AI’s next phase: Investing beyond the bottlenecks

Key points:

• Cheaper AI strengthens the adoption story. Falling model and inference costs should allow AI to spread across more businesses, employees and everyday applications. That is positive for overall usage, even if it puts pressure on the pricing power of model developers and some hardware suppliers.
• More intentional AI capex does not mean the end of hardware demand. Hyperscalers will keep spending, but they will increasingly optimise what hardware they buy, where they deploy it and what return it generates. The winners may shift from broad hardware exposure towards specialised bottlenecks and efficient infrastructure.
• The KOSPI can rebound after deleveraging, but cleaner positioning is not a new memory cycle. Korea’s market mechanics have improved, yet a return to its previous pace of gains requires further earnings upgrades. The main risk is that memory expectations peak before underlying demand does.

The first phase of the artificial-intelligence trade was built around scarcity.

There were not enough advanced chips, high-bandwidth memory, networking components, power connections or data-centre capacity to satisfy demand. Investors did not need to know exactly how AI would be monetised. It was enough to own the companies supplying the bottlenecks.

That phase is evolving.

AI infrastructure spending remains strong, and hardware will remain essential. But the next phase will be less about who can build the largest model or secure the most advanced chip—and more about who can make AI affordable, efficient and useful.

For investors, this is a shift from AI scarcity to AI economics.

Why efficiency is moving to the centre of AI

The AI industry is not becoming less important. Parts of it are becoming more commoditised.

A growing number of companies can now access capable models through open-source systems, cloud platforms and application-programming interfaces. Businesses do not always need to train their own model, build a data centre or buy the most advanced GPU.

Most enterprises simply need AI that is:

• sufficiently accurate;
• easy to integrate;
• secure and reliable;
• fast enough for the task;
• affordable at scale.

That changes the investment question.

The first phase was about maximising capability. The next phase will increasingly be about achieving the required capability at the lowest possible cost.

Two forces are accelerating this transition: Chinese competition and increasing scrutiny of hyperscaler spending.

China is competing through price and efficiency

China may not yet lead at every point of the advanced semiconductor frontier. But it is increasingly competitive in model architecture, efficiency, open systems and price.

Alibaba recently introduced its largest AI model to date, while DeepSeek released an ultra-low-cost model priced far below many competing systems. The releases reinforce China’s strategy of using open and affordable models to attract developers and accelerate adoption.

This matters because lower model prices can rapidly reduce the cost of deploying AI.

A retailer does not need the most powerful model in the world to improve product recommendations. A bank may not need frontier-level reasoning for every customer-service query. A manufacturer may require reliable anomaly detection rather than the most sophisticated general-purpose intelligence available.

As the performance gap between affordable and premium models narrows, fewer users will pay for maximum capability when a cheaper model is good enough.

That does not eliminate demand for advanced hardware. Frontier models, scientific research and hyperscale training will still require the most powerful chips available.

But it could change the mix of demand.

Growth may increasingly come from:

• smaller and specialised models;
• customised accelerators;
• efficient inference chips;
• conventional memory used intelligently;
• edge computing;
• software that reduces computing requirements;
• cloud services that offer a choice of models and hardware.

Research into AI inference economics suggests that many workloads do not use all the computing power bundled into leading GPUs. Some inference tasks are constrained more by memory capacity and data movement than by raw computational power.

The implication is not that advanced GPUs become obsolete. It is that not every AI workload requires the most expensive hardware configuration.

AI capex is becoming more intentional

Questions about AI capex are sometimes interpreted too simplistically.

Investors worry that hyperscalers becoming more disciplined will mean an immediate fall in semiconductor and data-centre demand. That is not necessarily the right conclusion.

Hyperscalers are still expanding infrastructure because AI usage is rising and existing capacity remains constrained. Consensus estimates for capex by Microsoft, Alphabet, Amazon, Meta and Oracle increased from around $485 billion at the beginning of 2026 to approximately $730 billion by July.

The shift is not necessarily from spending to not spending. It is from spending at any price to spending with clearer economic intent.

That makes AI capex more selective, not necessarily smaller.

The overall spending cycle may continue, while growth rates slow from extraordinary levels. UBS estimates cited by Reuters suggested hyperscaler capex could grow by 76% in 2026, before slowing to 25% in 2027 and 6% in 2028.

For hardware investors, that distinction matters.

A slowdown in capex growth is different from an outright decline in capex. But equity markets react to changes at the margin. Hardware companies priced for continuously accelerating spending can still fall when growth merely becomes less exceptional.

From bigger models to better economics

Lower costs could create a powerful feedback loop.

As AI becomes cheaper, businesses can use more of it. A company that previously restricted AI to a small group of developers may roll it out across customer service, marketing, research, logistics and internal operations.

The cost of each individual task falls, but the number of tasks performed increases.

This is the bullish counterargument to concerns about hardware demand: efficiency can stimulate so much additional usage that overall computing requirements continue rising.

But value may be distributed differently.

When models are scarce and expensive, the model itself captures much of the economics. When models become easier to access and substitute, value can shift towards companies controlling:

• distribution;
• proprietary data;
• customer relationships;
• cloud infrastructure;
• business workflows;
• trusted applications.

That is why digital platforms may become more important.

Alibaba, Tencent, Microsoft, Alphabet and Meta do not only provide models. They can embed AI into cloud services, advertising, commerce, workplace software, search, social media and existing customer ecosystems.

They do not have to find an entirely new customer every time they launch an AI product. They can distribute AI through services customers already use.

Is the memory cycle close to a peak?

The shift towards efficiency creates an important question for Korea.

