Key Points
- The narrative surrounding massive CAPEX has shifted from enthusiastic welcome to growing concern regarding free cash flow (FCF), return on investment (ROI), and mounting depreciation costs.
- Given the robust underlying fundamentals of these firms, we expect this FCF erosion to be temporary.
- Useful life of chips has extended. In the coming years, inference demand is expected to significantly outpace that of AI training. For inference tasks, while the latest hardware is beneficial, it is often not essential; consequently, older but still functional chips may remain useful for selected inference workloads
- Following the recent sell-off, the forward price-to-earnings (P/E) ratio for these top ten firms is now trading near the lows seen during the 2025 liberalisation sell-off. At these levels, valuations appear more balanced.
- Today’s hyperscalers are demonstrating significantly more prudence. Although they have outlined massive Capital Expenditure (CAPEX) plans for 2026, these outlays are underpinned by substantial order backlogs.
- Consequently, by weighing these risks against the prevailing opportunities, we maintain that earnings visibility remains relatively strong and fundamentals remain healthy; thus, we reaffirm our 3.5 Stars "Attractive" Rating for the digital economy Internet sector.
Since the launch of ChatGPT in late 2022, the digital economy has soared on the back of an "AI boost". However, since late 2025, the narrative has shifted meaningfully. The very catalysts that initially sparked the rally are now the primary drivers of share price volatility.
Big Tech firms announcing plans for increased Capital Expenditure (CAPEX) have seen their shares heavily sold off—a stark contrast to two years ago, when aggressive CAPEX was viewed as the essential "entry ticket" to leading the AI race. Furthermore, the sentiment has soured for other segments of the technology sector; software, once deemed a primary beneficiary of AI, is now increasingly viewed as being under threat from it.
Given this backdrop, we revisit the investment case for the digital economy. This analysis is divided into two parts—Big Tech and Software—with this first article focusing specifically on the outlook for the tech giants.
Here are the three issues the market is debating:
1. Can hyperscalers sustain this level of capital spending without damaging free cash flow?
2. Will the depreciation cycle of GPUs create an earnings drag before AI monetisation materialises?
3. Is the demand pipeline — driven by inference and agentic AI — strong enough to justify the infrastructure build-out?
Is AI CAPEX becoming too large for Big Tech balance sheets?
The recent sell-off in Big Tech was not triggered by disappointing earnings results, but rather by the fact that the major players—Meta, Alphabet, Microsoft, and Amazon—surpassed CAPEX expectations last quarter and significantly raised their projections for 2026. Consequently, CAPEX for 2026 is now expected to reach USD $650 billion, representing a 56% increase from 2025—a massive leap from our previous estimate of a 22% increase.
Beyond the ongoing concern that "AI is eating software", investors are increasingly questioning whether AI CAPEX is becoming large enough to materially reshape broader capital allocation across the economy. The AI build-out accounted for less than 0.1% of GDP in 2020; by 2025, it has surged to nearly 2%. Projections for next year are even more robust. Should this pace continue, AI-related infrastructure spending could become one of the most significant capex cycles relative to GDP in recent history.
Figure 1: AI spending in on track to largest infrastructure spending in US history
Figure 2: Big Tech firms beat and raise CAPEX projections
Revisiting our earlier concerns: Are conditions improving?
In our previous analysis, we noted that the narrative surrounding massive CAPEX has shifted from enthusiastic welcome to growing concern regarding free cash flow (FCF), return on investment (ROI), and mounting depreciation costs.
1) Increasing strains on Balance Sheet (Negative)
This aggressive "AI arms race" has fundamentally altered the FCF profile of Big Tech, shifting the model from high-margin stability toward a period of heavy, front-loaded investment. Indeed, free cash flow has trended downwards since these firms began their heavy CAPEX bets.
More importantly, AI spending is increasingly being supplemented by debt issuance rather than funded solely through internal cash flow.
Google’s recent announcement of a 100-year bond issue has further intensified these concerns, even though the bid-to-cover ratio remained impressively high at nearly 10x. While current balance sheets can sustain this level of investment, projections of an additional $630 billion in CAPEX this year suggest that much of this will be co-funded by debt. This could place greater pressure on balance sheets and reduce flexibility for other corporate allocation priorities.
However, given the robust underlying fundamentals of these firms, we expect this FCF erosion to be temporary. As future CAPEX eventually moderates, cash flow should recover; consequently, we view the current balance sheet pressure as more cyclical than structural, rather than a long-lasting impairment.
Figure 3: Share buybacks have slowed down in latest quarter
Figure 4: Debt to Ebitda suggest companies could stomach more bonds
2) Depreciating GPUs over 5-6 years is reasonable (Positive)
Investor anxiety largely stems from two primary risks: first, the substantial depreciation expenses that accompany massive CAPEX outlays; and second, the potential for a "depreciation trap." If GPUs become economically obsolete within three years but are amortised over six, companies could face significant future write-downs.
