I used to tell clients that what went in the data center mattered a lot more than the facility itself. Yes, a data center program represented a big capital request, but if you depreciated the shell over 25 years and the mechanical and electrical equipment over 15 years, the opex just didn’t matter that much compared to systems, software and labor. I made lots of slides back in the day showing that facility cost might make up ten percent of the TCO of a server image.
How times have changed! Companies (mostly hyperscalers) spent USD 260 billion on data center construction last year before they installed a single server.
Historically, data center programs have been bespoke efforts, where much of the work happens on site. Companies had to build even “modular” data center architectures largely on site. Both providers and consumers are pushing to build more of the data center in a factory and less on site.
This applies a very old economic dyanamic to a very modern domain. At the birth of the industrial revolution, British industrialists transformed textile manufacturing productivity by replacing the “putting-out” model (in which workers spu thread and wove cloth in their cottages) with the factory system that brought the workers together at one site. Yes, this allowed for greater automation -- but knowledge sharing and management control as well.
InfraPartners is in the middle of the transition of data center construction from a bespoke project to an industrial process. CTO Harqs Singh and I talked about how that works and what that means. A few things that stood out to me
Much of the productivity advantage from building data center modules in factories rather in on site derives from co-location rather than automation
Sometimes customers push to tweak data center designs, even when it could add months to lead times
Skilled labor from the trades is as much (or almost as much of) as bottleneck as chip availability
Building data center capacity is as much of a moral as an economic project in that it contributes to abundance
James Kaplan: James Kaplan here for another session of the Prosaic Times podcast. I’m recording this on the Brown University campus because the Watson School is hosting a great session on governance in China, which I’ll have the privilege of attending today. But that’s not what we’re talking about here.
We’re going to talk about data centers, data center architecture, and data center economics. So I’ve asked one of the smartest hosts I know on that topic — Harqs Singh, who I’ve known for years — to join us.
Harqs, do you want to introduce yourself and talk a little bit about your journey from being an infrastructure leader to being a data center entrepreneur?
Harqs Singh: Absolutely. Thank you, James, for having me. Hi, everybody. My name’s Harqs Singh. I’m currently CTO and co-founder of a company called InfraPartners. We build prefabricated AI factories. I’ve been in this industry over 20 years now in technology — cyber, data — started at British Telecom, then Thomson Reuters, then BlackRock.
Then I decided to jump into this largest deployment of capital in human history, as I was told, in the AI space and the excitement around what this could enable for the world — and jumped in with both feet, riding the rollercoaster of delivering AI across the world.
James Kaplan: Fantastic. So you’ve seen multiple generations of what a data center is — from the bespoke Tier 2, Tier 3 data centers we built ages ago to the more modular enterprise data centers or Tier 4 data centers that got built in the 2010-ish timeframe to now — something that feels almost entirely different.
I was wondering if you could reflect on that evolution and talk about what, if anything, has stayed the same and what has changed.
Harqs Singh: Of course. It’s changed a lot in the last 20 years.
Enterprise to colocation to cloud, and now to AI — we’re seeing some really big technical changes. So we’re seeing direct-to-liquid chip cooling, which has really been driven by the intensity of the AI loads.
We’re looking at racks that are now 140 kilowatts per cabinet, going up to 200, 320 — which is something we didn’t imagine in the past. And so this brings along a lot of engineering challenges and thinking about things differently.
NVIDIA and the next generation of equipment are looking at using DC power — 800-volt DC power — for their data centers. And so I think you’re going to have two types of data centers: you’re going to have the CPU-focused data centers, and then you’re going to have the GPU-focused data centers, essentially.
The growth has been significant. If you look in the past, people talking about having a gigawatt portfolio was quite a large data center estate — and now everybody is looking to build at least a gigawatt, and you’re looking at some of the largest hyperscalers looking at 20 gigawatts. The projections are anywhere between 100 and 200 gigawatts of additional incremental data center capacity that’s needed by roughly that 2030 period.
