From cyclical short term contracts to strategic long term partnerships
The conversation with Sumit Sadana confirms the call message that the memory industry has changed from a cyclical short term contract industry to a strategic relationship one, with long term take or pay contracts.
We believe this is very relevant for the discussion on Micron valuation, which we believe could be attractive if these new market trends confirm!
In this post we dissect the fireside chat, moderated by John Vinh (KeyBanc), with Sumit Sadana, EVP & Chief Business Officer at Micron (with the company since 2017
KeyBanc opened the conference citing Gartner forecasts of global IT spending up 13.5% for the year and data center spending growing over 55%, positioning Micron as one of the companies most directly exposed to that.
1. The genuinely new information
Two items were disclosed as updates since the last earnings call, and they are the most market-relevant parts of the whole session:
- Demand signals have gone up again. Aggregate demand signals from customers have increased further since earnings. As a consequence, Micron now expects calendar 2027 to be tighter than calendar 2026 — a step up from the prior message, which was only that tight conditions would persist beyond CY2027. The reasoning given is mechanical: industry demand growth for CY2027 is now running faster than industry supply growth.
- More SCAs have been signed since the earnings call, beyond the 16 announced there.
2. The central thesis: “this is not a cycle”
Sadana’s argument for why this period differs from prior memory cycles rests on four linked claims:
AI is on a self-accelerating path. He explicitly stated a belief that the industry is heading toward AGI and then artificial superintelligence, and that the timeline is compressing because AI is now being used to improve AI (”recursive self-improvement”). This is unusually strong language for an operating executive on an investor stage, and it underpins the rest of the demand case.
Demand is still at an early stage, not a peak. Three vectors of expansion were named: billions of users worldwide who don’t yet have AI access; Fortune 1000 / large global enterprises that are in the “very early innings” of deployment; and the shift from chat to agentic interaction.
The quantified point here is important — agentic workflows consume 5–30x the tokens of a chat interface for a comparable task, and deep-reasoning agents consume even more.
The digital/physical asymmetry. Demand grows in the digital domain; supply must be built in the physical domain. Sadana argued there is inadequate appreciation among investors for how long the supply vector takes to respond, and that this is a structural mismatch rather than a timing one.
The memory wall is a hardware-utilization problem, not just a capacity problem. In many AI workloads the processor — GPU, ASIC, or CPU — sits idle roughly 50% of the time waiting on data from DRAM. That reframes memory from a component cost into a determinant of return on the far more expensive compute asset.
3. The supply-side mechanics (the crux of the bull case)
This section carried the most concrete technical content:
- The HBM trade ratio. Producing 100 bits of HBM requires forgoing 300 bits of DDR supply — the 3:1 ratio at HBM3E. That ratio degrades to roughly 4:1 by HBM4E. So every unit of HBM growth cannibalizes conventional DRAM supply at an accelerating rate.
- Wafer supply must therefore rise disproportionately, and the industry has arrived at a point where most expansion has to be greenfield — building an entire fab cluster from empty land.
- Fabs are harder than data centers. Sadana was direct that building a leading-edge front-end fab is a considerably harder engineering and construction project than building a data center, and that the technology ramp afterward is also slow.
- The customer bottleneck has shifted. The consistent message from customers — described as a common theme, not one or two accounts — is that their #1 constraint is DRAM, ahead of power, real estate, data center capacity, or logic wafers.
- Micron cannot see the crossover point. Repeated three separate times: there is no line of sight to when the supply vector intersects the demand vector.
Investment response: U.S. investment plans raised from $200B to $250B; expansion also in Japan, Taiwan, Singapore, and back-end manufacturing in India; a $500M investment in GlobalWafers for raw wafers as part of an approximately $3B supply chain investment.
[See our Semiconductor Capacity Dashboard!]
4. Strategic Customer Agreements (SCAs) — the most detailed segment
Vinh framed these as potentially covering roughly half of Micron’s revenues, with upfront capital commitments north of $20B. Sadana’s account:
Origin. Micron says it pioneered the SCA construct, began working on them with customers many months ago, initiated the proposal itself (rather than reacting to inbound requests), and believes it has signed the most SCAs in the industry.
