Skip to main content

Cashing in our Chips

Over our careers we have seen several occasions when a new or disruptive innovation looks like it will upset the order of things. Artificial intelligence, like the internet before it, has the potential to bring about radical change but as with all new technologies predicting how and when this happens is exceedingly difficult.

cashing in our chips

Often in these periods animal spirits take over and the market becomes a popularity contest with bottlenecks to rapid short-term growth gaining particular attention. The underlying quality of the business and sensible pricing are replaced by crowding into a concentrated list of names, with momentum and fear of missing out driving decision making. We look at two companies we could have invested in over the past year and how our investment philosophy serves as a guide during times of irrational exuberance.

A Thought Experiment

We began by wondering what you might say, as an investor in our fund, if we had to seek your approval before we bought stocks. Imagine for a moment that we had come to you in September 2025 with two ideas we were considering.

The First Contender

The first company is struggling to keep up with demand and despite this looks cheap. Although its share price has doubled over the past year it is only on around seven times forward earnings or around two and a half times its book value. In the previous quarter, revenue was up 35% and gross margins in the 50s, high, but not unheard of. Had we spoken in October we would also have noted that the company’s entire output of DRAM and NAND (two major types of memory chips) was sold out for the next year.

There were, however, a few things worrying us.

First, the chairman. The company is controlled by a South Korean chaebol, a family conglomerate, the patriarch of which is the chair of this company. He has been convicted, not once, which could perhaps be a mistake, but twice of taking money out of the group’s own subsidiaries. The first time was in 2003, as part of an accounting fraud, for which he served seven months of a three-year sentence and was later pardoned; the second was in 2013, when the Seoul Central District Court found he had removed roughly $50m USD from two affiliates to invest on his own account1. That second sentence, four years, was upheld by the Supreme Court, which observed that ‘the ruling is significant for holding the chairman and vice chairman…responsible for using funds of the group’s affiliates for personal gain’2. The company was also found guilty in 2005 by the US Department of Justice of price-fixing3.

Second, the cycle. Revenue is growing very rapidly, but there have been cycles that look a little like this in the past.

When demand dries up there is usually inventory sitting in the channel that needs to be consumed first, so orders here do not slow; they stop almost entirely. In the prospectus this company issued in July 2026, it sold another $26 billion in equity while warning that current orders may reflect inventory accumulation rather than end demand.

In the past 20 years this company’s net income has fallen by more than 80% on three separate occasions4. In 2023 this business made a loss of over nine trillion won (approximately $7bn USD), and the gross margin was negative for the full year5. The company had to sell its product for less than it cost to make. When things go wrong, they go wrong fast and badly.

Finally, demand is increasingly being generated in a very circular way, which The Economist recently referred to as the ‘bank of AI’, with around a trillion dollars of deals where the dominant supplier provides cash or guarantees to the data-centre landlords and model builders so that they can buy its chips6. This company sits one layer down that same structure. It has signed letters of intent with an AI (artificial intelligence) lab to supply up to 900,000 wafers a month, around 40% of global DRAM7 output, and the lab is under no obligation to take any of it8. Two of the hyperscalers have simply placed open-ended orders, telling suppliers they will accept whatever is available at whatever it costs9. The Economist’s own conclusion is that if these bets misfire the cost falls chiefly on shareholders – exactly how capitalism is supposed to work. We agree. It is just that in this instance the shareholders would be us.

So that is the first company. We suspect you may already have some reservations. According to our philosophy of establishing the quality of a company based on alignment with owners, strength of the franchise and conservative financials, this business raises some serious questions.

The second company, by contrast, is a little less dramatic.

The Second Contender

This business is a mobile phone chip designer and, at the time of our exercise, was on a trailing price/earnings multiple of about 20 times and a forward multiple of about 17, at a market value of a little over two trillion Taiwanese dollars (around $63 billion USD)10. This is a reasonable price for a solid business that is primarily linked to the fortunes of the mobile phone industry – likely to grow a little into the future with steady margins. This price assumes no uplift from demand in AI, yet this is one of a small number of companies capable of designing key elements of the technology required.

