The logic of AI investment has changed: cloud providers have entered the rental era, and the infrastructure supply chain is under pressure.

In the same earnings season, Microsoft's market value surged by $450 billion in a single day, setting a record for the largest single-day market value increase in global stock market history, while Amazon entered the $3 trillion club within five days. On the other hand, SK Hynix's stock price was cut in half, Kioxia's stock price halved as well, and most global storage stocks fell by over 40%. Both are engaged in AI. One half is soaring, and the other half is plunging. This is not random fluctuation. The market is using real money to answer a core question: Who is making money and who is footing the bill in this AI feast? Over the past two years, the market rewarded "who spends the most"when CSPs (Cloud Service Providers) raised capital expenditures, GPUs, HBM, optical modules, switches, PCBs, power supplies, liquid cooling, and data centers all experienced a surge. Now, the market is starting to ask a stricter question: The trillion-dollar investments made by the tech giants, where have they landed in terms of cash flow, and who has accumulated depreciation from them? The answer may be surfacingAI money hasnt disappeared, its just changed pockets. Its flowing from those selling shovels to those turning shovels into businesses. From those building roads to those collecting rents on them. Three signals are flashing simultaneously, prompting the market to reevaluate. To understand this round of divergence, one cannot just look at the ups and downs in numbers. During this time window at the end of July, three things happened simultaneously. The first thing: CSPs proved they can turn computing power into revenue. Microsoft Azures year-on-year growth rate was 43%. Amazon AWS recorded a year-on-year increase of 37%, the fastest in 18 quarters. Google Cloud boasted an 82% year-on-year growth, the highest in three years. Together, the three generated an annualized cloud revenue of $389 billion, with a single quarter's increase in recurring revenue of $50 billiondouble the average of the past three quarters. More importantly, revenue is accelerating, but its not just a few AI labs buying computing power. Microsoft disclosed a key number: the commercial remaining performance obligations reached $678 billion, an 84% year-on-year increase, and all the quarter-on-quarter growth came from non-frontier lab customers. Banks, manufacturing, healthcare, and governmentthose traditional enterprises that the market previously deemed unrelated to AI are now signing contracts. AWS isn't just saying itit's doing it. On August 4, AWS head Matt Garman publicly confirmed that the company is signing five-year commitment contracts with clients, stating that market demand still far exceeds supply, and we are working hard to accelerate construction and investment to keep up with customer needs. This is not just a demand story; its a story of guaranteed orders. A five-year contract means that clients are not just testing AI but embedding it into their core business processesthe stickiness of this revenue is the most potent antidote to any capex anxiety. The second thing: the narrative model of the infrastructure chain has broken down. SK Hynix reported the strongest revenue and profit in its historyoperating profit surged by 557%. Yet, its stock price plummeted. Why? Because the markets expectation was not merely the strongest in history but strong enough to support the dozens of points in expectations built over the past two months. When actual growth doesnt meet long-term expectations, when the year-on-year growth rate of CSP capex slides from double-digit acceleration to high single digits, the valuation model must changefrom how much can you sell to how fast can you grow. This is a physical law, not a judgment. CSP capex cannot sustain year-on-year growth of 30% or 50% indefinitely. As the industry transitions from the accelerated expansion phase to the high growth phase, the second derivative of growth becomes negative, and valuations need to be repriced. This explains why Micron, Kioxia, and SK Hynix experienced the most severe selling despite reporting the strongest earnings. The third thing: the pricing power of models is collapsing. During the busiest earnings report week, OpenAI cut the price of its new model GPT-5.6 Luna by 80%. From $1 per million tokens, it dropped to $0.20. This was a price reduction forced by the emergence of DeepSeek. Over the past year, the proportion of Token calls for Chinese open-source models on the global API platform OpenRouter surged from 4.5% to 46%. DeepSeek V4 Flash is priced at $0.14, while Alibaba's Qwen3.7 Flash is down to $0.03. Putting these three things into the same temporal context CSPs can earn money + Infrastructure growth is peaking + Model pricing power is collapsing = Profits are flowing from hardware and model layers back to infrastructure and platform layers en masse. What gives CSPs the ability to charge rent? The ROI ledger of cloud providers: heavy assets low returns. To understand why CSPs suddenly became market darlings, its not just about feeling; its about doing the math. The first calculation: profit