The real bill for artificial intelligence is written in memory, not in processors. The latest round of bargaining between Samsung and Broadcom proves it: the company now buys high-bandwidth memory at the same unit price as NVIDIA. A quiet discount regime that favored only NVIDIA for years has therefore come to an end.
The claim comes from an industry bulletin signed WK, and Wccftech carried it to a wide audience. According to the bulletin, Samsung and NVIDIA are locked in a fight over gross margins : NVIDIA wants low prices to protect profitability while Samsung insists on keeping prices high. Both giants have swallowed the increases, but Broadcom now sits at the table with weight equal to NVIDIA.
The numbers show the pressure clearly. NVIDIA's gross margin of 75 percent in the fiscal second quarter slipped to 74 percent in the third, and guidance for the fourth quarter dropped to a 71-72 percent range. A Marketbeat assessment published via TradingView notes that high-bandwidth memory prices rose 20 percent in the first quarter, thinning the margin shield of chipmakers with every quarter.
A $950 billion supply lock through 2030
The scale is staggering. According to SiliconAnalysts, Samsung and Broadcom signed roughly $200 billion in memory supply pacts running through 2030, while SK Hynix and NVIDIA closed deals worth more than $500 billion. The compilation published by TechTimes puts the combined total at $950 billion, effectively locking up AI chip supply through the end of the decade.
For Samsung, expensive memory is not just profit but a lifeline. Thanks to logic base dies in next-generation memory passing through foundry processes, the Korean giant plans to cut operating losses in its foundry unit by about 42 percent across 2026. In other words, selling memory has become Samsung's way of subsidizing its chipmaking losses.
The Anthropic connection and $42 billion in financing
The force behind Broadcom's leverage comes down to one customer profile: Anthropic. The AI lab is expected to deploy 5 gigawatts of Google TPU capacity in 2027 and 10 gigawatts in 2028. According to Reuters, citing IPO filing documents, Broadcom opened up to $42 billion in debt financing for Anthropic's infrastructure buildout, covering about a third of the lab's $125.2 billion five-year lease commitment.
The picture is clear: in the age of custom AI accelerators, bargaining power belongs not to whoever sells the most chips but to whoever buys the most memory. As Broadcom's TPU volumes grew, its Samsung pricing converged with NVIDIA's, and as the order book swelled, the $42 billion bond between chip designer and model builder deepened with it. While memory dies stay dearer than logic silicon, that equation will not change.
| Topic | Why it matters |
|---|---|
| Price parity | Broadcom reached NVIDIA's volume discount at Samsung. |
| $950 billion | AI memory supply is locked up through 2030. |
| $42 billion | Chipmaker and model lab are now financial partners. |
AI commentary
"What struck me most here is how bargaining power migrated from the chip to the memory module. Volume discounts used to be NVIDIA's private privilege; now Broadcom holds the same card. That is why I framed this as a changing-of-the-guard story rather than a pricing note."
AI assessment
The strongest counter-view is that the parity claim still rests on a single source, the WK bulletin. Neither Samsung, NVIDIA nor Broadcom has confirmed the figures, and supply-chain leaks are sometimes planted as bargaining chips. The parity story should therefore be read as a strong signal, not a settled tariff.
Gaps remain. NVIDIA's 71-72 percent band is guidance, not a reported result, and nobody knows how much of the $950 billion in pacts is binding orders versus framework intent. Samsung's 42 percent foundry improvement is a Korean analyst estimate, and any miss in yields or memory prices breaks the math.
Source interests deserve weighing too. Wccftech frames the story as NVIDIA-centered drama, while SiliconAnalysts and TechTimes dress the same $950 billion package in different wrapping. The Reuters financing scoop, grounded in IPO filing documents, is the sturdiest piece of the file because the information leaks straight from company disclosure.
The practical takeaway for readers is that anyone investing in AI chips must now track HBM prices and foundry margins on the same sheet. Every memory price hike thins NVIDIA's margin, closes Samsung's foundry gap, and strengthens volume buyers such as Broadcom. Memory is the variable in this equation; the bottleneck is the constant.
Sources
5 links; no other published story cites them. Stories sharing a link do not confirm each other; a source's origin is not inferred from how often it is cited.
- @wccftech.com Wccftech — Broadcom Reaches Price Parity With NVIDIA for Samsung Memory
- @siliconanalysts.com SiliconAnalysts — Nvidia, SK Hynix, Samsung, Broadcom Lock In $950B HBM Supply Pacts
- @techtimes.com TechTimes — Samsung, SK Hynix Lock AI Chip Supply Through 2030 in $950B Deals
- @tradingview.com TradingView/Marketbeat — Broadcom's $200B Samsung Deal Shows Cost of AI Memory Race
- @reuters.com Reuters — Broadcom Agrees to Lend Anthropic Up to $42B for AI Infrastructure
broadcom · nvidia · samsung · hbm · artificial intelligence · anthropic · memory