Cloud Computing Update: Increasingly diverging

The cloud sector enters 2H26 with the strongest fundamental momentum on record and the weakest share-price momentum in two years. Global cloud infrastructure spending (by major CSPs) reached US$170bn in Q2 2026, up 80% YoY — the eleventh consecutive quarter of accelerating growth. Every major provider delivered results ahead of expectations: Google Cloud grew 84%, Microsoft Azure 43% (constant currency), Amazon AWS 37% (its fastest growth), and Oracle OCI 93% (1Q26).

Yet, since 1 July, neocloud stocks have fallen between 25% and 38%, while the SOX Index declined 6.7% in a single session. Neocloud providers are highly specialised cloud computing companies built entirely around delivering GPU-as-a-Service (GPUaaS). The trigger was a report suggesting that Meta is developing "Meta Compute" to resell surplus AI capacity. That single headline crystallised the bearish narrative that the market had largely ignored throughout the year: supply may eventually catch up with demand, and the marginal seller could become your own largest customer.

Entering the second half of the year, we expect inflationary pressures to remain elevated for longer, with the possibility of further rate hikes during the period.

Our view remains that the demand-side evidence remains compelling, supported by record backlogs, high utilisation rates, power constraints, and continued pricing strength. While demand remains robust, the more imminent risk now lies in capital structure and depreciation timing.

This distinction underpins our positioning: overweight the self-funded hyperscale cloud service providers (CSPs) such as Microsoft, Alphabet and Amazon, while remain selective on neoclouds, and avoid the highly leveraged middle tier.

Traditional CSPs: Growth Re-Acceleration path intact

The investment case for hyperscalers is increasingly driven by their ability to convert AI demand into sustainable shareholder value rather than simply delivering faster cloud growth.

Market share has remained relatively stable over recent years, with AWS accounting for approximately 28%, Azure 21%, and Google Cloud 14% (Synergy, Q1 2026). However, the composition of growth has shifted meaningfully. AI-related workloads now account for around 19% of total cloud spending, up from just 8% in 2023. More importantly, AI demand is expanding the overall cloud market rather than merely redistributing market share.

Collectively, the major CSPs (including Meta) are projected to spend more than US$800bn in capital expenditure this year. This investment cycle is supported by a substantial order backlog. Combined cloud order backlogs for the major CSPs reached US$2.09tn in Q1 2026, representing 185% growth from a year earlier. While both the size and growth of these backlogs have been remarkable, operating margins have come under pressure due to elevated capital expenditure. However, in the recent quarter, operating margins have shown some signs of stabilisation as companies achieve it through operational efficiencies and in-house cost-effective chips.

Among the major providers, we believe Amazon AWS and Google Cloud are best positioned to deliver further growth re-acceleration, with quarterly growth potentially exceeding 35% for AWS and 60% for Google Cloud in Q2. Microsoft still has room for acceleration, but more likely in a gradual manner. Amazon's Trainium and Inferentia chip stack could become an increasingly important competitive advantage, as it structurally reduces cost per token and, insulates AWS from GPU component inflation.

Alphabet is the few provider that owns the entire AI technology stack — spanning models (Gemini), silicon (TPUs), networking, and data centres — which we believed should be better positioned when component inflation becomes the dominant cost driver. Management attributed the increase in backlog to strong enterprise AI demand alongside TPU hardware sales, indicating that Google is now monetising its silicon externally rather than solely for internal use.

Compared with Amazon and Google, Microsoft remains the most exposed to component inflation. While the company is also developing its in-house AI chip, MAIA, it has yet to reach the level of vertical integration achieved by its peers. Microsoft increased its FY2026 capital expenditure guidance last quarter, with approximately US$25bn of incremental spending reflecting memory and GPU price inflation rather than additional computing capacity. As a result, margins could remain more volatile than those of its peers. Furthermore, as Microsoft continues allocating a portion of its computing capacity to internally developed AI models, we do not expect a meaningful re-acceleration in Azure's revenue growth, which we estimate will remain around 43% in Q2. Nonetheless, our positive stance remains : Readmore: Microsoft 2Q26 Earnings Update: Passing with Flying Colours

Oracle highlights an important distinction between operational momentum and investment attractiveness. Oracle has the most impressive backlog growth but also the weakest balance sheet dynamics. Unlike its peers, Oracle has adopted a "bring your own chip" strategy, whereby a significant proportion of the recent increase in Q3/Q4 remaining performance obligations (RPO) came from contracts in which customers either prepaid for or supplied their own GPUs, representing approximately US$75bn of hardware value. While this approach reduces Oracle's dependence on debt financing, it does not carry the same margin profile as a conventional infrastructure-as-a-service (IaaS) model, and management has guided for lower gross margins in FY2027.

Overall, we continue to favour the traditional CSPs, as we expect cloud revenue growth to accelerate further over the coming quarters. We also expect cloud revenue to grow at a faster pace than capital expenditure, which should gradually alleviate concerns over AI monetisation.

Oracle remains an exception. Although its cloud fundamentals continue to improve, we believe the combination of elevated capital requirements, weaker free cash flow generation and increasing leverage warrants a more cautious stance than its operating performance alone would suggest.

