Agentic AI Adoption & the Case for Samsung Electronics

This article is based on quarterly commentary from Contrarius’s global investment team.

Most people’s experience of artificial intelligence has been the chatbot: ask a question, get an answer. That use is still growing. What is changing—and what matters for investors—is the next layer of demand.

A task may now involve many steps—looking something up, calling a tool, receiving and analysing data, checking a result, trying again—sometimes on demand, sometimes on a schedule, and sometimes continuously in the background without waiting for a person between steps. 

That is a very different intensity of use from a one-shot answer (often 5–30 times the tokens of a single answer). It requires more compute, more memory, and more network traffic. 

Chat-style use does not disappear when this happens. Training ever-larger models, including the reinforcement-learning loops used to improve them after pre-training, does not get lighter either. The new layer sits on top of both.

As output quality and usefulness improve, more tasks become worth automating, and each task automated adds to demand. Earlier this year Cloudflare announced that automated systems had passed humans in internet traffic a year ahead of its own estimates. OpenRouter had already noted that token generation from agentic workloads overtook human-driven tokens around early February 2026; by August, agentic demand was about five times the human level. Token growth is unlikely to remain human-paced. 

Over time, automated systems should generate the vast majority of tokens, with people directly involved in only a thin sliver of that activity.

Efficiency gains are real and are expected to continue. The question is whether the volume, complexity, and rising usefulness of this next layer can outrun those gains. We think the early evidence points that way. For decades Moore’s law made compute cheaper even as demand rose; that deflationary cushion has slowed, while AI demand is compounding. The semiconductor industry may simply need to be much larger. 

While some investors believe the memory cycle has reached its end, we continue to hold positions in the sector.

SAMSUNG ELECTRONICS

Samsung is one of the world’s largest technology manufacturers; in the June quarter its operating profit briefly exceeded Nvidia’s then-record quarterly result, a reminder of the extent of the current cycle.

While memory is core to the investment case for Samsung, the rest of the group is large in its own right. The foundry business remains a distant second to TSMC, but it has a concrete customer win: a multi-year, roughly $16.5 billion chip-supply agreement with Tesla, through 2033, tied to Samsung’s Texas plant and Tesla’s next-generation AI silicon. The consumer franchise is still one of the world’s largest—Mobile and Networks alone generated revenue of about KRW 130 trillion ($90 billion) in 2025, on the order of 240 million smartphones a year. Visual Display ships about 35 million televisions a year and has led the global TV market by value for two decades, with close to a 30% revenue share. Harman adds a further leg: led by JBL, it has close to half the Bluetooth speakers market, and its audio or connected-car systems are already in more than 50 million vehicles. Together that is a direct window into how AI reaches handsets, homes, and cars. Its robotics effort is still early. Samsung’s robotics business is not currently a material earnings contributor, however its Rainbow Robotics subsidiary is building toward a humanoid prototype giving Samsung exposure to where AI goes next.

We also have a favourable view on other memory companies, including SK Hynix (and the associated SK Square) and Micron Technology. 

Samsung’s appeal to us is different. It offers deep exposure to memory, but also to businesses positioned for real-world AI, an area which we believe holds significant potential.

The memory cycle has always been violent in part because so much of the demand sat with consumer and enterprise buyers who were sensitive to price. When memory got expensive, they cut back. They could often afford to wait: Moore’s law kept delivering cheaper bits over time, so deferral felt rational. That elasticity shaped how investors thought about every upcycle—sharp, temporary, and capped by demand destruction when prices rose.

The mix looks different now. Buying is far more concentrated in data centres building AI capacity, with robotics and other real-world uses set to raise memory intensity further. These customers cannot afford the old deferral, as the returns from achieving AI leadership—commercial and geopolitical—make waiting for cheaper silicon a more expensive choice than paying for capacity now. Relative to the old consumer- and enterprise-led buyers, this newer cohort is expected to be both larger and less price-sensitive. 

Firming prices do not automatically bring the same cutbacks they once did. If that proves durable, the cycle may last longer, and be more profitable for the industry’s large manufacturers, than the classic cyclical view assumes.

A further difference sits in the product itself. High-bandwidth memory (HBM) is not just another generation of the DRAM that filled PCs and phones in prior cycles. Ordinary DRAM feeds a processor at a measured pace and can often be adjusted, substituted, or deferred when prices rise. HBM stacks memory chips on top of one another and connects them with short, dense pathways so that AI accelerators can move far more data in and out of memory at once—which is what training and running large models require. It is harder to manufacture, uses more silicon for each unit of useful bandwidth, and is typically sold into a concentrated set of accelerator customers rather than a diffuse consumer base. In the last upcycle the industry was largely shipping conventional memory into price-sensitive end markets. 

This time a growing share of industry profit sits in a scarcer, more specialised product tied tightly to AI build-outs—another reason the old elasticity playbook may understate how durable cash generation can be.

A third piece of evidence that this cycle may differ sits in how memory is being sold. Samsung has said it aims to put some 60–70% of memory production capacity under long-term agreements, leaving 30–40% on the spot market. As of mid-2026 it had completed deals with five major global data-centre customers and was close to finishing a further five large AI-related agreements. Typical terms are multi-year—often a five-year base that rolls forward—with advance payments and price floors on some contracts. This is very different to previous cycles.

It is fair to ask whether this is just another “this time is different” story. Memory has a long record of teaching investors that lesson the hard way, and we do not claim cycles have been repealed. Competition will come, including from Chinese capacity led by CXMT and YMTC. New capacity coming online has, together with demand destruction, been how past down cycles have arrived. 

The risk of looking foolish when prices turn is real—and it is precisely that risk which shapes today’s consensus. 

Many investors prefer to treat the upcycle as familiar, mark it to a cyclical fair value, and move on, rather than underwrite the possibility that the industry has changed in ways that matter for sustained and prolonged cash generation. Owning Samsung on the thesis above is, in our view, the contrarian position: not that memory never cycles, nor that supply additions will not arrive, but that end use, buyers, product mix, and harder contracts than the softer supply agreements of the past can make this upcycle more durable than the market is willing to pay for and that structural AI demand can outstrip new supply for longer than most expect. 

None of that makes the shares riskless. It does mean that dismissing the case as mere cycle enthusiasm may be the more comfortable error—and the more expensive one if the structural shift is even partly durable.

On consensus estimates, Samsung trades at about four times FY2027 earnings. The preference shares the Fund owns are cheaper still: about three times. The long-term agreements already described give more visibility on those earnings than a spot-market cycle would. 

That rating still looks to us like a cyclical semiconductor being marked to the old cyclical view. We expect Samsung’s rating to improve if the structural changes set out above prove durable.

If our estimates are right, Samsung can generate free cash flow equal to a meaningful portion of today’s market value over the next few years—and clear shareholder-return policies mean we expect a large share of that cash to come back to shareholders, as dividends or buybacks.

Memory holdings have already contributed significantly to the Contrarius Global Equity Fund’s performance, and we have trimmed the pure-play positions that have outperformed. Samsung has lagged its memory peers over the past year. That leaves it an attractive way, in our view, to remain exposed to memory’s central role, while also owning a broader set of businesses that are expected to benefit as AI moves beyond the screen. 

Summary

Samsung Electronics is a top holding in the Contrarius Global Equity Fund because we believe that the memory opportunity is real, the structural shifts set out above look more durable than the classic cycle playbook, and the free cash flow the business can generate is large relative to the price we are asked to pay.