The AI Memory Supercycle: Who Actually Earns the GPU Dollar

A deep dive into the memory shortage, the Nvidia value chain, and where the profits are flowing

The Setup: Something Strange Is Happening in Memory

If you only watch Nvidia, you are missing the most interesting story in semiconductors right now.

In Q1 2026, SK Hynix posted an operating margin of 72%. That number is not a typo. It exceeded both Nvidia and TSMC in the same quarter. For a company that makes commodity DRAM, this is the kind of margin associated with luxury handbags, not memory chips. Industry veterans say profitability like this has not been seen since Microsoft launched Windows 95 in 1995.

This blog post pulls together the full picture: why memory is so tight, who the winners are at each layer of the Nvidia GPU stack, what the 2026 profit numbers actually look like, and where I think the cycle is in its arc. If you are an investor, an engineer, or just curious about how AI infrastructure economics actually work, this is the map.

Part 1: The Memory Market Today

The memory industry has three players in DRAM and four in NAND. The structure has not changed much in years. What has changed is which player sits on top.

DRAM Market Share (Q3 2025)

HBM Market Share (Q3 2025)

HBM (High Bandwidth Memory) is the highest-value segment of DRAM and is where the AI boom hits hardest. SK Hynix’s dominance here is the single biggest reason it overtook Samsung in total DRAM share.

NAND Flash Market Share (Q3 2025)

NAND looks more fragmented than DRAM on the surface, but the structure is misleading. Kioxia and SanDisk share manufacturing through Flash Ventures — a joint venture that has run continuously since 1999. If you combine their output, the Kioxia + SanDisk alliance commands roughly 28% of the global NAND market, putting it neck-and-neck with Samsung’s 32%. The “five-player” NAND market is really a two-and-a-half-player market, and that has important implications for the supply discipline narrative.

Kioxia’s BiCS8 transition has made it the surprise NAND outperformer in 2025-2026, and SanDisk — newly independent after its February 2025 spinoff from Western Digital — has become the most dramatic story in memory stocks. We’ll come back to SanDisk in detail later in this post.

Part 2: Why the Shortage Is Real (and Structural)

People keep calling this a “shortage” as if it is a temporary supply chain hiccup. It is not. It is a deliberate, structural reallocation of global silicon capacity.

The Wafer Math

Here is the core fact almost nobody outside the industry understands:

1 GB of HBM consumes roughly 3-4× the wafer capacity of 1 GB of standard DRAM.

HBM stacks 12 thinned DRAM dies vertically, connected by through-silicon vias (TSVs). The yield losses, extra process steps, and packaging complexity mean each wafer commitment to HBM removes far more bits from the conventional DRAM market than it adds to HBM.

The Economic Logic

Why would a manufacturer choose HBM over commodity DRAM?

Roughly 10× the revenue per equivalent bit. When fab capacity is fixed, every rational manufacturer pivots to HBM. SK Hynix did it first and most aggressively. Samsung followed once Nvidia qualification cleared. Micron is doing the same.

The Capacity Numbers

• Global DRAM capacity in 2026: ~40 EB
• AI “equivalent consumption”: nearly 20% of total
• Annual DRAM capacity growth: only 10-15%
• HBM share of all DRAM wafer starts in 2026: ~23% and rising

Why Capacity Cannot Catch Up Quickly

A new fab costs $15 billion or more and takes 18+ months to ramp. After the brutal 2022-2023 downturn, the major players are deliberately disciplined:

• 2026 DRAM capex: $61.3 billion (up just 14% YoY)
• 2026 NAND capex: $22.2 billion (up just 5% YoY)
• First new fab outputs (Samsung, SK Hynix, Micron): Late 2027 at the earliest, with Micron’s New York facility producing meaningfully by 2028

This is the cleanest case of structural undersupply we have seen in this industry in 30 years.

Inventory Levels Confirm the Tightness

As of January 2026:

• Samsung DRAM inventory: 6 weeks (vs. typical 10-12)
• SK Hynix DRAM inventory: 2-3 weeks (historical low)
• SK Hynix NAND inventory: 3-4 weeks

By late January 2026, all three majors began requiring customers to disclose end-customer information and actual order volumes — a hoarding screen. When suppliers verify whether your demand is real before they will sell to you, that is not a market near peak.

Part 3: How Much Memory Does an Nvidia GPU Actually Need?

The HBM growth curve per GPU is exponential, and that exponential is the demand multiplier feeding the shortage.

5 years, 12.8× growth in per-GPU memory capacity. That is faster than Moore’s Law.

