The AI Memory Tax.Original analysisNot investment advice

In 2021, phone weakness was a demand shock. In 2026, consumer devices are being squeezed by a different force: AI servers are pulling memory supply and capital toward data centers, raising DRAM and NAND costs for smartphones, PCs, and low-end electronics.

In 2021, the smartphone supply-chain problem was easy to understand. COVID waves hit India and other developing economies. Phone sales slowed. Factories shut. Logistics broke. OEMs cut orders. The shock moved backward through the semiconductor supply chain. The uploaded SemiAnalysis piece captured that moment well. Phone expectations were cut, lower-end devices were under pressure, and consumer weakness forced Apple, Samsung, Xiaomi, OPPO, Vivo, and others to reassess production plans.¹

That was a demand shock. The 2026 version is different.

This time, smartphones are not being hit only because people suddenly stopped wanting phones. They are being hit because AI is changing the price and allocation of memory. AI servers are pulling DRAM and NAND toward data centers. Memory suppliers are prioritising high-value AI and server demand. Smartphone OEMs are facing tighter supply, higher memory costs, and weaker consumer price tolerance.

The result is a strange semiconductor split. AI infrastructure is booming while consumer electronics absorbs the cost.

That is the AI memory tax.

---

I. The 2021 thesis

In May 2021, Dylan Patel published a SemiAnalysis piece on smartphone weakness. The framing was a demand shock. COVID waves rolled through India and other developing economies. TrendForce had cut smartphone-production expectations. Chinese Android vendors were reportedly cutting orders. The piece pointed out that cheaper phones carry lower silicon content, so a mix shift toward weaker low-end demand fed back through SoCs, RF, cameras, memory, OSATs, and component suppliers.¹

I revisited that piece because the 2026 cycle looks nothing like that pattern, yet the headline (smartphones weak, semis nervous) sounds similar. The mechanism has changed.

---

II. The 2026 update

The 2026 cycle is not just a demand problem. It is a memory allocation and pricing problem. The chain runs in the opposite direction. Instead of weak consumers walking back through the supply chain, it starts at the top of the stack and pushes down.

Diagram 01 · Two cycles, different starting points

2021 cycle · demand-shock chain

COVID & income shock

Weak phone demand

OEM order cuts

Component & OSAT ripple

2026 cycle · memory-tax chain

AI server demand

Memory shortage & price spike

Smartphone BOM inflation

Price up / spec down / shipments cut

In 2021 the smartphone supply chain ran backward from a demand shock. In 2026 it runs forward from a memory-cost shock.

---

III. The smartphone market is being squeezed

IDC’s smartphone tracker for Q1 2026 paints the same picture. Shipments declined year over year, memory supply was tight, memory prices were at multi-year highs, and OEMs faced significantly higher bill-of-materials costs as a result. The mainstream price points moved up faster than usual, especially in emerging markets.²

This is not a normal demand cycle. A normal demand cycle starts with consumers pulling back. This one starts upstream, with memory allocation moving toward AI and data centers.

The pressure starts upstream and shows up downstream.

---

IV. The low end breaks first

TrendForce, Counterpoint, and Reuters all converged on the same picture going into and through 2026. Global smartphone shipment growth is expected to slow or decline this year, with most of the pain landing on devices under roughly $200. BOM costs for those devices have risen meaningfully, with memory as the largest single driver. A mainstream 8 GB + 256 GB configuration carries a noticeably higher memory cost than it did 12 months ago, and memory has become a much larger share of smartphone BOM at the low end.³⁴

Memory inflation is regressive. It hits cheap phones first.

A $25 cost increase is annoying in a $1,200 iPhone. It is brutal in a $120 Android phone.

Low-end OEMs end up with three real choices. Raise the price and lose demand. Reduce memory or storage and ship a worse product. Cut shipments and protect margin. Most pick a blend, all three are visible in the 2026 product cycle.

Card · Smartphone BOM pressure

< $200

Price band where memory inflation breaks unit economics first.⁴

8 / 256

Mainstream 8 GB DRAM + 256 GB NAND configuration; memory cost a much bigger share of BOM than 12 months earlier.³

3 options

Raise price · cut spec · cut shipments. Most low-end OEMs do all three.

Memory inflation is regressive. The bottom of the smartphone market absorbs the cost first.

