AI’s next phase: From memory winners to efficiency enablers

Key points:

• Memory remains a critical AI bottleneck, but the trade is no longer one-way. AI demand continues to support memory suppliers, but higher memory and storage costs are starting to affect customers across the value chain — from consumer devices to hyperscaler capex.
• The “AI tax” is becoming visible in consumer technology. Recent price increases from Apple, Xbox, Sony and Nintendo show that the memory squeeze is moving beyond semiconductor earnings into hardware pricing, affordability and upgrade cycles.
• The next phase of AI may favour efficiency over pure capacity. As AI becomes more expensive to build and run, market focus is likely to shift toward companies and technologies that reduce inference costs, improve data movement, lower power and cooling needs, optimise software usage, and enable more on-device AI.

Memory has been one of the strongest themes in the AI value chain.

That made sense. AI workloads require more high-bandwidth memory, DRAM, NAND, enterprise SSDs and storage capacity. Memory has moved from being a cyclical semiconductor input to a strategic constraint in the AI buildout.

But that story is now well recognised.

The more important question is: at what point does the memory bottleneck become a demand problem for the rest of the AI ecosystem?

Memory remains structurally supported, but expectations are now high. The next phase of the AI debate may shift toward efficiency — companies and technologies that help the ecosystem do more with less capex, less power, less memory and lower inference cost.

The memory trade is strong, but no longer one-way.

Memory suppliers still have clear tailwinds:

• AI demand remains strong.
• Supply remains tight.
• Pricing power has improved.
• Large customers are trying to secure allocation.
• Memory has become a critical AI infrastructure layer.

But the risk-reward is becoming more balanced.

The risk is not that AI demand disappears. The risk is that higher memory costs begin to change behaviour:

• consumers delay device upgrades;
• hardware makers raise prices or reduce specifications;
• PC, smartphone and console demand becomes more price-sensitive;
• enterprises become more selective on AI workloads;
• hyperscalers scrutinise capex more closely;
• AI software companies face margin pressure if compute costs rise faster than revenue.

This is where a bottleneck story becomes a margin story.

Consumers are starting to pay the AI tax

The AI memory squeeze is no longer only visible in semiconductor earnings. It is now appearing in consumer pricing.

Recent examples show the pressure spreading:

• Apple raised prices across parts of its MacBook and iPad range, citing higher memory and storage costs.
• Microsoft’s Xbox announced global console price increases from August, including higher prices for 512GB and 1TB models, while discontinuing the 2TB model.
• Sony raised PlayStation 5 prices earlier this year, with rising memory-chip costs among the key pressure points.
• Nintendo announced Switch 2 price revisions across major markets, citing changing market conditions rather than memory specifically, but still pointing to pressure on console economics.

Consumers may not always see “AI memory shortage” on the price tag. But they may feel it through:

• higher device prices;
• lower storage or memory in base models;
• delayed upgrade cycles;
• discontinued higher-spec products;
• weaker affordability in gaming, PCs and tablets.

This matters because AI is not only creating new demand. It is also raising the cost of existing demand.

If memory costs rise too far, customers will adapt. That does not invalidate the AI story, but it makes the next phase more selective.

Hyperscalers face a tougher return test

Hyperscalers remain among the strongest players in the AI economy. They have scale, balance sheets, cloud distribution and greater ability to secure supply.

But they are not immune to input-cost inflation.

The key question is becoming more disciplined: Can AI revenue scale fast enough to justify AI infrastructure spend?

Higher memory costs make that equation harder.

Potential responses include:

• absorbing higher costs and accepting margin pressure;
• raising prices for cloud or AI services;
• prioritising higher-return workloads;
• investing in custom silicon;
• improving inference efficiency;
• phasing capex more carefully if returns disappoint.

The AI trade is therefore moving from capacity at any cost to returns on AI capital.

The market is likely to remain supportive of AI leaders, but with greater focus on monetisation, efficiency and free cash flow.

Efficiency becomes the next key AI theme

If memory is expensive, the question is not only who supplies it.

The more important question is: who helps reduce the cost of using it?

That makes efficiency a central theme for the next phase of AI. The company references below are illustrative examples of value-chain exposure, not recommendations.

1. Compute efficiency

As AI shifts from training to inference, cost per query becomes more important. The focus moves from simply adding more chips to improving performance per dollar, per watt and per unit of memory.

Areas to monitor include:

• custom AI accelerators;
• inference chips;
• low-power processors;
• chiplets;
• advanced packaging;
• semiconductor design tools.

Illustrative company references:

• Nvidia — not just GPUs, but also the software stack around inference optimisation, including TensorRT and TensorRT-LLM.
• Broadcom — custom AI accelerators and ASICs for large cloud and AI customers.
• AMD — AI accelerators and AI PC processors, offering an alternative compute platform as customers seek more supply and better cost efficiency.
• TSMC — advanced process technology and packaging that enable more efficient AI chips.

The key point: compute efficiency may become as important as compute capacity.

2. Networking and data movement

AI clusters are only useful if chips, memory and storage communicate efficiently. If data movement is slow, expensive compute capacity can sit underused.

