The most important change in AI infrastructure in July is not that the hardware logic has failed.

Instead, the focus of research is shifting from "who lacks it most and who can rise the most" to "who will ultimately pay and whether cloud providers can recover their investments."

Step one: Hardware can no longer be treated as a monolith.

Hynix, Samsung, Micron, TSMC, along with capacity expansion in South Korea and the listing of ChangXin, remind us that AI training and inference will not reward all hardware equally.

HBM, high-value DRAM, enterprise SSDs, advanced manufacturing, and packaging are at different positions.

Currently, we need to look at bottlenecks, delivery, and customer certification; later, we must also see if new supply will dilute scarcity.

Step two: "Tenants" for cloud providers are beginning to appear.

Microsoft 365 Copilot paid seats have exceeded 30 million; approximately 90% of Microsoft Cloud's revenue comes from customers outside of frontier model companies.

AWS Q2 revenue was $42.2 billion, up 37% year-over-year, higher than the previous quarter's 28%.

At the current pace annualized, AWS's AI business and chip business each generate over $25 billion in annual revenue.

But cash recovery is still under construction.

Amazon's free cash flow remained negative over the past 12 months; management expects cash Capex to be around $220 billion in 2026, which includes not only AI but also investments in robotics, chips, satellites, etc.

Step three: Begin observing AI Work.

As models become cheaper and easier to deploy, whether value will shift to workflow entry points is a new issue worth tracking in the second half of the year.

Product launch does not equal revenue; what truly matters is whether usage leads to renewals, volume growth, and ROI.

In August, I am watching three things:

Macro: Will employment, inflation, PCE, and Jackson Hole change September interest rate expectations?

Hardware: Can orders, deliveries, gross margins, and next-quarter guidance in AMD, Supermicro, and NVIDIA earnings reports continue to prove demand diffusion?

Workflows: Who secures enterprise entry points, data, and permissions, and who can turn "helping people work" into recurring revenue.

This month's judgment: The protagonists of AI infrastructure have not changed, but the research focus is extending from supply bottlenecks to payers, workflow entry points, and cash recovery.

If hardware earnings simultaneously show weakening orders, deliveries, and next-quarter guidance, or if the next round of cloud provider earnings shows that enterprise payments and cloud demand are not continuing and operating cash flow continues to lag behind investment, I will tighten this judgment.

Which do you think will provide answers first in August: macro, hardware, or AI Work? 🤔

$SK Hynix(SKHY.US)$Microsoft(MSFT.US)$Amazon(AMZN.US)$Meta Platforms(META.US)$Alphabet - C(GOOG.US)$NVIDIA(NVDA.US)