---

V. AI is repricing memory

Micron made the structural shift explicit in its FY2026 Q2 prepared remarks. AI demand is driving DRAM and NAND data-center bit TAM above 50% of total industry TAM for the first time in calendar 2026, with AI and traditional server demand running into inadequate DRAM and NAND supply. The company also frames its product line as a portfolio for AI inference architectures: HBM next to accelerators, LPDRAM for memory-rich systems, DDR for general server compute, and SSDs (with vector databases and KV-cache offload built around them).⁵

For years, smartphones helped define the memory cycle. Now smartphones compete against AI servers for memory allocation.

Phones used to drive the memory cycle. Now phones compete against AI servers for memory allocation.

---

VI. Why AI needs so much memory

The reason memory has become the binding constraint is mechanical. AI training needs to keep model weights, activations, and optimiser states close to the accelerator, which means HBM, high bandwidth, and a dense networking fabric. AI inference adds another set of memory consumers on top: model weights, the growing KV cache that scales with context windows, batching state, SSD offload of long-tail data, vector databases for retrieval, and the latency budget needed to keep token economics workable.

ASML’s 2025 strategic report makes the same point at the macro level. AI requires leading-edge, high-performance processor chips and a significant increase in DRAM compared with traditional compute architectures. Advanced Logic and AI-related DRAM are the structural drivers of lithography demand.⁷

This is not a temporary phone issue. AI is shifting the centre of gravity of memory demand.

---

VII. Apple and Samsung defend. Low-end Android cannot.

This does not hit every vendor equally. Premium brands have tools that the low end does not.

Q1 2026 vendor commentary lined up with this split. Apple and Samsung performed relatively better, while several Chinese Android brands with heavy sub-$200 exposure showed sharper weakness.⁴⁹

Premium vendors can pass through memory inflation. Low-end Android vendors eat it.

---

VIII. TSMC shows the semiconductor cycle is splitting

The semiconductor cycle used to be unified. Strong phones meant strong semis, weak phones meant weak semis. That coupling is breaking. TSMC’s 2025 annual report makes the split explicit: robust AI-related demand throughout 2025, mild recovery in non-AI end markets, advanced-node and AI/HPC demand strong, smartphone and consumer markets more fragile.⁶

That is the unusual part. Smartphones are weak. Semiconductors as a whole are not. AI is pulling so hard on the other side of the industry that the aggregate index keeps moving up.

Diagram 02 · Three semiconductor cycles, one industry

Cycle 01

AI infrastructure

• GPUs / accelerators
• HBM, DRAM, NAND
• Networking and switching
• Advanced packaging
• Power and cooling
• Hyperscaler data centers

Booming

Cycle 02

Premium devices

• Apple, Samsung flagships
• Higher ASPs and trade-ins
• AI on-device features
• Premium PCs and tablets
• Wearables and XR
• Sticky replacement cycles

Holding

Cycle 03

Low-end consumer

• Sub-$200 Android
• Emerging-market income pressure
• Memory inflation in BOM
• Channel inventory caution
• Mass-market PCs and tablets
• Consumer electronics

Fragile

The new semiconductor cycle is AI up, premium holding, low-end consumer fragile. Three speeds inside one industry.

In the old cycle, weak phones meant weak semis. In the new cycle, weak phones can coexist with a booming semiconductor industry.

---

IX. Why the data looks confusing

Different research firms report different Q1 2026 smartphone numbers. Some show declines, some show mild growth driven by front-loading ahead of expected memory price increases, some flag a worse second half.⁸⁹ The headline numbers look noisy because they are catching three things at once.

Vendors are managing a memory shock in real time. They are front-loading shipments, protecting premium models, cutting low-end exposure, de-speccing devices, adjusting prices, and trimming orders. Every quarterly tracker captures a different slice of that activity. The underlying trend is consistent, even if the surface-level shipment numbers diverge.

Everyone is trying to manage the memory shock before it fully hits the P&L.

---

Quick terms

---

X. What could break the thesis

A serious piece needs counterarguments. The case that AI is repricing the smartphone market has plausible failure modes.

The correct claim is not that smartphones are permanently broken. The correct claim is that AI has changed the memory-cost floor under consumer devices.

---

XI. The AI memory tax

The smartphone is no longer the centre of the semiconductor cycle. AI is.

In 2021, phone weakness hurt chips because demand collapsed. In 2026, phone weakness is partly caused by chips becoming too valuable somewhere else. AI servers are pulling memory supply toward data centers. Memory prices are rising. Low-end phones are being squeezed. Premium vendors are better protected. The old semiconductor cycle is splitting.

That is the AI memory tax.

---

This is Essay No. 019. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.