invented the Future, Trapped by the Future

In 2017, Google published the landmark paper Attention Is All You Need, introducing the Transformer architecture that directly ignited this round of AI revolution. Even core concepts like "Hallucination" were first coined by Google researchers.

Yet it was OpenAI that turned Transformers into a disruptive consumer product (ChatGPT).

The root cause: Google feared AI hallucinations would damage its commercial reputation and worried AI direct answers would cannibalize search results, leading to indecisive, slow product execution. Exactly like Kodak inventing the digital camera but not daring to promote it.

3. Big Company Disease

Kodak had layers of middle management and bureaucratic processes; Google has been equally plagued by "big company syndrome" in recent years — countless overlapping internal projects (remember all those messaging apps that were built and killed?), with employees incentivized toward promotion-friendly micro-innovations rather than risky disruption of existing business lines.

Why Google Shouldn't Try to Be the Next OpenAI

OpenAI follows a classic "unicorn model" — high cost, closed source, monetized through premium subscriptions. If Google tries to compete head-on in this lane, it's fighting with its weakness against the competitor's strength, while also bearing the burden of its search ads being cannibalized.

Looking back at Google's rise, its core moat was never proprietary closed technology — it was ecosystem dominance through openness:

• Android: defeated Windows Phone and Symbian through open source, monopolizing global mobile traffic via bundled services
• Chromium: open-sourced the browser engine, destroyed IE, and made Chrome the internet's default gateway
• Kubernetes & TensorFlow: open-sourced cloud-native standards and deep learning frameworks, establishing Google's authority in the developer community

Open source is the antibody in Google's bloodstream. When you apply this logic to the AI era, the answer becomes clear.

The Breakout Triangle: Open Source × Small Models × Hardware Alliance

I. Open Source: The Symmetric Weapon Against Closed Walls

When every developer and enterprise in the world can access industrial-grade open-source models with zero barriers, OpenAI's expensive subscription and API model rapidly loses its justification.

More importantly — the endgame of software is ecosystem. As long as millions of developers worldwide build fine-tunes and applications on Google's open-source framework, Google retains control of AI's technical standards.

But Google isn't doing this well enough. Take the Gemma series as an example — its open-source strategy exposes several critical flaws:

• "Wants to open source, but won't give the good stuff": core technology and large-parameter models are all locked behind closed APIs; what's open-sourced is the B-team
• Safety reins pulled too tight: excessive alignment makes the model rigid and lifeless in creative tasks and open discussions
• Code capability cliff: same-size Gemma consistently gets outperformed by Qwen and even DeepSeek on code benchmarks
• Starved community ecosystem: Gemma-related derivative projects on Hugging Face lag far behind the Qwen and DeepSeek ecosystems in both quantity and quality

By contrast, the teams truly practicing the spirit of open source are Chinese organizations. DeepSeek built open-source models rivaling GPT-4 at a fraction of the training cost, winning global developer respect through sheer technical transparency. Alibaba Cloud's Qwen series covers dialogue, code, and multimodal tasks, with Hugging Face community activity and derivative ecosystems already far surpassing Gemma. A search giant invented the Transformer, yet the fruits of the open-source ecosystem were picked by startups and latecomers — this is the reality that should alarm Google most.

II. Small Language Models (SLM): The Volume Play for Device Dominance

OpenAI's paradigm led the industry toward a blind worship of "bigger is better" — trillion-parameter models are certainly smart, but their high latency, high cost, and extreme power consumption make mass adoption impossible.

AI's endgame isn't cloud-based compute dominance — it's ubiquitous on-device intelligence.

What can actually reach billions of users daily are 3B to 8B parameter models that run directly on devices. This is precisely the battlefield where Google holds an overwhelming advantage:

• Controls the ultimate entry points: Android and Chrome reach billions of users globally
• Local, instant response: embed competitive small open-source models natively into the OS — users need no internet, no API payments
• Privacy and security: data never leaves the device, naturally complying with privacy regulations

Win the devices with small models, and you win the broadest consumer market.

III. Ally with Intel and AMD: Build the Anti-Nvidia Compute Alliance

In the cloud, Nvidia has built an extraordinarily stubborn compute monopoly with GPUs and CUDA; the OpenAI model turns the entire industry into Nvidia's "tenants." Google fighting alone with TPU in the cloud is still playing by the opponent's rules.

The key to breaking through is shifting the perspective from cloud to edge, forming a deep hardware-software alliance with Intel and AMD:

• Intel and AMD control the world's dominant CPU and next-gen NPU ecosystem
• Google deeply optimizes small open-source models for Intel NPU and AMD Ryzen AI/ROCm architectures
• Achieving zero-barrier local inference, bypassing expensive GPU requirements

When "Google open-source model + Intel/AMD chip" becomes the factory default for AI PCs, billions of edge devices worldwide become the breeding ground for Google's ecosystem.

