What truly profits in AI cloud computing isn’t a single company, but the entire “computing power toll road.” AI cloud is becoming a critical intermediary layer: connecting upstream to GPUs, HBM, power, and data centers, and downstream to tech giants, AI labs, and enterprise customers who actually need computing power. What truly matters isn’t just “who is building AI,” but who keeps collecting fees at every layer of the AI cloud stack. Layer 1: Who provides the most scarce foundational resources? $NVDA GPUs remain at the center of the entire computing stack. The amount of AI computing power an AI cloud can deploy depends first on how many high-end GPUs it can secure—and when those GPUs can actually be deployed. $MU $SKHY HBM is one of the most critical resources alongside GPUs. As models grow larger and inference concurrency increases, GPUs require higher-bandwidth memory to continuously feed data—making HBM a key bottleneck in AI server expansion. $DLR $EQIX $CORZ $APLD These companies provide data centers, server rooms, power access, and physical infrastructure. AI clouds don’t need to own every building—just the ability to rapidly acquire pre-powered, GPU-ready data center capacity to continue scaling. Layer 2: Who turns this hardware into truly rentable AI computing power? $CRWV $NBIS $IREN This layer sits at the very heart of the AI cloud stack. What they truly do isn’t simply “buying GPUs,” but combining: GPU + Power + High-speed Networking + Storage + Software Orchestration into customer-ready AI computing platforms. This is Neocloud’s core business model: procuring scarce infrastructure upstream and selling computing power downstream via long-term contracts or on-demand usage. Layer 3: Who actually buys this computing power? $MSFT $META $GOOGL These three companies are unique because they are both customers and competitors of AI clouds. When their internal data centers and GPU supply can’t keep up with demand, they purchase capacity from external Neocloud providers—while simultaneously building their own data centers, developing proprietary chips, and launching their own cloud platforms. Thus, the future AI computing market won’t be a binary choice between “build vs. buy”—it will more likely be: Core capacity built in-house + Elastic capacity sourced externally. The next phase to watch is the expansion of enterprise customers. $SHOP $CRWD As AI applications gradually move beyond a few hyperscalers and top-tier labs into e-commerce, cybersecurity, finance, software, healthcare, and enterprise automation, Neocloud’s customer base has the potential to expand from a handful of super-clients to a broad enterprise market. OpenAI and Anthropic represent another category of high-intensity customers. Their demand characteristics are: Training requires massive compute; after model deployment, continuous inference is still needed. The stronger the model and the wider the use of agents, the more compute demand doesn’t stop after training—it shifts from one-time training expenditure to long-term, ongoing inference costs. Thus, the entire AI cloud stack can be simply understood as: Upstream: Selling resources $NVDA $MU $SKHY Middle layer: Selling computing power $CRWV $NBIS $IREN Foundation layer: Providing data centers $DLR $EQIX $CORZ $APLD Downstream: Buying computing power $MSFT $META $GOOGL OpenAI Anthropic Previously, the market focused most on: Who owns the most GPUs? The more important question in the next phase may be: Who can most rapidly combine GPUs, power, networking, and data centers into immediately deliverable, immediately billable AI computing power? AI cloud is becoming the “toll booth” of the entire AI value chain. As training and inference demand continue to grow, upstream chip sellers, middle-layer compute providers, and foundational data center operators will all continue to profit from the same wave of AI infrastructure expansion.