Ben Thompson's memory oligopoly bear case: the sharpest challenge to the MU thesis
The Micron and HBM bull case has been the consensus for a year. Ben Thompson made the sharpest bear case against it on his July 6 Sharp Tech podcast. Three long-tail threats, one Tim Cook quote, and a memory oligopoly quietly manufacturing its own competition.
On July 6, 2026, Stratechery's Ben Thompson devoted an entire Sharp Tech episode to the question every memory bull has been avoiding: what happens if the MU / SK Hynix / Samsung oligopoly pushed prices too far, too fast, and manufactured the conditions for its own disruption? His answer is the most credible bear case published on memory this year, and it comes from someone with an unusual credibility anchor: he owns none of the names in the debate. Source: BargoAI research.
| Memory price hikes | Apple 2027 memory spend | Thompson's position |
|---|---|---|
| 9x | 9x | Zero |
| since the 2023 trough | vs 2 years ago (Tim Cook) | no book, most credible bear |
What did Ben Thompson actually say?
The Sharp Tech episode was titled "Did Memory Makers Overplay Their Hand?" The core Thompson framing:
The memory guys not investing sooner, not moving faster to address this shortage, having this little three-three company club of scratching each other's backs and not competing. They are creating the conditions for everyone else in the world to endure pain to reduce their dependency on those big three.
Ben Thompson, Sharp Tech podcast "Did Memory Makers Overplay Their Hand?", July 6, 2026
Thompson then delivered the number that anchors the entire thesis:
Apple is spending 6x as much on memory today as they were a couple years ago, and Tim Cook said it's going to be 9x by 2027.
Ben Thompson, citing Tim Cook, Sharp Tech, July 6, 2026
When your single largest customer publicly states they will spend 9x more on your product within 18 months, you have either extraordinary pricing power or you have created an existential motivator for that customer to build alternatives. Thompson argues it is the second.
Why does his "no position" stance make this the sharpest bear case?
Most memory bear cases published in 2026 come from short sellers with active positions (Chanos, Einhorn) or from CEOs of competing products (Cerebras, Groq founders). Both categories are legitimate but book-flagged. Thompson explicitly opened the episode by disclosing that he does not hold any of the memory names, does not short them, and is not paid by anyone involved:
I specifically am not in this game.
Ben Thompson, Sharp Tech, July 6, 2026
That single line is why this bear case matters more than most. Thompson is a technology strategist who has been correct on multiple industry-cycle turns (mobile, cloud, streaming). When he identifies structural weakness in a moat, and has zero financial reason to be right, the take carries more weight than a short seller talking their book.
Why this matters for portfolio construction. Every Bargo memory article to date has been variations of the bull case (Micron blockbuster earnings, HBM bandwidth as the bottleneck, memory oligopoly benefiting from AI capex). Thompson's take is the first sharp counter-argument published by an unaffiliated voice. Own the memory bulls without engaging his three threats and you are running the thesis on incomplete information.
Why did the memory oligopoly create its own problem?
The "three-three company club" Thompson references is the DRAM and HBM oligopoly. Three companies control roughly 95 percent of global memory production.
| Company | Est. DRAM share | Est. HBM share | Recent gross margin |
|---|---|---|---|
| SK Hynix | ~34% | ~50% | ~55% (Q1 2026) |
| Samsung Electronics | ~40% | ~35% | ~48% (Q1 2026) |
| Micron Technology (MU) | ~22% | ~15% | 84.9% (Q3 FY26) |
Thompson's argument is not that the oligopoly does not work today. It clearly does. Micron just posted an 84.9% gross margin quarter, the highest ever for a memory company. His argument is that the discipline that produces those margins (limited investment, coordinated pricing, refusing to compete on capacity) is exactly what tips customers over from grudging acceptance into active substitution behavior.
A monopoly or tight oligopoly can extract 90%+ margins in the short term. Over a 3-5 year window, the customer base always finds a way out. Three specific escape routes are already forming.
What are the three threats collapsing the memory moat?
Thompson identifies three separate escape hatches. None alone would be enough. Together, over five years, they could rerate the entire memory complex lower.
Threat 1: Algorithmic memory efficiency
Every hyperscaler and phone OS vendor is now investing heavily in reducing the memory footprint of AI workloads. Techniques include quantization (running models at INT4 or INT8 precision instead of FP16), attention sparsification, mixture-of-experts routing, and speculative decoding. These are not laboratory ideas. Google, Apple, Microsoft, and Meta all have production systems running today at 2x to 4x lower memory footprints for the same output quality.
If a hyperscaler cuts memory required per inference in half, they buy half as much memory. That is a direct top-line hit on Micron and SK Hynix over 5 years. This threat is silent, cumulative, and already underway.
Threat 2: Architectural disaggregation (Apple + Nvidia)
Apple is spending 6x on memory today vs two years ago, going to 9x by 2027 per Tim Cook. That is not gratitude. That is a company screaming "we need alternatives" as loudly as any procurement team can. Expect Apple to accelerate:
- Custom on-package memory designs that bypass third-party HBM stacks
- Alternative memory hierarchies (LPDDR, GDDR, novel MRAM variants) for edge inference
- Deeper vertical integration with TSMC packaging that reduces standard HBM dependency
Nvidia is doing the same on the server side. Rubin memory racks (disaggregated memory pools connected via NVLink) are the first architecture where memory can be pooled and shared across GPUs rather than sitting captive to one accelerator. If disaggregation works at scale, the same total AI system needs less HBM overall.
