With AI-related stocks making up around 50% of US and emerging market (EM) benchmarks today, it is impossible to form a view on global equities without taking a view on AI. The tech sector has generated half of 2026’s global earnings growth and drove almost three quarters of global equity returns in the first 10 months of the year.
2027 looks set to be a pivotal period. The ‘build’ phase of the AI rollout has already seen many companies bank incredible gains, with eye-watering levels of investment from a small number of US hyperscalers generating incredible profits for the providers of technology inputs. As we now enter a period where markets focus less on the build-out, and more on the rollout, the opportunity set is evolving. We will start to have a clearer picture on the return ahead for all this capital expenditure (capex), and the relative winners and losers in the space. The path ahead is still likely to be up, but may be bumpier.
AI stocks
Ultimately, the destiny of all parts of the AI value chain ties back to two key questions. The most important issue for the broad ecosystem is whether corporates and consumers are willing to pay for new technologies. At a more granular level, the question is which companies will be able to deliver the category-killer products that result in the widest moats, in turn supporting sustainably stronger margins than their peers.
To address these questions, we find it helpful to break down the AI ecosystem into several groups (Exhibit 14). Tech inputs are involved in designing and building the chips, resource inputs provide energy and other critical resources, the hyperscalers buy those inputs as part of building AI infrastructure, and the model providers then rely on this infrastructure to train and run large language models. Finally, there are the customers, whether businesses or consumers, whose payments ultimately need to be sufficient to feed the entire value chain.
If the hyperscalers build it, will the customers come?
In recent years, five US hyperscalers – Alphabet, Amazon, Meta, Microsoft and Oracle – have greased the wheels for all parts of the AI ecosystem, providing an incredible revenue stream for the providers of tech inputs as part of data centre build-outs while also making huge investments in the AI model providers. Yet the bill for that spending is now starting to bite, with free cash flow turning negative and companies increasingly tapping the debt market for funding.
The challenge is that for every additional dollar of capex, the revenues required to generate a positive return on investment rise too. Expectations for 2027 capex have risen by an incredible USD 499 billion for the same five US hyperscalers since January 2026 (see Exhibit 15). So far, bond investors have shown that they are broadly willing to shoulder the burden, although the recent pressure on the valuations of the smaller US ‘neocloud’ providers highlights that the bond market is becoming more discerning, as we discuss in Bonds look attractive as pressures ease, but be active.
Data centre build-out
The strong performance from the Democrats at November’s midterms has raised fears that data centre build-out will now face major opposition in the form of new regulation, or even an outright pause. In our view, the election result may limit further upside to capex plans but it is unlikely to lead to major downgrades. Data centre build-out will have to be done more thoughtfully, and with potentially greater costs associated, but politicians will ultimately have to balance their campaign promises with the revenues and jobs that are created by housing data centres.
While some stabilisation in capex plans may be welcomed by hyperscaler shareholders, it would have the opposite impact on the providers of tech inputs for which the hyperscalers are the key customers. Investors are already wise to the fact that this year’s earnings boom across semiconductor and memory stocks cannot continue indefinitely, as evidenced by the sharp decline in valuations. Yet 2027 consensus forecasts still anticipate earnings growth of 78% and 46% for US semiconductors and Korean tech names, respectively. Any reduction in future capex expectations would likely lead to earnings downgrades.
New entrants to the sector should also be monitored carefully. The key reason that semiconductor and hardware companies have been able to deliver such strong margins is that they generally operate in an oligopoly. Take Samsung and SK Hynix as an example: the pair are estimated to account for 80% of the high-bandwidth memory (HBM) market and together represent close to 60% of MSCI Korea. When a Chinese rival memory maker made a technological breakthrough in October, the sharp falls in Samsung and SK Hynix share prices demonstrated how quickly the market’s perception of company ‘moats’ can shift. In highly concentrated markets, these shifts can lead to extreme volatility at the index level (see Exhibit 16).
Token volumes and AI adoption
It is not only memory makers that face rising competition from China. While the most popular US large language model providers, OpenAI and Anthropic, are still seeing steady growth, the arrival of cheaper or more efficient models (many of which are from China) has led to a huge increase in competition and an associated decline in corporate spend per AI token (see Exhibit 17). With chief financial officers (CFOs) increasingly cognisant of the rising bills associated with soaring token volumes, price is becoming a bigger factor when choosing the AI tool for the job.
Of course, it is quite plausible that there will ultimately be room for multiple model providers across different tiers, just as neither the Ferrari nor the Ford Fiesta dominates today’s roads. The challenge from an investment perspective, however, is twofold. First, if today’s largest players are ultimately going to have to accept a smaller share of corporate spend, it’s not clear that they will need all the computing capacity that they have already committed to, with downstream implications for the rest of the supply chain. Second, given the current ease with which companies can redirect their queries to different models, the risk of being leapfrogged by a competitor’s model release reduces the visibility of future earnings.
