Posts: 1,600   +47
Staff
The big picture: As TSMC and other semiconductor manufacturers prepare to transition toward bleeding-edge 2nm transistors, more reports regarding the Taiwanese giant's plans have emerged. Although unconfirmed, claims of higher prices would align with prior analysis and the company's trends from recent years.

Semiconductor industry sources recently told the Taiwan-based Commercial Times that TSMC plans to charge over $30,000 for wafers built on its upcoming 2nm N2 process node. The move would continue the price increases that have escalated for a decade.

The reason behind the hike is unsurprising. Manufacturing 2nm semiconductors is expected to be more complex and precise than the 3nm wafers that currently represent the cutting edge. The rising costs account for software, manpower, and increased reliance on expensive EUV lithography. However, an AI-assisted design process is expected to soften the blow.

If the reports prove accurate, they would support last year's analysis from consulting firm IBS. The group estimated that building a 2nm semiconductor manufacturing plant that produces 50,000 wafers per month would cost around $28 billion. That's about $8 billion more than a 3nm fab. Furthermore, the estimated per-wafer cost presents a 50 percent increase over 3nm and a doubling over 4 or 5nm.

Although prices will vary among TSMC's clients, they have come a long way from 2014's roughly $3,000 28nm transistors. In response to the trend, some have noted that, in some respects, the cost per transistor has stagnated since 28nm, effectively freezing the cost aspect of Moore's Law and driving the prices of advanced computer chips upwards.

With nanosheet gate-all-around transistors and other innovations, N2 promises a 25-to-30 percent reduction in power consumption compared to 3nm, a 10-to-15 percent performance boost, and roughly 15 percent more transistor density. TSMC plans to begin mass-producing N2 in the second half of 2025. Apple has reportedly claimed all of TSMC's initial 2nm supply, which is expected to power the iPhone 17 Pro.

Samsung, Intel, and Japanese firm Rapidus have similar semiconductor manufacturing plans but are unlikely to exert market pressure on TSMC. Samsung is seeing yields below 20 percent, putting the company far behind TSMC.

Meanwhile, Intel recently canceled its 20A (2nm) node – originally planned for this year – and will introduce an 18A process next year instead. The company will also use transistors from competitor TSMC for its latest generation of CPUs. Rapidus, collaborating with IBM, aims to introduce its 2nm process next year but won't begin mass production before 2027.

Permalink to story:

 
The last two paragraphs are the real reason for why the price increases would be so steep: lack of competition.
 
It's about 70% denser than 4nm. But only 15% denser than the latest 3nm.
So yes, prices will skyrocket... the lack of competition is the green light for this.
 
The marketing misleading node naming is irritating, a bit. 3nm node process is 28nm at the tightest path. I wonder what the real numbers are for the 2nm node.
 
Sooo, 4nm wafer costs $15000, you can cut around 100 RTX4090s from It, let's asume 25% defective, so ending up in deferent cards, another $72 for memory according to dramexchange, but let's double It to $144. $100 to $120 for the board. Now multiply that by 2x to cover all the developement costs and We end up with $927 for the RTX4900, but probably half of that. A sucker is born every minute. I can't wait till Chinese will rip that market apart one day.
 
I can guarantee this arrogance is on the back of Samsung's abject failure with their own fabs. Strong competition is urgently needed and now it seems we need Intel to succeed. TSMC are a monopoly that are holding all companies to ransom. Very similar to Nvidia and Intel of old.
 
So is the $30,000 price based on fab plants outside the U.S.? Didn't TSMC once say expect a 30% increase in pricing from fab plants in the U.S.? But like the article reads prices will vary with clients.
 
Last edited:
Oh wow.................at the end of the day -I wonder who will be paying for this??........./s
 
TSMC is in almost EVERYTHING nowadays, and that likely will continue going forward. And the manufacturers will pay whatever TSMC asks.
 
It will probably hit a tipping point at some point. TSMC can charge anything they want, which eventually gets passed to consumers, and will become too expensive for most consumers to buy on a regular basis. Looking at the situation now, some people whom used to change phones every year have slowed the upgrade to 2 or more years.
 
Wafers sowed are still very cheap compared to how many profitable chips it does harvest ?

I wonder how much percent of profits Nvidia makes, it goes to the fabs who chiseled them ?
 
Wafers sowed are still very cheap compared to how many profitable chips it does harvest ?

I wonder how much percent of profits Nvidia makes, it goes to the fabs who chiseled them ?
Quite a few things that affect that.
* Yield - How many chips turn out OK vs how many have a defect
** How the design handles defects - A failed RTX 5070 might make for a great RTX 5060 TI
* Size of the chip¹ - Chips are rectangular and wafers are round². The bigger the chip the bigger the percentage that is lost due to waste around the edge
* Clock speeds - if aggressive clock speeds are chased less chips will make the cut. iirc this is due the laser used for the etching having the best focus in the center, the further away it gets from the center the more distortion becomes an issue.

¹ This is (a small) part of the reason why Ryzens chiplets were a big win for AMD. The CPU dies chiplet is tiny. Although the real win was that only the CPU die had to be manufactured on an expensive cutting edge node whilst they get the I/O die made on a more mature process with much lower costs.
² Wafers being round is due to how they're made although TSMC is supposedly currently figuring out how to use rectangular wafers but not expecting it to happen for another 5 to 10 years

In the end expect both TSMC and NVIDIA to charge multitudes higher than the actual costs. There's a reason why they're respectively the nineth and third highest valued companies in the world.
 
Last edited:
Does do anyone much good, if you can't afford it.
 
Quite a few things that affect that.
* Yield - How many chips turn out OK vs how many have a defect
It is not a small pie. It's about millions of wafers. In my opinion TSMC should have it's value combined to the value of other businesses it helps, not less. Every time it creates a new node, others sell more. At one point you could tell, why aren't they slowing down ?

By the way. This is the magnitude of how many chips it does create for AMD, for example.

"For all of 2023 that includes Xeon and Epyc/TR, on a net basis Intel produces 309,296,024 units and AMD not including dGPU 79,342,897 units for a 20.42% share and with dGPU 104,053,115 units for a 26.3% share."
 
I'm not going to pay that much more for new chips. I'll just keep using what I've got if that's what it's going to cost.
 

Similar threads