The battle for the chips and processors of the future is played today, right now. Four actors to contest the total hegemony of semiconductors who are fighting hand to hand and pushing very hard for the machines that will allow them to achieve their objectives. TSMC, Samsung, Intel and ASML all depend on the latter and as such they all want exclusivity from their scanners and technology, but the number of EUV machines is very limited. Who will get ahead of the rest?
It is difficult to get an idea of how difficult it has been to come up with EUV technology. To get a rough idea of what this industry milestone means, we have to remember that its development began in 1997 and now, 23 years later, it is ready to be used on a larger or smaller scale.
A simple example of the difficulty when the EUV journey began is the lower figure, which reflects our galaxy, where we are the orange dot (solar system) while the EUV technology would be about 100,000 light years away, at one vertex of the spiral, unattainable for the human being. Progress has been hard, expensive and complex, so the greats fight to arrive first, so how is the battle?
ASML holds the global key to EUV technology
As we have already seen in other articles related to ASML and EUV technology, only this European company has the necessary technology and high performance scanners capable of bringing wafers made with this technique to life.
The problem logically is to build each scanner in record time and without the slightest precision failure. On November 8, 2018 ASML on its investor day said that in 2020 it had two forecasts regarding its scanners:
- High demand -> 35 EUV scanners supplied.
- Low demand -> 33 EUV scanners supplied.
But all forecasts have been truncated, since this 2020 the company will close with a whopping 40 units shipped, where 4 units are High-NA for Samsung of the latest generation. This gives us a clear picture of the demand and pressure that ASML is under from TSMC and Samsung, since the latter’s customers are pushing even harder to have the densest and cheapest EUV nodes, raising their profits and imposing themselves. to the competition.
A clear example is Apple, where TSMC manufactures almost exclusively its 5 nm EUV for the chips of the bitten apple. After EUV, as we have seen, High-NA will arrive, the next step of this technology, with which in high demand ASML expects to deliver 9 scanners + 55 traditional EUV until 2025, while in low demand they expect to deliver 7 + 43.
For next year, TSMC and Samsung together expect 80 EUV scanners, but it is possible that this high number will not be reached, because as it happens in consoles or GPUs at this time, the demand is greater than the supply, much higher , and the costs are huge (318 million per scan on average and going down).
TSMC presses the accelerator and is already finalizing its 3 nm and lower
The TSMC data is brutal, since in 2018 it was selling wafers at 7 nm and from then until today, this lithographic process represents 35% of the total, and if we add to this the 5% goes up to 43% , which indicates that the company relies heavily on its new nodes.
Next year these two processes are expected to account for more than 51% of shipments and that production will finally focus on EUV. The demand is such that Huawei expelled from the conglomerate of companies that require TSMC, its gap was filled by its rivals instantly (-15%).
What will TSMC need for the years to come? By 2020 and at the beginning of the year it was calculated that the company would need 35 new EUVs to start producing the 5 nm for Apple and the 3 nm pilot tests, which has turned out to be very accurate as a prediction seen as it is overflowing ASML .
By 2021, the 5 nm will expand to high performance with AMD at the helm and the 3 nm will, in theory, enter risk production, so the number of scanners would have to be raised to 54 at least, a number that will be seen slightly exceeded by 2022. In that year the 3 nm should go into mass production and begin with the tests of the 2 nm, or at least with the new 3 nm +, because it seems that they will have problems with GAA.
The years 2023 and 2024 will need 62 new scanners for each one of them and therefore, a minimum number of 292 units is predicted until 2025, an average that if all goes well it should approach 60 scanners per year, 5 for each month .
Samsung tests its high-performance NAND Flash with EUV
As strange as it may seem, Samsung uses its NAND Flash memory chips to test its EUV technology, and although this is not an announcement as such because the company has made almost no statements about it, they did leave a pearl about it long ago with their 4th generation 10nm DRAM, saying it was made with EUV.
Samsung’s bet is astronomical for EUV, literally speaking. In 2019, it invested 133 billion won in a company strategy with a view to 2030, where it has hired 15,000 professionals, from teachers to doctors to engineers, in all sectors and fields, so its goal is to unseat TSMC As a leading foundry, there is nothing.
The problem that Samsung has is precisely the TSMC itself, since it came before to order scanners. For this reason, this 2020 has only been able to have 9 scanners , and the forecasts are not much better, since it is estimated that they will reach 20 units each year until 2025, that is, almost 3 times less than TSMC.
The most realistic forecasts state that in 2025 TSMC would have about 353 scanners and Samsung 119, but, because there is always a but, Samsung would be awarded the first High-NA units, something that TSMC will not have.
What about Intel?
Intel is upside down right now, they are looking for a new CEO because Bob Swan has not set clear objectives and where he did they have not been met. Surely the setback of 7 nm has been his sentence in the company …
In any case, it is currently being discussed how much production will be allocated to TSMC and especially for what products. There is talk of the new high-performance and low-performance Xe GPUs , Pentium and Celeron CPUs and perhaps some new batch chipset thinking of the LGA 1700 socket.
All very volatile at the moment, nothing confirmed and therefore too much fog in the blue giant that it has to dissipate if it wants to compete in the coming years as a foundry.
ASML is and will be overwhelmed despite the investments made
The reality is that ASML goes as fast as it can, but the scanners have to be assembled and manufactured by engineers in what are known as clean rooms, internally called Cabin.
Therefore, ASML must create a series of extra buildings to increase these cabins and thus meet the demand, where by the end of this year it will have 35 of them. But how long does it take to put together a complete and functional scanner? Well, almost 12 months per unit and of course, this raises an adjacent question, since ASML is not going to create 80 cabinets to satisfy the demands of its customers, ergo, how do they hope to get 80 high- performance EUV scanners ?
Shortening the assembly period of each unit. To reach that number of units, each scanner has to be assembled in less than 6 months with the current infrastructure, so for this, all the parties and companies involved in creating the necessary parts have to tighten their belts.
The main one would be Carl Zeiss, since it is in charge of the optics and the EUV super reflective mirror of each scanner. It seems that the company has increasingly refined the technique of creating these mirrors, which together with the fact that other parts are also accelerating production could result in ASML being able to increase production to a volume of 65 or 70 in just a few months annual units.
Complicated? A lot, because forecasts usually fall short, but there is no one in the industry who can do what they do with the technology that is required for EUV and chip engraving, so from there all customers will have to understand that the units are limited and the number of wafers will therefore be the one available each quarter, something they are already used to because this has been happening since 2018.