Kai ChenIn April 2026, TSMC said CoWoS advanced-packaging capacity was growing at an 80% compound annual rate, while CNBC reported Nvidia had reserved most of its 2026 capacity. Memory-industry executives were also saying 2026 production slots were almost sold out.
Key takeaways
- AI-memory scarcity now centers on qualified system capacity: HBM, accelerator logic and CoWoS packaging must all be available on the same product calendar.
- TSMC’s CoWoS capacity can grow at an 80% compound annual rate while remaining constrained because Nvidia reportedly reserved most of its 2026 capacity.
- Samsung’s commercial HBM4 shipments and early HBM4E samples do not establish an Nvidia award; the cited reports name neither the shipment customer nor a qualified volume program.
- Pricing power should favor suppliers with qualified memory, packaging access and interface components tied to named accelerators—not manufacturers that merely add bit capacity.
- SK Hynix’s 192GB LPDDR5X-based SOCAMM2 production for Vera Rubin shows that a named platform role can be more commercially meaningful than an unnamed HBM sample.
When Nvidia announced the H200 in November 2023, it specified 6.5Gbps-per-pin HBM3E in place of the H100’s 5.3Gbps-per-pin HBM3 and scheduled shipments for the second quarter of 2024. The memory upgrade had to reach qualification and packaging on the accelerator’s calendar.
That coupling makes the scarce unit a qualified stack tied to a named accelerator and packaging window. Buyers reserve logic, memory and CoWoS against the launch rather than treating each component as an independent purchase.
Samsung’s shipment claims stop short of Nvidia approval
In February 2024, Samsung announced a 36GB, 12-layer HBM3E stack slated for mass production in the first half of that year. Bloomberg reported in July 2024 that Samsung anticipated Nvidia approval by November, citing sources; that report described an expected decision, not an approval.
On January 26, 2026, Bloomberg reported that Samsung had entered the final qualification phase for Nvidia’s HBM4 supply and targeted mass production in February. On February 12, Samsung said it had begun its first commercial HBM4 shipments to customers, while Bloomberg said the company was still seeking to supply Nvidia. Those reports do not identify Nvidia as the shipment customer.
On May 29, Bloomberg reported that Samsung had sent 12-layer HBM4E samples to major clients before SK Hynix’s stated second-half 2026 sampling target. Samsung sampled first, but the report did not name the clients or establish qualification and volume delivery.
Nvidia books capacity by accelerator generation
On March 5, 2026, the Financial Times reported, citing sources, that Nvidia shifted manufacturing capacity at TSMC away from H200 chips intended for China and toward Vera Rubin products. Five weeks later, CNBC reported that Nvidia had reserved most of TSMC’s 2026 CoWoS capacity. Nvidia was assigning capacity by accelerator generation and changing that assignment as product priorities moved.
CoWoS integrates logic and HBM in one package, so Nvidia’s reservation protects more than a packaging service. A completed memory stack cannot fill a launch if its logic wafer or packaging window has moved to another accelerator.
The reports do not disclose reservation terms, cancellation rights or how much capacity could return if demand weakens. A booking also does not prove that every reserved slot will become a shipped system.
Broadcom gives Samsung a long-horizon foundry anchor
On June 11, The Information reported, citing sources, that Google was discussing having Samsung manufacture the memory input-output die for Icefish, its tenth-generation TPU, while TSMC would make the computing engine. The talks could place Samsung at the interface between memory and accelerator architecture, but they do not establish a contract, qualification or volume production.
Samsung’s confirmed Broadcom manufacturing contract runs through 2030 and covers 2nm and more advanced processes. The agreement documents a long-horizon foundry commitment. It does not identify an HBM supplier, establish CoWoS access or show that Samsung has secured an HBM design-in.
SK Hynix entered the boom with HBM3E in production
SK Hynix’s rise past Samsung as South Korea’s most valuable company followed 14 years of investment in HBM. On September 26, 2024, SK Hynix said it had begun mass-producing a 36GB, 12-layer HBM3E stack for delivery by year-end.
That announcement documented production volume but did not identify an accelerator or customer. The 14-year investment record and the 2024 production date give SK Hynix operating history that a new capacity announcement cannot reproduce. They still do not settle HBM4 awards: Bloomberg’s January 2026 report described Samsung undergoing a separate Nvidia qualification for that generation.
Vera Rubin gives LPDDR5X a named role
By July 2026, SK Hynix said it had begun mass production of a 192GB SOCAMM2 module based on LPDDR5X for Nvidia’s Vera Rubin. Unlike an unnamed HBM4E sample, the announcement identifies the supplier, memory architecture, accelerator and production stage.
The announcement does not say SOCAMM2 replaces Vera Rubin’s HBM. It establishes that Nvidia assigned LPDDR5X a specific role in the same platform, leaving chip designers room to use different memory architectures for different workloads and power budgets.
A 19-fold profit jump leaves the customer unresolved
In July 2026, Samsung forecast second-quarter operating profit up 19-fold year over year and revenue up 129%. Investors nevertheless sent Samsung shares down more than 6% after the forecast. Those figures do not identify an HBM customer or quantify HBM’s contribution to the quarter.
Samsung’s next proof point is narrower than an earnings forecast: a named customer, qualified stack, accelerator generation and volume date. Its February HBM4 shipments establish commercial activity, but the public reports cited here do not establish repeatable Nvidia volume awards.
TSMC can grow CoWoS 80% a year and still have Nvidia reserve most of 2026. Samsung’s 19-fold profit forecast does not identify the customer behind its HBM shipments. TSMC can create room with more CoWoS; Samsung still needs a named program to claim it.
Samsung’s scale signals versus named commitments
| Evidence date | Signal | Specific figure | Identified scope |
|---|---|---|---|
| 2026-07-07 | Q2 operating-profit forecast | About $58.44B; up 19-fold year over year | Samsung quarterly operating profit |
| 2026-07-07 | Q2 revenue forecast | About $111.7B; up 129% | Samsung quarterly revenue |
| 2026-07-07 | Market reaction | Shares fell more than 6% | Samsung stock after the forecast |
| 2026-07-25 | Confirmed manufacturing agreement | Over $200B through 2030 | Broadcom chips using 2nm and more advanced processes |
Frequently asked questions
Why is AI memory becoming an allocation market rather than a commodity market?
An AI accelerator requires qualified memory, logic wafers and advanced packaging in the same launch window. Extra memory bits have limited value if the corresponding accelerator or CoWoS capacity is unavailable.
Has Samsung been approved to supply Nvidia with HBM4?
Not according to the public reports cited. Samsung began commercial HBM4 shipments in February 2026, but the customer was not named, and Bloomberg said Samsung was still seeking to supply Nvidia.
How can CoWoS remain scarce if TSMC is expanding capacity by 80% a year?
Demand and advance reservations can absorb new capacity before it becomes broadly available. CNBC reported that Nvidia had reserved most of TSMC’s 2026 CoWoS capacity.
What would constitute convincing proof of an HBM win?
A named customer, qualified memory stack, specified accelerator generation and volume-production or delivery date. Samples, shipment claims without a customer and earnings growth do not establish a repeatable accelerator program.
Does Vera Rubin’s use of LPDDR5X mean it no longer needs HBM?
No. SK Hynix’s announcement establishes a specific LPDDR5X-based SOCAMM2 role in Vera Rubin, but it does not say that SOCAMM2 replaces the platform’s HBM.