Digital Economy (Internet) Outlook 2H2026: Better Late Than Never; Maintained 3.5 Stars

Key Points

• We expect the second half of 2026 to focus more on enhancing AI agent capabilities, while also addressing some of the key challenges faced by earlier generations of agents, including high token consumption costs, data leakage concerns, and inconsistent performance. Big Tech and Software should benefit.
• Hyperscalers are now collectively expected to spend more than USD800 billion in 2026, and continue into 2027, with total CAPEX by these companies projected to reach USD1.1 trillion, representing a further 37% increase from our 2026 estimate.
• Users are more cost-conscious, the focus is shifting from maximising token consumption to optimising the value of each token used, which could potentially slow down CAPEX in 2027.
• We believe the market is overly concerned about the monetisation of AI CAPEX. The ratio of cloud revenue to CAPEX—which measures how much cloud revenue is generated relative to capital invested—has been trending upwards across all hyperscalers, indicating that revenue growth is catching up with capital spending.
• Maintain 3.5 Stars (Very Attractive) Rating for the Digital Economy (Internet) Sector with target price of USD 68

The financing wave is in full swing. We expect that, throughout the second half of 2026 and into 2027, an increasing proportion of AI-related CAPEX will be financed through debt. Debt issuance is increasing in both scale and geographic diversification, including large non-USD investment-grade issuances by hyperscalers and debut euro-denominated high-yield issuances by data centre operators.

Alongside traditional bond financing, hyperscalers are increasingly adopting off-balance-sheet financing structures such as SPVs and joint ventures to fund data centre expansion. Under these arrangements, third parties own and finance the assets, while hyperscalers commit to long-term capacity agreements, preserving balance-sheet flexibility despite elevated investment.

Overall leverage remains manageable across most hyperscalers, although smaller cloud providers such as Oracle could face greater investor scrutiny as borrowing requirements increase.

This financing shift also makes hyperscalers increasingly sensitive to changes in interest rates, reinforcing our view that AI CAPEX growth could become more disciplined over time.

Figure 2: Balance sheets remain fairly healthy for most

Figure 3: On/Off balance sheet financing by hyperscalers

Scope: Microsoft, Amazon, Alphabet, Meta, Oracle. Bulk of capex still funded by internal cash flow — items shown are external financing layered on top. Amounts are announced/headline figures; SPV debt-vs-equity splits and draw timing are private and approximate. Sources: company filings, Moody's (Feb 2026), Bloomberg / FT / press. Q2 2026 issuance not yet fully reported.

While we expect AI CAPEX to continue increasing over the next two years, largely driven by higher memory prices—a key component with very limited substitutes—we believe that CAPEX excluding memory will begin to slow at a faster pace, which could potentially lead to CAPEX being lower than expected, for the following reasons:

a)      Increasing likelihood of interest rate hikes. The probability of further rate hikes has increased significantly following a series of events, including the US-Iran war, which disrupted global oil supplies, and more hawkish comments from the US Federal Reserve. As hyperscalers, once flush with cash, gradually deplete their cash reserves and increase leverage to finance ambitious data centre build-outs, they are becoming increasingly exposed to higher borrowing costs.

b)     Shifting in AI demand. As shown in Figure 4, the price per token has fallen by more than half. However, token usage has increased by approximately 4.5 times, resulting in the overall bill more than doubling. Another illustration can be seen in Figure 5, where the decline in the Silicon Data LLM Expenditure Index—which measures the price and mix of LLM token usage—contrasts with the trend observed on OpenRouter, where overall LLM usage continues to increase. One possible explanation is a shift towards lower-cost models.

This is supported by recent developments, including Amazon removing its token leaderboard, Microsoft cancelling Claude Code subscriptions, and multiple reports of unexpectedly large token bills. As users become more cost-conscious, the focus is shifting from maximising token consumption to optimising the value of each token used. A draft email does not require frontier-level reasoning, whereas a customer-facing financial model does. Organisations that match model complexity to the task at hand will have fundamentally different cost economics from those that run every workload on the most powerful model available.

Importantly, this changes the nature of AI infrastructure demand. As enterprises increasingly optimise model selection according to task complexity, incremental AI workloads are likely to shift towards smaller or lower-cost models that require less compute intensity.

As a result, not every AI workload will require frontier-level AI infrastructure. Over time, this could moderate demand for the most expensive compute clusters and slow the pace of AI CAPEX growth outside the most critical workloads.

c)      Bottleneck remains in construction & power. We continue to believe that power availability and construction capacity remain the most critical constraints on future data centre expansion. More than half of the planned data centre pipeline for 2026 could face delays. While on-site generation may ease power shortages, transformer supply constraints and labour shortages are likely to remain key bottlenecks.

Read more: Data Centre Infrastructure 2026: Where are we now, and Where are we heading to

If AI infrastructure spending becomes more disciplined rather than indiscriminately expansionary, we expect to market performance to be uneven. Companies previously penalised for elevated AI CAPEX—including Microsoft, Meta, Amazon and selected software names—could see meaningful valuation recovery. 

Figure 4: Average cost per token is falling, but token use is rising faster

Figure 5: Spending on token decreases

Figure 6: % of memory on capex

Free cash flow (FCF) may deteriorate further but should remain manageable. CAPEX is consuming an increasing share of operating cash flow across the major hyperscalers. CAPEX as a proportion of cash flow from operations has been rising across the hyperscalers, with companies such as Amazon and Oracle both reporting negative FCF in the latest quarter. We see limited scope for a meaningful recovery in FCF over the next few years, given their commitments to continued AI investment. However, we expect FCF generation to gradually improve towards late 2027 as the pace of CAPEX growth moderates.

