HBM and packaging
The chips can be financed. They still have to be stacked and packaged. High-bandwidth memory is booked through 2027, TSMC’s back end is sold through 2027, and a gas from Qatar is a second choke on the same line.
In June, Micron’s chief executive told the market what the order book already knew. HBM3E and HBM4 — the stacked memory that sits next to an AI GPU — are fully booked through calendar 2027, with demand stretching into 2028. Sumit Sadana said he does not see when supply will catch demand. The credit story is about who guarantees the rent. This one is about wafers, stacks, and gas.
High-bandwidth memory is not ordinary DRAM. It is DRAM stacked with through-silicon vias and bonded to a logic die. The bonding step, at TSMC, is CoWoS — Chip-on-Wafer-on-Substrate. A finished accelerator is a sandwich. If either slice is missing, the sandwich does not ship.
The extra stack shows up as cache, not FLOPs. SemiAnalysis’s late-August AgentX bench, under 384 concurrent agentic traces, had B300’s vLLM configuration hitting 91 percent of its HBM cache. B200, at a lower concurrency, hit 73 percent and leaned on DRAM offload. Working set roughly 43 million tokens against 22 million. B300 carries about 50 percent more HBM than B200. That is why the sold-out book is a capacity step, not a die-shrink story.
DRAM spot went from $3.76 a gigabyte in 2025 to $9.71 in 2026 — up 158 percent. Contract prices printed about 90–95 percent quarter-on-quarter in the first quarter of 2026 and another 58–63 percent in the second. The same pass puts the supply-demand gap at about 4.9 percent for DRAM, 4.2 percent for NAND, and 5.1 percent for HBM, the widest since 2011. Micron’s third-quarter print was the signature of a choke point, not a volume cycle: $41.5 billion of revenue, up 346 percent year on year, at an 84.9 percent gross margin, with a fourth-quarter guide of $50 billion and about 86 percent.
DRAM spot
Linear in dollars per gigabyte. Same cited pair as the HBM sold-out book.
Mark Murphy, Micron’s CFO, pulled the HBM market’s 2027 target forward a year — “easily crossing $100 billion,” he said — and listed the contracts that make the book look like a utility: sixteen take-or-pay agreements, about $100 billion of remaining performance obligations, about $22 billion of customer deposits. Take-or-pay means the buyer pays whether the bits ship or not. Gavin Baker, who is long the trade, put DRAM at 30 to 40 percent of next year’s hyperscaler capital budget. “Hundreds of billions of dollars… going straight to dram.” Amazon’s July raise put a buyer-side number on the same squeeze: cash capex moved from $200 billion to about $220 billion, and the company named memory-cost inflation as the reason.
Deposits, not a fab check
SK Hynix’s own second-quarter print, posted July 29, is the other sold-out book speaking first-party. Revenue KRW 79.3187 trillion. Operating profit KRW 60.5426 trillion, a 76 percent margin. Cash KRW 88 trillion. The company said it had finalized long-term agreements with around ten customers, typically about five years, with volume commitments and deposits. HBM4 mass shipments began in the quarter. The M15X mass-production schedule was pulled forward. It did not say take-or-pay. It did not disclose deposit amounts. It named Nvidia in the HBM partnership sentence and did not name Alphabet, Meta, or Microsoft.
That matters because the May rumor was a different mechanic. Trade press said those three had offered to bankroll the Cheongju expansion — fab equity, a dedicated line, EUV on someone else’s dime — in exchange for allocation. Reuters left the proposers unnamed. ChosunBiz named them and said management declined the equity form. Alphabet’s and Microsoft’s filings show large unnamed supply-agreement balances and do not mention SK Hynix or HBM. Deposits confirmed is not Big Tech owning the plant. The question stays a hypothesis.
The wafer-start number is still contested. A May pass put about 350,000 wafers a month and a 900,000 total on M15X. Later write-ups, including the piece that started the bankroll rumor, put that increment on Yongin. Official Q2 does not adjudicate. Micron remains the only of the two with a first-party $22 billion / $100 billion / take-or-pay disclosure.
The other half of the sandwich
Morgan Stanley’s allocation gives Nvidia about 60 percent of 2026 CoWoS — roughly 595,000 wafers — with Broadcom around 15 percent and AMD around 11 percent. TSMC’s own expansion target is 130,000 wafers a month by late 2026, from about 35,000 in late 2024, nearly four times in two years. The advanced-packaging lines are still booked through 2027. Analysts who watch the gap expect only slight easing. The leftover wafers do not all stay inside TSMC, and they do not all go to the same assembler.
our packaging capacity is so tight that now it's limiting my customers' growth… Today, we're working very hard to shorten the gap between the demand and the capacity.