There were three elements that caught our attention.

The first is the chief executive. Rick Tsai ran TSMC from 2005 to 2009 and had spent a career before that building and operating semiconductor fabs. He joined this business in 2017 to fix a company that, on his own account, had strong engineering but could not get manufacturing, sales and marketing to pull in the same direction11. The advice he keeps repeating, that you should ‘do the right things, right’, is not the language of a short-term booster.

The second is a history of adaptation. This is a company that was spun out of a foundry in the 1990s and made its name by removing friction for its customers, bundling hardware and software into turnkey chipsets so that anyone could build a phone. It made a strategic mistake and missed 4G, but admitted it, wound the programme down early and moved the engineers onto 5G12. In November 2023 it published a five-year plan built around custom silicon, automotive, broadband and geographic diversification, arguably not a particularly exciting-looking roadmap13 but demonstrating its ability to diversify away from its core and to evolve with the times.

The final element is the culture change, and specifically the shift to being genuinely client-centric. The company co-designs with its customers rather than selling to them: it designed the CPU (central processing unit), memory subsystem and high-speed interfaces in Nvidia’s GB10 Grace Blackwell superchip14, which will allow more powerful local AI, and was at the time already working with Google on custom chips15. Tsai has said openly that the point of the Nvidia work is partly that his engineers learn how to improve on innovation and product development16. This, unlike the first company, is not a company extracting the last dollar from a captive buyer. It is a company making itself easier to buy from, which is one of the few durable forms of pricing power we know.

This was a company that passed our quality test and was available for a price that had little expectation for any development beyond its existing mobile business despite already working with some of the biggest players in AI. We hope we might have convinced you that the second company was the one that felt more like investing than speculating and, in the absence of having to seek approval, it was the one we bought. 

The companies, if you haven’t guessed by now, are the Korean company SK Hynix and Taiwanese MediaTek.

Now the uncomfortable bit

In the short term the market suggests we made the wrong decision.

In the year that followed, Hynix gained nearly 350%, versus a more modest 260% in MediaTek. If you were blessed with perfect timing. you could have booked a gain in the Korean company of just over 600%. In less than a year. In Q2 of this year, Hynix posted an operating margin of 76% and a net margin of 118% with net income higher than revenue thanks to gains in its investments17.

Structural or cyclical?

These remarkable results, however, raise some further questions. Although we like businesses with pricing power, we are wary of a company that takes advantage of a shortage to extract everything it can from its customers. A gross margin above 80% against a through-cycle average nearer 30% is not sustainable; it is clearly a windfall, and customers know the difference. Even TSMC commented on it, with its CEO, C.C. Wei, pointing out, ‘our customers [need] to be successful. I don’t want to squeeze them out from the market. And besides, we are a very trustable company with our customers. So we don’t suddenly increase our price by…4x or 5x. You cannot survive…that kind of a price increase’18.

HP has told the market that memory went from 15–18% of its materials bill to 35% in a single quarter19. Dell’s CEO said the company had ‘never witnessed costs escalating at the current pace’20. Apple put through price rises across most of its hardware line in June and then suffered its worst day in the stock market for over a year. Gartner now expects the PC market to shrink 10–11% this year and smartphones 8–9%, which is the demand destruction showing up in volumes21.

Separately, and as we write, the CEO of SK Hynix is the subject of a live criminal investigation, formally assigned to two police divisions in August, over the way employee bonuses have been linked to profits. The charge is breach of trust, which in Korea carries a five-year minimum sentence22. All of these subsequent events are more red flags that could lead to a bad outcome for minority shareholders. 

Finally, the fight-back is happening. Elon Musk has announced that Tesla will have to build its own memory fab23. Intel’s chief executive has talked about a new memory architecture and a return to the business24. Apple is even considering Chinese suppliers25. A consumer class action alleging price-fixing across Samsung, SK Hynix and Micron was filed in the Northern District of California in June26. And in March, Google Research published a compression technique that claims to cut memory consumption in large language models by a factor of six, at which point memory stocks fell, which tells you the market believes this is a significant risk27. Raising prices by 400% potentially teaches your customers three things: to need less of your product, to make it themselves or to sue you.