At the same time, we believe the market is overly concerned about the monetisation of AI CAPEX. The ratio of cloud revenue to CAPEX—which measures how much cloud revenue is generated relative to capital invested—has been trending upwards across all hyperscalers, indicating that revenue growth is catching up with capital spending. In the latest quarter, every major hyperscaler generated higher cloud revenue for each dollar invested in CAPEX.

The bigger issue, in our view, and one that has contributed to market disappointment, is the sharp slowdown in share buybacks. The more immediate headwind is weaker shareholder returns. Historically, strong FCF supported consistent buybacks, but capital is now being redirected towards AI infrastructure, resulting in slower repurchase activity and removing an important source of valuation support.

We do not expect this to be a structural shift. As AI infrastructure spending normalises over time, stronger FCF generation should provide room for share buyback programmes to recover.

Figure 7: CAPEX to Cash Flow from Operations

Figure 8: Cloud Revenue vs CAPEX

AI IPO – The next reality check

As the IPO window remains highly receptive following SpaceX's historic debut, several multi-billion-dollar technology and fintech leaders are actively advancing their listing preparations. Among the most anticipated are OpenAI (the company behind ChatGPT) and Anthropic (the company behind Claude), the two largest AI large language model (LLM) companies in the US.

Both OpenAI and Anthropic have recently filed for IPOs, with their expected listing timelines in the second half of 2026 (OpenAI reportedly postponed to next year). Their latest valuations have reached USD862 billion and USD965 billion, respectively, potentially surpassing the trillion-dollar mark before listing. These are among the most highly anticipated IPOs in recent years, given that both companies are at the forefront of the AI revolution. Both companies have continued to deliver exceptional growth in users and recurring revenue, supporting valuations of USD862 billion for OpenAI and USD965 billion for Anthropic.

However, these IPOs will also put both companies to the test. As privately held companies, investors have had limited visibility into their financial positions. Consequently, market attention has largely focused on publicly disclosed metrics such as weekly active users and ARR. Beneath the surface, however, neither company has yet achieved profitability, with OpenAI continuing to report substantially deeper losses, according to the latest Financial Times reports. Anthropic appears to be on a much stronger trajectory. The company is expected to more than double its quarterly revenue to approximately USD10.9 billion in the second quarter, an exceptional pace of growth that could enable it to achieve an operating profit for the first time.

In addition, based on their latest valuations relative to ARR, OpenAI is trading at a significantly higher multiple than Anthropic (33x versus 19x), as well as the broader software sector average of around 8x.

Implication for the digital economy sector. These companies are closely intertwined with the broader digital economy, particularly the major hyperscalers. A significant proportion of hyperscalers' cloud revenue is generated from AI companies such as OpenAI and Anthropic. As a result, any developments affecting these companies could have meaningful spillover effects across the broader AI value chain.

At present, we believe companies with greater exposure to OpenAI—such as Oracle, SoftBank, and Microsoft—face relatively higher risks than those with stronger ties to Anthropic.

Figure 9: OpenAI Financials (per FT)

Software’s defence

The software sector has experienced another round of sell-off following the strong rebound in May, as concerns over AI disruption resurfaced.

While our investment thesis for the software sector remains intact—namely that share prices are oversold relative to the current visibility of AI disruption. Recent partnerships among several major software companies have highlighted how incumbents are working together to defend their market share. On 18 June, Google, Microsoft, Salesforce, Snowflake, ServiceNow, and other companies jointly announced their support for a new software standard called Agentic Resource Discovery (ARD). The agreement aims to enable corporate employees to access all of a company’s AI tools and capabilities through a single application, thereby establishing a unified entry point for enterprise AI usage. Notably, neither Anthropic nor OpenAI appeared on the initial list of supporting parties.

Under ARD, AI applications such as GitHub Copilot, Gemini and Salesforce CRM can automatically discover and invoke compatible enterprise AI services, reducing the need for users to switch between multiple applications.

This initiative is designed to defend incumbents’ market positions against LLM providers such as Anthropic and OpenAI, which aim to position Claude and ChatGPT as the primary gateway through which employees access AI tools and enterprise applications, rather than as merely one component within another ecosystem. Incumbents are pursuing this strategy by leveraging governance, security, and reliability advantages that current AI models lack.

Rather than being displaced, leading enterprise software vendors are increasingly positioning themselves as the orchestration layer through which AI is deployed inside organisations.

2H2026 Outlook: Sentiment to remain weak; Bargain hunting opportunities emerge

Overall, the first half of 2026 has been uneven due to weak sentiment. However, we believe that fundamentals—particularly for large technology companies—remain largely intact. While cash flows may deteriorate further, most big tech firms retain sufficient financial capacity to sustain elevated levels of investment.

Sentiment is likely to remain weak in 2H26 as hyperscalers continue investing aggressively, but we believe markets have become overly pessimistic relative to earnings fundamentals, which creates opportunities for selective bargain hunting.

Selectivity is required. We continue to prefer companies with strong balance sheets, namely the large hyperscalers (Alphabet, Microsoft, Amazon, Meta), as well as leading software companies and others within the data infrastructure and cybersecurity sectors.

In our view, current valuations already discount much of the near-term financing and free cash flow concerns, while earnings expectations remain intact. 

For product recommendation, we continue to prefer the Invesco NASDAQ Internet ETF (NASDAQ:PNQI) for passive approach and Fidelity Global Technology A-ACC-USD and  Eastspring Investments Unit Trusts – Global Technology SGD   for active approach.

We expect earnings for the Digital Economy (internet) sector to grow by 11% and 13% in the next 2 years, which translates into the target price of 68 (54% upside potential) by applying our fair P/E ratio of 25x. 

Table 1: Valuation Table

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.