James Kaplan: I remember when we used to talk about square feet rather than critical power, and we got really excited when we started talking about 75 watts per square foot. That was a reasonably sophisticated data center at some point.
Harqs Singh: Yeah, absolutely. The intensity has grown significantly. So these data centers fundamentally have changed and evolved. And I think to the point you talked about earlier: we need to change; we need to industrialize; we need to be able to deploy at a scale we haven’t deployed in the past.
We did some analysis on the supply chain to support the growth by 2030. There needs to be something like a 10× increase in supply chain. So we need to think about doing things differently. In the past, we used to build every data center differently. Every data center was almost like a snowflake.
But now we need to standardize and accelerate deployment by having really standard components and units that we can deploy very quickly wherever needed. The biggest issue has been — in the past, to your point — space was the biggest thing we optimized around. Now it’s power and available power. If you think about where the data centers are now going — whether it’s West Texas or the Nordics — a lot of the data centers are going where there’s spare capacity and spare power. We’re seeing a lot of issues with the grid. So a lot of data centers at the moment are looking at on-site generation using gas engines or other forms of technology to be able to essentially bring their own power behind the meter.
James Kaplan: It sounds like what you’re describing is much more of an industrialized process. It used to be the case that if you were a big bank, you might do a major data center program maybe once a decade — maybe once every 15 years — and you were sort of relearning it with each generation. What you’re describing is we’re adding — or the world is adding — tons of capacity every year, and therefore it’s much more of a continuous industrial process rather than an episodic project.
Harqs Singh: I think this wave that’s come with this sort of growth of AI and what it’s going to enable — capital isn’t a concern at the moment, which is a good thing — but it’s really things like supply chain.
You really need five things to be able to deliver AI capacity. First is land, power, and permits — that’s where we’re going where available power is. The second is supply chain — that’s where we’re really focused, because skilled labor shortages are a real problem. Being able to find skilled labor in West Texas or the Nordics is very difficult, and that’s why we look to build most of what we do in the factory. Being able to get chips from NVIDIA or from other chip manufacturers. The third thing you need is chips. Fourth is capital, and the fifth is an end user.
Only if you have all of those ingredients are you able to actually deliver AI capacity. And so it is complex. If you think about it, James — one gigawatt of data center capacity, including the chips, is about $50 billion. Which is obviously a significant investment. So being able to pull that together — although we hear a lot about these gigawatt announcements — think about that as $50 billion that’s been deployed.
James Kaplan: What is, to your thinking, the rate-limiting factor? Especially if you drill down one level — what type of labor is the biggest constraint? Is it electricians? Is it something else? And what type of equipment is it?
Harqs Singh: The biggest skills gap is electrical, mechanical.
There are shortages, especially in those more remote locations.
If you’re in demand as part of the skilled labor force, you may not choose to live in West Texas or the outer reaches of the Nordics, because you have choices. And so that’s the biggest issue on the skill side. If you look at the other side around supply chain, the longest-lead items are essentially generators — backup generators. They can be nearly two years now. And also transformers. We’re seeing some innovations in the design around looking at designs that don’t need generators — being able to design them out or just design them for a portion of the facility — 20% of the facility related to sort of the network and management pieces. But transformers are the other big one that we look to optimize around.
Either being able to find smaller transformers or reuse them if you can, if you’re not in the queue already to be able to buy them.
James Kaplan: Are demand signals starting to work on labor supply? Obviously this all happened relatively quickly, but you have a specialized skill set around the electrical and mechanical. Where did those people come from? What’s required to expand the flow of people into those jobs?
Harqs Singh: I think it’s really being able to encourage more people to come through the vocational side — the skilled labor side, apprenticeships, those kinds of things. It’s really interesting because AI is going to create an abundance of intelligence and agents and those kinds of things — sadly, it’s not going to create an abundance of electricians or plumbers or mechanical skills. So I think that’s something we’re going to have to focus on — getting people through colleges and universities and really trying to broaden that skilled labor pool. Jensen recently mentioned that if you look at the salaries now that some of these skilled trades for data center construction are pulling, it’s becoming really attractive.
