-->

Big tech and frontier AI labs are buying up GPUs at a pace that requires issuing bonds. The combined capex estimate for four hyperscalers (Microsoft, Google, Meta, Amazon) in 2026 is roughly $725 billion, up 77% from the prior year. At this scale, it is natural to ask whether this qualifies as overinvestment. In an era when latecomers can close the performance gap through distillation at a fraction of the cost, does burning hundreds of billions of dollars for a model that is only marginally better for a few months make sense?

Conceptual image depicting a large-scale GPU data center gateway and an asymmetric scale

This article does not answer that question with “bubble” or “not a bubble.” Instead, it examines two structural logics driving big tech spending and considers what they mean for infrastructure providers and enterprise customers. The starting point was an analysis that circulated widely on X (@Tesla_Teslaway thread), and key figures were verified against primary sources.

Why It Looks Like Overinvestment

Distillation is the technique of collecting outputs from expensive frontier models and using them to train cheaper in-house models. For latecomers, it means replicating capabilities at lower cost that the frontier player already paid to develop. The logic holds: no matter how much money the leader pours in, the gap closes quickly. And it is true that open-source and latecomer models have rapidly narrowed benchmark gaps against top-tier models.

Read this far and overinvestment seems obvious. But if what the leader is buying is not “a few months of model advantage,” the math changes.

Asymmetric Insurance: What the Leaders Are Actually Buying

The real reason big tech buys GPUs is not a three-to-six-month performance lead. It is insurance: the right to be a player with direct leverage when a large capability jump occurs. The loss asymmetry between the two scenarios is extreme.

If a jump happens and you are not there, your trillion-dollar core business (search, cloud, office) can unravel overnight. Think Google becoming Yahoo. If the jump never happens and you turn out to have overinvested, your core business survives intact and the GPUs and data centers you bought do not go to zero. One tail is “the reason your company exists disappears.” The other tail is “depreciation losses.” When the loss magnitudes are that asymmetric, the rational choice for a company operating under uncertainty tilts toward overinvestment. This is not a bubble. It is a rational response to an asymmetric payoff structure.

“The Jump” Is Reliability x Task Horizon, Not a Smarter Chatbot

The key question then is what that jump actually means. It is not a chatbot that scores a few more points on a benchmark. It is the product of reliability and task horizon: the ability to carry multi-step work through to completion without collapsing, even without human intervention.

METR measured this task horizon directly. The original thread cited Anthropic, but the accurate source is METR’s “Measuring AI Ability to Complete Long Tasks” study. METR reported that the length of tasks an AI can complete at 50% reliability, measured in human-equivalent time, has doubled roughly every seven months from 2019 to 2025. More striking, from 2024 to 2025 that doubling period compressed to roughly four months. The trend is accelerating.

On the reliability side, simple arithmetic makes the threshold visible. An agent with 95% per-step reliability completing a 20-step task succeeds with probability 0.95^20, roughly 36%. A human must verify every step, so there is no labor saving. The same task at 99% reliability yields about 82% success; at 99.9% it reaches roughly 98%. Reliability climbs linearly, but the moment it crosses the threshold where you can remove humans from the loop, economic value jumps discontinuously. That jump is what big tech is betting on.

The Big Four and Pure Labs Have Different Spending Motives

Even when buying the same GPUs, the motivations differ in kind. For Microsoft, Google, Meta, and Amazon, GPU capex is comparatively cheap insurance protecting trillion-dollar core businesses. Against the risk of missing a capability jump, the capex is a manageable premium. For pure AI labs like OpenAI and Anthropic, GPUs are the core business. There is no separate business to fall back on, so the spending is not insurance but survival. The same numbers carry different meaning.

One more note: more than 60% of this capex goes not to chips but to power and data center construction. Numbers that look like “GPU shopping” are actually closer to bets on power infrastructure, which further complicates a simple bubble verdict.