The KOSPI has been one of the clearest beneficiaries of the AI hardware and memory cycle. Samsung Electronics and SK Hynix have provided investors with exposure to booming demand for server memory and high-bandwidth memory.

After the recent correction, many investors are asking whether the market can return to its previous style of gains now that crowded positioning has been reduced.

There is a credible tactical rebound case.

J.P. Morgan analysts recently estimated that the leveraged ETF unwind was complete and that hedge funds were roughly 90% through the deleveraging process.

Cleaner positioning reduces the probability that further forced selling overwhelms fundamentals. It also gives investors greater confidence to rebuild exposure after a sharp correction.

But deleveraging resolves a market-structure problem. It does not automatically restart the earnings cycle.

The previous KOSPI rally was powered by several forces working together:

• rising memory prices;
• repeated earnings upgrades;
• growing AI demand;
• heavy foreign inflows;
• increased leverage and momentum buying.

The leverage and momentum components can rebuild. The harder question is whether earnings expectations can continue rising at the same pace.

This is where the risk of a cyclical peak emerges.

It is important to define what “peak” means. It does not necessarily mean memory prices are about to collapse or that AI demand has peaked.

TrendForce still expects the DRAM market to remain undersupplied, with demand growth potentially continuing to exceed supply growth into 2027.

The greater risk is a peak in:

• the rate of memory price increases;
• the pace of earnings upgrades;
• investor expectations;
• valuation expansion;
• the willingness to pay for distant profits.

Equities often peak before industry revenues or profits do. A company can continue reporting record earnings while its share price struggles because those records were already expected.

The memory outlook is also becoming more differentiated.

High-bandwidth memory and advanced server DRAM could remain structurally tight because AI systems require vast quantities of fast memory close to the processor.

Commodity DRAM and NAND are more cyclical. They are more exposed to capacity additions, Chinese competition and weaker demand from smartphones, personal computers and consumer electronics.

This makes it difficult to treat memory as one uniform investment theme.

How to position for AI’s next phase

The conclusion is not simply to sell hardware and buy software. Both categories contain potential winners and losers.

The stronger relative investment themes are:

• AI applications and digital platforms over broad hardware exposure. Companies that can distribute and monetise AI may benefit more directly as model costs fall.
• Profitable software over generic software. Prefer businesses with proprietary data, critical workflows, security, high switching costs and visible AI revenue. Avoid assuming every software company benefits from AI.
• AI adopters over AI storytellers. Look for businesses already using AI to increase sales, improve margins or reduce operating costs—not those merely discussing future potential.
• Efficient infrastructure over maximum computing power. Custom silicon, inference optimisation, networking, power management and cooling could become increasingly important as buyers focus on cost per task.
• Advanced memory over broad commodity memory—but with greater valuation discipline. HBM and server memory remain structurally attractive, while commodity DRAM and NAND are more exposed to cyclical and supply risks.
• China platforms and applications over export-dependent Chinese hardware. Lower-cost models favour companies with domestic distribution, cloud ecosystems and large customer bases. Hardware exporters face higher geopolitical risk.
• Selective Korea over a broad KOSPI overweight. Cleaner positioning supports a tactical recovery, but sustainable gains require continued earnings delivery and broader market leadership.
• Monetising hyperscalers over capex-dependent suppliers. Prefer cloud and platform companies demonstrating revenue growth and improving returns from AI investment over suppliers relying solely on another upward revision to capex.
• Non-technology AI beneficiaries over concentrated semiconductor exposure. Financials, healthcare, industrials, logistics and consumer platforms could benefit as cheaper AI improves productivity.

Risks to the view

Hardware demand could reaccelerate

Lower costs may produce a much larger increase in AI usage. If token consumption rises faster than efficiency improves, demand for chips, memory and networking could strengthen again.

Memory shortages could persist

Advanced-memory supply remains constrained. HBM qualification, manufacturing complexity and limited packaging capacity could sustain stronger pricing and profits for longer than expected.

Software valuations may already reflect the rotation

Investors could move too quickly from semiconductors into expensive software stocks. Strong underlying growth does not protect investors who pay an excessive valuation.

AI may undermine traditional software pricing

Cheaper models can help software companies, but they also lower barriers to entry. Companies relying on easily replicated features or per-user subscription pricing may face pressure.

China’s price war may destroy margins

Affordable models can accelerate adoption while making it difficult for model developers and cloud providers to earn attractive returns.

Hyperscaler spending may continue to surprise higher

Demand remains strong, and companies may conclude that underinvestment presents a greater strategic risk than overspending. That would favour the existing hardware leaders.

Geopolitical restrictions may widen

Further US controls or Chinese retaliation could affect chips, optical components, materials, cloud access and overseas sales. Policy can overwhelm company fundamentals.

KOSPI deleveraging may not be fully behind us

Positioning has improved, but volatility can return if memory shares fall sharply, foreign investors reduce exposure or leveraged products rebuild too quickly.

Bottom line: The next AI trade is about returns

The AI theme remains structurally powerful. But it is becoming broader, more competitive and more demanding.

The first phase rewarded companies that controlled scarce infrastructure. The next phase will reward those that lower the cost of intelligence and turn it into revenue, productivity and cash flow.

Hardware remains necessary, but investors should no longer assume that every additional dollar of AI spending flows to the same suppliers—or that the most advanced hardware is required for every task.

The investment framework is therefore shifting:

Applications over AI promises; efficient infrastructure over hardware at any price; advanced memory over commodity exposure; China’s digital ecosystems over export-dependent suppliers; and companies generating returns from AI over those merely spending on it.

AI’s next phase is not about the end of the infrastructure buildout.

It is about moving beyond bottlenecks—and deciding who can make intelligence economically useful.