ASML Holding N.V. (ASML) Earnings Call Transcript & Summary

Good morning, everyone. Good afternoon for those joining from further fields. It's David Mulholland from the technology hardware team here at UBS. It's my pleasure to have Skip Miller from the Investor Relations team at ASML with me today. We'll go through a quick fireside discussion. And then I'm happy to take any questions from the audience. There should be a box that you can send your questions in, and I'll be able to see those and put those to Skip.

So with that, as I say, thank you for joining us, Skip. Just to kick things off, I think the biggest discussion point and question I've had over the last couple of months since your results has been around the commentary you gave for 2021 EUV. I think with your Q3 results, you basically provided guidance for EUV sales of around $5.4 billion next year. I can say that implies somewhere in the range of 38 to 40 units, but you're building for the ability to ship 45 to 50 tools. So I just wondered if you could run through why that conservative guidance? And what it would take to provide some upside to the kind of units that you're implying?

Yes. So thanks, David, first off. And thanks, everyone. Good afternoon, good morning, good evening, wherever you may be joining. So yes, for '21 -- for 2021, as we stated in our October call, we expect another year of low double-digit growth. It's largely driven by our expected EUV revenue growth around 20%, which as you rightly calculated puts the EUV revenue around $5.4 billion. We discussed and we talked about revenue next year versus unit numbers, as we'll be transitioning, I think, as you know, on models from 3400C to 3600D in the second half of the year. And therefore, you have to manage both units and ASPs. And obviously, in the end, revenue was most important -- most important. So that's what we'll actually talk about. I think if you look at the drivers of the EUV demand next year, the growth is still predominantly Logic, although we do have, obviously, starting memory for DRAM next year as well. But Logic, I think, continued demand for these advanced nodes, there's growing EUV layers supporting some of these end markets. I think if you look at the AI 5G HPC, high-performance computing, for example, these are all really strong and I think require resilient in this pandemic we're going through. And I think you'll see that continue as you add more applications to this advanced node space. So I think that's one, and then the start of the ramp of the one alpha node in DRAM obviously contributes to this EUV demand as well. Likely from a memory perspective, the plan for EUV hasn't changed much over the course of the year, meaning we expected one alpha we talked -- I think our customer in that space has been quite vocal on their plans there, so I think that's been pretty consistent. But I think in Logic, we've seen some customers recalibrating basically the timing and slope of their node ramps. On one, you have a customer that's talked about delay in terms of timing of the road map, this translated to tools that were needed later in time. So this reduced some of the demand next year, which we talked about in October. Then earlier in the year, we had call an impact from the geopolitical environment where we had a customer of a customer who was impacted and took away, if you will, some of the demand on the advanced nodes. So these 2 items were the primary drivers for our adjustment as it relates to our original view of 2021. And I think if you look at the void that was created at the advanced nodes at foundry customers, that appears to be based on our customers' commentary that it's filling faster than maybe they had expected. And this is fueled by, obviously, the end market's strength and other customers moving -- customers of theirs moving into these advanced nodes. And currently, this development, the way its progressing, this momentum could continue to translate to a potential additional EUV demand next year. So we are continuing to drive our manufacturing cycle times down. We talked about coming from 30 to 20 weeks, that's still our plan. 30 at the start of the year, 20 weeks by the end of the year in terms of our manufacturing cycle time. And we're planning some buffer capacity to accommodate this potential upside, if you will, should it develop.

So just to follow-up on that and be very clear about what you've assumed in 2021 is obviously going to push at a large North American customer of yours in terms of the time line. You -- have you embedded anything in your 2020 assumptions in that $5.4 billion for any capacity being built elsewhere, potentially a foundry for that customer?

No, I don't want to -- yes. So I won't comment too specific on the customer. But to your question, I think, yes, we adjusted the demand, as I mentioned, because of the timing of this road map. We don't yet -- we haven't made any adjustments with respect to the potential scenario that could evolve in terms of outsourcing versus internal manufacturing. And I think it was wait and see as we get into January, which is when this customer said they will provide an update on their manufacturing strategy as it relates to going forward. And so we'll have to see how that deal plays out. I think, Peter on the call referenced this communicating vessels in the sense that if demand were to come down, in this case, think of, could it be some internal demand reduction that it would -- in a true communicating vessel scenario go up somewhere else instantaneously. Well, in fact, there's probably some separation in time, whether it be 1 to 3 quarters. So we'll see, depending on scenario, if there were to be some external and this goes up in time, we need to be prepared for that and see if it's later in next year or early 2022. Again, these are scenarios that we'll have to see as we get clarification, it sounds like in January from this particular customer. We at least be prepared to have some buffer capacity in place depending on the scenarios for this or as well as the other one we talked about in terms of the backfill [ to be hedged ] in the foundry.

