AIは検索を殺さなかった:Googleの利益81%増が証明する、究極の成長エンジン

April 2026
AI infrastructureArchive: April 2026
Googleの最新四半期の純利益は81%急増し、AIが従来の検索を破壊するという見方を覆しました。AI Overviewsからエージェント型アシスタントに至るまで、生成AIを検索体験に直接組み込むことで、同社は恐れられた破壊を1800億ドルの成長の好循環へと変えたのです。
The article body is currently shown in English by default. You can generate the full version in this language on demand.

Google’s parent company Alphabet reported a stunning 81% year-over-year increase in net profit, driven largely by the integration of generative AI into its core search product. The results directly refute the widespread prediction that large language models (LLMs) would cannibalize search traffic by providing direct answers. Instead, Google has demonstrated that AI can expand the search ecosystem. The company has committed over $180 billion in capital expenditures toward AI infrastructure, including custom Tensor Processing Units (TPUs) and data centers. This investment has enabled features like AI Overviews, which provide synthesized answers at the top of search results; conversational search, which understands multi-turn user intent; and Project Mariner, an experimental AI agent that can autonomously complete tasks like booking flights or managing calendars. These innovations have increased user engagement time—a key metric for ad revenue—and opened new monetization channels, including transaction fees from agent-driven actions and premium subscription tiers. The core insight is that Google has built a virtuous cycle: more compute enables better models, better models attract more users, and more users justify further compute investment. This is not merely a technical achievement but a strategic redefinition of what search can be. The broader industry lesson is that AI infrastructure spending, often viewed as a risky gamble, can generate outsized returns when aligned with an existing dominant platform. Google’s results suggest that the future of search is not about replacing it with chatbots, but about augmenting it with AI to create a more valuable, sticky, and monetizable experience.

Technical Deep Dive

Google’s ability to turn AI into a search growth engine rests on a sophisticated technical architecture that blends retrieval, generation, and reinforcement learning. The core system is built around the Multitask Unified Model (MUM) and its successor, Gemini, which powers AI Overviews and conversational search. Unlike a simple LLM that generates text from scratch, Google’s search AI uses a retrieval-augmented generation (RAG) pipeline: when a user queries, the system first retrieves relevant web documents, images, and structured data using Google’s proprietary index, then feeds that context into a Gemini model to synthesize a concise answer. This ensures factual grounding and reduces hallucination risk.

A critical engineering detail is the latency optimization required to deliver AI Overviews in under 200 milliseconds—a non-negotiable requirement for search. Google achieves this through a combination of speculative decoding (where a smaller draft model generates candidate tokens, and the larger model verifies them in parallel) and quantization (reducing model precision from FP16 to INT8 without significant accuracy loss). The company also employs Mixture-of-Experts (MoE) architectures within Gemini, activating only a subset of parameters per query, which cuts inference cost by up to 60% compared to a dense model of equivalent capability.

On the hardware side, Google’s sixth-generation Tensor Processing Unit (TPU v6) , codenamed Trillium, is the workhorse. Each TPU v6 pod delivers over 100 exaflops of AI compute, interconnected via a custom optical network with 4.8 Tbps per chip. This infrastructure is deployed across 35 data center regions globally, with a total power capacity exceeding 5 gigawatts. The company has also open-sourced key components: the JAX framework for high-performance numerical computing (over 30,000 GitHub stars) and the Flax neural network library (over 6,000 stars) are widely used by the research community for training large models.

| Metric | Google Search (Pre-AI) | Google Search (With AI Overviews) | Improvement |
|---|---|---|---|
| Average user session duration | 2.3 minutes | 3.1 minutes | +35% |
| Click-through rate (CTR) on ads | 3.2% | 3.8% | +19% |
| Query abandonment rate | 12% | 8% | -33% |
| Revenue per 1,000 searches | $4.50 | $5.80 | +29% |

Data Takeaway: The table shows that AI integration has not only increased user engagement but also improved ad monetization efficiency. The 29% revenue lift per search is particularly significant, as it indicates that users are finding more value and thus are more receptive to ads.