Areas to monitor include:

• high-speed switching;
• optical networking;
• AI interconnects;
• retimers;
• connectivity chips;
• data-centre network optimisation.

Illustrative company references:

• Broadcom — switching, connectivity and custom silicon exposure.
• Arista Networks — Ethernet networking for large AI clusters.
• Marvell — data infrastructure, optical connectivity and custom silicon.
• Astera Labs — connectivity products that help reduce bottlenecks between processors, memory and accelerators.

The key point: as AI systems scale, utilisation matters as much as raw capacity.

3. Power and cooling

AI is increasingly constrained by power and heat. More GPUs and more memory mean more electricity demand, more cooling needs and more pressure on data-centre infrastructure.

Areas to monitor include:

• liquid cooling;
• power management;
• electrical equipment;
• grid infrastructure;
• thermal systems;
• backup power.

Illustrative company references:

• Vertiv — thermal management, power systems and liquid-cooling infrastructure for data centres.
• Schneider Electric — data-centre power, cooling, energy management and infrastructure systems.
• Eaton — electrical equipment and power-management exposure.
• Siemens — grid, electrification and automation exposure tied to data-centre infrastructure.

The key point: AI infrastructure is becoming an industrial-efficiency theme, not just a semiconductor theme.

4. Software optimisation

The most underappreciated efficiency layer may be software.

Enterprises do not only need bigger AI models. They need cheaper, measurable and more controlled AI deployment. As usage scales, the focus shifts to reducing token waste, improving model routing, monitoring costs and proving return on investment.

Areas to monitor include:

• model compression;
• quantisation;
• inference optimisation;
• workload routing;
• AI observability;
• data governance;
• retrieval-augmented generation;
• agent orchestration.

Illustrative company references:

• Nvidia — TensorRT-LLM and related tools help optimise inference on Nvidia GPUs.
• Cloudflare — AI Gateway provides caching, rate limits, model fallback and cost visibility across AI providers.
• Datadog — LLM observability helps monitor latency, token usage, cost and performance.
• Microsoft — Azure AI and enterprise tools can help customers manage model usage and deployment, although Microsoft also remains a major AI capex spender.
• ServiceNow — workflow automation and AI agents are tied to measurable enterprise productivity rather than just model capability.

The key point: AI adoption will increasingly depend on cost control and return on investment, not just model performance.

5. Edge AI and on-device AI

If every AI task runs in the cloud, the cost curve becomes harder to manage. Moving more inference onto devices can reduce cloud costs, improve latency and support privacy.

This is where selectivity matters. Device sellers such as Dell, HP and Lenovo may benefit from an AI PC cycle over time, but they also face memory and component-cost pressure first. They are not the cleanest efficiency exposures.

Areas to monitor include:

• smartphone AI chips;
• AI PC processors;
• low-power compute;
• neural processing units;
• on-device inference.

Illustrative company references:

• Apple — tight integration of silicon, software and on-device AI could reduce dependence on cloud inference over time, although near-term memory-cost pressure remains a headwind.
• Qualcomm — Snapdragon platforms and NPUs are directly tied to on-device AI processing.
• Arm — power-efficient architecture across mobile, edge and embedded AI devices.
• AMD — AI PC processors and local inference capability, with execution dependent on adoption of AI PC use cases.

The key point: edge AI becomes more valuable when cloud AI becomes more expensive.

Where caution is warranted

The AI theme is not over. It is becoming more disciplined.

Memory suppliers remain important, but the trade is no longer early. It is also no longer safe to assume that higher prices can continue indefinitely without a demand response.

Areas where more caution may be warranted include:

• Memory stocks where expectations already discount sustained pricing power, including Micron, SK Hynix, Samsung Electronics, Western Digital and SanDisk.
• Consumer hardware companies facing component-cost pressure, including Apple, Sony, Nintendo, HP, Dell Technologies and Lenovo.
• Low-margin electronics and device makers where price increases can quickly affect demand, especially in PCs, smartphones and gaming hardware.
• Hyperscalers where AI capex is rising faster than visible monetisation, including Microsoft, Alphabet, Amazon, Meta and Oracle.
• AI software and application companies with rising compute bills but unclear pass-through, particularly where AI features are bundled into existing products rather than separately monetised.
• Highly valued AI infrastructure names where expectations may already assume smooth execution, strong pricing and no meaningful demand pullback.

The useful distinction is no longer simply “AI exposure” versus “no AI exposure.”

The better framework is:

• who benefits from the bottleneck;
• who absorbs the cost of the bottleneck;
• who helps reduce the cost of the bottleneck;
• who can monetise AI fast enough to justify the spending.

That distinction matters because the next phase of AI may be less about scarcity and more about productivity.

Bottom line

Memory suppliers were the first winners of the AI bottleneck.

Efficiency may define the next phase.

The AI trade is moving from scarcity to productivity. The next stage will not only reward capacity. It will place more emphasis on technologies that make AI cheaper to run, easier to scale and faster to monetise.

The key investor takeaway is straightforward:

AI exposure needs to become more selective. The next opportunity set may be less about chasing the bottleneck itself, and more about identifying the technologies that reduce the cost of that bottleneck.