Why Google Hesitates

If open-source ecosystem is in Google's DNA, why is it moving so cautiously? Three reasons:

1. Financial sunk cost: training top-tier hundred-billion-parameter models costs hundreds of millions of dollars. Under Wall Street's quarterly earnings pressure, management struggles to commit to giving away top models for free
2. Safety liability: as a megacorp, if Google open-sources a top model that gets misused, the regulatory accountability and PR risk far exceed what a startup would face
3. Internal strategic tug-of-war: Google still hasn't truly chosen between "use closed models to defend the search ad base" and "go full open source to rebuild the cloud ecosystem"

Conclusion

OpenAI is an episode in AI's early evolution, not the industry's final paradigm.

From "blindly chasing closed-source APIs" to "open-source top models + small models to capture devices + hardware alliances to empower the ecosystem + GCP to harvest the compute infrastructure" — this isn't just a return to Google's historical DNA; it's the only path away from its Kodak Moment.

Drop the ego, break the old paradigm, re-embrace developers and hardware allies — and the Google that once rewrote the rules can do it again, on its own terms.

This article is a personal observation on Google's strategic dilemma. If you're interested in full-stack development or AI application deployment, feel free to get in touch.

重回 Android 时代：Google 摆脱「柯达魔咒」的破局之道

📖 English version above — scroll up for the English translation.

今天的 Google（Alphabet）正伫立在命运的十字路口。它深陷于"传统搜索广告被侵蚀"的焦虑，又在生成式 AI 的浪潮中陷入了追赶 OpenAI 的误区。这种"发明了未来却被现有现金牛困住"的姿态，像极了当年发明数码相机却最终倒下的柯达。

但 Google 最大的失误，并不是技术落后，而是盲目地将 OpenAI 视为 AI 的唯一范式。

柯达的教训：发明了未来，却死在了过去

1975 年，柯达工程师 Steve Sasson 发明了世界上第一台数码相机。管理层看到这项技术后的第一反应不是兴奋，而是恐惧——"这玩意会杀死我们的胶卷业务"。

柯达的胶卷帝国太赚钱了：高毛利、庞大的冲印生态系统、全球化的渠道网络。管理层出于保护核心现金牛的本能，将数码相机束之高阁，最终被自己发明的未来彻底颠覆。

这就是克莱顿·克里斯坦森所定义的 "创新者的窘境"（Innovator's Dilemma）——成功企业被自身的成功所困。

Google 的"柯达式隐患"

对比今天的 Google，它在多个维度上表现出了极其相似的症状。

1. 现金牛的枷锁

当用户获取信息的方式从"在链接列表中筛选"变成"直接获取答案"，Google 最赚钱的赞助商链接广告位就失去了转化价值。这和当年数码照片让冲印店关门是同一个逻辑。

2. 发明了未来，却被未来困住

2017 年，Google 发表了划时代的论文 Attention Is All You Need，提出了 Transformer 架构，直接开创了这一轮 AI 革命的技术基础。甚至"Hallucination（幻觉）"等核心概念也是 Google 研究员率先提出的。

然而，将 Transformer 转化为颠覆性消费级产品（ChatGPT）的却是 OpenAI。

根源在于：Google 既担心 AI 幻觉破坏商业声誉，又担心 AI 直答取代搜索结果，最终在产品化上举棋不定、步子太慢。和柯达发明数码相机却不敢推广，如出一辙。

3. 大企业病

柯达有庞大的中层管理体系和官僚流程；Google 近年来同样饱受"大公司病"诟病——内部项目众多但缺乏统筹（回忆一下那些被反复重做又被砍掉的通讯应用），员工倾向于做能"升职加薪"的微创新，而不是承担风险去颠覆现有业务。

为什么 Google 不应该成为第二个 OpenAI

OpenAI 走的是典型的"独角兽模式"——高造价、高闭源、靠卖高价订阅变现。如果 Google 跟着在这条赛道上死磕，不仅要背负搜索广告被侵蚀的包袱，更是用自己的短板去拼别人的长板。

回顾 Google 的崛起史，它的核心壁垒从不是封闭的独占技术，而是 开放生态的统治力：

• Android：通过开源击败了 Windows Phone 和 Symbian，靠内置服务垄断了全球移动端流量
• Chromium：将浏览器底层开源，彻底瓦解了 IE，让 Chrome 成为互联网的绝对入口
• Kubernetes & TensorFlow：将云原生标准和深度学习框架开源，奠定了 Google 在技术社区的宗师地位

