Overpricing a Bottleneck Can Backfire on AI CapEx

Disclaimer: This article represents my personal reflections and review of public industry materials and business logic. It does not constitute investment advice, nor does it involve any individual stock recommendations or implied positions. Markets carry risk; make your own independent decisions.

While reading *My Grudge Against Memory*, I noticed a reverse constraint. I used to look at bottlenecks through a simple lens: the tighter the bottleneck, the higher the profit. Now I add a prerequisite to that framework: a supply chain bottleneck only continues to translate into excess returns while customers remain both willing and able to absorb the price. Once a bottleneck segment raises prices too far, customers have to adjust their CapEx, free cash flow, and architectural choices, and the bottleneck itself shifts from a source of profit to a demand risk.

Memory (DRAM / HBM) makes this very concrete. Supply is tightly concentrated among a handful of memory makers, and AI demand has handed them temporary pricing power. I agree with that part. But if memory costs keep taking up a larger share of cloud providers' AI CapEx, those cloud providers may move to optimize memory usage, delay purchases, renegotiate long-term agreements, seek financing, or slow down their CapEx growth rate. Once price hikes start altering customer capital expenditures, I stop automatically equating the bottleneck with excess returns. Market expectations for memory makers might also turn before the financial numbers do.

The Bullish Narrative Is Missing Its Second Half

The old bullish narrative went like this:

AI needs more GPUs → GPUs need more HBM / DRAM → memory supply is tight → memory makers keep raising prices → memory makers keep benefiting.

The framework I use now adds a second half:

Memory prices keep rising → cloud providers' unit AI infrastructure costs rise → FCF and CapEx flexibility decline → customers push memory optimization / pooling / offload and CPU-side memory reductions → the market starts worrying about demand and industry profit margins peaking.

So when looking at the memory bottleneck, you cannot just ask whether demand is strong enough; you also have to ask whether pricing has damaged customers' ability to bear the cost.

What I Can Use from *My Grudge Against Memory*, and What I Cannot Use Yet

This piece compiles several types of secondhand leads:

• Market research firm charts: three major memory makers together control most of the DRAM market.
• Industry research and financial media charts: HBM is formed by stacked DRAM dies, TSVs, and microbumps, and advanced AI chips require more HBM layers and bandwidth.
• Investment bank and research institute charts: DRAM / HBM ASPs, along with memory's share of AI hardware and cloud provider CapEx, may continue to rise.
• Cloud provider earnings reports or management commentary: multiple cloud providers have mentioned component pricing or memory cost pressure.
• Memory maker management interviews: meaningful new capacity might not come online until 2028.

These leads work well as research entry points. The note does not go back item by item to verify them against company financial reports, earnings call transcripts, original reports from industry research firms, cloud providers' periodic filings, or official disclosures. I treat them strictly as entry points, not established facts.

The same material also contains a layer of inference, likewise unverified item by item: profit margins in the memory segment might reach abnormally high levels in 2027; memory costs as a share of AI CapEx might approach 40%; once cloud providers' cash flows are squeezed by capital expenditures, talk of slowing down or optimizing will show up on earnings calls; market expectations for memory makers might turn before profit margins actually decline because the room for upward revisions disappears; and after 2028, as new capacity is released, memory might revert to cyclical commodity logic.

The shakiest part of this inference is the timing. For the call of a mid-2027 peak to hold, several things need to happen at the same time: memory prices keep climbing, customer cost pressure clearly worsens, AI revenue or strategic demand fails to cover the costs, and the market starts worrying about supply coming online in 2028. If even one of these conditions is missing, the timing does not hold up.

Connecting to Caching and Rewriting Bottleneck Research

When memory prices rise enough to affect full-system BOM and CapEx budgets, model providers and chipmakers look more aggressively for caching, offload, memory pooling, flash-as-DRAM, smaller models, quantization, and architectural optimization. I wrote about this separately in Long-Term Bottlenecks in AI Inference Costs May Shift from Compute to Cache.

This also reminds me that the "find the AI supply chain bottleneck" framework hides a risk: the tighter the bottleneck, the higher the profit; yet high profits easily push customers to route around the bottleneck, suppliers to expand capacity, competitors to enter, and the market to price it in early.

So to do a complete piece of supply chain research, I have to answer two sets of questions at the same time:

1. Is the bottleneck still there: Is this segment genuinely hard to expand, hard to substitute, and pulled along by rigid customer demand?
2. Is pricing backfiring: Have prices already risen high enough to make customers change their purchasing cadence, architectural plans, or CapEx guidance?

Five Types of Signals, and How I Weigh Them

These five types of signals are the checkpoints I use to discount the idea that "price hikes equal benefits." I do not treat them as an action checklist.