Tech exposure requires careful management
While each part of the AI value chain faces different challenges, the key factor connecting all AI-related companies is the rate of economy-wide adoption. Ultimately, if demand for AI from individuals and businesses disappoints, all parts of the ecosystem could be hit. Our base case for 2027 remains positive, with adoption metrics on the rise and corporate spend accelerating. Forecasts of end-state adoption remain highly speculative, but with penetration rates still low, companies still getting to grips with a variety of use cases, and agentic AI gaining momentum, we struggle to see why adoption would decline significantly over the coming quarters without a shock to the economy.
We therefore remain constructive on the ability of AI spending to drive further gains across equity markets in 2027. Yet with several major equity benchmarks exhibiting incredibly high levels of AI-related concentration, there are three key actions for investors to consider.
First, investors must manage single-stock risk very carefully. Exhibit 18 highlights how several major equity benchmarks allocate substantial index weights to individual companies whose prospects are closely tied to the AI build-out. Passive investors are therefore banking on tomorrow’s leaders being the same as today’s, when perhaps the most important lesson of the dot com boom and bust is that the ultimate winners of big investment cycles can take a very long time to emerge.
Second, given ongoing uncertainty about where company moats will be strongest, diversification across the different parts of the AI value chain is also critical. Volatility across the various tech subsectors has risen sharply in 2026, and we expect this to remain a feature of this trickier phase of AI’s expansion (Exhibit 19). One implication is that the more diversified mix of AI companies in US benchmarks is relatively appealing compared with the narrower, hardware-focused exposure in emerging markets, although we recognise that EM valuations already embed a lot of ‘bad news’ in the price.
Finally, investors must assess whether their portfolios’ overall exposure to AI is consistent with both their long-term goals and their ability to stomach short-term volatility. Given today’s concentrated markets, a passive global equity allocation likely has close to 50% of assets in companies directly related to AI, before we even consider AI’s influence in areas such as energy, materials and industrials. Fortunately, we see ample opportunities for diversification in some of the less-loved regions around the world, as we turn to next in European stocks can take advantage of AI adoption and diversify risk.
Frequently asked questions
Investors are scrutinising the prospects for the return on AI-related investment increasingly carefully. The article highlights four key items to monitor.
- The revenue hurdle rises with every capex dollar. The key question for the market is whether hyperscaler revenue growth is rising fast enough to justify the higher capex base.
- Watch free cash flow direction, not just absolute spend. For the five US hyperscalers (Alphabet, Amazon, Meta, Microsoft, Oracle), the bill for their spending is now starting to bite, with free cash flow turning negative and companies increasingly tapping the debt market for funding.
- Use the bond market as an early warning signal. Bond investors have so far been broadly willing to shoulder the burden, but recent pressure on the smaller US 'neocloud' providers highlights that the bond market is becoming more discerning.
- Assess how the US regulatory backdrop evolves following the Democrats’ big wins at the midterm election. The direction of travel for AI capex remains unchanged, but greater regulation that forces a reduction in capex plans is highlighted as a risk.
- More than USD 1 trillion in spending is anticipated in 2027, 2028 and 2029 from the five US hyperscalers alone.
- Expectations for 2027 capex have risen by an incredible USD 499 billion for the same five US hyperscalers since January 2026.
- Direction is a two-sided risk for the market. Any stabilisation in capex plans could be welcomed by hyperscaler shareholders (easing the revenue hurdle) but that would hit chipmakers harder (vice versa).
- Capex plans are ultimately underpinned by rising adoption rates and end-user demand.
The article identifies four specific risks to hyperscaler cash flows:
- Rising debt dependency. Hyperscalers are increasingly tapping the debt market for funding.
- A more selective bond market. The sell-off in US 'neoclouds' shows bond investors are becoming more selective about AI-related credit.
- Downstream demand risk from LLM customers.
Grouped by where in the AI value chain the risk sits:
- LLM-provider risk feeds back into infrastructure demand. Cheaper and more efficient models (many from China) are driving token prices down even as volumes rise, compressing corporate spend per token. Easy switching between models means the risk of being leapfrogged by a competitor's model release reduces the visibility of future earnings and future compute demand.
- Hyperscaler capex-vs-revenue risk. This includes more than USD 1trillion per year of planned spend on projects, concerns around negative free cash flow, rising debt issuance and a more discerning bond market.
- China competition across the whole stack. Chipmakers, model providers and data centre builders all face stiff competition from Chinese rivals. The October 2026 volatility in Korean memory stocks after a Chinese technological breakthrough is cited as a prime example of how quickly perceptions of "moats" can shift.
- Regulation, with investors wary that Democrat state governors could look to slow or even pause data centre buildout.
- Adoption shock. The key factor connecting all AI-related companies is the rate of economy-wide adoption. Ultimately, if demand for AI from individuals and businesses disappoints, all parts of the ecosystem could be hit. The article's base case is still positive, barring an economic shock.