In MediaTek, things are thankfully a little less dramatic. The first AI accelerator ASIC goes into mass production this quarter, data-centre ASIC revenue guidance for this year has doubled from one billion dollars to two and a second programme is on track for 2028. The board has approved five billion dollars of financing to secure capacity, and management is targeting 10–20% of a market it thinks will be worth 70 to 80 billion dollars by 202728. Nvidia has announced a 3.5-billion-dollar investment in the company alongside a role in its NVLink Fusion ecosystem29.

Keeping our discipline

We now have a problem, however. MediaTek is on 64 times trailing earnings, which means the free option has been priced and our next return has to come from delivery rather than revaluation10. The Nvidia investment means that MediaTek is now also partially fed by the circular financing that worries us elsewhere. Nor is it immune to the impact of Hynix’s egregious pricing – its own customers are cutting handset volumes because of memory prices, with smartphone shipments now estimated to be down around 15% this year30. If demand is cyclical for memory, it will be cyclical for chips, too, and Broadcom, Marvell and AMD are all pushing into the same custom silicon market31. This ebullience on one hand and growing risk on the other is why we have been trimming our position.

As responsible stewards of your money, have we let you down? Over the past 12 months, against the benchmark, you could argue yes, but as long-term investors, owners of our business and investors in our funds, we believe that avoiding SK Hynix has been the right long-term investment decision. A company whose customers are actively trying to avoid their products combined with a history of poor governance is not sustainable in our view. At 5.5% of the index, a correction will have an impact, especially given that Samsung, another company with very similar characteristics and therefore subject to the same swings and roundabouts, accounts for another 7%. In fact, the top 10 in the index, with the exception of China Construction Bank, are all AI-related (and for the sake of fairness this includes MediaTek at 1.5%) and they together make up 37% of the index (excluding CCB)32.

We have been around for long enough to have seen times like this before, although never to some of the extremes we see now, which we believe creates a greater risk of permanent loss. We built Skerryvore to give us the freedom to be able to avoid these excesses by not having to own companies simply because they are in an index. This discipline can be painful, and the past year has been challenging but successfully compounding capital over the long term requires making these kinds of decisions when price and value diverge. We continue to believe that it is preferable to explain why we don’t own some companies than to have to admit that we abandoned our investment philosophy because it felt uncomfortable. Comfort has never been the point.

Thank you, as always, for your continuing trust.