Regarding the latter, Nvidia’s latest earnings report offers a strong rebuttal. Management emphasised that six-year-old Ampere-generation chips remain in active use, validating the five-to-six-year useful life currently recorded on hyperscalers' balance sheets. Demand for older chips remains robust, driven by the surge in AI inference.
In the coming years, inference demand is expected to significantly outpace that of AI training. For inference tasks, while the latest hardware is beneficial, it is often not essential; consequently, older but still functional chips may remain useful for selected inference workloads. Indeed, Google continues to run seven-to-eight-year-old TPUs (Tensor Processing Units) at full utilisation.
Nevertheless, the first concern remains a valid headwind. Huge CAPEX spending inevitably brings heavy depreciation charges that will weigh on performance. If returns generated from AI workloads fail to outpace the associated depreciation burden, this could become a persistent drag on earnings quality.
Figure 5: We see a total of 5-6 years of useful life for AI accelerators/GPUs today
Source: Bofa Global Research
Figure 6: Older chips are being used for AI inferencing
3) The ROI on current CAPEX remains unproven (Neutral)
Investors remain understandably anxious regarding the Return on Investment (ROI) of current Capital Expenditure (CAPEX). By prioritising heavy infrastructure spending, companies are inevitably diverting funds that could otherwise bolster immediate shareholder wealth through dividend payouts or share buybacks.
We expect these concerns to remain a key overhang on the investment case for some time, as a definitive ROI is unlikely to materialise in the near future. However, historical precedent suggests that periods of aggressive infrastructure investment can appear excessive before monetisation becomes visible.
A prime example is the fibre-optic boom of the 1990s. During this era, telecommunications firms spent an estimated $1 trillion—reaching 1.5% of GDP—laying millions of miles of undersea and terrestrial cables. Initially, this resulted in a "capacity glut" and a subsequent market bust. Yet, that "excessive" infrastructure eventually became the indispensable foundation of the entire modern digital economy, enabling the rise of high-bandwidth platforms like YouTube, Netflix, and the App Store decades later.
Furthermore, historical projections have often proved overly cautious. In 1995, analysts predicted the web would reach 100 million users by the turn of the century; in reality, the figure exceeded 360 million—surpassing even the "bull case" projections of the mid-90s by more than threefold. What originally appeared to be an unrealistic forecast proved, in hindsight, to be a massive underestimation.
Consequently, while it remains difficult to estimate the future Total Addressable Market (TAM) for AI, history suggests that early projections for transformational technologies are often too conservative rather than too ambitious.
The AI Boom: Same Scale, Different Substance
In contrast to the prior bubble episodes, investors today are responding with scepticism rather than euphoria, even where earnings remain robust.
Crucially, the risk of a near-term supply glut currently appears minimal. During the "Fiber Glut" of the late 1990s and early 2000s, telecommunications giants such as Global Crossing and WorldCom poured billions into laying vast networks of fiber-optic cable, driven by the erroneous belief that internet traffic was doubling every 100 days. By 2002, however, up to 90% of the fiber laid over the preceding five years remained "dark"—installed but never activated with transmission equipment. The current landscape stands in stark contrast. While technology firms are spending heavily on data centres to support AI workloads, vacancy rates of data centres currently sit at less than 1%.
Today’s hyperscalers are demonstrating significantly more prudence. Although they have outlined massive Capital Expenditure (CAPEX) plans for 2026, these outlays are underpinned by substantial order backlogs. In the latest quarter, cloud demand backlogs for all major hyperscalers accelerated; notably, both Microsoft and Alphabet saw year-on-year growth exceeding 100%. By bringing this capacity online, they are not simply spending aggressively; they are bringing online capacity backed by visible demand and large order backlogs.
Figure 7: Massive spending backed by strong order backlogs
Figure 8: Cloud Growth
Valuations are becoming more reasonable
The leading technology companies – Meta, Microsoft, Alphabet, Amazon, in aggregate have declined by -7.4% year-to-date, significantly underperforming the broader market (S&P 500 index: -0.9%) as of 9 March 2026. For broader historical comparison, we use the top ten US companies as a proxy for "Big Tech."
Following the recent sell-off, the forward price-to-earnings (P/E) ratio for these top ten firms is now trading near the lows seen during the 2025 liberalisation sell-off. At these levels, valuations appear more balanced relative to their five-year average and remain substantially below the extremes witnessed during the dot-com bubble.