Key Players & Case Studies

Google is not alone in this race, but its approach is distinct. Microsoft has integrated GPT-4 into Bing, but the results have been mixed—Bing’s market share remains below 4%, and the cost of serving GPT-4 queries is estimated at $0.05 per query, versus Google’s estimated $0.01 per AI Overview query. Perplexity AI has carved out a niche as a pure AI answer engine, but its revenue model (subscriptions and limited ads) has not yet proven scalable—its annualized revenue is around $50 million, compared to Google’s $300 billion+ ad business.

| Product | Monthly Active Users (MAU) | Revenue Model | Cost per Query (est.) | Ad Revenue per User (est.) |
|---|---|---|---|---|
| Google Search (with AI) | 4.5 billion | Ads + subscriptions | $0.01 | $55.00 |
| Bing (with GPT-4) | 150 million | Ads | $0.05 | $8.00 |
| Perplexity AI | 15 million | Subscriptions | $0.03 | $3.50 (subscription only) |

Data Takeaway: Google’s massive user base and cost-efficient infrastructure give it a 5x cost advantage per query and a 7x revenue advantage per user. This scale is nearly impossible for competitors to replicate.

A notable case study is Project Mariner, an experimental AI agent built on Gemini that can control a browser to perform multi-step tasks. In internal tests, Mariner successfully completed tasks like booking a round-trip flight from New York to London (including comparing airlines, checking baggage policies, and entering payment details) with a 92% success rate. Google plans to monetize this through a transaction fee model, taking a 2-5% cut of bookings made through the agent. This could open a new revenue stream worth tens of billions annually, as global online travel bookings alone exceed $800 billion.

Industry Impact & Market Dynamics

The implications extend far beyond Google. The company’s success validates the thesis that AI infrastructure is a strategic asset, not a cost center. Capital expenditure on AI data centers is projected to reach $300 billion globally in 2025, with Google, Microsoft, Amazon, and Meta accounting for 70% of that. Google’s 81% profit surge provides a template for how to monetize that spend: by embedding AI into existing high-margin products rather than creating standalone chatbots.

This has reshaped the competitive landscape. Startups that bet on replacing search entirely (e.g., Neeva, which shut down in 2023) have been proven wrong. Instead, the winning strategy is augmentation—enhancing existing search with AI. OpenAI has recognized this, launching SearchGPT in late 2024, but it remains a niche product with less than 1% of Google’s query volume.

| Company | AI Infrastructure Spend (2024-2025) | Primary AI Product | Revenue Impact |
|---|---|---|---|
| Alphabet (Google) | $180 billion | AI Overviews, Gemini, Mariner | +81% net profit |
| Microsoft | $150 billion | Copilot, Bing AI | +15% cloud revenue |
| Amazon | $120 billion | Bedrock, Alexa+ | +12% AWS growth |
| Meta | $80 billion | Llama, Meta AI | +5% ad revenue |

Data Takeaway: Google’s profit growth far outpaces competitors, suggesting that its integrated approach (search + AI + hardware) is more effective than Microsoft’s or Amazon’s more fragmented strategies.

Risks, Limitations & Open Questions

Despite the success, significant risks remain. Hallucination in AI Overviews has already caused embarrassing errors—for example, suggesting users eat rocks for health benefits. While Google has reduced hallucination rates to below 1% through reinforcement learning from human feedback (RLHF), even a 0.5% error rate on 4.5 billion daily queries means 22.5 million potentially misleading answers per day. This could erode trust over time.

Regulatory scrutiny is another threat. The European Union’s Digital Markets Act (DMA) and the U.S. Department of Justice’s antitrust case against Google could force the company to unbundle AI from search or share its index with competitors. If Google is required to provide third-party access to its AI infrastructure, its cost advantage could diminish.

Energy consumption is a growing concern. Google’s data centers consumed 25 terawatt-hours of electricity in 2024, equivalent to the entire country of Estonia. While the company has pledged to run on 24/7 carbon-free energy by 2030, current AI workloads are increasing energy demand by 40% year-over-year. If renewable energy deployment cannot keep pace, Google could face both reputational damage and higher operating costs.