Threat 3: CXMT and YMTC climbing the ladder faster than expected
CXMT (China's DRAM champion) and YMTC (China's NAND leader) have been dismissed by Western analysts as generations behind on process nodes. That framing is outdated. Both companies have closed the process-node gap dramatically in 2024-2026, and the US export controls that were supposed to slow them down have instead concentrated Chinese demand and capital into their fabs.
Thompson's specific prediction: non-US OEMs (HP-type Taiwan/Korea/EU-based) will start qualifying Chinese memory for their non-US SKUs within 24 months. Once that qualification happens for one major OEM, the domino falls. HBM3 and HBM4-class Chinese memory becomes a real supply-chain option for anyone selling into non-US markets.
Why does Apple's 9x memory spend matter here?
The Tim Cook 9x number is the single most important datapoint in Thompson's argument. Here is the math:
| Year | Apple memory spend (est.) | Multiplier vs 2023 |
|---|---|---|
| 2023 baseline | ~$3-4B annually | 1x |
| 2026 (Thompson quoting Cook) | ~$18-24B annually | 6x |
| 2027 (Cook forecast) | ~$27-36B annually | 9x |
Apple is not spending 9x more because they love the memory oligopoly. They are spending 9x more because:
- Prices went up (roughly 3-4x on HBM and premium mobile DRAM since 2023)
- iPhone / MacBook memory content per unit is going up (from 6-8 GB to 16-24 GB baseline)
- Apple is building bigger Neural Engines for on-device AI, requiring more HBM-adjacent memory
Every dollar of that $30B+ 2027 spend is an argument in Cupertino for building alternatives. Historically, when Apple decides a supplier has too much pricing power, they act. They designed their own SoCs to escape Intel. They designed their own modems to escape Qualcomm. They built their own Neural Engine to escape Nvidia. Memory is the next logical target.
Who wins if Thompson is right?
The bear case on the memory oligopoly is bullish for a very specific set of tickers.
| Name | Direction | Mechanism |
|---|---|---|
| Micron (MU), SK Hynix, Samsung | Bearish (3-5 year) | Share and margin compression as algorithmic efficiency + disaggregation + Chinese memory bite |
| Apple (AAPL) | Bullish | Eventual custom-memory savings + input-cost relief as competition emerges |
| NVIDIA (NVDA) | Bullish | Rubin memory-rack disaggregation reduces per-GPU HBM dependency, improves margins |
| ACM Research (ACMR) | Bullish | Only US-listed pure-play on Chinese memory equipment demand (CXMT and YMTC customers) |
| Cerebras (CBRS) | Bullish | Wafer-scale architecture uses no HBM. Every architectural disaggregation trend helps. |
| CXMT, YMTC (private) | Bullish | Direct beneficiaries of forced localization + non-US OEM qualification |
What this means for your portfolio
Every Bargo memory piece to date has argued the bull case. This is the honest counter. Here is how to hold both ideas at the same time:
- The memory bull case works over the next 4-6 quarters, minimum. Micron's $22B customer cash deposits and 5-year take-or-pay contracts are real. HBM tightness beyond 2027 is confirmed by the CFO. Do not exit the memory trade on Thompson's take.
- Thompson's thesis matters for exit timing. If you own MU or SK Hynix, the exit signal is not weak earnings. It is any of the three threats becoming visible in actual customer behavior. Watch Apple procurement announcements, hyperscaler MoE (mixture-of-experts) deployment scale, and any non-US OEM qualifying CXMT product.
- ACMR is the direct read-through. The Chinese memory buildout that Thompson identifies as threat 3 is exactly what makes ACM Research the "US-listed way to play China's semiconductor decoupling." Own ACMR as your hedge against the memory bulls being wrong.
- Nvidia is a two-way win. If HBM stays tight (bull case), NVDA benefits. If HBM gets disaggregated (bear case), NVDA GPU margins expand. Rubin memory racks are a hedge Nvidia is building in real time.
- Anthropic and OpenAI IPO valuations depend on this. If Thompson is right about memory costs eventually falling, hyperscaler + frontier lab inference economics improve, which supports Anthropic's $1T IPO math. If Baker is also right on his frontier-margin thesis, Anthropic loses either way.
This is not investment advice. All live financials, options positioning, and signals are on bargo.ai.
Sources
- Ben Thompson, Sharp Tech podcast "Did Memory Makers Overplay Their Hand?", July 6, 2026, via stratechery.com
- Tim Cook (Apple CEO) commentary on 2027 memory spend, cited by Ben Thompson
- Micron Technology (MU) Q3 FY2026 earnings release and CFO Mark Murphy commentary, June 24, 2026, investors.micron.com
- SK Hynix Q1 FY2026 earnings materials
- SemiAnalysis reporting on CXMT and YMTC process node progression, June to July 2026
- BargoAI live SIP-tape market data, via bargo.ai
Reviewed by the Bargo editorial desk. Ben Thompson quotes verbatim from Sharp Tech podcast episode "Did Memory Makers Overplay Their Hand?" published July 6, 2026. Memory market share estimates per SemiAnalysis and industry consensus. Apple memory spend estimates are derived from Tim Cook public commentary. This is research and educational content, not investment advice.