Sure. So obviously, you can't be too specific on some of the customer elements, but there is a scenario out there where one of your large customers where Intel may now start to outsource to foundries going forward. What do you think that might mean? And what impact could it have on your business and on EUV demand? I guess, there's 2 potential puts and takes in this that I have in discussion with investors. One, what would be the intensity of EUV of that node if manufacturing was to shift to someone else? And what would be the potential headwind in terms of efficiency benefits of that production being done elsewhere? And how might that then impact like-for-like production out of foundry instead of at Intel?

Yes. Yes. I think this communicating vessel scenario we talked about, you can get into lots of detail about the utilization, layer scenario demands that all could come into play here. So I don't -- because we don't know yet, I don't want to get into a speculation. Again, I don't want to be too customer-specific either. But I think it's fair to say that we see it as -- if it does, in fact, come down, it will come up somewhere else in a, you want to call it a net zero-sum game or could it be a plus or minus, want to see. But I would think, in general, we view it as a net zero-sum game, and it's just this communicating vessel scenario we spoke of in October.

Okay. I guess we'll have to wait and see what happens in January to go through in more detail. In terms of the -- one of the areas where you have been a bit more optimistic on the spending environment has been memory. You've talked for a while about seeing utilization rates starting to increase. And the potential impact that might have on, therefore, demand for litho. How do you see that trend today? And how do you see that trending into 2021?

Yes. I think if you listen to our customers, I think they're clearly stating that they're seeing positive signs as it relates to memory. In particular, I would say, the DRAM, as also DRAM is a bit more litho intensive -- quite a bit more litho intensive with respect comparing it to say NAND. The -- I think 2 comments there: One, there was a concern as we went into this year about inventory levels. And I think if you listen to what they're saying, they expect to have most of this normalized or sorted out by year-end. And then maybe the second item is looking out into next year. I think the -- the bit growth we see strengthening next year. So again, 2 items that maybe a bit more uncertain seem to be becoming a bit clearer as we move through this year and start looking into next year. In terms of our -- if you look at the litho tool utilization in DRAM, we spoke of those in Q2 and Q3, we said coming into the year, we started to see these utilization levels moving up. And I think if you look at where we're today, we're seeing that they're already reaching these maximum levels such that additional demand needs to come from new systems. And if you look at what we said at the start of the year, we saw this year starting to see a recovery in the second half of the year, I think that's still the case. And in particular, we see a stronger weighting in Q4. But looking out beyond Q4 -- or maybe the first comment about the Q4 shipments, that doesn't translate obviously immediately to bit output. You have to take into account that it takes, you could say, a couple of quarters before system shift in a given time frame will be actually producing qualified bits out. So what we're shipping in Q4 in the quarter right now will actually be translated more towards bits in the middle of the year. But now if you look into next year and you talk about DRAM bit growth, say, around 20%, which we consider more normal bit growth, we expect they'll need not only the technology translation, but you'll need additional capacity. So we expect this continued memory -- this continued recovery in memory to continue next year. And we'll provide -- we plan to provide more color around the degree of that. We provide this directional commentary in our October results with respect to next year. We'll hopefully provide more color around some type of -- how strong do we expect it to be in relation [ relational view ] of our next year, I will give our update in January.

Okay. And so if we move back on to the EUV topic, looking a little bit further beyond 2021, there's obviously been starting to be a lot more discussion around the next EUV node and TSMC'S N3 development progress. They've already been quite public around the time lines on which they plan to build for both N4 and N3. Is there any comment you can give on, if we even just, call it, Industry 5 nanometer, what you think will be the EUV intensity in terms of the layer count? And whether the ecosystem is ready for that node?

