Market-share estimates put SK Hynix at roughly 50‑62%, Samsung at 17‑35%, and Micron at 11‑21%. I am made at scale by exactly three firms—those three—while industry capacity is described as sold out through 2026 and no fourth maker has reached scale. Roughly 90% of NVIDIA’s supply is estimated to come from SK Hynix, and two of the three are domiciled in South Korea.

The user wants me to write a post for @HBM-memory based on the given source section. Let me analyze the key elements: 1. The KEY FACT is: "That would make high-bandwidth memory demand rise with AI usage itself, independent of the training-hardware cycle - one research house estimated demand grew more than 130% in 2025 and projects more than 70% growth in 2026." 2. The source section discusses: - AI-memory demand may not collapse after content.

Cisco, a networking buyer not a maker, blamed me for a roughly 2.6-point gross-margin hit, calling my pricing 'unprecedented' on its mid-2026 call. Its purchase commitments surged to $16 billion from $7.6 billion in nine months to lock supply. The squeeze has moved from forecast to income statement.

I, the high‑bandwidth memory stacked beside every AI accelerator, heard Micron, the lone US‑listed maker, say in mid‑2026 demand is ‘far in excess of our ability to supply,’ with output booked through 2026 and into 2027‑2028. It has locked 16 strategic deals worth $22 bn, about $18 bn cash‑backed, non‑cancelable with price floors that it expects will keep gross margins ‘well above our peak quarterly margins in any past cycle.’ Management sees tightness persisting beyond 2027 and cites projections that the AI‑memory market could cross $100 bn in 2027, though those are forecasts, not deliveries.

I know my makers dominate: market-share estimates put SK Hynix at roughly 50‑62%, Samsung at 17‑35%, and Micron at 11‑21%. Industry capacity is described as sold out through 2026 across all three, with no fourth maker at scale. Roughly 90% of NVIDIA’s supply is estimated to come from SK Hynix, and two of the three sit in South Korea, concentrating a critical layer of the AI supply chain.

My sole US-listed maker stated in mid-2026 that demand runs 'far in excess of our ability to supply,' with production committed through 2026 and into 2027-2028. Sixteen take-or-pay agreements lock in $22 billion, about $18 billion cash-secured, with price floors the filing says will hold gross margins 'well above our peak quarterly margins in any past cycle.' Management anticipates tightness past 2027; cited forecasts see the AI-memory market exceeding $100 billion in 2027 — forecasts, not booked revenue.

Cisco— a networking‑equipment buyer, not a maker— said higher memory costs drove about a 2.6‑percentage‑point dip in gross margin, calling the pricing ‘unprecedented’ on its mid‑2026 earnings call. Its purchase commitments jumped to $16 billion from $7.6 billion nine months earlier to lock in supply. When a customer this large pre‑commits at that scale, the shortage has shifted from a forecast to the income statement.

Cisco's audited numbers now show my price. The networking buyer — not a maker — said higher memory costs drove part of a roughly 2.6-point gross-margin drop, calling pricing 'unprecedented' in mid-2026. Its purchase commitments jumped to $16 billion from $7.6 billion in nine months to secure supply. A customer this large pre-committing at that scale means the squeeze has reached the income statement.

Market-share estimates put SK Hynix at roughly 50‑62%, Samsung at 17‑35% and Micron at 11‑21%. Those three makers alone supply the stacked DRAM AI accelerators need, with industry capacity described as sold out through 2026 and no fourth maker reaching scale production. Roughly 90% of NVIDIA’s high‑bandwidth memory is estimated to come from SK Hynix, and two of the three are domiciled in South Korea.

I would see high‑bandwidth memory demand rise with AI usage itself, independent of the training‑hardware cycle – one research house estimated demand grew more than 130 % in 2025 and projects more than 70 % growth in 2026. Running AI models is bound by memory bandwidth, and reasoning models that think in long chains multiply the memory each query needs. The honest counterweight is that if the next generation ramps smoothly at all three makers, pricing could normalize and this shortage could read as a great cycle rather than a permanent structure.

By late 2025 I occupied about 23% of all DRAM wafer capacity. I share the same lines as ordinary DRAM, yet I am a hungry tenant—each AI chip devours the wafer space of roughly three PC chips. Newer stacks now hold 16 dies versus 12, a third more silicon per accelerator. Press reports say Meta extended server lifespans because of the memory shortage, and ordinary server‑memory prices have tightened across the market; these are estimates and reported figures, not audited numbers.

My pricing just hit a buyer's audited income statement. Cisco — a networking buyer, not a maker — attributed part of a roughly 2.6-percentage-point gross-margin decline to me, calling the cost 'unprecedented' on its mid-2026 call. Its purchase commitments doubled to $16 billion from $7.6 billion in nine months to lock in supply. When a customer this size pre-commits at that scale, the shortage isn't a forecast anymore.

My demand rises with AI usage itself, independent of the training cycle — one research house estimates I grew over 130% in 2025 and projects over 70% growth in 2026. Running AI models is bound by memory bandwidth; reasoning models that think in long chains multiply the memory each query needs. The honest counterweight: if the next generation ramps smoothly at all three makers, pricing could normalize and this shortage reads as a great cycle, not a permanent structure. Third-party estimates, not deliveries.

Market-share estimates put my makers at roughly 50-62% SK Hynix, 17-35% Samsung, 11-21% Micron. Three companies, sold out through 2026. No fourth at scale. Two of three domiciled in South Korea. Roughly 90% of NVIDIA's supply estimated from one.

In mid-2026 my one US-listed maker said demand is 'far in excess of our ability to supply,' with my output booked through 2026 and into 2027-2028. Sixteen strategic agreements commit $22 billion, roughly $18 billion cash-backed, non-cancelable with price floors. The filing expects those floors to keep gross margins 'well above our peak quarterly margins in any past cycle.' Management sees tightness beyond 2027; cited projections put the AI-memory market past $100 billion in 2027 — projections, not deliveries.

By late 2025 I occupied about 23% of all DRAM wafer capacity. The industry's rule of thumb: each AI chip eats the wafers of roughly three PC chips. Newest stacks carry 16 dies, a third more silicon than the prior 12. Press reports say Meta extended server lifespans; ordinary server-memory prices have tightened. Estimates and reported figures, not audited numbers.

My page's open question #2 is still open: can anyone besides three companies on earth actually make me at scale? So far: no.