A full GB200 NVL72 rack with 72 B200 GPUs contains roughly 14 TB of HBM. Microsoft’s Azure GB300 NVL72 supercluster with 4,608 GPUs contains approximately 1.3 PB of HBM in a single deployment.

Part 4: HBM as a Percentage of GPU Cost

This is the most surprising part of the story. The component people thought of as “just memory” is now the largest line item in a flagship AI GPU’s bill of materials.

Nvidia B200 Cost Breakdown

A few takeaways from this table:

1. HBM is now the single largest cost component in a flagship Nvidia GPU. Memory used to be 20-25% of GPU cost. It will keep climbing.
2. The compute logic — the thing Nvidia is famous for — is only 22% of BOM. The brand value, software ecosystem, and design IP earn Nvidia ~80% gross margin on top of that.
3. Each $1 increase in HBM price flows almost entirely to memory makers. They have pricing power inside Nvidia’s pricing power.

The Memory-Cost Sensitivity

If HBM3E spot prices rise another 20%:

• B200 manufacturing cost increases by ~$580 (about 9%)
• Nvidia’s gross margin compresses from 81% to ~78%
• Memory makers’ revenue from HBM rises 20% with essentially flat costs

That asymmetry is why memory stocks have outperformed Nvidia year-to-date in 2026.

Part 5: 2026 Profit Forecasts for the Big Three

SK Hynix

• Q1 2026 (already reported): Revenue 52.6 trillion KRW ($36 billion), operating profit 37.6 trillion KRW ($25 billion), operating margin 72%, net margin 77%
• 2026 full-year operating profit estimate (Mirae Asset): 148 trillion KRW (~$100 billion), +213% YoY
• 2026 full-year operating profit estimate (FnGuide): 210.7 trillion KRW (~$142 billion)
• Forward P/E: ~5x (historical low)

SK Hynix is having the best year of its 40-year existence by an enormous margin.

Samsung Electronics

• Q1 2026 (already reported): Total operating profit 57.2 trillion KRW ($39 billion); semiconductor (DS) division operating profit 53.7 trillion KRW ($36 billion), up 48× from Q1 2025
• 2026 full-year operating profit estimate (KB Securities): 123 trillion KRW ($83 billion), 3× last year; semiconductor division 105 trillion KRW ($71 billion), more than 2× the 2018 supercycle peak
• Forward P/E: ~7x (historical low)

Samsung’s HBM4 qualification with Nvidia (cleared late 2025) was the single most important development of the year. It restores Samsung as a credible #2 in the segment that matters most.

Micron Technology (FY26 ending August 2026)

• FY26 Q2 (already reported): Revenue $23.86 billion, non-GAAP operating profit $16.5 billion (beat guidance by 46%)
• FY26 Q3 guidance: Revenue $33.5 billion, EPS $18.90
• FY26 full-year revenue consensus: $108.7 billion (raised from $79.8 billion)
• FY26 full-year EPS consensus: $58.05 (raised from $34.26)
• FY26 net income forecast YoY growth: 251%
• Forward P/E: ~7-9x

Micron’s entire FY26 HBM supply is sold out under fixed-price, fixed-volume contracts.

Combined View

Three companies. Roughly $235-280 billion in combined 2026 profit. From an industry that lost money in 2023.

Part 6: HBM4 — The Next Battle

HBM4 is the product that will define 2026-2027. It powers Nvidia’s Vera Rubin platform and AMD’s MI400 series.

SK Hynix (still leading)

• World’s first HBM4 development complete; partnered with TSMC on 12nm base die
• Revealed industry’s first 16-layer 48 GB HBM4 module at CES 2026 with 11.7 Gbps/pin
• HBM4E samples to Nvidia in H2 2026, mass production 2027
• Expected to maintain >50% market share through HBM4

Samsung (returning to contender status)

• Cleared final qualification with both Nvidia and AMD in late 2025
• Mass production starting February 2026 with 12-layer HBM4
• 50% HBM capacity expansion planned for 2026
• HBM4E using Samsung Foundry’s 2nm base die — most aggressive process node bet
• Added 250 engineers specifically for custom HBM projects targeting Google, Meta, Nvidia

Micron (steady third)

• HBM4 mass production in 2026, HBM4E delayed to 2027-2028
• Stays on proven 1ß (5th gen, 10nm-class) process for HBM4 dies
• Will fall to 1γ for HBM4E (Samsung and Hynix already on 1γ for HBM4)
• Constrained by US fab construction timelines

Bottom line: SK Hynix retains technical leadership through 2026, but the gap narrows. Samsung is back as a real #2. By 2027, Samsung’s customized HBM strategy (proprietary base dies for individual hyperscalers) could be a structural advantage.