Yes. I think we commented on our Q3 call we expect to see a roughly doubling of our EUV layers at the next node called Industry 5 node, which translates to over 20 EUV layers. So that gives you a feel for the intensity, which we've said before industry 7 were 10 or more layers. So that gives you a feel for the growth of the node transitions here. And I think that's, again, a reflection, not only of the confidence in EUV, but also that as you continue to scale the opportunity the EUV provides there. I think the eco -- to your ecosystem question, I think, the EUV ecosystem continues to develop and mature as we obviously move up the learning curve and more wafers run, and major challenges. I mean if you look at the ecosystem discussion that's went on for the past several years, that really were around the resist, the mask, pellicles, the inspection capability. And I think that's something -- these major challenges within the ecosystem are something that you see when you make a [indiscernible] transition. But if you look at, as we progressed here, some of our customers who had communicated at the different conferences throughout the year, they would provide a stoplight chart in terms of how they scored the different aspects of the ecosystem. As I just mentioned, the resist, the mask, pellicles and so on and so forth. And there's a lot of red, I think, in the past few years. But if you look at this past year, I think majority of that is green. And so that obviously is a positive checkmark. But I think it's also sign that the industry with more suppliers, more resources are put into these different activities to drive that towards green. And as you look to the next node, there will be improvements, but they're more evolutionary improvements required. And I think the broader industry is working on these and feel that we'll be ready at any time to the next step.

I guess one topic on and we've talked a while part of the point of -- or the benefit of EUV deployment for customers has been moving away from multi-patterning approaches back to single patterning with the adoption of EUV. There's been a discussion at a few industry events I've been at, that we might need to start seeing multi-patterning again with EUV before we get to High-NA, which will come to in a second. Is that something you're starting to see resolved? Are we going to need multi-patterning for Industry 5, TSMC N3? Or can we get by without it? How do you see that playing out at this point?

Well, I think as we mentioned earlier, we expect EUV layers to be on this Industry 5 to increase in a way where you're above the 20 layers for the Industry 5. We're planning High-NA for volume production in the 2025 time frame. Obviously, the value driver there being not only continuation of shrink, but also the avoidance of multi-pattern. I think as customers are still in development of their future nodes and how they'll actually be pattern, it's a bit too early at this point in time to comment what the patterning strategies will be beyond the Industry 5 nanometer node. So we'll have to wait and see all that get a bit further out in time before we have clarity exactly how we plan to pattern this.

Okay. And in terms of the DRAM adoption, you've already mentioned that there's quite positive trends in terms of the likelihood of adoption at one-off and Samsung has been very vocal about how aggressively they're pressing forward with it. How much have you accounted for already within your EUR 38 million to EUR 42 million or your EUR 5.4 billion for 2021 in terms of DRAM? Could there be any incremental demand there if Samsung press this harder? Or is that as fast as they can go next year? And then how do you see this trending for Hynix and Micron in terms of adoption?

Yes. So the short answer is there's definitely EUV systems planned next year that are targeted for DRAM. And maybe again, in support of that, we've said that they are a first mover in memory to use EUV in HVM, has communicated their plan that they'll adopt EUV on one layer as a learning node at the 1z, which is currently being produced, but not for the full node, so it was a way they can manage learning a risk as a broad EUV into memory. They then stated they plan to fully deploy EUV in HVM at the one alpha mode, which would start, be able to plan ramp timing. If you look at what they communicated would mean that EUV systems would need to start shipping next year. This is obviously incorporated, I mentioned, into our 2021 view. The value is really about process simplification. That's the value proposition that EUV provides in the sense that you reduce or eliminate multi-pattern. This translates, you can first calculate cost advantages of making this move. But there's also a device or performance -- device performance or yield advantage as well as cycle time, maybe that's even more compelling than just the straight cost of ownership equation. But either way, I think because of this value proposition and in addition to the maturity of EUV in a sense where we are with productivity, it makes it much more compelling for other customers to move into EUV as well. We expect they will in due time. And most have communicated some public plan for use of EUV in the next 2, 3 years.

I guess the one where some of the commentary has changed I say, that have been a bit more vocal was around Micron, where in the past, they have been quite reluctant, if I can put it that way, in terms of the pace at which they've been moving to EUV. Are you seeing any increase in the pace of interest in EUV beyond -- both of the customers beyond Samsung? Or is it still kind of the -- I guess, has there been any change in the last 6 months or the last 12 months in the engagement you're seeing elsewhere given how aggressively Samsung is pressing ahead?

