AIバブルの階層構造
The structure of AI Bubble
M7ハイパースケーラーのROI不足問題は、膨大なバックログにより存在しないかに見えます。ただ、クラウド貸しビジネスを主とするM7ハイパースケーラーは、実のところ一段上のツルハシ屋で、AIビジネスの真の起点ではありません。実際のビジネスの起点はこれらクラウドのお客で、彼らのビジネスが成功しないと膨大なバックログは蜃気楼となり、AIバブルは崩壊します。それらクラウドの客の何割がキラープロダクトで今後成功し、収益をあげられるのでしょうか?今の市場は、一段上のツルハシ屋に目を奪われ、AIビジネスの本質的な成功の可能性は見えない振りをしているようです。
- 本稼働の割合: 2025年の調査では、企業のAIユースケースのうち実際に本稼働(プロダクション)に到達したのは約31%に留まっています。
- ROIの達成状況: 期待したROIを達成できている取り組みは、成長性の面で4件に1件(25%)、効率性の面ではわずか7%という厳しいデータもあります。
- 投資の偏り: 企業のAI予算の約3分の2(66%)がいまだにインフラ(データ・ストレージ・処理能力)に割かれており、ソフトウェアやアプリケーションへの投資は33%に過ぎません。これは、依然として「ツルハシの購入」が先行している状態を裏付けています。 [1, 2]
- コーディング・開発支援: エンジニアの生産性が55%以上向上した例もあり、OpenAI(1.4.1)のAPIを利用する企業が急速に増加するなど、「最初のキラーユースケース」として定着しています。
- カスタマーサポート: 2026年の予測では、AIエージェントによる自動対応が最も速くROIを生む分野とされており、Google Cloud(1.3.5)の早期導入者の74%が1年以内のROI達成を報告しています。
- 収益貢献: AI導入企業の56%がビジネス成長を実感しており、そのうち71%が6〜10%程度の売上増を報告し始めています。 [1, 2, 3, 4, 5, 6]
- 蜃気楼のリスク: スタートアップへのVC投資の6割以上(2,580億ドル以上)がAIに集中していますが、HubSpot(1.5.7)のデータによるとAI系スタートアップの黒字化率は約61%で、非AI系(54%)を上回っているものの、まだ盤石とは言えません。
- 本質的な課題: 技術自体の不足よりも、企業のガバナンス欠如、データの質の低さ、スキルの不足が「成功を阻む壁」となっています。 [1, 2, 3, 4]
当面のツルハシ屋のROIと言う意味では1段目も2段目も満足していると言える。ただ、2段目のツルハシ屋も最終需要の長期ガイダンスを出す能力は持ち合わせていない。知っていても言わないと思うが、多分宗教的に信じているところもあるかもしれない。
- 投資の論理: 彼らにとっての最大のリスクは「投資のしすぎ」ではなく、「投資を控えて競合に覇権を奪われること」です。この恐怖心が、最終需要の不透明さを無視してでも巨額投資を正立させてしまう宗教的土壌になっています。
- 在庫の積み増し: 供給不足を懸念した「仮需」や、自社クラウドの優位性を保つための「抱え込み」がバックログを膨らませており、最終需要との乖離(ブルウィップ効果)が発生しているリスクがあります。
- ガイダンスの限界: 二段目のツルハシ屋(NVIDIA等)は、ハイパースケーラーの注文書があれば決算をまとめられますが、その先の一般企業が「AIでいくら稼げたか」までは捕捉できていません。
- 現在はまだインフラ構築フェーズであり、アプリケーション層での「キラープロダクト」による収益化の成否が判明するのは、物理的なGPUがデータセンターに設置され、稼働し始める1〜2年後になります。
- このタイムラグの間は、ツルハシの売買だけで経済が回るため、バブルの頂点まで「蜃気楼」は消えない構造になっています。
市場は1段目、2段目のツルハシ屋の好調で満足し、最終顧客がどうなっているか肝心かなめのAIビジネスの成否は無視しているようだ。全体を理解しないで投資している人がいかに多いかを示している。
The ROI shortfall problem for M7 hyperscalers appears not to exist, thanks to their massive backlog. However, hyperscalers—whose core business is cloud leasing—are, in reality, just a higher-tier “pick-and-shovel” layer and not the true starting point of the AI business. The real starting point lies with their customers. Unless those customers succeed, the enormous backlog will prove to be a mirage, and the AI bubble will collapse. What percentage of these cloud customers will actually create killer products, succeed, and generate sustainable profits going forward? The current market seems captivated by this higher-tier “pick-and-shovel” layer, while pretending not to see the fundamental question of whether AI businesses themselves can truly succeed.
Hyperscalers such as the “Magnificent Seven (M7)” are, in effect, a higher-tier version of pick-and-shovel providers in a “digital gold rush.” Whether their backlog translates into real profits ultimately depends on whether their customers—enterprises and startups buying those “tools”—can produce practical killer products and generate sustainable ROI.
Based on market data and projections from 2025 to 2026, the “success rate” and reality look as follows:
1. The gap between AI product “success rates” and reality
There is a mix of optimism and pessimism regarding how many cloud customers have transitioned AI from experimentation to monetization.
