Micron's Record Profits, Apple's CXMT Plea: AI is Eating All the Memory

Summary

Austin and Vik frame the current memory market as an AI-driven crisis and supercycle, with Micron posting extraordinary revenue and margin expansion while supply is effectively booked out for years. They argue that AI/datacenter demand is largely inelastic, so any consumer pullback from higher prices will likely be absorbed by hyperscale and enterprise buyers. The episode also covers Apple’s unusual pricing actions and reported attempt to source DRAM from CXMT, a Chinese supplier on the U.S. entity list, as a sign of how tight supply has become. The hosts dig into why HBM is so memory-hungry, why agentic AI expands DRAM needs beyond GPUs into CPUs and pooled memory systems, and why the next battleground may be model routing, token cost, and on-prem enterprise AI deployment.

Notes

• Micron is presented as the poster child for the memory supercycle, with a single quarter around $41.5 billion and gross margin near 85%.
• The hosts treat AI memory demand as the core driver of the current price spike.
• Consumer electronics inflation is not limited to phones; even niche hardware is affected because almost everything now embeds memory.
• The episode’s thesis is that memory inflation is now visible to ordinary consumers worldwide, not just enterprise buyers.
• Micron says supply is effectively spoken for through at least 2027.
• Long-term agreements and prepayment clauses are described as locking in demand even if customers do not fully take delivery.
• GoPro is used as an example of a discretionary device maker squeezed by memory costs and weak willingness to pay.
• Apple’s price increases are framed as unusual because they happen mid-cycle rather than at a normal product refresh.
• Apple may be trying to protect gross margins by passing through higher memory costs.
• Apple reportedly asked CXMT for DRAM supply, despite CXMT being on the U.S. entity list.
• The hosts describe the CXMT move as extraordinary because it would mean a major U.S. company seeking an exception to source from a blacklisted supplier.
• Samsung’s vertical integration is treated as a strategic advantage because it can make memory and phones.
• Lower-tier phones may avoid sticker shock by cutting memory specs instead of raising prices.
• Korean memory investment is described as enormous, with the government putting north of half a trillion dollars behind DRAM capacity expansion.
• SK hynix is said to have overtaken Samsung as Korea’s most valuable company, signaling how much memory has re-rated.
• Kioxia is highlighted as a NAND-flash winner, even surpassing Toyota in domestic market value.
• Micron’s margin expansion is shown as a rapid step function over four quarters.
• The hosts characterize current pricing as effectively a bidding war among desperate customers.
• Micron’s turnaround is portrayed as dramatic: from losses to a massive quarterly profit.
• The market is still cyclical in structure, but the hosts think AI is pushing it into an unusually strong phase.
• HBM is described as a major DRAM sink because stacked packaging reduces effective bit density.
• The technical reason given is through-silicon vias, which require drilling through dies and reduce usable density around the vias.
• Agentic AI is said to need not just HBM but also plain DRAM for CPUs and long-context workloads.
• The hosts imagine pooled memory appliances as a way to serve KV-cache workloads through CXL.
• Data-center buyers are described as effectively price-insensitive and willing to buy all available wafers.
• Consumer demand is the opposite: elastic, more likely to slow when prices rise, and therefore a shock absorber for the market.
• The hosts explicitly call this an AI memory supercycle, not just a temporary shortage.
• They also note that supply is unlikely to catch up until 2027 to 2029.
• Restrictions on frontier model releases are discussed as one factor that could slow training-side demand.
• Austin cautions that slowing American frontier releases could also help Chinese labs catch up.
• The second half shifts from training to inference, which they see as the larger and longer-lasting source of memory demand.
• Vikram argues companies now need an agentic-AI posture or risk appearing obsolete.
• Token economics, not just model quality, are becoming the main competitive dimension.
• Austin expects both frontier and cheaper models to grow, with task-specific substitution between tiers.
• He also expects some workloads to move to open-source or on-prem token generation to avoid API costs.
• Frontier models still matter because they unlock new capabilities that lower-cost models do not.
• Companies are already fine-tuning proprietary models for domain data that general models do not know.
• Model routing is identified as a likely standard layer in enterprise AI stacks.
• Austin says he wants visibility into which agents use which models, because orchestration itself is becoming an operational problem.
• Vikram draws an analogy to internal simulation farms to argue that large firms can run local AI systems on-prem.
• He says corporate AI adoption is constrained by fears of data leakage into cloud models and by the risk of future distillation.
• Austin closes by saying AI adoption is gated by money, security, and latency, not just raw model quality.