calculationCSPs' ROI ledger, heavy assets low returns. For the past two years, the market has presumed that cloud providers are the scapegoatsthe money is all going to Nvidia and Hynix. However, Morgan Stanleys calculations from late July overturned this judgment: the benchmark ROIC (Return on Invested Capital) for pure GPU leasing is about 31%. Under the assumption of a 1GW data center, 410,000 GB300 chips, and a 75% utilization rate, if leasing prices rise from $7 to $10 per hour, ROIC can jump from 23% to 39%. Moreover, dont forget that the main components of data centersland, power, and facilitieshave a lifespan exceeding 30 years, capable of spanning five to six generations of servers. Each generation of servers can be updated without restarting infrastructure. This means that CSPs that entered the market first will have better economic margins in subsequent generations, not worse. The market has finally recognized this. Morgan Stanley analyzed three types of AI business models from a bottom-up ROIC perspective: Pure GPU leasing (IaaS): The benchmark ROIC is around 31%. In a scenario where GPUs are in high demand and utilization is 75%, if leasing prices rise from $7 to $10 per hour, ROIC can climb from 23% to 39%. Model APIs on proprietary infrastructure: The benchmark ROIC is around 46%. The premise is that model companies maintain product differentiation and pricing powerbut considering the current price wars, that premise is wavering. Using third-party computing power for model APIs: ROIC is only 25%. This requires additional payment of middle-layer profit to cloud providers. The core conclusion of these calculations directly challenges a preconceived judgment that heavy assets low returns. The second calculation: revenue calculationdemand has spread from labs to enterprises. Over the past two years, the biggest fear the market had about AI capex was "circular financing": CSPs pouring money into model companies like OpenAI and Anthropic, and those companies buying back computing power from CSPs. Wheres the real end demand? This fear is being refuted by this quarters data. Microsofts commercial remaining performance obligations showed that all quarter-on-quarter growth came from non-lab clients. AWS's backlog has reached $496 billion, with triple-digit growth. Google Clouds backlog is $514 billion, with a single quarter increase of $50 billion. Nearly 90% of the Fortune 100 companies are already using Gemini Enterprise. Wall Street Insights previously used a vivid analogy: CSPs were previously building for the frontier labs, the only tenant; now the whole office building is filling up. This isnt just an isolated signal from one company. Three companies are showing the same trendAI demand is spreading from a few startup AI labs to traditional industry enterprise clients. As the customer structure shifts from concentrated to broad, the narrative of circular financing loses its foundation. The third calculation: model calculationfour companies, four rental methods. CSPs are not a unified concept. The interesting part of this round of earnings reports is that four giants have formed four completely different AI return paths. Microsoftthree-tier rental model. Azure provides the most basic GPU leasing and cloud services, which is the first tier. The Foundry platform offers model calls, fine-tuning, and deployment, which forms the second tier. Copilot directly embeds into Office 365 and GitHub, charging by seat (with over 30 million paid seats), making it the third tier. The three tiers channel into each other: Azures enterprise clients can upgrade to Foundry for model usage; models on Foundry can penetrate into Copilot applications; and Copilot, in turn, drives consumption of Azure. This is not a single revenue line, but an income matrix. Amazonthe simplest, most transparent rental model. AWS itself is a business of renting computing power and storage, running for 18 years with a mature customer, billing, and retention system. This quarters revenue reached $42.2 billion, up 37% year-on-year, marking the fastest growth in 18 quarters. Googlethe most impressive numbers, the most ambiguous books. Google Clouds 82% growth rate and the improvement in profit margins appear to be the strongest among the four. However, the markets feedback was a pattern of falling then rising. The reason lies in the gray area of cost attribution. Gemini serves search, advertising, Workspace, YouTube, consumer applications, and cloud services, sharing the AI research costs at the group level. The segment profit margin of Google Cloud cannot be equated to the real economic profit margin of Googles chip-model-cloud stack. Some research institutions pointed out that if all Gemini expenses were factored in, Google Cloud's profit margin would drop to just over 10%. Google isnt unprofitable; the path to profitability is just too complex. Metashifting from "pure spending" to "starting to think about renting." Over the past two years, Meta has been seen as the most awkward player among the four in the eyes of investors: AI indeed improved ad recommendations, content rankings, and material generation, but these gains were hidden in the existing advertising business, making it impossible to match directly