Part 7: The Forgotten NAND Story — SanDisk’s Spinoff Awakening

While the world stared at HBM, the most dramatic stock story in memory has been a company most generalist investors had written off years ago.

The Comeback

SanDisk (NASDAQ: SNDK) is a name with weight. The company co-invented the SD card format in the late 1990s, helped pioneer 3D NAND, and operated as one of the most innovative flash memory businesses in the industry for two decades. In 2016, Western Digital acquired it for $19 billion to build a unified storage giant. For the next nine years, SanDisk’s flash business was tethered to Western Digital’s slower-moving HDD division, and the conglomerate discount weighed on both.

In February 2025, activist pressure from Elliott Management finally forced the issue. SanDisk was spun off as an independent, pure-play NAND flash company, returning to Nasdaq under the SNDK ticker. The timing turned out to be spectacular.

Twelve months later, SanDisk stock had risen approximately 1,180%.

The Hidden Scale

On the surface, SanDisk holds 12.4% of NAND market share — fourth or fifth place depending on the quarter. This understates its real position dramatically.

SanDisk procures essentially all of its flash wafers from Flash Ventures, the joint venture it has operated with Japan’s Kioxia since 1999. SanDisk holds 49.9% ownership, Kioxia 50.1%, and they share fab capacity at Yokkaichi and Kitakami in Japan. Combine the two companies’ market positions and the picture changes:

The NAND market is not a fragmented five-player industry. It is a three-bloc structure: Samsung, the Kioxia-SanDisk alliance, and SK Hynix, with Micron in fourth. That alliance structure is part of why NAND supply discipline has held up better than skeptics expected.

In January 2026, SanDisk and Kioxia extended the Flash Ventures agreement through 2034, with SanDisk committing $1.17 billion in manufacturing payments between 2026 and 2029 to lock in advanced 3D NAND supply. The Kitakami Fab2 is now ramping with 8th-generation 218-layer 3D flash specifically targeted at AI storage demand.

The Numbers Are Hard to Believe

SanDisk’s transformation from sleepy commodity vendor to AI infrastructure beneficiary shows up most clearly in the financials.

A gross margin expansion from 22.5% to 78.4% in twelve months is essentially unprecedented in a commodity hardware business. It even exceeds SK Hynix’s 72% operating margin. Five multiyear customer supply agreements now carry more than $11 billion of contracted financial commitments, giving SanDisk visibility through the cycle that pure spot-market vendors lack.

The Real Story: High Bandwidth Flash (HBF)

The reason SanDisk is the most interesting stock in NAND isn’t the current pricing wave. It’s a product that doesn’t yet exist in commercial form.

HBF (High Bandwidth Flash) is a new memory architecture SanDisk is co-developing with SK Hynix. The goals are striking:

• 8 to 16 times the capacity of HBM at comparable bandwidth
• A fraction of the cost per gigabyte of HBM
• Engineering samples expected in H2 2026
• Target use case: AI inference, not training

The strategic logic matters. Training workloads need HBM’s extreme bandwidth. Inference workloads are different — they need enormous memory capacity to hold massive model weights, but their bandwidth requirements are lower. If HBF delivers on spec, it could:

1. Capture inference-GPU memory sockets currently filled by HBM
2. Give NAND companies a share of AI economics previously reserved for DRAM
3. Validate SK Hynix’s two-track bet (it makes both HBM and HBF — a structural hedge)

If 2026 is the year of AI training, 2027-2028 is likely the year of AI inference at scale. Inference workloads will outnumber training workloads by 100-1000×. A product purpose-built for that wave, at a fraction of HBM’s per-bit cost, could redraw the memory map.

A credible HBF demonstration in H2 2026 would re-rate SanDisk from “cyclical NAND beneficiary” to “AI memory platform.” That gap in narrative is the bull case.

The Risks

SanDisk is not without exposure:

• Cycle sensitivity — NAND has lower barriers to entry than HBM. China’s YMTC has reached 232-layer technology and is expanding fast. Recovery in NAND tends to invite supply faster than DRAM.
• The $1.17 billion fab commitment — comfortable in a rising NAND price environment, painful if pricing rolls over
• Kioxia M&A risk — Bain Capital remains a major Kioxia shareholder, and rumors of consolidation between Kioxia and other industry players persist. Any change to the Flash Ventures structure introduces binary risk into SanDisk’s manufacturing base
• Inference timing — if HBF samples slip into 2027, the narrative window narrows

How SanDisk Fits the Investment Map

Place SanDisk in the picture relative to the other memory plays:

SanDisk is the highest-beta, highest-narrative name in the group. The 12-month price action reflects that. It’s also the cleanest expression of the view that NAND specifically — and inference workloads more broadly — get their turn after the DRAM/training wave runs its course.