Figure 1: Margins for Google Cloud and AWS is more stable than Azure due to in-house chips

Figure 2: RPO reaching $2trn USD

Figure 3: market share for cloud computing

Neoclouds – first layer impact

While traditional hyperscalers (AWS, Azure and Google Cloud) offer thousands of IT services, ranging from general-purpose CPU compute to managed databases and legacy enterprise software, neoclouds focus almost exclusively on high-performance computing (HPC) for artificial intelligence workloads. They optimise their entire technology stack—including bare-metal performance, high-bandwidth memory (HBM), lossless Ethernet/InfiniBand networking, and advanced liquid cooling—to maximise Model FLOPS Utilisation (MFU) for AI training and inference.

Over the past few years, the traditional CSPs have de-risked their infrastructure expansion by adopting a hybrid build-and-lease strategy, whereby they both construct their own data centres and lease capacity from neocloud providers. This approach reduces the risk of overbuilding while allowing them to scale capacity more flexibly in response to AI demand.

Like the major CSPs, neocloud providers have benefited significantly from the AI investment cycle. CoreWeave currently has a contracted backlog of US$99bn, a substantial increase from just over US$40bn in Q1, with approximately 36% of the backlog expected to be recognised within the next 24 months. The company currently operates more than 1GW of active capacity, with contracted capacity reaching 3.5GW. Management is targeting more than 1.7GW of active capacity by 2026 and 8GW by 2030. Nebius also has a sizeable, contracted backlog, with an order book of approximately US$46bn, including a Meta contract worth up to US$27bn over five years and a separate Microsoft agreement valued at US$17.4–19.4bn.

CoreWeave demonstrates why strong revenue visibility alone is insufficient. The investment outcome increasingly depends on whether future cash flows are sufficient to support the capital structure. The distinction between CoreWeave and Nebius therefore lies not only in their growth profiles, but also in how that growth is financed. .

CoreWeave operates as a leveraged GPU infrastructure provider, leasing facilities, financing GPU purchases with debt, and relying on long-term take-or-pay contracts to service that leverage. Nebius, by contrast, is a vertically integrated owner-operator that owns its data centres, designs its own infrastructure, and funds expansion primarily through customer prepayments and equity rather than debt. Consequently, Nebius generates positive operating cash flow, whereas CoreWeave's interest expense consumes nearly half of its adjusted EBITDA.

However, Nebius has greater near-term exposure to spot and short-duration pricing than CoreWeave, whose revenue is predominantly derived from long-term take-or-pay contracts. Under take-or-pay agreements, customers commit to multi-year contracted capacity, whereas spot and short-tenor capacity is repriced continuously in line with market conditions.

It is important to distinguish between Nebius's contracted order book and its near-term revenue recognition. The company has a substantial multi-year backlog of approximately US$44–50bn, anchored by the Meta contract (approximately US$27bn over five years) and Microsoft's agreement (US$17.4–19.4bn over five years). Relative to its market capitalisation of approximately US$53bn, this represents one of the largest contracted books in the sector. However, unlike CoreWeave—which expects to recognise around 36% of its US$99.4bn backlog within the next 24 months—much of Nebius's contracted revenue will only begin contributing from early 2027, when the dedicated Meta deployments based on NVIDIA's Vera Rubin platform come online.

As a result, despite its substantial contracted backlog, Nebius's near-term revenue remains disproportionately dependent on spot and short-tenor contracts. In other words, Nebius is under-contracted from a revenue recognition perspective today but becomes heavily contracted from 2027 onwards, leaving an exposure window over the intervening 18 months before those long-term contracts begin generating meaningful revenue.

Should AI token consumption temporarily slow—for example, due to delays in major model releases, AI laboratories reducing training activity following large-scale deployments, or enterprises postponing AI pilot programmes—Nebius would likely be affected through two channels simultaneously. Firstly, utilisation could decline as unrenewed capacity sits idle. Secondly, pricing could come under pressure as the marginal GPU-hour clears at lower market rates. In our view, this dual exposure represents the key downside risk for the company over the next 18 months.

Figure 4: Order books are strong

Figure 5: Margins still very low

Balance Sheets and Operating Leverage set them apart

We expect greater divergence across the cloud sector in the second half of the year. Consistent with our broader macro outlook, our base case is that inflationary pressures will remain persistent, increasing the risk of further interest rate hikes. In such an environment, companies with (1) higher financial leverage and (2) more demanding valuations are likely to face greater downside risk.

Traditional CSPs are relatively well positioned under this scenario. Large hyperscalers such as Microsoft, Alphabet and, to a lesser extent, Meta maintain strong balance sheets with ample financial flexibility. In addition, both Microsoft and Meta are currently trading below their historical valuation averages, providing some valuation support even if our downside scenario materialises. While Amazon carries higher leverage than its peers, its balance sheet remains healthy, and leverage is still within manageable levels. By comparison, Oracle has the weakest balance sheet among the major CSPs. To finance its rapid data centre expansion, the company has recorded consecutive quarters of negative free cash flow and has been increasingly reliant on both debt and equity financing.