Yes. I don't -- there -- I think -- so first, I think the -- there is continued interest for other customers in the memory space with EUV beyond the first mover. So I think that's a fair statement. In terms of the U.S. customer that, it has pushed deep UV longer. I think in one of their technology updates last week, those that attended were of the opinion that they were a bit more positive on use of EUV, and they communicated their timing, I think, in the 2023 time frame. I think that has to do with the fact that we have accelerated our productivity road map. We continue to make progress going from the C now, next year we'll have the D which will be 15% to 20% faster productivity. We continue to move availability up, which not only helps the productivity but also helps in terms of the predictability in using EUV and manufacturing. The maturing ecosystem, as we talked earlier, obviously, will give them more confidence to employ EUV. And I think the other thing is they were scaling and planning to expand their fabs to accommodate EUV. So I think a number of those things probably are all in support of their view, it's a bit more positive around use of the EUV in memory.

And to the point you just made in terms of the progress you're making on the productivity road map and what that can mean. As we started seeing the D tool launch next year, and I think you've pointed to somewhere in the range of a 10% to 15% productivity increase that we'll see with that, we've well understood what that can mean for ASP. But can you help us understand what that might mean for gross margins on that tool as well? Is there much incremental cost of manufacturing? Or is it the increment mostly something that drops through to the bottom line?

Yes. So maybe a comment on just on EUV system gross margin. But again, coming out of 2019, we said we would be around 30%, and I think we executed nicely that to above 30. I think, this year, we said we expect the EUV system gross margin to be around 40%, and we're tracking nicely to this and likely will slightly exceed that target. So that's, again, positive development. The -- to your question, the 3600D, which we'll start shipping in the middle of next year, predominantly have a majority of our shipments in the second half will be Ds. This will provide value for our customers in both form of improvements in not only productivity, but also [indiscernible]. We'll share that value, which translates to higher ASP with improved margin. And I think we've said in the past that with 3600D, we can move towards the corporate gross margin. And if you look, though, out the next 2, 3 years, I think we said by 2022, 2023, we expect our EUV system gross margins to be similar to deep UV-type margins. And so we'll continue to move on that path, and we'll plan on producing EUV machines roughly on a 2-year cadence. So it's probably reasonable to assume we'll see another machine beyond the D, could be called an E that would come out mid-2023, which will provide additional customer value as well as, obviously, higher value to ASML.

That's great. And in terms of the next step beyond that, obviously, then we'll be starting to get towards High-NA. Where are we in terms of the readiness of the High-NA tools? It's -- from what I understand you've already produced some of the initial optics platforms. There's been a lot of heavy R&D engagement. We're seeing press reports about customers looking to invest, if I can use that word, which wasn't very descript. What still remains to be developed on the tool? And how confident are you on the time line of introduction maybe for High-NA?