Sources

  1. FT (31 January 2013). S Korea court jails prominent tycoon. https://www.ft.com/content/d77db72e-6b7f-11e2-a700-00144feab49a
  2. The Korea Times (27 February 2014). SK head’s 4-year jail term upheld. https://www.koreatimes.co.kr/southkorea/20140227/sk-heads-4-year-jail-term-upheld
  3. NBCNews (21 April 2005). Chipmaker Hynix admits to price fixing. https://www.nbcnews.com/id/wbna7589977
  4. CompaniesMarketCap. Earnings history for SK Hynix. https://companiesmarketcap.com/sk-hynix/earnings
  5. SK Hynix press release via Nasdaq (24 January 2024). SK hynix reports financial results for 2023, 4Q23. https://www.nasdaq.com/press-release/sk-hynix-reports-financial-results-for-2023-4q23-2024-01-24; Investing.com. SK Hynix financial summary. https://www.investing.com/equities/sk-hynix-inc-financial-summary
  6. The Economist (3 September 2026). Nvidia is driving the AI boom. Good. (The phrase is one the paper reports rather than coins, and it also draws the comparison with dotcom-era vendor financing); see also The Intelligence podcast, The Economist (4 September 2026). Bargaining chips: Nvidia is the bank of AI. https://shows.acast.com/theintelligencepodcast/episodes/bargaining-chips-nvidia-is-the-bank-of-ai
  7. A type of high-bandwidth memory.
  8. Tom’s Hardware (1 October 2025). OpenAI’s Stargate project to consume up to 40% of global DRAM output. https://www.tomshardware.com/pc-components/dram/openais-stargate-project-to-consume-up-to-40-percent-of-global-dram-output-inks-deal-with-samsung-and-sk-hynix-to-the-tune-of-up-to-900-000-wafers-per-month; The Deep Dive (31 March 2026). OpenAI locked up 40% of global RAM with no obligation to buy any of it. https://thedeepdive.ca/openai-locked-up-40-of-global-ram-with-no-obligation-to-buy-any-of-it
  9. Reuters (3 December 2025). The AI frenzy is driving a memory chip supply crisis. https://www.reuters.com/world/china/ai-frenzy-is-driving-new-global-supply-chain-crisis-2025-12-03
  10. Yahoo Finance (30 September 2025). MediaTek Inc. valuation measures. https://finance.yahoo.com/quote/2454.TW/key-statistics
  11. Wikipedia. Rick Tsai biography. https://en.wikipedia.org/wiki/Rick_Tsai; Digital Trends (26 November 2019). MediaTek failed to win at 4G, but it won’t do the same with 5G. https://www.digitaltrends.com/phones/MediaTek-rick-tsai-interview
  12. Nomad Semi (14 May 2025). MediaTek Deep Dive. https://www.nomadsemi.com/p/mediatek-deep-dive
  13. RCR Wireless (16 November 2023). MediaTek debuts five-year plan for globalization, diversification. https://www.rcrwireless.com/20231116/chips/MediaTek-debuts-five-year-plan-for-globalization-diversification
  14. MediaTek press release (15 October 2025). Newly launched NVIDIA DGX Spark features GB10 superchip co-designed by MediaTek. https://www.barchart.com/story/news/35463525/newly-launched-nvidia-dgx-spark-features-gb10-superchip-co-designed-by-MediaTek
  15. CommonWealth Magazine (13 April 2026). How MediaTek quietly became a core player in Google’s AI infrastructure. https://english.cw.com.tw/article/article.action?id=4707
  16. Benzinga via Yahoo Finance (10 March 2026). Rick Tsai shares advice once given to him by Jensen Huang. https://finance.yahoo.com/news/rick-tsai-shares-advice-once-220201356.html
  17. CapitalIQ (29 July 2026). Q2 Earnings Call Transcript for SK Hynix Inc.  
  18. TSMC Earnings Call (16 July 2026). https://www.fool.com/earnings/call-transcripts/2026/07/16/tsm-tsm-q2-2026-earnings-call-transcript
  19. Tom’s Hardware (26 February 2026). HP says memory costs doubled in one quarter, now account for 35% of PC build materials. https://www.tomshardware.com/tech-industry/hp-says-memory-costs-doubled-to-35-percent-of-pc-build-materials-in-one-quarter
  20. Bloomberg News (26 November 2025). Tech firms from Dell to HP warn of memory chip squeeze. https://news.bloomberglaw.com/tech-and-telecom-law/tech-firms-from-dell-to-hp-warn-of-memory-chip-squeeze-from-ai