Figure 9: Valuation are getting more reasonable
Individual big tech outlook
|
Company |
Investment Thesis |
Target Price (USD) |
|
Meta |
Revenue increase 25% y/y. Ad impressions delivered increased by 18%, Average price per ad increased by 6% for the fourth quarter. Capital expenditures for 2026 nearly double from 2025, aimed at expanding AI infrastructure through Meta Superintelligence Labs. This could translate into future ads growth. |
1016 |
|
Alphabet |
Alphabet’s AI driven growth in cloud revenue (48% y/y) and search remained resilient at 17% y/y. Although capex is expected to roughly double in 2025, the elevated investment in AI compute, frontier model development at DeepMind, Cloud capacity, and advertiser ROI appears strategically necessary and is already generating tangible returns, reinforcing confidence in the company’s long-term competitive positioning and earnings trajectory. |
360 |
|
Microsoft |
Cloud growth was 39% y/y, slightly below previous quarter, however it isn’t a sign of AI demand slowdown, but the reallocation of some Azure capacity towards Copilot. If not, growth figures should be even higher. This investment could translate into growth in its productivity software over the longer term. The beat and increase in CAPEX were driven by a huge backlog (commercial bookings surged 230% and commercial remaining performance obligations grew 110%, with the balance nearly tripling over a two-year period), providing long-term revenue visibility. |
555 |
|
Amazon |
Slight miss in earnings, but cloud growth was impressive (24% y/y), along with backlog that has surged to USD 244bn. On the other hand, Advertising continues to emerge as a key secondary growth driver, expanding 23% y/y (above estimates). CAPEX projection for 2026 is USD 200bil, though cloud growth was below peers, but given the largest growth in absolute value, we are glad to see growth picking up. Trainium and Graviton (Amazon own custom chips ) together now represent a USD 10bn+ annual revenue run-rate, growing at triple-digit rates. |
285 |
Figure 10: Earnings Estimates for Big Tech remains resilient
Embracing the era of Agentic AI
In recent developments, AI models have increasingly facilitated the rise of autonomous digital agents capable of independently managing complex computational tasks.
Recent advances in frontier models and open-source agent frameworks signals that the industry is rapidly entering the "agentic AI" era. This transition has catalysed an exponential surge in inference, the primary process by which models and agents are deployed within the cloud. The shift from reactive chatbot experiences to proactive agentic execution has have contributed to a sharp rise in token usage and cloud inference demand.
Within this agentic landscape, token consumption scales at a significantly higher magnitude than traditional single-turn interactions. This is primarily due to the iterative reasoning loops required by autonomous agents, which must continuously cycle through stages of perception, planning, tool execution, and self-reflection. The resulting multiplication in token consumption has led to a tremendous spike in usage over recent weeks; consequently, hyperscale cloud providers are poised to benefit directly from this unprecedented growth in token demand.
Figure 11: Tokens usage surged in recent weeks
Source: Openrouter. Data as of 3 March 2026.
Downside Risk
1) Private company concentration
The order books of major hyperscalers, particularly Microsoft and Amazon, have become increasingly concentrated in a handful of private AI labs, most notably OpenAI and Anthropic.
Recent financial disclosures indicate that Microsoft’s Remaining Performance Obligations (RPO) are heavily weighted toward OpenAI, which represents nearly 45% of its total backlog. Similarly, Amazon’s AWS has seen its future revenue commitments significantly bolstered by multi-billion dollar compute deals from these same entities. However, these private AI firms often lack the organic financial capacity to sustain such massive obligations independently. Their ability to fulfill these contracts is frequently dependent on a "circular economy" of funding, where the hyperscalers provide the capital through strategic investments that the startups then commit back to the hyperscaler for cloud infrastructure.
This creates a risk that reported backlog strength may overstate the degree of underlying self-funded demand .
Figure 12: Microsoft and Amazon’s backlog tied to OpenAI
2) Regulatory Concerns
Beyond financial concentration, the rapid displacement of entry-level and routine-based roles by AI systems is precipitating a new wave of regulatory intervention aimed at maintaining social stability. As generative AI automates tasks previously handled by human labor, many companies are downsizing, such as Block announced to reduce its workforce by -40%, identifying AI advancements as the primary catalyst. Meanwhile, Amazon eliminated roughly 30,000 corporate roles (about 10% of its corporate workforce) between late 2025 and January 2026, pointing to the integration of generative AI across internal processes as a key driver for this overhaul. This introduces a significant legislative risk, which could slow adoption, increase compliance costs, and lengthen the path to AI monetisation.
Volatility likely to persist in the near term; Dollar Cost Averaging as the strategy to navigate current situations
Big tech firms currently find themselves at a point of divergence, with many expanding aggressively underpinned by robust order books. While anxieties regarding a potential AI bubble and the timeline for monetisation will likely persist for some time, structural tailwinds—specifically substantial cloud order backlogs and the transition to Agentic AI—are reinforcing the revenue path for major technology companies, particularly hyperscalers.