Yes. So maybe a bit -- I'll comment on each of these in terms of progress. First, we were making good progress on High-NA. I think the major areas of innovation, if you look at High-NA, again, lots of engineering there, but we're trying to summarize into what are the 3 major areas: I think it's the source, it's the stage and it's the lens. I maybe comment a bit on each of these. I think first on the source, we're basically using the same source design architecture, if you will, which is called LPP, which is laser-produced plasma, that we use in the low NA or the current 0.33 NA systems that are shipping today. Obviously, we have to scale the power. But we feel we've already demonstrated some of this feasibility that gives us confidence that we can deliver, I think, on the source front. So I think that's, again, you could say, progressing nicely and confident there that we can provide the source scalability that we need. Next is on the stage. I think we have to not only -- the key is when you go to High-NA, you need to, obviously, maintain the productivity. And we need, therefore, to scale speed and acceleration. Stage technology, mechatronics is one of our key areas of expertise, and we've already been working on protos that give us confidence that we know how to do this. So again, I think that's positive on the ability to deliver on that front. And the third, which is potentially the biggest change, at least as it relates to High-NA in terms of the development is really around the lens, the optics. The challenge, obviously, on this front is scaling, but scaling in a way where you also need to keep the mirrors very, very flat over those larger areas. So you're talking something on the word of 1 foot in diameter going to something closer to 3 to 4 feet in diameter, but the flatness level across this obviously becomes more and more challenging. And you're talking picometer level here. So we -- to do this, first, you have to have the manufacturing capability, but you also have to measure it, which is quite a challenge. And so we had to deliver and develop with our partner, Carl Zeiss, these interferometers that can accommodate very large optics that are in a vacuum. And I'm talking very large vacuum chambers that are on the scale of 10 feet tall. And we had to build a facility to accommodate this. And the good news is that last year, we met one of the key milestones in terms of the ability to demonstrate the production of a mirror and utilizing the interferometer. So I think that, again, moving along quite nicely on that front. On the ecosystem, I think which we discussed earlier, I think the nice part with High-NA is that you can reuse or it's more evolutionary in terms of the developments required around the resist, the mask and pellicles. Just like I said, from going from one node to the next, we talked earlier, this would be yet another node. And you can really reuse the base technology, but you do have to make evolutionary improvements in terms of performance compared to what's required because the features are getting smaller and smaller, the capabilities are getting more and more challenging. So you need to scale accordingly. So -- but you don't have to reinvent and restart, and it's not a revolutionary-type development that we saw when we went for the first low NA introduction. So I think that takes some of the risk and concern away. In terms of timing, we still plan to ship the initial development systems in the '22-'23 time frame and HVM systems in the 2025 time frame. So I think that kind of gives you an idea of the progress of the program as well as the timing.

There's questions coming from the audience. But in terms of the productivity gains, and I guess the best way to answer this is maybe if you have a customer in 2025, that's trying to decide between do I adopt High-NA or can I get by with using whatever it will be, the 3600E by that stage, how will they think through that trade-off? Because as far as I've seen before, there might be another situation where High-NA is slightly lower productivity again in terms of wafers per hour throughput, but you then maybe avoid multi-patterning. So can you help us understand a little bit about the value discussion and the productivity benefit that High-NA can give to your customer?

Yes. I think what -- and we provided some of that at least 1 graph on our Investor Day material 2018, like we can talk a bit more about it in next year at our Investor Day. But I think the message there was around just showing cost per layer, obviously, you can translate that in cost per wafer, what High-NA could offer in terms of a single-exposure High-NA process versus a double or triple-patterning EUV low NA process. And I think even for scenarios we're talking about here in terms of productivity in both High-NA and low NA, that we can show a cost compelling advantage of just EUV single-exposure High-NA compared to low NA double patterning. The other areas where the benefits are clearly there, but maybe even more compelling are around when you simplify a process, and I think a lot of this is learning, you see having gone through, we went through with the multi-pattern deep UV comparing to EUV low NA single exposure, is that the process complexity really also hits you a lot around the device performance. The device performance obviously can translate to yield. And the way to think about that is you're tightening the distribution of a given device parameter. And so that, in fact, translates in the end to variation, which, in turn can be yield fallout. So I think that's one big factor that motivates our customers to move to the High-NA, this single-exposed process and minimize the use of multi-pattern. The other cycle time. Cycle time has multiple advantages. One, it's just getting wafers out the door faster. But the second is the more time the wafer sit in the fab, higher risk of a particle landing on the wafer and causing a defect-related yield hit. So there's multiple advantages there, I think. It's more than just cost, but cost is typically the first step that you show and what we have shown in the past to be the advantage of going to single exposed.

That's great. There's obviously a couple of areas where we're going to see some gross margin impact of High-NA. You called out with the last set of results, there'll be some impact next year. Can you just talk through a little bit what are the actual cost drivers of that percentage point impact next year? What is it you're doing that's leading to that cost? And then I'll come back on the longer-term afterwards.

Yes. Yes, I think so -- I'm not an accountant, but -- and Peter and Roger are. They have training there. And they said that this was a rule, accounting rule that was applied. It didn't fully satisfy exactly to them, at least the explanation. But it's a case where we're bringing some of the normal, let's say, cost in terms of building the modules forward in terms of time as compared to shipment. As we mentioned, we'll be shipping layer, but we're already taking cost in 2021, which we indicated that will be roughly a 1% headwind corporate margin -- corporate gross margin. Now the -- you could say the plus of that is that some of these costs that we are planning on taking later, I think, in '22-'23 time frame, you have already taken some of that in 2021. So it just moves the timing of that cost, but it does impact us [indiscernible] we'll have as opposed to taking all at the time of revenue [indiscernible] which make not so much difference in terms of what we expected in cost.