with new capEx. When capEx guidance shifted from $130 billion to $145 billion, free cash flow fell to just $784 million (down 91% year-on-year), and the markets patience reached its limit. However, Meta is undergoing a subtle change. The company is reported to be planning a cloud business, considering providing model access services for developers and offering surplus AI computing power for rent. This indicates that Meta's computing power assets are no longer a binary choice of self-use or idle. They can shift to external rental. Price war over models: The biggest boost for landlords. The most easily overlooked trigger variable in this round of divergence is not the earnings report but the price collapse at the model layer. Back to that day on July 30. Besides Microsoft's earnings report, OpenAI did another thing: slashing the price of GPT-5.6 Luna to one-fifth of its original price. This wasnt a price reduction after a year of release, but just three weeks post-launch. This is driven by a larger trendopen-source models are dismantling the toll booths of closed-source models. Over the past two years, the profit distribution logic of the AI industry chain has been: Nvidia earns money from GPUs, CSPs earn hard-earned money by renting computing power, and model companies (OpenAI, Anthropic) earn premiums via technological barriers. The model layer is the toll booth built on the cloudregardless of whose computing power you use or what applications you run, accessing the strongest model requires passing through it. Now that toll booth is being dismantled. The breakthroughs of Chinese open-source models are the greatest wall-demolishing hammers. DeepSeek V4 trained with Huaweis Ascend 950 chips boasts 1.6 trillion parameters, performing comparably to GPT-5. Alibaba's Qwen3.7 Flash is priced as low as $0.03 per million tokens on the largest model API aggregation platform OpenRouter, where the usage of Chinese models surged from 4.5% to 46% within a year. When 46% of the usage shifts to inexpensive open-source models, closed-source models can no longer charge one-of-a-kind premiums. OpenAI is forced to follow suit with price cuts not because it wants to, but because it has no other choice. What does this mean for CSPs? Barclays analyst Raimo Lenschow wrote an astute assessment in his August 3 report: When models transition from scarce goods to commodities, bargaining power shifts back to platforms that own computing power, customers, and distribution channels. CSPs no longer need to be tied to any particular model company. They can simultaneously offer GPU leasing + self-hosted open-source models + self-developed model APIs. Clients dont need to select only one model; CSPs can act as routersusing top closed-source models for high-complexity tasks, inexpensive open-source models for routine tasks, and deploying self-developed ASIC chips for high throughput, low-cost steady-state loads. Who benefits most from this multi-model architecture? Its not the companies that are best at training models but those that excel at linking different models together, allocating computing power as needed, and charging based on usage. Thats the CSPs. Morgan Stanleys July 27 report even broke this logic down into mathematics: If model API prices drop by 50%, but Token call usage increases fivefold as a result, its a net gain for CSPs self-built infrastructure reasoning revenue. Their cost side is following the performance of chips and the efficiency curve of software optimization, while their revenue side is propelled by an explosive demand usage curve. As long as the second curve runs faster than the first, its a good business. But for independent model companies, a 50% drop in prices directly slashes revenues in halfunless their Token usage can double, but the substitution effect from open-source models makes that unfeasible. Who is footing the bill: the cruel mean reversion of the infrastructure chain. If CSP stocks rising are a rental story being priced in, then the drastic drop in the infrastructure chain is a case of the growth myth being exposed. Morgan Stanley strategist Michael Wilson used an accurate analogy: semiconductor stocks, especially storage, behave very similarly to silver. There are two lines of logic. First, both have experienced parabolic risesemotions and funds pushed prices to levels unsustainable by fundamentals. Second, both possess commodity attributesstorage chips are among the categories in the AI hardware complex most like commodities, exhibiting high price elasticity and quick reversals. This round of adjustment led by storage stocks, with SK Hynixs stock cut in half from its peak, Kioxia halving, and Micron dropping over 30%, and of the 14 global storage concept stocks, 10 have fallen over 40% from their highs. This aligns perfectly with Wilson's predictions made in early July. However, storage is merely a leading sample. The deeper issue is that all companies valuated based on linear extrapolations of total CSP capex now face the same questionif the year-on-year growth rate of capex drops from 30% to 15%, can your valuation still hold? This question can't be answered by claiming AI demand remains strong. Demand is still there, but growth rates are another matter. For