Digital Economy (Internet) 4Q26 outlook: Sailing through the tides; Maintain 3.5 Stars

3Q2026 recap: Sentiment turned, fundamentals delivered

In our 2H2026 outlook, we argued that markets had become overly pessimistic relative to earnings fundamentals, creating opportunities for selective bargain hunting. That call has played out faster than we expected. Since our last update at end-June, the digital economy (internet) sector has gained approximately 16%, as measured by the Invesco NASDAQ Internet ETF (NASDAQ: PNQI), while the iShares Expanded Tech-Software Sector ETF (BATS: IGV) has gained about 18%, reaching our target price.

The turning point was the 2Q26 earnings season, which provide evidence that AI CAPEX is translating into revenue and supporting margins at scale. Software staged a strong recovery on the back of solid results from Salesforce, Autodesk and CrowdStrike.

The macro backdrop, on the other hand, moved against the sector. With oil prices above USD100 per barrel amid the Iran conflict, the US Federal Reserve raised rates by 25 bps to 3.75%–4.00% on 16 September, alongside upcoming US midterm election. Rising bond yields hit the most leveraged corners of the AI trade hardest. In short, the market is no longer rewarding AI exposure indiscriminately; it is rewarding visible returns and punishing stretched balance sheets. We expect this divergence to widen in 4Q26.

AI-led growth remains intact, with clearer visibility on returns

Demand indicators, including rising weekly Token consumptions on OpenRouter, elevated rental prices for older GPUs and reported AI compute price increases at Nebius, suggest that AI demand remain robust.

Increasing Open-weight models and agentic AI pushes further demand for AI. The Muse app overtook ChatGPT as the leading iOS free app in the US on Friday, and racked up 730,000 downloads over about five days after it launched on Sep. 8.

Read more: Agentic AI: Where are the opportunities

We are seeing two particularly encouraging signals from the latest quarter.

1.      Custom chips are protecting margins. The latest quarter demonstrated how hyperscalers can continue to invest heavily in CAPEX while maintaining operating margins, with customised chips helping to improve compute economics. Amazon's custom silicon business, comprising Trainium and Graviton, has surpassed US$25 billion in annualised revenue and continues to grow at a triple-digit rate. Trainium commitments are also expanding beyond anchor customers such as Anthropic and OpenAI to enterprises such as Uber.

Despite being in the heaviest investment cycle in its history, Google Cloud's operating margin expanded to 35.6% from 20.7% a year ago, while AWS's margin increased by more than 500 bps year-on-year to around 39%. The cost advantage can be further illustrated in the figure below.

Both Meta and Microsoft are also increasingly focusing on developing and deploying their own customised chips, which we expect to see further developments expected in the coming quarters.

Figure 1: Estimated Cost of Data centre per GW

2.      Payback periods are becoming visible. While free cash flow among hyperscalers continues to decline as expected, we now have clearer visibility on the payback periods for CAPEX investments compared with the previous quarter. These typically stand at around three to four years for the higher-margin hyperscalers.

Amazon disclosed that its AI servers break even in just under three years, compared with useful lives of five to six years. Moreover, most of its capacity is contracted for five years or more before the chips are even purchased. Meanwhile, Google has indicated a two-year payback period for its AI servers, while its TPUs are expected to have a much faster payback period compared with GPUs.

From here, we dig a little deeper into the various estimates.

CSPs primarily monetise compute through GPU rentals (Infrastructure as a Service, or IaaS) and model APIs (Model as a Service, or MaaS). Morgan Stanley estimates that, using a 1GW NVIDIA Blackwell Ultra GB300 data centre as an example, renting out GPUs (IaaS) at US$7–10 per GPU-hour with 75% utilisation could generate median annual revenue of US$23 billion, an incremental EBIT margin of 67%, and a return on invested capital (ROIC) of 31%. This implies a payback period of slightly more than three years.

On the other hand, selling tokens through model APIs using the same infrastructure could lift median ROIC to 46%, assuming 65% of capacity is used for inference and token pricing remains at US$1.75 per million tokens. The downside is that if API prices fall faster than CSPs can reduce their internal compute costs, this would put pressure on their gross margins.

Nevertheless, the current revenue mix of major CSPs remains heavily weighted towards GPU rentals. Under these assumptions, estimated payback periods are around three years, with average FCF turning positive in 2H2027.

Figure 2: Estimated ROIC for Model Enabled API 

Figure 3: Estimated ROIC for GPU Rental

Revising CAPEX projections higher for 2027; downside surprise remains possible

Given clearer visibility on demand, driven by trends such as agentic AI and cheaper models, combined with increasingly visible payback periods, we believe hyperscalers have greater justification to increase CAPEX. As such, we have revised our CAPEX projection for 2027 upwards to US$1.15 trillion, representing a 39% increase from our 2026 projection, with growth moderating to 4% in 2028.

However, the broader trend remains clear: we believe CAPEX growth will be highest in 2026, with growth starting to moderate in 2027. Nevertheless, the possibility of a downside surprise remains, particularly if AI demand or the returns on incremental CAPEX fall short of current expectations.

We continue to see the potential for lower-than-expected capital expenditure next year, as users shift from maximising token consumption to optimising the value of each token used. Not every AI workload requires the most advanced semiconductor chips, which over time could help moderate CAPEX requirements per workload, although lower compute costs may also stimulate greater usage.