  21. Gartner (26 February 2026). Gartner says surging memory costs will reduce global PC and smartphone shipments in 2026. https://www.gartner.com/en/newsroom/press-releases/2026-02-26-gartner-says-surging-memory-costs-will-reduce-global-pc-and-smartphone-shipments-in-2026, CNBC (25 June 2026). Apple posts worst day in over a year after MacBook and iPad price hikes. https://www.cnbc.com/amp/2026/06/25/apple-macbook-ipad-price-hike-memory.html
  22. Seoul Economic Daily (5 August 2026). Police open probe into Samsung, SK Hynix chiefs over shareholder complaint. https://en.sedaily.com/society/2026/08/05/police-open-probe-into-samsung-sk-hynix-chiefs-over; Tech Times (5 August 2026). Criminal probe of Samsung, SK Hynix chiefs invokes law Seoul wants abolished by year-end. https://www.techtimes.com/articles/323210/20260805/criminal-probe-samsung-sk-hynix-chiefs-invokes-law-seoul-wants-abolished-year-end.htm
  23. Bloomberg (15 February 2026). Rampant AI demand for memory is fueling a growing chip crisis. https://www.bloomberg.com/news/articles/2026-02-15/rampant-ai-demand-for-memory-is-fueling-a-growing-chip-crisis
  24. WCCF Tech (13 August 2026). Intel CEO hints Chipzilla’s return to making memory. https://wccftech.com/intel-ceo-hints-that-chipzilla-might-return-to-making-memory
  25. Reuters (9 August 2026). Apple tests China's CXMT memory chips for iPhones and MacBooks, WSJ reports. https://www.reuters.com/business/retail-consumer/apple-tests-chinas-cxmt-memory-chips-iphones-macbooks-wsj-reports-2026-08-09/
  26. India Today (29 June 2026). Samsung, SK Hynix and Micron sued over alleged price-fixing amid RAM crisis. https://www.indiatoday.in/technology/news/story/samsung-sk-hynix-and-micron-sued-over-alleged-price-fixing-amid-ram-crisis-2936357-2026-06-29; Rain Intelligence (1 July 2026). When the shortage is the strategy. https://www.rainintelligence.com/blog/dram-price-fixing-lawsuit-samsung-sk-hynix-micron
  27. Google Research (24 March 2026). TurboQuant: Redefining AI efficiency with extreme compression. https://research.google/blog/turboquant-redefining-ai-efficiency-with-extreme-compression; MarketWatch (25 March 2026). Micron’s stock is dropping. Is Google partly to blame? https://www.marketwatch.com/story/microns-stock-is-dropping-is-google-partly-to-blame-d4cad72e
  28. The Register (3 August 2026). MediaTek lines up $5B war chest for AI datacenter push. https://www.theregister.com/ai-and-ml/2026/08/03/MediaTek-lines-up-5b-war-chest-for-ai-datacenter-push/5282304; TrendForce (30 April 2026). MediaTek reportedly doubles 2026 AI ASIC revenue target to $2B, aims 10–15% market share. https://www.trendforce.com/news/2026/04/30/news-MediaTek-reportedly-doubles-2026-ai-asic-revenue-target-to-2b-aims-10-15-market-share; see also TrendForce (5 February 2026). MediaTek forecasts $1B in ASIC sales for 2026, custom AI chips set for 20% revenue share in 2027. https://www.trendforce.com/news/2026/02/05/news-MediaTek-forecasts-1b-in-asic-sales-for-2026-custom-ai-chips-set-for-20-revenue-share-in-2027
  29. NVIDIA newsroom (31 August 2026). NVIDIA and MediaTek deepen long-standing partnership to build AI edge to cloud computing platforms. https://nvidianews.nvidia.com/news/nvidia-and-MediaTek-deepen-long-standing-partnership-to-build-ai-edge-to-cloud-computing-platforms; see also CXOVoice.com (1 September 2026). Nvidia and MediaTek expand partnership on AI computing platforms from edge to cloud. https://cxovoice.com/nvidia-and-MediaTek-expand-partnership-on-ai-computing-platforms-from-edge-to-cloud
  30. TrendForce (30 April 2026). MediaTek reportedly doubles 2026 AI ASIC revenue target to $2B, aims 10–15% market share. https://www.trendforce.com/news/2026/04/30/news-MediaTek-reportedly-doubles-2026-ai-asic-revenue-target-to-2b-aims-10-15-market-share
  31. TrendForce (20 August 2026). Marvell, AMD reportedly shake up Google TPU race, putting Broadcom, MediaTek under pressure. https://www.trendforce.com/news/2026/08/20/news-marvell-amd-reportedly-shake-up-google-tpu-race-putting-broadcom-MediaTek-under-pressure
  32. MSCI Emerging Markets Index Factsheet, August 2026. https://www.msci.com/documents/10199/c0db0a48-01f2-4ba9-ad01-226fd5678111