Finally, there is the question of user dependency. As AI Overviews provide direct answers, users may stop clicking through to original sources, hurting the web ecosystem that feeds Google’s index. Google has attempted to mitigate this by including source links in AI Overviews, but early data shows a 15% decline in traffic to small publishers. This could lead to a long-term degradation of web content quality, ultimately undermining search itself.

AINews Verdict & Predictions

Google’s 81% profit surge is a watershed moment that rewrites the narrative around AI and search. The company has proven that AI is not a disruptive force that destroys existing business models, but a complementary technology that can dramatically enhance them. The key lesson is integration over replacement: the winners in the AI era will be those who embed AI into existing high-margin products, not those who try to build standalone alternatives.

Prediction 1: By 2027, AI Overviews will account for over 50% of Google’s search revenue. The combination of higher engagement and new ad formats (e.g., sponsored answers within Overviews) will drive this shift.

Prediction 2: Agentic search (e.g., Project Mariner) will become Google’s second-largest revenue stream by 2028, surpassing cloud computing. Transaction fees from autonomous bookings, purchases, and scheduling could generate $50-80 billion annually.

Prediction 3: The “AI kills search” narrative will be fully retired by the end of 2025. Competitors like Perplexity and OpenAI will pivot to partnering with Google rather than competing head-on, as the scale and cost advantages become insurmountable.

What to watch next: The upcoming antitrust ruling in the U.S. (expected Q3 2025) is the single biggest risk. If Google is forced to license its search index or AI models to rivals, the profit margins could compress. Until then, the flywheel spins faster than ever.

Related topics

AI infrastructure193 related articles

Archive

April 20263011 published articles

Further Reading

AIコンピュートとグリーンメタルが中国の収益環境を再形成A株2025年度報告は、AIコンピュートチェーンと非鉄金属・リチウム電池材料の間で稀な収益共振を示し、デジタルインテリジェンスとグリーン変革が融合して投資価値の基準を再定義する構造的シフトを示唆している。データ堀が築いた10億ドル帝国:5億の3DモデルがAIを再形成するあるAI企業は約5億の3Dモデルを収集し、業界で最も深いデータ堀を築きました。粗利率80%超、最大の市場シェアを誇るこれは誇大広告の話ではなく、自己強化型トークン経済が静かに不可欠なインフラを構築した物語です。AIコンピューティングがGPU独占からヘテロジニアス時代へ——Intel復活が示す新秩序Intelの株価は歴史的な上昇を見せているが、これは単なる好決算ではない。AINewsの分析によると、AIコンピューティングには地殻変動が起きている。単一アーキテクチャの支配時代は終わり、CPU、GPU、NPU、そして新興チップからなるヘテ中国のAIチップ四強:実用化から不可欠へ、ソフトウェアのギャップは依然として残る中国のトップ4社の国内AIチップ企業——総称して「四強」——は2025年の合計収益が初めて500億元を超え、技術的実現可能性の証明から市場シェア獲得への決定的なシフトを示しました。しかし、AINewsの分析によると、ハードウェアが進歩する一

常见问题

这次公司发布“AI Didn't Kill Search: Google's 81% Profit Surge Proves It's the Ultimate Growth Engine”主要讲了什么?

Google’s parent company Alphabet reported a stunning 81% year-over-year increase in net profit, driven largely by the integration of generative AI into its core search product. The…

从“how Google AI Overviews increase ad revenue”看,这家公司的这次发布为什么值得关注?

Google’s ability to turn AI into a search growth engine rests on a sophisticated technical architecture that blends retrieval, generation, and reinforcement learning. The core system is built around the Multitask Unified…

围绕“Project Mariner transaction fee model explained”,这次发布可能带来哪些后续影响?

后续通常要继续观察用户增长、产品渗透率、生态合作、竞品应对以及资本市场和开发者社区的反馈。