C.C. Wei, TSMC, July 2026
That is the foundry saying the back end is capping the front end. Wei also said he welcomes a competitor — Intel’s EMIB-T — offering packaging flexibility so TSMC’s wafers can still ship. He was not declaring the gap closed. He was protecting the larger wafer business sitting behind the packaging gate. Front-end wafer is a majority of TSMC revenue; every package TSMC cannot build strands wafer it cannot bill. Ceding back-end share to Intel or ASE can be a good trade if it unlocks front-end volume — Wei said the two businesses are separable, not a gateway where packaging loss becomes foundry loss. Glass-substrate alternatives are about a year from production maturity. They are not a 2026 substitute. Whether Intel is actually winning EMIB-T volume at scale is still open; the monopolist validated demand, not Intel’s booking.
The overflow has a name, and it is not Amkor this year
In August, trade press said TSMC had started handing the CoW step — the front end of CoWoS — to outside assemblers. The first name in that sentence was ASE, not Amkor. ASE’s 2026 capital budget is $10.5 billion, up 98 percent, as the delegation began early in the month. Its SPIL unit broke ground on August 11 on a nearly TWD 100 billion Douliu plant that does not open until 2028. One street note, from Mizuho, not from TSMC, puts ASE at 20,000 wafers a month this year and Amkor at 20–25,000 by the end of 2027.
TSMC did name Amkor, on June 16: a ten-year deal to buy packaging and test from Amkor next to the Arizona fabs. Production is the end of 2027 or early 2028. Nvidia separately prepaid $1.5 billion for Amkor’s US capacity, cash arriving in 2027. That is a US supply-chain story. It is not this year’s Taiwan residual. Amkor’s own second-quarter print is computing and HDFO — $1.9 billion of sales, no TSMC CoWoS line item.
Meanwhile TSMC is still building its own back end. Ten packaging facilities. Street talk of 140,000 wafers a month moving toward 220,000. An August 11 board that put 10–20 percent of a $29.4 billion capex vote into packaging. Overflow and internalization run together. The sandwich still has a missing slice. The open question is who plates it in 2026, and how much of Arizona will be captive TSMC once both campuses are up.
The bonders are bought, not built
Internalizing the process is not the same as building the tools. Richard Blickman, who has run Besi since 1995, told the second-quarter call that TSMC’s new Chiayi plant — AP7 — is filling with his systems. The last plant, AP6, took about sixty bonders. The published plan for AP7 is roughly double that, and in the first quarter he said the number might run to a hundred, or higher.
Besi’s book kept compounding after the first-quarter surge. Second-quarter orders were €292.9 million, more than double a year earlier. Trailing twelve-month orders hit a record €987.6 million. Twenty-one customers now buy the hybrid-bonding tools, up from fifteen at the end of 2025. That is not the same as saying every new CoWoS line is a hybrid bonder. Blickman said CoWoS-like work is already a significant part of the orders — flip-chip, thermo-compression, and hybrid together. The second-quarter hybrid bookings were for CPO, logic, and probably wearables, not HBM volume. Hybrid bonding for high-bandwidth memory is still on the evaluation bench at one of the three memory makers, with another just starting.
The €400 million backlog that circulated in May was a Morgan Stanley calculation, not a Besi number. The company does not print a backlog or a unit count. ASE’s 2026 CoW overflow, the residual that is not staying inside TSMC, is talking to Korean dicing and bonding suppliers. Applied Materials owns nine percent of Besi and sells the front end of a shared die-to-wafer line. It is a partner, not a replacement. Blickman will not call a super-cycle. A typical up-cycle, he said, runs six to eight quarters, then overcapacity.
A cheap gas that stops the line
Helium is the inert gas fabs use to cool extreme-ultraviolet lithography tools, leak-test vacuum chambers, and carry deposition gases. It is a small line item until it is missing. Qatar supplies about a quarter of the world’s helium. After the Ras Laffan halt, spot doubled from about $500 per thousand cubic feet to $1,000–$1,200. A mid-September SK Hynix recheck still did not describe Ras Laffan as restarted. QatarEnergy repair talk remains three to five years, with about 14 percent lower annual output in secondaries. Sanjiv Lamba, Linde’s chief executive, had already put normalization into early 2027. The official Hynix print began HBM4 mass shipments in the second quarter and did not disclose a helium-forced production halt. No retrieved Commerce report imposes a Korean-HBM tariff. The January 25 percent plus a data-center exception is the live text.