So if we leave the accounting and exactly when you have to recognize it to one side, obviously, we all remember very well what's happened over the last 3 or 4 years as you've introduced EUV and the headwind to gross margins and then eventually shipments pick up and gross margins pick up, what will the story be like if we ignore accounting? And just look at on a product basis, are we looking at barely any or negligible gross profit on the initial R&D systems, and then we get the profit with the high-volume tools in 2025? Or what's the profile of that gross margin for the High-NA tools? Because there's a lot of costs you've already taken upfront with investment in Zeiss and everything else as well?

Yes. I think so. What I can say now, we'll hold the rest for Investor Days. And we think the gross margin on High-NA, we expect it will be better starting point than the level we started low NA. As you know, it was pretty bad on the low NA front. And so -- but we can provide more details in terms of the profile, what that may look like as we get closer, obviously, to Investor Day next June and we dig a bit more into the High-NA program. So I'd like to push that till that time.

If I can at least try 1 more on that. And so because I guess, if we think of some of the dynamics around your HVM tools by, say, 2025, ASP is potentially doubling from where you are today. Is it fair to say that cost doesn't double? Or how do you think of that?

Sounds like it would be fair.

Okay. So it can potentially be a higher-margin tool, certainly than you have today, higher compares to what low NA is by then we'll see. So moving on to the other side of the business. One of the big stories this year has been the installed base management business. And obviously, there's 2 sides to that. There's the upgrade side of the business and then your service and maintenance side. In the upgrade side of the business, first of all, field upgrades have been very strong this year, as far as I understand. I know you don't give exact growth rate numbers. But it sounds like you've had a very strong year this year for field upgrades. Can that continue into 2021? What are the drivers of that? And do you still have more upgrades to sell next year that you can still grow that side of the business?

Yes. So we -- I think -- yes, like you said, we comment on the installed base as a whole. But I think we expect the overall installed base business to grow next year. The degree of growth, obviously, will largely depend on upgrades. Service, I think, is safe to say will grow pretty linearly as a function of install, call it, increased installed base. But the upgrades are a bit more, you'd say, variable or lumpy. Not that we don't have the products. We do have plenty of products in the upgrade pipeline, if you will, but it also depends on customer readiness. So it has to be the timing that fits their technology, their capacity ramps, and they also have to be able and willing to take the tool down to the upgrade. So -- and this creates, you can say, the variability we see from year-to-year in terms of the growth. I think installed base this year, we expect to be up over 20% versus, say, 2019 compared to 2018, which was less than 10%. So this is the upgrade variability. So to your question, yes, there could be -- again, it was another strong year. There's plenty of upgrades in the pipeline. But we'll have to see how -- what the customers conclude. Using the forecast on upgrades are a bit shorter in terms of lead times versus system forecast. And so we'll better -- we'll have obviously better insight into next year, as we get into next year, so we'll provide more commentary on our installed base view next year in January.

But is it fair to say on the upgrade side of the business with D tool being launched in the second half of the year, a new EUV system profile, essentially, can you move to upgrade tools as quickly as you launch a new tool? Or is there normally a bit of a delay and potential offering of upgrades to D is more of a 2022 story?

Yes, I think there's usually somewhat of a delay using -- if the new machines out the door used the hardware for that, and then you come back with some separation and time to start doing upgrades in the field. So I would say that your current view that more of the D-type upgrades would be in 2022 is probably fair relative to '21.

Okay. And the EUV service business has obviously been starting to pick up quite heavily, particularly towards the end of 2020. As that carries on over the next 2 years, how quickly can that grow because there's quite a lot of different inputs to it that are hard for us to directly track? Obviously, we understand the volume and the installed base growth that we do know or hopefully have a track of, at least. But the -- in terms of what's happening in your ability to generate revenue per tool is dependent on warranty cycles, it's dependent on throughput that the tools give out. How do you think about in the next couple of years, how that -- how quickly that service business can grow? And where do you think the endpoint is in terms of kind of your revenue per tool opportunity in EUV?