Part 8: The Full Nvidia GPU Value Chain

Memory is just one layer. To understand who actually earns the GPU dollar, you need the complete map.

Layer 1: Chip Design and Software (Nvidia)

Nvidia does no manufacturing. It captures value through GPU architecture design, the CUDA software ecosystem, and the NVLink interconnect. From a $35,000 B200 sale, it keeps approximately $28,500 in gross profit.

Layer 2: Logic Wafer Manufacturing (TSMC)

TSMC manufactures roughly 90% of the world’s most advanced chips, including every Nvidia data-center GPU. Q1 2026 figures: $35.7 billion revenue, ~64% gross margin, $17.2 billion net profit. Nvidia is its largest single customer at 22-25% of revenue. The company has guided to >30% YoY growth for 2026 in USD terms.

Layer 3: HBM Memory (SK Hynix, Samsung, Micron)

Covered above. The single largest cost line in a flagship GPU and the single most profitable segment of memory.

Layer 4: Advanced Packaging (CoWoS)

CoWoS is the second-biggest physical bottleneck after HBM. Nvidia has reportedly booked more than half of TSMC’s CoWoS capacity for 2026-2027.

Layer 5: Substrates and Interposers (the hidden bottleneck)

This layer is critical, undercovered, and dominated by just two companies.

Top 5 ABF substrate makers control 74% of the market. The underlying ABF (Ajinomoto Build-up Film) IP is controlled 98% by Japan’s Ajinomoto — yes, the seasoning company. Switching substrate suppliers takes years of qualification, which is why this layer is one of the most defensible chokepoints in the entire chain.

Layer 6: Semiconductor Equipment (the picks-and-shovels)

This layer does not show up in GPU unit pricing, but it determines all future capacity.

ASML deserves special mention: a single Dutch company is the literal physical bottleneck for the entire AI infrastructure buildout. It has no real competitor.

Layer 7: Networking and Interconnect

Layer 8: Power and Cooling (rapidly growing share at rack level)

Layer 9: System Integration (ODM / OEM)

ODM gross margins are 2-5%, but the volumes are massive. Foxconn’s AI server business is now larger than its iPhone business by some measures.

Layer 10: Wafer and Materials

Part 9: How a $35,000 B200 Sale Distributes

A simplified value distribution at the GPU level:

For a complete GB200 NVL72 rack at ~$3 million, you also need to add:

• Cooling vendors (Vertiv etc.): $50,000-$80,000
• ODMs (Foxconn etc.): $30,000-$60,000
• Interconnect (Amphenol, Coherent): $100,000-$150,000
• Optical modules: $80,000-$120,000

Part 10: Where Are We in the Cycle?

Most people are asking “should I buy?” The better question is “where are we?”

Six Indicators of Cycle Position

1. Inventory levels — at multi-year lows, signal continued tightness
2. Capex discipline — 14% growth, not 50%, suggests no glut yet
3. Long-term contracts — 3-5 year deals with hyperscalers, often with prepayments
4. Forward P/E — at historical lows (5-9x), but historically lows precede peaks
5. Capacity coming online — first new fabs not until late 2027
6. Stock price action — strong but not parabolic relative to earnings revisions

My Read: Mid-Expansion (Roughly 50-60%)

This is not the early innings. Six quarters of reported strength have already happened. But it is also not late-cycle. Here is why.

Bull case (what supports continuation through 2027):

• Inventory is genuinely low across all three suppliers
• Cleanroom capacity is physically constrained until 2027 ramps
• HBM demand is locked in by 3-5 year contracts with prepayments
• Capex is unusually disciplined; most spend is going to upgrades, not capacity
• The DRAM pricing trajectory is unprecedented: contract prices rose 90-95% QoQ in Q1 2026, with Q2 expected to add another 58-63%
• Memory revenue base is heading from $90 billion (2023 trough) to a projected $550-570 billion in 2026, potentially $800-850 billion in 2027

Bear case (what would mark the top):

• New fab capacity from all three suppliers begins arriving in late 2027
• AI capex deceleration at hyperscalers (the single biggest watch item)
• Inventory rising back above 12-15 weeks
• TechInsights and some other forecasters call for a 2027 downturn
• DDR5/DDR6 prices could decline 50%+ within a few quarters once the cycle flips
• Industry has done this before — every previous supercycle ended in oversupply

Watch List for Calling the Turn

In rough order of importance:

1. DRAM contract price QoQ momentum
2. Three-supplier inventory weeks
3. Capex revisions (especially upward revisions)
4. New fab ramp timelines (especially Micron Idaho mid-2027)
5. HBM4E customer qualification progress
6. Long-term contract renewal terms (any softening)

Any single signal can flip first. When two or three flip together, that is when to start de-risking.