I think the future of the distribution of skills we need on AI is going to be very different than where we are today.
James Kaplan: If you’re planning to deploy $50 billion, basic economics would tell you: if someone’s planning to deploy $50 billion of capital and the constraint is an electrician, they’ll pay real money for that electrician, right?
Harqs Singh: Absolutely. And that’s what you’re seeing, James. If you look at the labor costs of data centers that are being built in some of these remote locations, you’re looking at them paying 20, 25% more on the overall cost — even double, triple the salaries you would normally expect. And that’s really where our model is that we build 80% in the factory, 20% on site.
What that means is the skilled labor gets to still live where they want to live, in the cities they want to live in, and only for a short period of time are they in a remote location. So it solves for that skilled labor issue. But you’re right — there is a premium to pay to have somebody leave their family for a year or two and move to a remote location.
So there’s an increase in the running costs, but there are also large retention bonuses that are being paid to really help push people to stay out in those remote locations to get them done. But the growth is so much that they’re struggling to find, even at the premium rates, the skilled labor to go out and do the work.
James Kaplan: Who’s going to train these master electricians, as an example?
Harqs Singh: Yeah, absolutely. And I think, James, the challenge is going to be: can we get enough of them to market in the next — the conversations I’m having a lot right now with clients, they want data centers online next year.
There’s a demand-and-supply imbalance at the start of 2027, and so it’s not going to solve for the short term — but I think this growth curve will go out to 2030, so hopefully we can solve for some of that in the longer term.
James Kaplan: I find your investment thesis around productivity in the factory as opposed to productivity on the site incredibly compelling, and we’ve seen that in many different sectors of the economy. One of the reasons housing costs are high in the United States is that, for a bunch of reasons including regulatory ones, it’s been hard for the housing sector to do more in the way of manufactured housing as opposed to housing built on site.
How does the skill mix differ when you build something in one of your factories versus when someone builds something on site? Is it the same set of skills but you just get more out of them? Or is it a different mix of skills? How do you think about that?
Harqs Singh: It is the same set of skills, but we get more out of them. That’s the way to think about it — DFM principles that shave time off the schedule. I think the other thing the factory solves for is it standardizes, it gives you a level of QC, and it also accelerates timelines. We’ve found there’s less waste. If there’s something left over, we reuse it for the next set of modules and blocks that we build. So we’ve found that from a sustainability story there are some positives there as well.
James Kaplan: What’s your labor pipeline look like? What’s the ideal person working in an Infra Partners factory? Is this someone who spent 20 years at a big bank working on a data center? What’s the profile for that type of person?
Harqs Singh: It’s mostly engineers — mechanical engineers, ideally from different backgrounds. We have quite a standardized way that we deploy. We have a standard reference design, which we start with, and that saves time as well. We have a standard reference design that aligns with the NVIDIA DGX lean framework. What that means is we need specific electrical, mechanical, construction skills. We find that in Houston and in major cities we can find those skills quite easily. We’re not asking them to move somewhere where they’re going to be away from their families.
James Kaplan: If you think of that factory as existing somewhere on a continuum — and one end of the continuum is a big room with people in it, and at the other end of the continuum is a lights-out, fully automated factory — where are you guys now, and how do you see that evolving in terms of the construction of components?
Harqs Singh: Yeah — I think right now we are more manual than we’re automated.
Physical AI and more of the robotic stuff that we’ll see — more autonomous, essentially, in terms of being able to build the building blocks and prefabricated solutions we have. But at the moment, because in prefab — to your point, James — this is something that’s happened across other industries as well as industries have industrialized to be able to deliver product at scale — we’re not [fully there yet]. To your point about the prefabricated concept in general construction, it’s already been used in a lot of countries to be able to build commercial buildings and also homes — stadiums as well.