We have not seen this happening yet – 2026 CAPEX guidance was raised again last quarter, although part of the increase reflects higher memory prices rather than additional capacity. However, stronger competition from Chinese AI labs, which rely on less advanced domestic chips, suggests the risk is still there.

At the same time, while delay in data centre build-ups remain manageable up until now (Est 10-20% of total planned capacity), we are also highly focused on the build up to the election and political risk to data centre expansion from the executive branch if the Democratic Party wins. Potential delays in key data centre clusters such as Project Jupiter are also under our close monitoring.

Figure 4: CAPEX estimates

Challenges ahead leads to greater divergence

While the demand story is intact, funding activities through debt and equity, as well as free cash flow health, will be the key differentiating factors moving forward. Most big tech firms are facing greater pressure on free cash flow after massive spending over the past year. Alphabet raised USD80 billion in equity in June on top of its debt issuance, Meta has taken on new debt and paused share buybacks, and Amazon's trailing free cash flow has turned negative.

At the weaker end of the spectrum, Oracle's CAPEX of USD28.5 billion last quarter exceeded its total revenue, driving free cash flow to negative USD5.4 billion, with total debt of roughly USD125–130 billion. Neoclouds are more stretched still – CoreWeave carries net debt of more than 10x EBITDA.

Off balance sheet commitments add another layer. Though many of these are largely supply contracts funded from operating cash flow, these commitments are growing increasingly larger and many have exceeded total liabilities of the companies, which warrants caution.

We expect increasing fund-raising activities through debt and equity in the coming quarters, which means more dilution and higher interest costs just as rates are moving up. Within cloud, we therefore prefer the large hyperscalers with strong internal cash generation over neoclouds, Oracle and newer entrants such as SpaceX, whose AI compute ambitions remain capital-intensive and more dependent on external financing.

Figure 5: Big tech’s off-balance sheet commitments.

Software: Time to be more selective (Downgrade)

IGV has gained about 18% since our last update, reaching our target price. After this round of re-rating, we believe the valuation gap has narrowed and it is time to be more selective. We prefer software companies that act as a system of record – owning the authoritative enterprise data that AI agents need to read and write – over workflow-oriented software, whose tasks may be more easily replicated by AI agents where proprietary data, governance and integration provide limited protection.

Recent share price moves illustrate this divergence. Salesforce jumped 22% on 27 August – its second-best day on record – after announcing an expanded partnership with Anthropic (“Claudeforce”) alongside its 2Q FY27 results. The tie-up brings live Salesforce CRM data and workflows directly into Claude, positioning Salesforce's governed customer data as the layer AI models plug into rather than a system they bypass. Autodesk, by contrast, fell almost 17% in the first nine days of September despite beating and raising guidance, after OpenAI's GPT-6 Astra – launched on 3 September – showed it can operate design software and reconstruct objects as editable CAD models from a handful of photos.

As such, with frontier models continuing to iterate at a very fast pace, dark clouds above the sector have yet to fully clear. With valuations having recovered over the past quarter, we believe that the road ahead requires more selective approach despite the sector still trading at a discount to history levels. We like companies such as NOW, ADBE, CRM.

Figure 6: Valuation 

Cybersecurity: AI adoption supports demand for stronger protection

We remain positive on cybersecurity, as AI adoption increases the need for stronger protection. Every AI agent deployed can introduce a new digital identity with access to sensitive systems, widening the attack surface that enterprises must defend. We view cybersecurity spending as among the least discretionary within IT budgets, which should provide resilience if the macro environment deteriorates.

Read more: Cybersecurity earnings are strong. The bigger opportunity may still lie ahead

Downside risks to monitor

Rate hikes are now a global regime. The Fed, ECB and Bank of Japan have all raised rates in September, the latter for the third consecutive month, while the Bank of England held at 3.75% with three members voting for a hike. Central banks are responding to inflation data that has already arrived rather than to forecasts, and markets are pricing in further tightening. Higher rates weigh on long-duration growth valuations and raise financing costs for an increasingly debt-funded AI build-out. We advise positioning in names with strong balance sheets, resilient earnings and high returns on equity.

Slower-than-expected data centre build-out. Power availability, transformer supply and construction labour remain the key bottlenecks, and delays would push out revenue recognition across the AI value chain.

Circular financing. A growing share of AI demand is funded by suppliers investing in their own customers – chipmakers and hyperscalers taking stakes in AI labs, which in turn commit to spending on their chips and cloud capacity. This can magnify drawdowns if any link in the chain weakens. The upcoming AI IPOs will be an important reality check: Anthropic's run-rate revenue reached USD65 billion in July and its listing is now expected in late October or November, while OpenAI, tracking above USD40 billion, has reportedly pushed its IPO to 2027.

Sub-sector	View	Rationale
Cloud	Positive	Benefits from strong demand for AI compute. Prefer big tech firms over neoclouds, Oracle and SpaceX.
Cybersecurity	Positive	Sustained demand as AI adoption increases the need for stronger protection.
Software	Selective	Sector has re-rated significantly; AI disruption risk remains. Prefer system of record over workflow software.
Broadline retail	Neutral	Higher-for-longer inflation dampens consumer appetite.