After the Maga7 earnings reports last week, if I were allowed to summarize the market's attitude toward cloud providers in just one sentence, I would say this: The market is no longer pricing whether AI demand exists, but rather who will bear the capital expenditures for whom..

The earnings reports of all three cloud providers were not bad; in fact, they could be described as very good. However, the degree of divergence in stock price reactions was rare in the AI rally over the past three years. Once you clarify the underlying pricing logic behind this divergence, Neocloud's position in this cycle becomes very clear: they are not the "little brothers" or "budget alternatives" of cloud providers, but rather the only pure targets naturally positioned on the receiving end of this capital expenditure arms race.

Earnings Review: Why the Treatment Differs So Drastically Despite Both Beating Expectations

Google: Historically Strong Cloud Growth, Yet a Bearish Candle

Alphabet's earnings report, looking solely at its cloud business, was almost impeccable: Google Cloud generated $24.8 billion in quarterly revenue, an 82% year-over-year increase, accelerating from the 63% growth in Q1. The backlog swelled to $514 billion, and operating profit from the cloud business reached $8.8 billion, roughly triple that of a year ago. Two years ago, such figures would have been enough to send the stock price jumping 10%.

But the market reacted with a decline.

The reason lies in the cash flow statement: The 2026 capex guidance was raised from $180-190 billion to $195-205 billion. Quarterly capital expenditures exceeded operating cash flow, causing free cash flow to turn negative for the first time in years.

This is a landmark moment.

The market voted with real money to tell all cloud providers: Revenue beats are merely entry tickets, not rewards; the "accounting nature" of capital expenditures is the core of pricing.

When capex consumes all FCF and forces companies to borrow, it ceases to be viewed by investors as an "investment for the future" and begins to drift toward being a "cost of maintaining competitive position." When an expense shifts from optional to unavoidable, its valuation attribute degrades from a growth option to an operating liability.

Microsoft: The Only One Rewarded for Disciplined Growth

Microsoft delivered the most comfortable combination among cloud providers this earnings season: Azure growth accelerated to 43%, exceeding the ~40% consensus estimate, with Intelligent Cloud revenue reaching $39.3 billion. More importantly, capex guidance remained relatively flat, and free cash flow stayed positive.

It is worth noting that Microsoft's situation before the earnings report was not good. Since the beginning of the year, it had fallen nearly 20%, and the FY26 full-year capex guidance of $141-143 billion had once caused significant unease in the market.

The reason this earnings report managed to reverse the narrative was not because of Azure's absolute growth rate (Google Cloud's 82% was right next door), but because Microsoft proved one thing: Growth acceleration and capital discipline can coexist.. In a market where everyone is questioning capex returns, "growth with discipline" is a scarce asset.

Of course, risks have not disappeared: The concentration related to OpenAI in commercial RPO is the weakest link in the backlog narrative. But at least this quarter, Microsoft held onto its position as the only payer in the capital expenditure camp that could still demonstrate returns.

Amazon: Why Did It Surge 10% After Also Raising Capex?

This is the most intriguing of the three earnings reports. Amazon's Q2 revenue of $200.6 billion and EPS of $1.97 both beat expectations. Jassy described AWS as "booming," with self-developed chips (Trainium/Graviton) and AI business both breaking through to an annualized revenue run rate of $25 billion. Meanwhile, the company raised its 2026 capex to $220 billion.

By Google's script, raising capex should have led to sell-offs. But Amazon surged more than 10% after hours.

The difference lies in attribution. Amazon explicitly attributed the capex increase to rising memory costs. This means this incremental spending was not an escalation of the arms race, but rather that the same capacity came at a higher procurement price.

The former is faith charging on the demand side; the latter is cost inflation on the supply side. The market punishes the former, but interprets the latter as upstream supply so tight that even hyperscalers must rewrite their budgets for it, meaning the narrative of pricing power across the entire supply chain is true.

Incidentally, this is also the highest-level demand confirmation heard by memory bulls this year: When the world's strongest buyers write DRAM/NAND price hikes as the primary attribution for capex increases in their earnings reports, this cycle in storage needs no more evidence.

Consensus and Divergence Among Institutions

● No controversy on the demand side. Evercore's Mahaney said cloud demand "looks relentless," with the divergence lying only in whose growth rate can catch up with Google.

● All controversy is in capital returns. Wolfe Research's pre-earnings comment that "investor concerns need to decrease first before the stock price can have a positive reaction" is the key to understanding the entire earnings season; Cowen's Derrick Wood directly warned that "another capex raise would trigger sell-offs"—this played out with Google, while with Amazon, it was resolved by "attribution differences."