A lot of stadiums are prefabricated in the way they’re designed. So it’s taking that concept really to the data center space. If you look at the value of prefabrication, it is a 20% to 40% improvement in timelines, as you can centralize it. So it’s taking that concept, bringing it to a data center — well, bringing it to an environment which is changing quite significantly over the next few years as we drive these transitions: going from air cooling to liquid, from AC power to DC power, from racks that had a handful of servers in them to probably having a lot more GPUs in the future that are going to be much heavier.
Once those transitions play out over the next few years, aligned to NVIDIA’s roadmap and generally the industry’s roadmap around what’s happening with chips and the extreme co-design they’re doing, that will allow us to get to a more stabilized point in terms of technical design.
I think that will then increase — to your point about the spectrum — the amount of automation and robotics versus human input as well.
James Kaplan: I think what you’re suggesting is that some of the toughest problems in AI live at the intersection of the physical and virtual world.
Harqs Singh: Correct. And I think that’s where a huge opportunity lies. You’re seeing a lot of great companies that are focused on that. It’s exciting in terms of what’s going to happen in that space.
James Kaplan: If you think about the complexity of moving from a language model to a vision model to a world model — my God, how many orders of magnitude is that in terms of conceptual complexity? Moving from a language model, which interacts with text — as fascinating as that is — versus the world model, which conceptualizes objects interacting in three dimensions in accordance with the laws of physics.
Harqs Singh: Look — we’re going to have to train a lot of these models and these humanoids and whatever technology we end up using, because I think the best outcomes of AI are when you have trained it with reinforcement learning and human input to be able to get the outputs you want.
When people think “out of the box,” it’s going to make a huge difference — but that’s not enough on its own. You actually need to invest, train, and build the capabilities and the outcomes you want, just like you would train any employee.
James Kaplan: I was wondering if you could speak a little bit about the level of abstraction versus integration between what I might archaically think of as the system level versus the facility level. As context, I spent a lot of time working with banks on modular data centers in 2010, 2011, 2012 — and you always ran up against the constraint of, well, the systems don’t quite work that way. Is there more abstraction? Is there less system diversity? To what extent is the construction of the system, for want of a better word, a constraint on the degree to which you can modularize the physical infrastructure?
Harqs Singh: I think there has been some progress in that space. There’s been some maturity, especially on the IT side. If you look at a lot of the NVIDIA DGX design, it is based on a reference architecture that has become the default way to deploy.
James Kaplan: Mm-hmm.
Harqs Singh: What we’re starting to see — we just released our partnership with Emerald AI — what they’re looking at is working to make data centers grid-interactive, making it an asset where they take grid signals in. What they’re able to do is flex either the IT side or the facility side to be able to work with the grid when it’s under stress, essentially. So you’re seeing a lot more orchestration between the whole stack. One of the things that’s been driven over the last couple of years is: look at the whole thing from land, power, permits to data center to chips to models to data — how it all comes together as a full ecosystem across all of it.
In the past, James, to your point, there was this IT–facilities divide — but I think we’re starting to see this view, and this concept of extreme co-design that NVIDIA have driven — we need to do that across everything, especially when you’re trying to optimize something that’s going to be on the order of $50 billion for a gigawatt. Everybody’s throwing gigawatt deployments out every week. When you’re talking about that level of investment, there needs to be extreme co-design and it needs to work together as one unit. The revenue that’s on offer and the investment that’s taking place means we have to do it that way. There’s no other way to do it. You’re starting to see that in the industry with partnerships that are taking place and the companies that are working together.
James Kaplan: Back in the nineties, I think we used to see buttons and pins from Sun Microsystems that said, “The network is the computer.” It sounds like what you’re saying is: the data center is the computer.
Harqs Singh: Well, the GPU — the network is also very critical. If you look at the training cycle — computation, communication, updating weights and biases — the communication phase: you could have billions of dollars worth of assets sitting idle if you don’t optimize for network.
So that’s really critical as well, especially in the inference space where response rates to prompts are going to be important — obviously some of the more real-time use cases as well. We’re seeing data center designs being optimized and GPUs being positioned and optimized for east-west traffic but also north-south traffic and response times — because if you think about the investment that’s there, the more time it’s idled, you’re losing revenue essentially.
