The AI industry is moving from a race to spend more into a competition to turn computing power into sustainable revenue. Microsoft monetizes AI through Azure, Copilot, and enterprise workflows; Amazon sells years of future computing capacity through AWS; while Apple controls the consumer gateway through devices, operating systems, and its ecosystem. The latest earnings do not suggest that AI investment is peaking—they show that AI business models are beginning to diverge.

**Price is only the result. Variables determine the stage. We do not predict prices; we observe cycles.**

Google’s $811 billion in future contractual obligations, Intel’s three-to-five-year CPU agreements, and multi-year memory supply deals in China reveal a major shift in the AI cycle: shortages are spreading beyond GPUs and HBM to CPUs, memory, storage, data centers, and power.

Competition is no longer only about buying more chips. AI companies are now trying to secure the delivery capacity of entire computing systems years in advance. These contracts can create supply advantages, but they also lock in costs, capital spending, and cash-flow risks.

Dazi Special Report 13 examines how AI is moving from a race for chips to a race for future supply—and which companies truly possess scarcity, demand visibility, and pricing power.

The latest earnings from Google, ServiceNow, and Tesla reveal a key contradiction in the AI cycle: enterprises are beginning to pay for AI agents, cloud revenue is growing, yet the infrastructure required to support that demand is becoming even larger.

AI adoption does not mean the industry is moving away from hardware. As AI shifts from training toward real workflows, it creates more demand for chips, memory, storage, networking, data centers, and power.

Dazi AI Watch 025 examines this new loop: application revenue drives more usage, rising usage creates capacity shortages, and those shortages lead to another round of hardware investment.

Price is only the outcome; variables determine the stage. We do not predict prices—we observe cycles.

Over the past two years, AI competition was defined by who could secure more GPUs, build more data centers, and expand compute capacity faster. But as Kalshi plots forward curves for GPU compute, CME prepares compute futures, and CoreWeave explores hedging memory-price exposure, compute is beginning to behave less like a scarce resource and more like a commodity whose future price can be discovered, locked in, and managed.

This marks an important shift in the AI cycle. The competition is no longer only about who can buy the most chips; it is increasingly about who can operate compute at a lower sustainable cost and turn massive capital spending into durable returns. From capacity expansion at TSMC and ASML to profitability tests facing cloud providers, a more mature market for AI compute is beginning to take shape.

**Price is only the result. Variables define the stage. We do not predict prices. We observe the cycle.**

Despite Samsung’s surging profits and ongoing supply shortages at Hynix, their share prices have dropped sharply; meanwhile, Micron and ABF substrate makers continue to be rewarded by the market due to new orders and supply-demand gaps.

Dazi 024 unpacks the misalignment between industry, earnings, and stock cycles: AI demand remains intact, yet the market has shifted from "buying into the boom" to "buying into new, unpriced expectations."

Price is merely the outcome; variables determine the stage.

Do not predict prices—simply observe the cycles.

The challenge for US AI is not a lack of computing power; rather, with massive investments in data centers, GPUs, HBM, power infrastructure, and debt, the industry must prove it can generate revenue and cash flow. Conversely, the challenge for Chinese AI is not a lack of demand; instead, constrained by limited supplies of high-end computing power, the industry must maximize the utility of every single chip, inference task, and industry application.

In this episode, we move beyond abstract comparisons of US and Chinese AI data centers to break down the economics: the US focuses on recouping investment, while China focuses on efficiency.

Price is merely the outcome; variables determine the stage of development. We do not predict prices—we observe cycles.

Meta’s reported plan to sell excess AI compute marks a new turning point in the AI cycle. The market is no longer asking only who can secure GPUs, HBM, power, and data center capacity. It is now asking whether these AI factories have real tenants, real workloads, and real monetization. Meta’s move is not just a cloud business story. It is the first visible sign that the AI infrastructure boom is shifting from capacity expansion to utilization verification.

Dazi AI Watch Special 10 focuses on one of the most rigid physical costs behind the AI boom: memory and storage. As AI moves from chatbots to agentic workflows, HBM, DRAM, NAND and enterprise SSDs are no longer just cyclical components. They are becoming critical memory resources for AI factories.

This article examines Micron’s strategic customer agreements, the HBM crowding-out effect, NAND/eSSD recovery, Apple’s reported CXMT lobbying, DRAM litigation risks, and the 2027–2028 supply overhang. The key question is not simply whether memory can keep rising, but where the next inflection point may appear.

Price is only the outcome; variables determine the stage.

We do not predict prices—we observe cycles.

This article breaks down the real cost structure behind one Agent task, including token usage, Codex-style multi-agent workflows, model routing, OpenAI capacity commitments, Micron memory contracts, human review, safety audits and rollback costs. The goal is to understand whether AI tasks can become cheaper than human labor, cheaper than traditional software, and valuable enough for enterprises to keep paying.

**Price is only the outcome; variables determine the stage.

We do not predict prices—we observe cycles.**

From Vera CPUs and Brazos cooling to compute contracts and model routers, decoding the real infrastructure bills behind the agent era.

### English Blog Description

**Dazi Special 09** follows up on the core thesis of Episode 021: AI compute is no longer just about GPUs.

GPUs remain the engine of AI factories, but agent-driven inference is creating demand for CPUs, power components, electrical equipment, data center retrofits, long-term compute capacity and model routing software. This article explains who is actually receiving orders beyond GPUs, which bottlenecks may have real pricing power, and which areas may become commoditized as supply expands.

**Price is only the outcome; variables determine the stage.

We do not predict prices—we observe cycles.**