Semiconductor Watch: Apple Tests Chinese-Made Memory as Chip Shortage Reshapes the Global Supply Chain

The CODEW Semiconductor Watch | August 14, 2026

When Apple Buys Chinese Memory: How Severe the Chip Shortage Has Actually Become

Executive Summary

The Shortage Is Now Bending the Procurement Behavior of the World's Most Powerful Buyer

Global semiconductor sales hit $403.3 billion in Q2 2026, up 35.1% quarter over quarter and on pace to exceed $1.5 trillion for the full year, according to the Semiconductor Industry Association. TSMC's July revenue grew 44.7% year over year. Those are the headline growth numbers. The more telling story broke the same week: Apple is reportedly testing memory chips manufactured in China for use in iPhones and MacBooks, because it cannot secure enough of the memory it actually wants from its traditional suppliers. When the company with arguably the most leverage in the entire electronics supply chain starts qualifying alternate-country memory out of necessity, that is a far more reliable signal of shortage severity than any industry forecast.

This edition tracks that shortage alongside two other structural shifts: Anthropic's move to co-design custom AI inference chips with Samsung as manufacturing partner, explicitly to reduce dependence on Nvidia GPUs, and a wave of new government-backed capacity commitments in South Korea and Japan. Semiconductor capacity constrained here becomes the physical hardware bottleneck covered in today's Hardware Watch, which in turn determines how much bandwidth the networking layer in today's Networking Watch actually has to move.

The Semiconductor Lead

Apple Tests Chinese-Made Memory as Its Traditional Supply Chain Runs Dry

Apple has historically been able to use its scale and premium pricing to command priority allocation from Samsung, SK Hynix, and Micron, insulating it from the component shortages that hit other device makers first. Reports that it is now testing Chinese-manufactured memory chips for iPhones and MacBooks mark a genuine break from that pattern. It is not a cost-driven decision — Apple pays premium prices specifically to avoid this kind of supply-chain compromise — it is a scarcity-driven one, forced by a memory market where, as Micron's own CEO Sanjay Mehrotra has told investors, the gap between DRAM and HBM demand and supply is the widest the company has ever measured.

The timing lines up with a broader pattern documented on August 10: South Korea creating a $3.52 billion government fund to strengthen its domestic semiconductor industry, Sony and TSMC pooling $6.32 billion to build a new Japanese fab for image-sensing chips, and TSMC's July revenue growth of 44.7% confirming that even the industry's most capacity-constrained leader is still finding room to grow. Four separate signals in a single week, from four different parts of the world, point to the same underlying condition: AI-driven demand is growing faster than the industry can build new capacity, and every actor in the chain — governments, foundries, and now even Apple — is adjusting procurement and industrial policy in response.

AI Chip Competition

The most consequential new development in custom silicon this week is Anthropic reportedly co-designing custom AI inference chips with Samsung serving as manufacturing partner, explicitly framed as a way to reduce reliance on Nvidia GPUs for inference workloads. This puts Anthropic in the same category as Google, Amazon, and Microsoft — a frontier lab building leverage over its own unit economics by owning silicon rather than renting it — but it is a more striking move coming from a pure AI lab rather than a hyperscaler with decades of chip-design infrastructure already in place, and it signals that even well-capitalized labs increasingly see custom inference silicon as unavoidable rather than optional at their current scale.

Elsewhere in the custom silicon race, AMD disclosed it plans to deploy Taalas chips alongside its own GPUs for AI inference, likely functioning as a dedicated LLM decode accelerator — a sign that even merchant GPU vendors are beginning to pair their general-purpose accelerators with narrower, purpose-built chips for specific stages of the inference pipeline rather than relying on a single chip architecture to do everything. AMD's own results underline the split playing out across the industry: data center revenue roughly doubled year over year, while gaming revenue fell 31%, with CEO Lisa Su directly attributing softer consumer demand to pricing pressure "weighing on" the client market — the AI chip boom and the memory-driven consumer squeeze are now visibly the same story told from opposite ends of AMD's business.

HBM & Advanced Packaging

Apple's memory sourcing shift is the clearest available proxy for how tight HBM and conventional DRAM capacity remains: Samsung, SK Hynix, and Micron have redirected the overwhelming majority of DRAM output toward HBM for AI accelerators, and Micron has already locked in price and volume agreements covering its entire calendar-2026 HBM output, leaving essentially nothing in reserve for conventional consumer or even premium-device memory. CoWoS advanced packaging remains the tighter structural constraint behind the memory story, and GlobalFoundries' data-center-driven growth is now being read across the industry as the leading edge of a broader U.S. photonics manufacturing buildout — packaging and photonics increasingly discussed in the same breath as memory itself as capacity bottlenecks that determine how fast finished AI systems can ship.