The Next Tollgate: Intent Router

As important as the spending logic is the question of why that position must be held at any cost. Every era has had a tollgate controlling the critical chokepoint. Windows in the PC era. Google Search in the internet era. App stores in the mobile era. In the AI agent era, the tollgate is likely to be the agent that receives user intent and routes it to the appropriate services: the intent router.

Picture it concretely. A user tells an agent: “Set up dinner plans for tonight and make the reservation.” The agent decides which restaurants to surface as candidates, which booking platform to route through, which delivery service to call. At that moment, restaurants and platforms are no longer exposed directly to the user. If you are not in the agent’s candidate list, you effectively do not exist. The structure shifts from “fall off page one of search results and lose traffic” to “fall out of agent recommendations and lose transactions.” Whoever holds the chokepoint sets the toll.

There is something to be honest about here, though. If agents become tollgates, infrastructure providers are not entirely insulated from that dynamic either. And the premise that “enterprises will resist dependency on external agents” is closer to a trend forming than a settled demand. This article does not claim otherwise. What we observe are signals: data sovereignty, regulatory compliance, and sovereign AI requirements converging to make that trend increasingly legible.

The ThakiCloud Angle

Why this picture matters for ThakiCloud is not because we need to track big tech trends. It is because we sit in the middle layer of that tollgate competition.

The harder big tech competes to control the intent-router layer, the more valuable the alternatives become for enterprises that do not want to hand their data and models to external agents. What those enterprises need is an execution environment for running their own agent infrastructure on-prem or in a private cloud. ThakiCloud’s Kubernetes-based AI/ML workload infrastructure and Kueue-based GPU workload scheduling sit exactly in that space. On the path from GPU cloud MSP to enterprise AI adoption partner, we are targeting this “tollgate internalization” demand.

The task horizon threshold logic also connects directly to product strategy. If an agent’s economic value jumps discontinuously the moment per-step reliability crosses the threshold where humans can leave the loop, then how reliably and stably we carry a customer’s AI workloads is not just an operational quality metric. It is the variable that determines whether the customer crosses the threshold that lets them remove humans from the verification loop. Infrastructure stability is a nonlinear lever on customer ROI. That is why we are obsessive about reliability, isolation, and scheduling quality.

When This Logic Breaks Down

For balance, here are the counter-scenarios. The asymmetric insurance logic can unravel in clear ways.

First, the reliability curve could plateau before it reaches the threshold. If task horizon keeps extending but per-step reliability stalls before 99.9%, the economic value of long autonomous tasks never clears the step. Second, distillation and open-weight models could become good enough that owning the frontier directly stops mattering. The insurance the leaders bought would depreciate. Third, power, land, and grid constraints could prevent capex from translating into actual operational capacity. Spending the money without being able to draw the power keeps the GPUs idle. If any one of these materializes, “rational insurance” becomes “expensive misjudgment.”

The point is not to pick a side. It is to know which indicators separate the outcomes. Does reliability cross the threshold? Do open-weight models substitute for the frontier? Does power keep pace with capex? These three are the metrics to watch over the coming quarters.

Summary

Big tech’s GPU overinvestment may be a bubble or it may be rational insurance. Reading it as “a rational response to an asymmetric payoff structure, combined with a race to capture the next tollgate” reveals something far more structured than mere mania. And on the other side of that race, enterprise demand to avoid tollgate dependency is forming. ThakiCloud is the infrastructure built for that side.

Sources

참고

빅테크 GPU 과투자의 진짜 논리: 비대칭 보험과 다음 세대 톨게이트

 

2026년 하이퍼스케일러 캐펙스가 약 7,250억 달러로 전년 대비 77% 늘었습니다. 거품처럼 보이는 이 지출을 비대칭 보험과 의도 라우터 톨게이트라는 두 구조적 논리로 읽고, METR의 태스크 호라이즌 데이터와 신뢰도 임계선 수학으로 검증한 뒤, 톨게이트에 종속되기 싫은 엔터프...