Consequently, by weighing these risks against the prevailing opportunities, we maintain that earnings visibility remains relatively strong and fundamentals remain healthy; thus, we remain positive on the sector. Given the significant weighting of these firms within the Invesco NASDAQ Internet ETF, we reaffirm our 3.5 Stars "Attractive" Rating for the digital economy Internet sector.
Nevertheless, in light of current geopolitical tensions in the Middle East and the lack of a clear path toward a ceasefire, we anticipate that market volatility will persist in the short term. We therefore advise investors to utilise Dollar-Cost Averaging (DCA) or Regular Savings Plan (RSP) strategies to mitigate the impact of this volatility on their portfolios.
For Unit Trust recommendations, we continue to recommend Fidelity Global Technology A-ACC-USD and Eastspring Investments Unit Trusts - Global Technology SGD.
Table 1: Valuation for PNQI ETF
|
|
2025 |
2026 |
2027 |
2028 |
|
EPS |
60.6 |
66.2 |
72.2 |
76.1 |
|
EPS growth |
|
9.2 |
9.1% |
5.4% |
|
PE |
|
22.2 |
20.3 |
19.3 |
|
Upside |
|
|
|
55.6% |
|
Fair PE |
30 |
|
|
|
|
Target Price |
1467 |
|
|
2283 |
|
ETF Target Price |
46 |
|
|
72 |
Views on Geopolitical Impact
Supply-side constraints represent the most critical risk. Leading hyperscalers—specifically Amazon, Microsoft, and Alphabet—either already operate or have plans to establish data centre facilities within the Middle East. A meaningful share of key industrial inputs used in semiconductor processing passes through the Strait of Hormuz.
While this is not our base case, should these data centres be physically impacted by the conflict or should sulfur production be halted, the resulting disruptions could severely hinder cloud growth for hyperscalers who are already grappling with existing capacity constraints.
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) and the analyst who produced this report hold a NIL position in the abovementioned securities.
All materials and contents found in this site are strictly for general circulation and informational purposes only and should not be considered as an offer, or solicitation, to deal in any of the funds or products found/identified in this site. While iFAST Financial Pte Ltd ("IFPL") has tried to provide accurate and timely information, there may be inadvertent delays, omissions, technical or factual inaccuracies and typographical errors. Any opinion or estimate contained in this report is made on a general basis and neither IFPL nor any of its servants or agents have given any consideration to nor have they or any of them made any investigation of the investment objective, financial situation or particular need of any user or reader, any specific person or group of persons. You should consider carefully if the products you are going to purchase are suitable for your investment objective, investment experience, risk tolerance and other personal circumstances. If you are uncertain about the suitability of the investment product, please seek advice from a financial adviser, before making a decision to purchase the investment product. Past performance is not indicative of future performance. The value of the investment products and the income from them may fall as well as rise. Opinions expressed herein are subject to change without notice. In respect of any matters arising from, or in connection with the said research analyses or research reports, recipients of the report are to contact IFPL at 10 Collyer Quay, #26-01 Ocean Financial Centre Building, Singapore 049315, or by telephone at +65 6557 2853. Where the report contains research analyses or research reports from a foreign research house and if the recipient of such research analyses or research reports is not an accredited investor, expert investor, institutional investor or an ex-accredited investor, IFPL accepts legal responsibility for the contents of such analyses or reports to such persons only to the extent as required by law. Please note that only certain security(ies) herein are available to all investors, while the rest are only available for certain persons to invest in, such as Accredited Investors (as defined in the Securities and Futures Act) or one who invests at least S$200,000 (or its equivalent currency) per transaction. To qualify as an Accredited Investor, one needs to submit a declaration form and certain relevant supporting documents, according to iFAST’s prevailing policies and procedures.
Please read our full disclaimers on the website at ( https://secure.fundsupermart.com/fsmone/policies/328125/investment-account-terms-&-conditions).
iFAST Financial Pte Ltd (IFPL) (registered address: 10 Collyer Quay #26-01 Ocean Financial Centre Singapore 049315, Telephone: 6557 2000) holds the Financial Advisers Licence issued by the Monetary Authority of Singapore ('MAS') to conduct regulated activities of advising on securities, marketing of collective investment schemes and arranging of any contract of insurance in respect of life policies, other than a contract of reinsurance and the Capital Markets Services Licence issued by the MAS to conduct regulated activities of dealing in securities and providing custodial services for securities. While IFPL has made every effort to ensure the independence of the report's contents, IFPL's nature of business is such that IFPL and its connected and associated entities together with their respective directors, officers and staff may be involved in providing dealing or investment-related services in the abovementioned securities, and have taken or may take positions in the securities mentioned in this report, and may also act as the principal for any buy or sell trades.