Wording on Customer Earnings Calls

I pay close attention to whether the main drivers of AI CapEx, such as cloud providers and AI cloud service providers, repeatedly use the following phrases on their earnings calls:

• memory pricing
• component pricing
• HBM availability
• AI infrastructure cost pressure
• CapEx efficiency
• optimizing memory usage
• supply constraints
• long-term supply agreements

Only when "memory pricing" shifts from a passing remark to a core risk item on earnings calls do I count it as an early signal of pushback. If it only shows up once, I will not jump straight to a conclusion.

The Slope of CapEx Guidance

Absolute CapEx still rising and the growth rate slowing down are two different things. I watch whether the slope is flattening rather than waiting for absolute numbers to drop before acknowledging risk. Risk in the memory segment does not necessarily come from a drop in CapEx; it may also come from several other shifts:

• Smaller upward revisions to CapEx;
• 2028 guidance failing to keep accelerating;
• Management starting to emphasize capital discipline;
• AI revenue growth failing to cover infrastructure cost growth.

Booms in cyclical industries often peak when things are "still good, but no longer getting better."

The Slope of ASPs and Industry Profit Margins

This item mainly tracks the ASPs of DRAM, server DRAM, and HBM, along with overall industry profit margins. When prices keep accelerating, it certainly benefits memory makers, but at the same time it gives customers stronger reasons to change their designs:

• Prices keep accelerating: positive for memory makers in the short term, while the risk of customer pushback is also growing.
• Prices flatten out at high levels: market expectations may start reflecting a peak in profit margins.
• Prices turn downward: the market will go back to viewing memory as a traditional cyclical commodity.

Whether Optimization Paths Have Become Product Constraints

The material mentions that accelerator makers might cut back on CPU-side memory modules in rack-scale solutions, and that other accelerator makers and cloud providers are also looking for memory optimization. Whether these paths have turned into actual products and procurement needs to be checked item by item; looking at a list alone does not prove anything:

• Cutting CPU-side DRAM;
• memory pooling;
• Memory expansion interconnects and flash-as-DRAM;
• KV cache offload;
• prompt / prefix cache;
• Smaller models, quantization, speculative decoding;
• Rebalancing HBM capacity and bandwidth requirements.

As soon as an optimization path lands in hardware configurations, it shows that customers are already treating memory cost as a system-level constraint.

When Supply Comes Online

The material treats 2028 as an important milestone. I will not treat that year as a confirmed inflection point; instead, I break it down into several supply questions that still need to be examined separately:

• HBM / DRAM capex at major memory makers;
• Timeline for new capacity coming online;
• Conversion yields from HBM3E to HBM4;
• Whether standard server DRAM continues to be crowded out by HBM;
• Whether customers' long-term agreements lock in high prices or leave customers with a buffer.

When looking at memory, enterprise SSDs, optical interconnects, or AI CapEx, I prefer to use the same question as a check rather than writing it up directly as advice: is the price increase at this bottleneck still boosting suppliers' excess returns, or has it already started destroying customer demand? You can line up the bullish and bearish clues side by side: rising ASPs versus flattening at high levels, continued upward revisions to profit margins versus approaching historical extremes, customers signing long-term agreements versus customers complaining about component pricing, delayed supply release versus new capacity clustering after 2028, upward revisions to AI CapEx versus slowing growth or financing pressure, and stronger HBM4 demand versus customers pushing memory optimization.

Where This Judgment Would Break Down

If any one of the following six points holds true, I will dial back my view that "pricing will backfire" by a notch.

1. AI revenue grows fast enough. If cloud providers can prove that AI revenue or productivity gains cover memory costs, CapEx may continue to be revised upward.
2. Supply remains tight over the long term. Even if customers complain about prices, if the release of supply after 2028 is still insufficient, memory makers' pricing power will persist.
3. Long-term agreements lock in profits. If memory makers use 3–5 year long-term agreements to lock in high prices and shipments, short-term profit margins will not necessarily pull back quickly.
4. Optimization paths have limited effect. If memory pooling, flash-as-DRAM, and CPU-side memory reductions only touch marginal costs, primary demand for HBM still cannot be bypassed.
5. The market has already priced in cyclical risk early. If expectations no longer bake in long-term high profits, "cyclicality is still there" will not necessarily bring a large shift in expectations.
6. Constraints show up in other segments first. If power, networking, packaging, or GPU supply gets bottlenecked earlier than memory, the memory backlash will not be as prominent.

The sweet spot for an AI supply chain bottleneck is when customers cannot do without it, customers can still absorb the price, and the market has not fully priced it in. Once pricing starts altering customers' CapEx and architectural choices, excess returns gradually turn into cyclical risk.

References

• *My Grudge Against Memory*: Memory supply concentration, HBM structure, ASP and CapEx share, cloud provider cost pressure, new capacity timelines, and the layer of inferences around 2027 / 40% (which have not been verified item by item) all come from the secondhand compilation in this material.
• Subsequent verification should go back to company financial reports, earnings call transcripts, original reports from industry research firms, cloud providers' periodic filings, or official disclosures.
• Long-Term Bottlenecks in AI Inference Costs May Shift from Compute to Cache