开源才是 Google 血液里自带的抗体。 当你把这个逻辑套用到 AI 时代，答案就清晰了。

破局三角：开源 × 精简模型 × 硬件联盟

一、开源：瓦解闭源高墙的对称武器

当全天下的开发者与企业都能无门槛获得工业级的开源模型时，OpenAI 式的高昂订阅与 API 模式将迅速失去合理性。

更重要的是——软件的终局是生态。只要全球上百万开发者基于 Google 的开源框架做微调和应用开发，未来的 AI 技术标准就依然掌握在 Google 手中。

但 Google 目前做得还不够好。以 Gemma 系列为例，它的开源策略暴露了几个致命问题：

• "既想开源，又舍不得给好料"：核心的优秀技术和超大参数量全都锁在闭源 API 后面，开源的只是"次等马"
• 安全缰绳套得太紧：过度对齐（Alignment）导致模型在创意任务和开放性讨论中生硬死板
• 代码能力断崖式落后：同尺寸的 Gemma 在代码基准测试中往往被 Qwen 甚至 DeepSeek 完虐
• 社区生态匮乏：Hugging Face 上围绕 Gemma 的二次开发项目数量和质量远落后于 Qwen 和 DeepSeek 生态

相比之下，真正在践行开源精神的是中国团队。DeepSeek 以极低的训练成本打造出性能比肩 GPT-4 的开源模型，直接用技术透明度赢得了全球开发者的尊重；阿里云的 Qwen 系列则从对话、代码到多模态全面开花，在 Hugging Face 上的社区活跃度和二次开发生态已经远超 Gemma。一个搜索巨头发明了 Transformer，开源生态的果实却被创业公司和后来者摘走——这才是 Google 最应该警醒的现实。

二、精简模型（SLM）：占据终端的走量杀手锏

OpenAI 的范式将行业引向了"越大越好"的盲目崇拜——千亿参数的模型固然聪明，但高延迟、高成本、极度烧电，根本无法支撑大规模普及应用。

AI 的终局绝非云端的算力霸权，而是无处不在的端侧智能。

真正能走进亿万用户日常的，是 3B 到 8B 参数量级、能够直接跑在设备上的精简模型。这恰恰是 Google 拥有压倒性优势的绝对主场：

• 掌控绝对入口：Android 与 Chrome 触达全球数十亿用户
• 本地化秒级响应：将精简开源模型原生植入系统，用户无需联网、无需支付 API 费用
• 隐私与安全：数据不出设备，天然满足隐私法规

精简模型赢得了终端，就赢得了最广阔的消费级市场。

三、联合 Intel 与 AMD：组建反 Nvidia 算力联盟

在云端，Nvidia 靠 GPU 和 CUDA 打造了极其顽固的算力垄断；OpenAI 模式让全行业都沦为 Nvidia 的"租客"。Google 仅靠 TPU 孤军奋战，依然是在对方的规则里打仗。

破局的关键在于把视角从云端投向端侧，与 Intel 和 AMD 结成深度的软硬一体战略联盟：

• Intel 与 AMD 控制着全球主流的 CPU 和新一代 NPU 生态
• Google 将精简开源模型针对 Intel NPU 与 AMD Ryzen AI/ROCm 做底层深度优化
• 实现零门槛本地运行，绕过高昂的 GPU 门槛

当"Google 开源模型 + Intel/AMD 芯片"成为 AI PC 的出厂标配，全球数以亿计的终端设备就会变成 Google 生态的繁衍土壤。

为什么 Google 现在如此犹豫

既然开源生态是 Google 的基因，为什么它走得束手束脚？三个原因：

1. 财务上的沉没成本：训练千亿级顶尖模型需要数亿美元。在华尔街季报的压力下，管理层极难下定决心把最顶级的模型免费开源
2. 安全的责任枷锁：作为巨头，Google 一旦开源了被滥用的顶级模型，面临的监管问责和舆情风险远高于创业公司
3. 内部的战略撕扯：在"用闭源模型守护搜索广告基本盘"和"彻底开源重构云生态"之间，Google 至今没有做出真正的抉择

结语

OpenAI 只是 AI 演进初期的一段插曲，绝非产业发展的终局范式。

从"盲目追赶闭源 API"转向 "开源顶级模型 + 精简模型占领终端 + 联合硬件巨头赋能生态 + GCP 收割算力底座"——这不仅是 Google 历史基因的复归，更是其告别"柯达时刻"的唯一通途。

放下面子，打破传统范式，重新拥抱开发者与硬件盟友——那个曾经无往不胜的 Google，就能再次按自己的节奏改写时代规则。

这篇文章是对 Google 战略困境的个人观察与思考。如果你对全栈开发或 AI 应用落地有兴趣，欢迎聊聊。