● Confirmation of supply bottlenecks is escalating. Bloomberg Intelligence explicitly stated: High-end AI computing power remains supply-constrained, and news of NVIDIA negotiating approximately $250 billion in financing backstops for OpenAI data center leases was interpreted as the industry chain willing to build its own financing closed loop for the next wave of expansion.

Institutions do not doubt demand, nor do they doubt supply tightness; they only doubt whether hyperscaler shareholders can get a share of the meat from this feast. And this skepticism itself is the starting point of Neocloud's opportunity.

The Real Bottleneck: Not GPUs, But Grid-Connected Power and Rising Memory Costs

To understand Neocloud's position, one must first answer a question: Where exactly are cloud providers stuck when spending $700 billion a year (the magnitude of combined capex for the four hyperscalers in 2026)?

The answer is no longer GPUs. NVIDIA's production capacity is ramping up, and the real hard constraints have shifted to two areas:

First, power and racks—"grid-connected capacity" has become the asset with the longest lead time.

Data centers, from land acquisition and substations to grid connection permits and mechanical/electrical delivery, take years, and this cannot be compressed simply by throwing money at it. This is why Google's CFO said on the earnings call the most important words for the Neocloud sector: Q3 will expand third-party cloud capacity procurement because internal infrastructure cannot keep up with demand..

Pichai's supplementary logic is more straightforward—better to outsource computing power and sacrifice some margin than to turn customers away due to lack of capacity. When the world's third-largest cloud provider publicly admits it cannot build fast enough, this is no longer a business choice, but a physical constraint.

Second, memory—supply-side inflation is rewriting everyone's budget.

Amazon's attribution of the capex increase to memory costs forms a complete closed loop with the storage chain we have been tracking (Kioxia beating estimates across the board, SK Hynix's record-breaking profit margins, Micron's hundred-billion-dollar RPO floor): HBM squeezes out conventional DRAM capacity → NAND/DRAM prices rise across the board → hyperscaler procurement costs rise → capex passively increases → market further questions capex returns → outsourcing motivation strengthens.

Note the final link in this chain; it connects memory price hikes to Neocloud demand.

Putting these two together, the conclusion is: The scarce resource in this cycle is computing capacity that is "grid-connected, deliverable, and financed," not the technology of computing power itself.

Whoever holds this capacity holds the pricing power.

Neocloud: The Natural Payee in the Capital Expenditure Arms Race

Position: Capex Payer vs. Capex Payee

We have always used a framework to divide the AI infrastructure food chain: capex payer (the payer) and capex payee (the payee).

Hyperscalers are payers: Every dollar of their capex must first pass the judgment of the depreciation schedule and FCF before it can become revenue; storage, equipment, and optical interconnects are payees: Hyperscalers' budgets are their orders.

Neocloud's position in this framework is special: They are both payers (buying their own GPUs, building server rooms) and payees (hyperscalers pay them rent).

And what happened this earnings season is rapidly pushing them toward the payee side:

When capital markets punish hyperscalers' self-built capex (Google's lesson) and reward capacity acquisition (Amazon AWS's "booming"), the most rational choice for hyperscalers is: To outsource part of the construction risk, converting capex on the balance sheet into opex on the income statement..

Renting Neocloud capacity is essentially hyperscalers purchasing a "financial statement beautification service" from the capital market: They still get the capacity, FCF doesn't look ugly anymore, and depreciation risk is left to others.

After getting slapped by the market, Google will instead more firmly expand third-party procurement. hThe greater the stock price pressure on hyperscalers, the stronger the order certainty for Neocloud.. This is a rare reflexive structure: Sector-wide bad news is precisely good news for this sub-sector.

Compared to Cloud Providers, Where Exactly Lies Neocloud's Opportunity?

From a supply perspective, I believe there are three points, in descending order of importance:

First, they sell the scarcest thing in this cycle—grid-connected power..

CoreWeave and Nebius have each locked in contract power in the range of 3.5GW. CoreWeave aims for 1.7GW activated by the end of 2026 and over 8GW by 2030. GPUs will depreciate with iteration, but a gigawatt-scale park that is already grid-connected and permitted is an irreplicable asset in reality, where grid queue times often stretch to five years.

Neocloud's valuation anchor should be the replacement cost per MW of grid-connected capacity + discounted contract cash flows, rather than the PE of the GPU leasing business. Looking at them via the latter makes them forever expensive; looking via the former, second-tier power asset-type targets (IREN, WULF, CIFR) still have room for revaluation.

Second, contract structures are rapidly de-risking..

IREN is the best sample: Of its year-end AI cloud revenue target of over $4 billion, about 85% is already covered by contracts, and 45% of GPU capital expenditures are financed by customer advance payments. This means nearly half of the expansion risk has been transferred to the demand side. When customers start paying your capex for you, you transform from a storytelling growth stock into a contract cash flow business somewhat akin to midstream energy assets.

