Has the Memory Cycle Broken? Plus: Nvidia & Broadcom’s…

PodcastThe CircuitBen BajarinOct 5

The discussion weighs whether memory supply could catch up with demand by 2028, as CXMT expands capacity and memory makers balance legacy products against higher-value HBM. Speakers also examine how NVIDIA and Broadcom use financing and balance sheets to compete, while noting the risks of debt in a cyclical industry. Longer device upgrade cycles may constrain non-data-center demand, and the timing of any memory-market downturn remains uncertain.

Semiconductors Technology Stocks Corporate Earnings

Preview · 21 more points in the app

Read the Full Summary

Key Takeaways

Agentic AI is lifting CPU demand and, with it, demand for LPDDR and other non-HBM memory, while longer PC and phone upgrade cycles could weigh on non-data-center demand. The speaker says AI phones could eventually spur upgrades, but only if they become compelling enough.

Memory supply and pricing remain uncertain: the speaker says CXMT can bring fabs online in about 18 months versus two and a half to three years for producers in the U.S., Korea, or Japan, and expects substantial memory availability by 2028. The speakers discuss a possible 2028 supply-demand equilibrium if DRAM bit growth stays in the low twenties and NAND in the mid twenties, but they do not agree on the timing or outcome, and the market’s future ASPs and pricing power remain unresolved.

Constrained wafer capacity creates an economic tradeoff between HBM and other memory: the speakers illustrate that capacity for four consumer chips could instead produce one leading-edge HBM chip selling for 100 times more. The speaker says consumer-price relief may therefore be limited if new capacity and R&D favor premium memory, although continuing DDR4 production in older fabs could help normalize supply as new factories come online.

The speaker argues that new memory products, processes, transistor designs, and stacking could preserve value in some segments, but does not expect margins to return to the 30s and is uncertain how large those segments will be. That leaves open where premium pricing power will persist even as the speaker describes memory as a structurally larger market whose demand could normalize or remain static for a period.

Balance-sheet financing has become a competitive tool in AI semiconductors: the speaker cites NVIDIA’s roughly $60 billion in cash and quarterly cash generation 20 times AMD’s, while AMD has about $10 billion in cash. NVIDIA framed its balance sheet as a way to fund ecosystem growth, and Broadcom is committing substantial financing for Google TPUs going to Anthropic; the speaker also notes the risk that debt payments could constrain R&D if a company misses a product cycle.

TECHNOLOGY

Broader Memory Demand — The speaker says agentic AI is driving CPU demand and, in turn, demand for lower-cost LPDDR and other memory used with CPUs—not just HBM.

“There's lots of things that will smooth it, and I think, you know, I think for many people, what matters most is the cost of memory, is just like everything…”—

CXMT Capacity — The speaker says China’s CXMT is pulling forward capacity targets and can bring memory fabs online faster than producers in the U.S., Korea, or Japan: about 18 months versus two and a half to three years. The speaker expects substantial memory supply by 2028 but says the pricing effect is uncertain.

“CXMT is pulling forward all its capacity targets. They're gonna have a lot more capacity next year. They're building fabs very quickly…”—

Memory Margins — The speaker argues that new memory products, processes, transistor designs, and stacking could preserve value in some segments. The speaker does not expect margins to return to the 30s, but acknowledges uncertainty about the size of those segments.

“Like, that's that's not slowing down. Like, there's memory innovation coming that will change this as well…”—

MARKETS

Consumer Memory Prices — Reports that consumer-memory prices have plateaued could reflect a deliberate decision by memory makers, but the speaker stresses this is only an interpretation; Micron’s call did not clearly confirm it. The speaker worries new capacity and R&D may be directed toward premium memory, limiting relief for consumers.

“There are some reports that like memory prices for consumer memory have plateaued, which I would read as a deliberate decision by the memory makers, but I'm not sure that's…”—

HBM Capacity Tradeoffs — The speakers say producing 16-high HBM takes significant wafer capacity and, while capacity is constrained, can come at the expense of other memory categories. They illustrate the incentive conditionally: capacity for four consumer chips could produce one leading-edge HBM chip that sells for 100 times more, making HBM economically attractive.

“You start making, you know, 16 high HBM memory, the effective yield on that, the amount of wafers that it takes to produce that is significant, and it all comes…”—

FORECASTS

Memory Equilibrium — The speakers discuss a bear case in which DRAM bit growth stays in the low twenties and NAND in the mid twenties into 2027–2028, with supply and demand reaching equilibrium in 2028 and markets turning more pessimistic as early as the following year. One speaker expects parity sooner as memory makers add capacity, but says the timing and shape of any downturn are unclear; another says they are not taking a stand on the outcome.

“I think the company said DRAM and NAND are growing sort of in the twenties, low twenties for DRAM and mid twenties for NAND into '27, '28, and so I…”—

Memory Cycle Outlook

The hosts debate whether memory demand and pricing can sustain a larger market or whether capacity additions will bring a downturn before 2029.

Supply And China Capacity

Jay weighs competing cycle forecasts, rising memory prices, and faster Chinese fab construction, arguing that supply parity may arrive sooner than the most bullish outlook assumes.

HBM Versus Legacy Memory

Ben distinguishes price-sensitive standard memory from design-in HBM and argues that persistent bit growth could let manufacturers keep monetizing older generations.

Consumer Memory Constraints

Jay questions whether memory makers will preserve legacy capacity when HBM offers stronger returns, while the hosts discuss shortages and longer device replacement cycles.

Micron Results And HPE

The hosts call Micron's quarter and guidance strong but note that investors remain uncertain about the cycle, then turn to HPE's networking strategy after Juniper.

HPE Networking Strategy

Ben describes enterprise security concerns driving network refreshes and HPE's engineering work on AI infrastructure, while Jay questions whether HPE can capture lasting value.

NVIDIA Financing Strategy

The hosts examine NVIDIA's use of its balance sheet to fund ecosystem growth, viewing it as vendor financing that can help in a capital-constrained market but raises risk.

Broadcom's $42 Billion Loan

Broadcom's proposed financing for Anthropic prompts a comparison with foundries; Jay says chip designers are closer to end demand and better positioned to assess and monetize it.

AI Infrastructure Competition

Jay argues NVIDIA is pushing rivals to expand their offerings and use their balance sheets, while warning that debt adds risk in a cyclical, high-fixed-cost industry.

AI Token Economics

Jay argues that token pricing and compute costs can make AI economics more measurable, including how reasoning features, input and output tokens, and inference costs are priced.

OpenAI Speed Pricing

The hosts discuss OpenAI's faster model tier and whether speed or specialized models could command higher prices, while emphasizing that token pricing will likely depend on product and customer needs.

Enterprise AI Pricing

Jay says enterprises will need better visibility into AI providers' costs to negotiate for specialized products, and the hosts close by previewing future episodes.