- Production deployment rate: As of 2025, only about 31% of enterprise AI use cases have reached full production.
- ROI achievement: Only about 25% of initiatives achieve expected ROI in terms of growth, and just 7% in efficiency.
- Investment imbalance: Around 66% of AI budgets are still spent on infrastructure (data, storage, compute), while only 33% goes to software and applications—evidence that “buying the tools” still comes first.
2. Emerging areas where “killer products” are taking shape
Not all customers are failing. In certain domains, clear winning patterns are emerging, and ROI is improving early:
- Coding and development support: Some cases show productivity gains of over 55%, and companies using OpenAI APIs are rapidly increasing—making this one of the first established killer use cases.
- Customer support: AI agents are expected to generate the fastest ROI by 2026, with 74% of early adopters reporting ROI within one year.
- Revenue impact: 56% of AI-adopting companies report business growth, with 71% of them seeing revenue increases of around 6–10%.
3. Bubble collapse or just the “trough of disillusionment”?
According to Gartner’s 2025 Hype Cycle, generative AI has passed the “peak of inflated expectations” and is heading toward the “trough of disillusionment” (the monetization barrier).
- Mirage risk: Over 60% of VC funding (more than $258 billion) is concentrated in AI startups. While about 61% of AI startups are profitable (higher than 54% for non-AI), this is far from robust.
- Core challenges: The biggest barriers are not the technology itself, but poor governance, low-quality data, and lack of skills.
Conclusion
To prevent the backlog from becoming a mirage, investment must shift away from infrastructure-heavy spending toward AI agents and industry-specific applications deeply embedded in business workflows. Over the next 1–2 years, whether the success rate of these “cloud customers” can exceed the current ~30% threshold will determine the true sustainability of the AI economy.
In terms of ROI for the “pick-and-shovel” providers, both the first and second tiers are currently satisfied. However, even second-tier providers lack the ability to give long-term guidance on end-user demand. Even if they know, they likely won’t say it—though in some cases, there may even be an element of quasi-religious belief.
As you pointed out, the current market is sustained by a self-reinforcing cycle in which “first-tier pick-and-shovel providers” (hyperscalers) keep ordering from “second-tier providers” (such as NVIDIA and TSMC), allowing apparent ROI to hold up on the surface.
The unease of this situation—and the reason they don’t (or can’t) provide forward guidance—can be understood from three perspectives:
1. A “religious” belief in compute scaling
Statements by hyperscaler leaders (e.g., Satya Nadella, Sundar Pichai) often reflect a near-religious faith in scaling laws: the belief that if enough compute is投入ed, intelligence will continue to emerge and solutions will follow.
- Investment logic: Their biggest risk is not over-investing, but under-investing and losing dominance to competitors. This fear justifies massive spending despite uncertain end demand.
2. Demand obfuscation by higher-tier suppliers
The strong backlog presented by hyperscalers may not reflect true end-user ROI.
- Inventory build-up: “Speculative demand” driven by supply concerns and strategic hoarding may be inflating backlog, creating a disconnect from real demand (a bullwhip effect).
- Limits of visibility: Second-tier suppliers like NVIDIA can rely on hyperscaler orders for earnings, but they cannot see how much end customers are actually earning from AI.
3. The time lag created by “looking the other way”
The market may be deliberately ignoring the fundamental question because it’s “too early to tell.”
- We are still in the infrastructure build-out phase.
- The success or failure of application-layer monetization will only become clear 1–2 years after GPUs are deployed and operational.
- During this lag, the economy can function purely on tool sales, allowing the “mirage” to persist.
As you noted, it’s not that they don’t understand the uncertainty of end demand—it’s that they have no choice but to keep believing and pushing forward, creating something akin to a high-stakes game of chicken.
If this mirage begins to fade (e.g., through backlog cancellations), the first signal will likely be a slowdown in hyperscaler CapEx growth—but so far, that brake has not yet been applied.
The market seems satisfied with the strong performance of both first- and second-tier “pick-and-shovel” providers, while ignoring the critical question of whether end customers are actually succeeding. This reflects how many investors are operating without understanding the full picture.
Exactly as you said, the market is currently fixated on visible secondary indicators—massive backlogs and flashy earnings—while ignoring the primary reality: the ROI of end customers.
From a broader perspective, three key blind spots emerge:
1. Confusing investment with consumption
Investors interpret hyperscaler CapEx as growth, but it is fundamentally expenditure. It must ultimately be justified by end-customer spending—something that has not yet materialized.
2. Circular capital flows between lenders and borrowers
Some AI startups receive funding from hyperscalers and then spend that capital on the same hyperscalers’ cloud services. This creates a loop that may not generate real external economic value.
3. “Productivity gains” as an escape narrative
While AI-driven efficiency gains are real, very few companies can quantify how much this translates into profits or new markets. The market is pricing in a narrative as if it were already proven reality.
As you pointed out, many investors appear to be repeating a classic bubble-era mistake: assuming that because the pick-and-shovel sellers are thriving, there must be gold in the mine.
When the moment of truth arrives—when the market collectively asks, “Are the customers actually making money?”—the adjustment could be significant.