Third, purity.

Buying Microsoft, you get Azure plus Office, gaming, and a pile of depreciation; buying Google, you get 82% cloud growth, but also negative FCF. Neocloud is the purest expression of this supply scarcity: High beta is its flaw, but also the meaning of its existence.

But Risks Must Be Fully Explained

Bullish does not mean brainless. The July crash in this sector (CoreWeave once retraced 35% from its yearly high) reminds us that its vulnerabilities are equally structural:

Financing dependence is the first fatal flaw.

CoreWeave's long-term debt has reached the level of $35 billion, with 2026 capex guidance of $31-35 billion corresponding to $12-13 billion in revenue.

This is a model where the speed of borrowing determines the speed of growth. In July, we repeatedly discussed the transmission of credit market pressure; for every 50bp widening in high-yield spreads, Neocloud's expansion economics must be recalculated.

NVIDIA's financing backstop alleviated this concern, but chain-master backing and revolving financing are separated by a thin line; the market will inevitably return to scrutinize this line sooner or later.

Customer concentration is the second fatal flaw.

In July, Meta released rumors about renting out its own idle capacity, causing CRWV and NBIS to plummet: When the largest customer can theoretically become a competitor at any time, the duration of the backlog must be discounted.

Hyperscalers renting Neocloud are essentially buying "time"; once they catch up with self-built capacity, renewal pricing power will shift. This determines that Neocloud's golden window has a deadline, likely the next 2-3 years, which represent the deepest supply gap.

Depreciation is the third fatal flaw, merely delayed..

The five-to-six-year accounting life of GPUs clashes with two-year technological iterations. This problem is currently covered by contract periods, but after 2028, when the first batch of clusters enters the residual value verification period, it will become the next major divergence point for the sector.

Conclusion: Buying not AI demand, but pricing power over supply bottlenecks

This earnings season confirmed three things: Demand side is secure (all three cloud providers saw accelerated or above-consensus growth), Hyperscaler self-built capex shareholder returns face systemic questioning (Google's FCF penalty), Supply bottleneck shifted from GPUs to power and memory (Google's outsourcing confession, Amazon's memory attribution).

These three points collectively point to a structure: Hyperscalers have a strong motive to convert capex into opex, and Neocloud is the sole recipient of this reporting arbitrage.

Therefore, Neocloud's opportunity relative to cloud providers does not lie in superficial narratives like "they grow faster," but in position: Cloud providers are payers of capital expenditures, needing to prove themselves to shareholders on the depreciation schedule; Neocloud is the payee, holding the cycle's scarcest grid-connected capacity and long-term contracts signed with the most visible buyers.

In a market that punishes payers and rewards payees, this position alone warrants a round of revaluation.

On an operational level, I prefer to view the sector in two layers:

Tier 1 (CRWV, NBIS) trades on the conversion rate of backlog to revenue and financing costs; earnings reports should focus on the ramp-up pace of activated power (GW), not absolute revenue;

Tier 2 power asset types (IREN, WULF, CIFR) trade on the revaluation of capacity per MW; catalysts are new contract closures and credit enhancements (if the model Google gave WULF is replicated, it will be a sector-wide positive).

The tracking list is also simple: The actual scale of Google's Q3 third-party capacity procurement (to confirm or falsify the outsourcing logic), the final form of NVIDIA's financing backstop implementation (determining the pricing of financing risk), and high-yield bond spreads (determining the discount rate for the entire sector).

As long as these three signals do not deteriorate, the July drawdown is more like a washout before this emerging sector's second pricing, rather than the end of the narrative.

Risk never lies in whether AI is a bubble, but in what you hold when the tide goes out: contracts, or stories.

$NVIDIA (NVDA.US)$Tesla (TSLA.US)$Meta(META.US)$Amazon (AMZN.US)$Gold ETF - SPDR(GLD.US)$Silver Trust ETF - iShares(SLV.US)$Meta(META.US)$TSMC (TSM.US)$Grayscale Bitcoin Mini Trust ETF(BTC.US)$Circle(CRCL.US)$Coinbase(COIN.US)$Intel (INTC.US)$AMD(AMD.US)$Alphabet-A(GOOGL.US)$Russell 2000 Index ETF - iShares(IWM.US)$Nasdaq 100 ETF - Invesco(QQQ.US)$Micron Technology (MU.US)$Broadcom (AVGO.US)$Eli Lilly (LLY.US)$Novo Nordisk A/S (NVO.US)$IREN(IREN.US)$Applied Digital(APLD.US)