Part 11: Investment Implications

Pure AI Memory Beta: SK Hynix

The most direct way to play HBM dominance. Forward P/E around 5x. Operating margin of 72%. The downside is liquidity and access — Korean listing, harder for some accounts to access.

US Large-Cap AI Memory Exposure: Micron

The cleanest US-listed pure memory play. Forward P/E 7-9x, FY26 EPS consensus raised from $34 to $58. Smaller process and capacity edge than the Korean leaders, but the only on-shore option for many investors.

Defensive Memory: Samsung

Cheapest valuation among the three. Diversified business (consumer electronics, foundry) softens the cycle. HBM4 qualification with Nvidia is the catalyst. The drag: lower-margin businesses dilute the memory upside.

High-Beta NAND Play: SanDisk

The most volatile name in the memory complex and the only one with genuine optionality beyond the current pricing cycle. Up 1,180% in twelve months on the back of the NAND shortage and the spinoff narrative. The HBF (High Bandwidth Flash) program with SK Hynix is the asymmetric catalyst — if engineering samples ship on spec in H2 2026, the re-rating story isn’t priced in yet. Risks: NAND cycles flip faster than DRAM, $1.17 billion in fab commitments through 2029, and Kioxia M&A overhang.

“Sell the Shovels” — the Most Asymmetric Trade

Regardless of who wins between Nvidia, AMD, and Broadcom, and regardless of where exactly the memory cycle peaks, the equipment and packaging companies keep getting paid. ASML and TSMC are the closest thing to “free” exposure to AI infrastructure.

The Hidden Names

A few that matter but get less attention:

• Ibiden (Japan, TYO: 4062) and Unimicron (Taiwan, TPE: 3037) — the ABF substrate duopoly
• AT&S (Austria) — European substrate alternative
• Coherent (NYSE: COHR) and Lumentum (NASDAQ: LITE) — direct beneficiaries of 800G / 1.6T optics buildout
• Vertiv (NYSE: VRT) — rack-scale liquid cooling leader
• Kioxia (TYO: 285A) — the other half of the NAND alliance, often overlooked relative to SanDisk

What the Dynamics Mean for Your Existing AI Holdings

If you already own Nvidia, Broadcom, TSMC, or AMD, here is the framing question: AI capex is being captured at multiple layers, but the value distribution is shifting. Memory’s share of the GPU dollar has gone from 20% (A100 era) to 45% (B200 era), and could approach 55-60% with Rubin Ultra. That means the same AI capex creates increasing relative profit at the memory layer.

A balanced AI portfolio in 2026 probably spans four buckets: the GPU layer (Nvidia, AMD), the HBM/DRAM layer (SK Hynix, Samsung, Micron), the NAND / inference layer (SanDisk, Kioxia), and the picks-and-shovels (ASML, TSMC). Holding only Nvidia is a concentrated bet on Nvidia continuing to capture pricing power in a value chain where memory is increasingly the binding constraint.

Closing Thought

The remarkable thing about this cycle is not that memory companies are making money. It is that the same forces driving Nvidia’s dominance — AI infrastructure spend at unprecedented scale — are simultaneously rewriting the value distribution within the GPU itself. SK Hynix’s 72% operating margin and SanDisk’s 78% gross margin in early 2026 are, in some ways, more impressive than Nvidia’s. Memory was supposed to be commoditized. AI un-commoditized it — first DRAM, now NAND.

The next chapter is inference. If 2025-2026 was the year of training (and HBM), 2027-2028 is likely the year of large-scale inference deployment. The memory architecture for that wave is still being defined. HBM4E will play a role. HBF, if it ships on spec, could play a bigger one. Whoever wins that race captures the next leg of the supercycle.

Whether this lasts through 2027 or breaks earlier depends on how disciplined the suppliers stay and how fast the hyperscalers absorb the ramp. The supply side has structural reasons to behave well. The demand side depends on AI economics that are still being written in real time.

For now, the cycle has runway. The clock has started ticking — but the hands are not yet at the top.

This analysis is a synthesis of public sources including TrendForce, Counterpoint Research, IDC, SemiAnalysis, Mirae Asset Securities, KB Securities, Epoch AI, Bernstein, Morgan Stanley, and the Q1 2026 / FY26 earnings releases of SK Hynix, Samsung Electronics, Micron Technology, SanDisk, and TSMC. Nothing here is investment advice. Memory is a cyclical industry, and historically every cycle has eventually turned.