SK Hynix sourced 64 percent of its 2025 helium from Qatar and carried a buffer dated to a mid-2026 cliff. Anas Alhajji, an oil economist, not a chip analyst, put the concentration higher: 75 percent of the helium going to South Korea and Taiwan comes from Qatar. “If there is no helium, you cannot produce the semiconductors.” The figure is his, and it has not been independently checked. The direction matches C.C. Wei’s earlier warning that the Middle East situation may hit chemical and gas pricing.
Samsung has historically been the HBM supplier Nvidia would not qualify. That test has already run. Samsung cleared Nvidia’s 12-layer HBM3E tests, began shipping, and, on the June 2026 certification, all three HBM4 vendors — SK Hynix, Samsung, Micron — passed for Vera Rubin. Counterpoint’s third-quarter 2025 HBM revenue split was SK Hynix 57 percent, Samsung 22 percent, Micron 21 percent. The sold-out book did not open. The exclusive SK Hynix read did.
Neil Movva, who is building custom silicon at Sail, wants to walk around the shortage rather than wait it out. The ideas, he said on Invest Like the Best, are “about sidestepping the HBM shortage and focusing on more extreme offload to other forms of memory such as Flash,” at a serving target of one to ten tokens per second. That is an attempted architectural workaround. It is not evidence the constraint has eased. The sold-out book is still the sold-out book.
OpenAI’s Jalapeño part is an early HBM4 adopter, not a second book. SemiAnalysis says the stack runs 15.4 terabytes a second and the memory is likely Samsung’s. Production ramps in 2027. The wiki attaches that to the sandwich. It does not re-rate Micron or TSMC on it.
Nvidia’s own 10-Q, as of July 26, raised supply and capacity commitments from $119 billion to $279 billion “to meet future demand.” The newsroom named a multiyear SK hynix memory partnership in the same print. That is customer-side attach for the sandwich. It is not a Micron or TSMC re-rate. A VLSI program title on a sixth-generation 11-nanometer DRAM cell for HBM4 is a title, not a result.
A mid-September SemiAnalysis note argued the sandwich is getting shorter, not taller. Rubin Ultra carries 192 gigabytes against 288 on Rubin and B300. The industry, in that telling, is moving from twelve-high stacks toward eight-high as the standard, and four-high as the best dollars-per-bandwidth — and wafers per token — for inference. HBM wafers are scarce; a shorter stack yields more cubes. A later note on Engrams made the same point from the model side: table offload to host DRAM, and even SSD, lets an iso-quality model use less HBM capacity. Bandwidth still matters more than capacity. Idle KV should leave the accelerator for network-attached DRAM. Modeled Kimi K3 frontiers stay below about 80 gigabytes per GPU. Less capacity per quality is not less HBM demand. Mix color on the same sold-out book. It does not re-rate Micron or TSMC, and it is not a second chain. Micron’s fiscal fourth-quarter call is dated Wednesday, September 30, 2:30 p.m. Mountain time. That is a calendar day, not a print.
Linde’s contracted book reprices quietly
The same Qatar shock that hits fabs hits the industrial-gas suppliers from the other side of the pipe. Ras Laffan’s force majeure after March 2026 strikes removed roughly thirty-five percent of global helium for planning purposes; QatarEnergy’s CEO cited a three-to-five-year full-repair horizon. Spot doubled from about five hundred dollars per thousand cubic feet toward eleven hundred to twelve hundred. Contract prices — what Linde actually bills — rose about forty percent.
Linde carries eighty-five to ninety percent of its business on long-term contracts, which insulates day-to-day spot swings while creating renewal upside as each deal rolls. CFO Matthew White said pricing will continue rising through 2026 as contracts renew; FY2026 EPS guidance of seventeen dollars sixty to seventeen ninety explicitly excludes helium upside — management’s word for sandbagged optionality. Samsung and SK Hynix signed long-term helium agreements with Linde and Air Products in April amid the squeeze — fab customers locking supply under scarcity is the counterparty-leverage thesis printing. Linde’s Beaumont strategic storage cavern holds more than eighty-five million cubic meters, roughly six months of global demand. JP Morgan upgraded the stock citing repricing power in inflationary supply shocks. Distinct from Micron’s HBM sold-out story: this is the gas vendor leg, not the memory maker leg, of the same Qatar shock.