Foundry & Manufacturing

TSMC's July revenue growth of 44.7% year over year confirms the foundry leader is still finding incremental capacity to sell even amid a historically tight supply environment, and its new joint venture with Sony — a $6.32 billion Japanese fab dedicated to image-sensing chips for machine vision — extends TSMC's manufacturing footprint into a new specialized category rather than simply adding more general-purpose leading-edge capacity. That geographic and product diversification mirrors a broader trend: rather than concentrating exclusively on Taiwan-based leading-edge expansion, the industry's largest players are increasingly building specialized, geographically distributed capacity for specific AI-adjacent applications.

The most striking new manufacturing commitment this week comes from outside the traditional foundry roster entirely: Elon Musk's Terafab chip-manufacturing facility, backed by an initial $16.8 billion in capital and spanning roughly 100 million square feet, is reportedly beginning to take shape. A facility at that scale, built by a company without prior semiconductor manufacturing experience, is a genuine wildcard in the foundry landscape — evidence that the capital available for new entrants chasing AI chip capacity is now large enough to fund attempts at greenfield manufacturing scale that would have been considered implausible outside TSMC, Samsung, and Intel just a few years ago.

Supply Chain

Government intervention in the semiconductor supply chain is accelerating on multiple fronts simultaneously. South Korea's new $3.52 billion fund is explicitly designed to strengthen its domestic industry "from the bottom up" — targeting smaller suppliers and equipment makers rather than only the memory giants at the top of the chain. Japan's Sony-TSMC image-sensor joint venture reflects a similar logic: rather than compete head-on for general-purpose leading-edge capacity, Japan is building out specialized capacity in a category where it retains genuine technological leadership. Both moves reflect the same underlying anxiety driving Apple's memory-sourcing shift and Anthropic's custom-silicon push: no single company or country wants to remain fully dependent on a supply chain that is this concentrated, and this stretched, and every actor with sufficient capital is now hedging by building or securing alternate capacity wherever it can.

Three Semiconductor Signals

1. The shortage has broken even Apple's historical supply-chain leverage. A company that has spent two decades using its scale to secure priority component access is now qualifying alternate-country memory out of necessity — the clearest available signal of how little slack remains anywhere in the DRAM and HBM supply chain.
2. Custom silicon has spread from hyperscalers to pure AI labs. Anthropic's Samsung-manufactured inference chip program shows that owning silicon is no longer a hyperscaler-only strategy — any sufficiently well-capitalized AI company now views custom chips as necessary insulation against Nvidia dependency and pricing power.
3. Capital available for new semiconductor capacity now exceeds what traditional industry players alone can deploy. Government funds in South Korea and Japan, plus a $16.8 billion greenfield fab bankrolled outside the traditional foundry industry entirely, show that AI-driven capital availability is now large enough to fund manufacturing bets that would have been unthinkable just a few years ago.

The CODEW Take

The semiconductor layer gaining the greatest strategic importance right now is still memory — but the Apple story adds a sharper edge to that conclusion than any capacity statistic could. When a company with Apple's scale and negotiating leverage starts testing alternate-country memory suppliers rather than simply paying more for its usual sources, it confirms that memory scarcity has become a genuine allocation problem, not a pricing problem that money alone can solve. That is precisely why Anthropic is now designing its own inference silicon and why governments in Seoul and Tokyo are committing fresh billions to domestic capacity: control over memory and packaging capacity has become more strategically valuable than control over any individual chip design. Everything downstream — the physical hardware in today's Hardware Watch and the network fabric in today's Networking Watch — is still waiting on this layer to loosen.

---

THE CODEW · SEMICONDUCTOR WATCH

Editorial Note

The CODEW Semiconductor Watch examines the developments reshaping the global semiconductor industry, including AI chips, memory, advanced packaging, foundries, semiconductor equipment, supply chains, manufacturing capacity, and the growing geopolitical importance of critical chip infrastructure.

Analysis is based on company disclosures, product developments, industry research, public reporting, and other publicly available information. Semiconductor demand, supply conditions, technology roadmaps, and geopolitical developments can change rapidly.