different segments of this chain, the degree of injury varies. It can be understood in three layers: Layer one: Direct pressurestorage and commoditized hardware. The most typical victims. Price cycles peaking + capex growth slowing + momentum-driven selling under silver attributes. After the drop, prices may be cheap, but cyclical stocks often have cheaper following drops. Layer two: Valuation adjustmentoptical communication, network equipment, and other capex beta areas. Although absolute capex continues to grow, slowing growth implies that these companies' valuation multiples need to be adjusted from high growth to stable growth. Stock prices may not necessarily plummet further, but returning to previous highs becomes significantly more challenging. Layer three: Relatively resilientsegments with structural upgrade logic. Advancing from 800G to 1.6T optical modules, from air cooling to liquid cooling, and from traditional power to higher power densitythese from 0 to 1 technical upgrades can still gain alpha through increasing market share and unit value even if capex total growth slows down. But the premise is that the upgrade logic is robust enough and that valuations havent been overly speculated. This is also why Nvidia has been relatively resilient (with about a 10% drop) during this downturnstructural demand remains in place; the market is simply no longer providing the linear pricing of the more CSP invests, the more I rise seen in the last six months. On August 26, Nvidia will announce its Q2 earnings. This is the next core verification window for the entire AI industry chain. If results and guidance exceed expectations, the infrastructure chain may catch a breath; if not, there will still be significant downward space. Looking ahead: High returns are not given freely; they require continuous verification. This is not the end of the AI cycle. At least not yet. Demand at the industry level is still surging. Microsoft says Azure will continue to accelerate in the second half, AWS claims that most of its capacity is booked through 2027, and Google says 90% of the Fortune 100 are on Gemini Enterprise. The combined annualized cloud revenue of the three CSPs is $389 billion, and all are accelerating. However, the market has changed its approach to problem-solving. From last two years' multiple-choice questionsCSPs raising capex identifying all beneficiaries buy, buy, buy; it has turned into proof questionscan your revenue cover depreciation? Is your utilization rate high enough? How diversified is your customer structure? Who bears the cost of your models? The requirements of these proof questions vary greatly depending on where they sit in the industry chain. For CSPs, proof questions are relatively friendly. Their business essentially revolves around transforming computing power into recurring revenueonce clients sign long-term contracts, run production-level loads, and integrate them into their core business processes, the stickiness of this revenue is very high. As long as enterprise-level AI adoption continues to expand, CSPs' rental logic can persist. However, differentiation is also occurring among CSPs. Currently, Microsoft and AWS are the two companies faring best in this proof questionthey both have multi-layer revenue streams from Copilot and Foundry, while AWS has the most transparent rental model and the most mature customer base. Google still needs to address the transparency of cost attribution. Meta needs to transform computing power leasing from a news headline into reportable revenue. For the infrastructure chain, proof questions are much more challenging. Over the past two years, companies in this chain have enjoyed a kind of "double dividend": CSPs continually raising capex total beta + investors linear extrapolation valuation expansion. Now that CSP capex may be close to or nearing peaking in growth, total beta is contracting, and valuations need to return to tracking only what can be earned. This isnt the fault of any single company; its a switch in the valuation model. But that doesnt mean there arent opportunities along the entire chain. The key depends on three factors: First, whether product iterations can outpace the cycle. From 800G to 1.6T optical modules, from air cooling to liquid cooling, from HBM3E to HBM4companies that benefit from technological upgrades and increasing unit value can still gain share and profit even when capex growth slows. This is alpha, not beta. Second, whether there is excessive reliance on forward linear extrapolation. If a stock's valuation has already priced in the extreme optimistic scenarios for 2027-2028, regardless of fundamentals, its risk is greater than that of already corrected peers. Third, whether revenue, profit, and cash flow can mutually verify each other. This is the most important. Those with orders but no revenue, with revenue but no profit, and with profit but no cash flow will suffer greatly in the proof question era. The market previously could overlook these mismatchesAfter all, CSPs are still increasing capex; lets look at the ceiling before addressing the floorbut that is no longer the case. This article is reposted from "Wall Street Insights", author: Long Yue; GMTEight editor: Chen Siyu.