The pricing mechanism, in detail. For the 16 SCAs announced at earnings, the ceiling is set at CQ2 pricing. SCAs signed later will have ceilings consistent with market prices at the time of signing. The floor of the band (transcribed as “flow pricing,” almost certainly floor) is set at a level that delivers a gross margin well above any prior peak in industry history — this is arguably the single most important disclosure in the session, because it converts a cyclical business into one with a contractually underwritten margin floor across a large share of revenue.
Capital discipline caveat. Micron will build out the clean rooms but will fit equipment into them only in line with its medium-term view of demand. This is a deliberate signal that the company is not pre-committing to a full-capacity equipment ramp.
5. The pushback: will customers break the SCAs?
Vinh raised the strongest bear objection directly, citing the post-COVID precedent where legally binding LTAs were broken once supply normalized. Sadana gave three answers:
- Contractual. The old LTAs had no binding terms at all; the SCAs have extensive binding terms with no contractual exit.
- Strategic. Customers now view memory on a 10–20 year horizon because of AI, so the mindset behind the agreement is different.
- Reputational/allocation. This is the most candid answer: shortages will recur, and customers who behave “adventurously” on SCA compliance may find themselves poorly allocated next time. He noted that customers who historically treated memory tactically are struggling more than others to get allocation right now — an explicit statement that Micron is allocating on the basis of relationship history.
6. Customer relationships and differentiation
Sadana said engagement intensity is at a level he has not seen in his career, with three consequences:
- Roadmap visibility now extends beyond 2030 — versus an industry where it was historically rare to know a customer’s next-year product. He claims Micron now has longer-term visibility than most of the broader semiconductor industry.
- A repeatable innovation pattern: be first to innovate with a customer → win sole-source → move to dual-source. Examples cited: HBM3E with 30% lower power consumption than the next-best product; LPDRAM brought into the data center for the first time with NVIDIA (sole-sourced with strong ROI, becoming multi-sourced only because Micron couldn’t keep supply up); first to market with Gen 6 data center SSDs; the industry’s highest-capacity 245TB SSD.
- Supply, not price, is the binding constraint on customers. In data center, Micron often cannot meet more than half of customer demand. Customers are signing SCAs while telling Micron they are leaving considerable opportunity on the table.
7. Q&A — the three audience questions
Q1 — Does U.S. production remain a differentiator given Samsung in Texas and SK Hynix in Indiana?
Sadana’s answer was a clean distinction: Samsung’s Texas investment is logic foundry, not memory; SK Hynix’s Indiana investment is back-end assembly and packaging. Micron is therefore the only company investing in front-end fab manufacturing for memory in the U.S.
The footprint: Idaho 1 online mid-2027, Idaho 2 at end of 2028, a New York cluster beyond that, and Virginia for long-life-cycle technology with 1-alpha DRAM. Crucially, he stated Micron expects to earn a pricing premium for U.S. supply, and that premium is already baked into the SCAs.
Q2 — When does custom HBM become material, and what does it do to share?
Custom HBM SKUs begin at HBM4E. But the more consequential point was structural: HBM co-design and qualification is so R&D-intensive and expensive that customers cannot practically run all three suppliers across all projects. Most platforms and many customers will use one or two HBM suppliers; only a small number will use all three. Sadana extended this to the broader portfolio — memory is becoming “ASIC-like,” with sole- or dual-source positions persisting for long periods. This is the highest-leverage claim of the Q&A: it implies HBM share is decided at design-in and is sticky, which changes how HBM market share should be modeled versus commodity DRAM.
Q3 — Could dynamic tiering of cold data from DRAM to flash reduce DRAM demand?
Sadana inverted the premise. Rather than DRAM being displaced by NAND, he described the opposite happening today: because customers don’t have enough DRAM, things like KV cache are already spilling over into NAND. As reasoning and agentic workloads lengthen context windows, DRAM requirements rise further. His framing of the two gates: first, processor-to-memory bandwidth (which drives HBM-style architectures); second, raw DRAM capacity in the system. He saw no threat here — though note he answered about the current supply-constrained state rather than about a future in which tiering software matures.
- Overall read
The message being sold is that memory has changed category — from a cyclical commodity to a strategic, contracted, partially design-in-locked business, with a margin floor written into multi-year take-or-pay contracts, a U.S.-supply price premium, and demand vectors (agentic AI, then robotics) stacked behind the current one. The closing statement was consistent with this: a materially different company with a structurally stronger business model, tremendous tailwinds over the next 5–10 years, and SCAs as the transformational element.
Thank you,
CoreValue Research