Neocloud providers face considerably greater financial risk. Companies such as CoreWeave and Nebius have contracted order backlogs that exceed their respective market capitalisations, providing strong long-term revenue visibility. On the other hand, CoreWeave carries US$24.9bn of total debt, with annual interest expense of US$536m—equivalent to nearly half of its adjusted EBITDA. Nebius is in a healthier financial position, generating positive operating cash flow, but the funding gap remains a key concern. Management has guided for US$20–25bn of capital expenditure in 2026 against cash balances of approximately US$9.3bn. The company intends to fund around 60% of this investment through customer prepayments from strategic partners such as Meta and Microsoft, with the remaining 40% financed through a combination of debt and equity. However, that remaining funding requirement amounts to several billion dollars that have yet to be raised, with financing terms likely to depend on a share price that has already fallen by around 20%. Consequently, both companies remain exposed to higher financing costs should interest rates increase further.

The risks extend beyond balance sheet considerations. Neocloud providers are also more vulnerable in an oversupply scenario because a significant proportion of their revenue is derived from hyperscalers, which are simultaneously their largest customers and direct competitors. Microsoft accounted for 67% of CoreWeave's revenue in 2025, while 68% of its year-end net accounts receivable was owed by Microsoft. Nebius's early revenue base is similarly concentrated around Microsoft. Should the hyperscalers find themselves with excess computing capacity, they could reduce leased capacity by scaling back or allowing contracts with neocloud providers to expire upon renewal. Furthermore, CoreWeave's contracted revenue is increasingly concentrated among AI laboratories, with OpenAI estimated to account for around one-third of its contracted revenue over the next seven years. As these AI laboratories also maintain direct relationships with the major hyperscalers, any slowdown in AI demand would likely see customers prioritise hyperscale cloud providers over neocloud operators. This customer concentration and competitive overlap leave neocloud providers significantly more exposed in a weaker demand environment.

Figure 6: Scatter plot of total debt to ebitda and interest coverage ratio

Figure 7: Illustration of relationship of Neoclouds and traditional CSPs

Other Key Debates Heading into 2H26

1)      Meta’s compute question. Meta increased its FY2026 capital expenditure guidance to US$125–145bn (from US$115–135bn), compared with approximately US$72bn in FY2025, and is reportedly exploring the option of reselling surplus AI compute capacity. While this would allow the company to better monetise excess capacity, it also introduces another well-capitalised competitor into an increasingly crowded cloud computing market.

Our base case is that this should not be interpreted as evidence of industry overbuilding. A company subleasing a few hundred megawatts of capacity while simultaneously committing tens of billions of dollars to infrastructure is optimising its asset base rather than attempting to become a full-scale hyperscaler.

We believe the incumbent CSPs remain well positioned to defend their market share through their differentiated competitive advantages—AWS through the breadth of its service offering, Azure through its enterprise distribution network, and Oracle through its pricing competitiveness.

Neocloud providers, however, are likely to face greater competitive pressure. Even so, we do not expect meaningful near-term cannibalisation, given the unique relationship between Meta and providers such as CoreWeave, where Meta is both a major customer and a potential competitor.

2)      Data centre buildup delays. As discussed in our previous article, Data Centre Infrastructure 2026: Where are we now, and Where are we heading to, we continue to see the risk of delays to the global data centre build-out due to ongoing structural constraints, including power availability, grid connections, equipment bottlenecks and construction timelines. Should these delays materialise, they would further constrain available computing capacity and could limit the pace of cloud revenue growth despite robust underlying demand.

We believe this risk is likely to have a greater impact on companies that have provided more aggressive capacity expansion and growth guidance, as any delay in bringing new infrastructure online would postpone revenue recognition while capital expenditure continues to be incurred.

Positive on big hyperscalers; Neutral on Neoclouds

The market is increasingly distinguishing between companies that participate in AI spending and those most likely to capture its economic value. While structural AI demand remains robust, balance-sheet strength, financing flexibility and execution discipline will increasingly determine shareholder outcomes. This reinforces our preference for large hyperscalers over more leveraged infrastructure providers.

Accordingly, we remain selective on companies with elevated funding requirements and weaker balance-sheet flexibility, including Oracle, neocloud providers and pre-contract miners, despite pockets of strong operating momentum.

Our preferred exposures remain Alphabet, Microsoft and Amazon. For investors seeking diversified exposure, they can consider the Invesco Nasdaq Internet ETF (NASDAQ: PNQI), which provides broad exposure to these technology leaders.

Declaration:

This research report was prepared with the assistance of artificial intelligence (AI) tools. iFAST Financial Pte Ltd does not rely exclusively on AI for content generation; the content of this report – including all investment theses, ratings, price targets and conclusions – has been independently reviewed and verified by the research analyst(s) to ensure accuracy and professional integrity.

For specific disclosure, at the time of publication of this report, IFPL (via its connected and associated entities) holds a NIL position in the abovementioned securities. The analyst who produced this report holds a position in Microsoft.