Key points
- Reported increases span foundry wafers, components, and AI systems.
- Nvidia's reported increase targets deliveries in 2027.
- Nvidia reportedly cited memory costs, not wafer prices.
- TSMC guided third-quarter gross margin lower anyway.
Nvidia (NVDA) told some of its largest customers to expect price increases above 15% on servers built around its AI chips, Bloomberg reported on August 22, citing people familiar with the matter. The increases would land on systems shipping early in 2027. Nvidia sits at the end of a chain of reported increases that runs back through the memory makers and the foundries.
The notices covered both the Vera Rubin and Grace Blackwell generations, and the increases would vary by chip generation and memory configuration. They went to the server makers assembling systems under contract for Microsoft, Google, Oracle, Amazon, and Meta. Nvidia has not commented publicly on the report, and some coverage has since described the company as having announced the increase itself, which Nvidia has not.
The reason attributed to Nvidia was memory costs rather than foundry wafer pricing. That is consistent with what the memory market has been doing all year. Contract prices set records in August, and spot HBM has traded at roughly five times contract price.
SK Hynix (SKHY) is the largest HBM supplier. On its second-quarter earnings call, Corporate Center president Song Hyun-jong said the company had closed long-term supply agreements with about 10 customers, generally running about five years and carrying purchase commitments and deposits, THE ELEC (디일렉) reported. The company said the price structures vary by customer but are designed to damp sharp swings, so spot moves do not reach SK Hynix the way they reach a spot seller.
Memory is also a large share of what a finished AI rack costs, which is why a move in it is visible at the system level at all: SemiAnalysis put memory at about 40% of rack cost on Rubin Ultra. That 40% is how much of a rack's cost memory accounts for, and it says nothing about how far memory prices themselves rose.
The foundry increases are real but smaller
TSMC (TSM) told customers in June to prepare for increases across its whole advanced portfolio, not just the newest nodes, according to a June 23 Culpium report. The increases reportedly run 5% to 10% and vary by customer, node, and product, and they extend past 3nm and 2nm to 5nm and 7nm. TSMC declined to discuss pricing but provided this statement: "TSMC does not comment on pricing. Our pricing strategy is strategic, not opportunistic. We will continue to work closely with customers and sell our value to them."
That reaches most of what TSMC sells. In its second quarter, reported to the SEC on July 16, 3nm was 30% of wafer revenue, 5nm was 33%, 7nm was 11% and 2nm was 3%. Technologies at 7nm and below, which is how TSMC defines advanced, came to 77% of wafer revenue. Revenue was $40.20 billion and gross margin was 67.7%. The same filing guided third-quarter gross margin to 65% to 67%, below the 67.7% TSMC had just posted. The guidance shows that reported price increases do not necessarily translate into higher overall margins.
Samsung raised its own foundry prices from July, on new orders at 4nm, 5nm, and 8nm, Reuters reported on August 19, citing two people familiar with the matter. Quotes on its 4nm SF4 process rose 10% to 15% for customers in China and the United States against June, while Taiwanese customers saw 5% to 10%, and 8nm went up close to 10%. Its 4nm line at Pyeongtaek has run at full capacity since late last year. Samsung's foundry division has lost money since 2022. Lee Min-hee, an analyst at BNK Investment & Securities, told Reuters that if Samsung keeps raising prices, "its foundry business could potentially become profitable as early as next year."
Where the AMD report fits, and how far to trust it
Advanced Micro Devices (AMD) is the newest name in the chain, and the thinnest sourced. Wccftech reported on Thursday, September 17 that AMD has notified partners of a roughly 10% increase tied to higher TSMC wafer costs, planned for the fourth quarter and covering AI accelerators, consumer graphics processors, and motherboard chipsets. The source is a post by @harukaze5719, a leaker on X, relaying unnamed supply-chain contacts. AMD has not commented. One detail has already been lost in the retelling: CPUs are not in the list, and several outlets have run headlines saying Ryzen processors are getting more expensive anyway. The report does not say that.
All three stocks were higher on Thursday, September 17, and we could not establish whether these reports contributed to the gains. AMD traded at $550.09 as of 12:28 p.m. Eastern, up 7.3% from its September 16 close of $512.50. Nvidia was at $218.85, up 2.3%, and TSMC at $427.16, up 2.3%.
Who ultimately absorbs the increases is not something the notices establish. The hyperscalers could face higher system costs, which is the layer Nvidia's notices reach, though they went to contract manufacturers rather than to the hyperscalers directly. Fabless designers could absorb higher wafer costs or pass them on, and passing them on is what the AMD report claims. Consumers could also face higher prices if those costs are passed through to finished products.