Yes. So as you noted, the EUV services already started contributing more meaningfully to the service business and the installed base as EUV continues to ramp in volume. I think this is due to the fact that the EUV service revenue business model, the change compared to the EUV, we had more of a break fix model to now we have more of a value sharing model and that the more wafers that are run, the higher value to our customers, and then we share that value. We -- in terms of your question on revenue, we provided a revenue value of around EUR 5 million to EUR 6 million per machine per year when the machines are running in, call it, high-volume manufacturing post warranty -- warranties around 2 years. And as our installed base obviously grows, this higher volume and more wafers are being run, that will contribute to higher revenue, but also as more machines come off warranty, that will also increase this contribution to service revenue. There's obviously the more machines, the faster the machines, there's an upside, obviously, that revenue per machine per year opportunity. And so that obviously will continue to contribute more and more meaningfully to the service side, even though we don't break out that level of detail in our service or our installed base number. Maybe around that same front would be come gross margins. If you look at gross margin of our EUV services negative coming into the year, then turned positive in Q3. And we expect that margin to continue to increase over the coming years.

Just a follow-up on that last point. Part of the point in the new business model that you're coming with was service, where the customer essentially pays for [indiscernible] and it's your responsibility now to deliver that and service the tool and maintain the tool. Where in the past, it was more their own risk. Have you got good visibility today on what service and maintenance cost will be that you know what the gross margin outlook will be? Or is there still some kind of understanding needed over the next 12 months as we get more tools up in volume production? Because when I asked this a year or 2 ago, there was still some uncertainty, as we've now got through Q4 this year, is that clear too? Or is there still some work to -- things to be worked on?

Yes. Well, I think -- so the drivers, I think, are clear and the direction is clear in terms of increase, the ultimate value, obviously, and time, you can have lots of discussion on and maybe reserve that for another day at a later time. But I think if you look at the improvements in gross margin that we went from just negative at the start of the year to positive at the end of the year, and has continued to increase over time, it's really a combination. One, we talked a lot about the revenue side. The more wafers you run, the more revenue you get. The more machines come out of warranty, the more revenue that will kick in as well. But there's also the cost side. And I think on that side, it's important is that you had to put all this support infrastructure in place at all of these customer sites across the globe when you had very few tool sets initially. Then you ramp more and more units in the field, you gain some efficiency there in terms of think of man to machine ratio and just our ability to support, you call it improve support efficiency. So you have revenue going up and you're trying to bring your cost down. So those combos, I think the combo of those 2 is really giving us the improvement going forward in terms of the ultimate number, though, I think, we can debate on the time and the exact number, but hold that for now.

Keeping things with the Capital Markets Day next year, Skip, make sure we all attend.

June 23. And yes, I would like to thank you, but [I rolled out at ] kind of right on the cusp of COVID, but hopefully, the vaccine will [indiscernible]

I think we all do. One final question I'll just throw in, because one of the topics that we've had over the last 12 months has been, I guess, around cash flow because, obviously, as EUV has started to get towards more maturity, some of the working capital challenges we've had, customers placing orders without deposit now placing with deposit, we would hope to see more stable cash generation from the business. Can you just comment on how you see that playing out over the next, I guess, 12 to 18 months, what we should expect from a cash flow perspective?

Yes. Yes. I think we commented in our Q3 call we expect to see an improvement in free cash flow in Q4. And then further improvements ahead in 2021. I think there's -- maybe first to comment on Q4. We did have some systems that shipped late in Q3 that will result in cash generation falling into Q4. But as we also mentioned, we're in the process of transitioning contracts from all extended payment terms to down paying to prepayments. And some of these extended payments will come due in 2021. And we'll also see increases in our down payments on future orders. So I think this basically is what gives us the confidence that you will see significant free cash flow generation ahead. This in addition to our assessment of the current COVID environment in the sense of our ability to operate and the supply chain assessment and why we basically resumed our share buyback program starting this quarter.

That's great. I think with that, we've run out of time. So thank you very much, Skip. It's been a pleasure, as always. And look forward to seeing you in person again sometime soon. And thank you to everyone for joining us. Have a good conference.

Thank you, David. Thank you, everyone. Take care. Stay healthy. Happy holidays. Look forward to seeing you soon.