Tokens equal revenue in the future. And so it’s an —
James Kaplan: I am still waiting.
Harqs Singh: I would —
James Kaplan: Tell me it already exists.
Harqs Singh: I would not be surprised if in the next 12 to 18 months you see some token exchanges where tokens are being sold. But I’m a big fan — we’ve been in this industry for a while — efficiencies. I was having this debate with colleagues just last week: we’ve done some great stuff where we’ve created incremental efficiencies and then Jevons Paradox comes in and everyone consumes more.
So I sort of said: we really need to flip this whole thing on its head a little bit. What we need to do is create an abundance of everything — we want an abundance of food, an abundance of energy, an abundance of compute, an abundance of AI.
If we think about it in that way, hopefully some of these young, smart people will be able to take this abundance of everything and create something magical out there.
And that’s what I like about AI: the boundaries and rules that we historically have lived by don’t exist in this world. If you look at what some of the AI natives are doing and how they’re thinking about building companies differently and solving problems differently and getting data from different sources — there’s so much opportunity, and that’s what’s so exciting: the opportunity is untapped.
James Kaplan: I don’t think anybody in this industry is brave enough to say what the future looks like, because we just don’t know. That’s why it’s so exciting, right?
Harqs Singh: That’s why it’s so exciting, right? I’m an optimist in that sense — I think if we can unlock an abundance of everything, then some really smart people — probably young people around the world — will think about problems differently and create some magic.
I look forward to seeing what that looks like. But I’m very optimistic about what this can enable.
James Kaplan: This is a little bit of a tangent, so we shouldn’t pursue it too far — but I happen to agree that you’re touching on something very important: I think we in the technology world need to talk differently. We need to talk about abundance with more conviction, because I agree.
I think you’re saying that what we’re all jointly engaged in is a moral project that will contribute to human thriving and human flourishing — and we should be. I don’t think any of us should be apologetic about that. There are risks, there are guardrails, there are things we can do better versus worse — but I don’t think any of us should apologize for the larger project.
Harqs Singh: Agreed. And I think we all need to think at scales we’ve never thought about in the past. I think that’s the lesson from history. If you look at sizes of data centers — they’ve gone from megawatts to hundreds of megawatts to gigawatts, single digits to double-digit gigawatts — we should think at scales we haven’t thought about in the past to create that abundance.
Put that abundance in some very smart, intelligent people’s hands to see what we’re able to deliver and create for humanity. I think that should be the legacy of what we do in this industry: that we enabled that abundance of everything — and we are the backbone upon which digitization and technology and AI have permeated every industry.
In the past, you remember when we were a cost center — and now we’re fundamentally a revenue generator.
James Kaplan: Some people still think of it as a cost center, sadly — but we’re working on that.
Harqs Singh: Yeah — we haven’t won; we won’t declare victory there yet, James.
If you think about businesses that don’t have a core business that has a technology strategy and the revenue and products it’s making — or a data strategy or an AI strategy — even security: these things have gone from being something that happens at the back end of an organization to being key to the product you sell. Whatever product you sell in the future — whether it’s to humans that are probably going to be AI natives, or whether it’s to machines, because that’s who your consumer is going to be in the future — you’re going to need all those things around digitization, data, security, and AI. If you’re not thinking about putting that into your core business, you’re more likely to be left behind or disrupted.
James Kaplan: It’s fascinating what you guys are doing from a historical view. As someone who is fascinated by history, you could observe that there was a fundamental change when certain parts of the world in the late 18th century went from making cloth in cottages to making it in factories.
That sounds somewhat prosaic — but all of a sudden large numbers of people could afford two sets of clothes instead of one. I would say this project of moving things from a job site to a factory is incredibly important in terms of the way it increases productivity and reduces unit costs.
Harqs Singh: That’s exactly it, James. That’s a great analogy, and exactly what we think this should help drive for the industry.
James Kaplan: Just going back to this point about standardization a little bit — I presume, given who’s building data centers or using data centers, the large majority of your customers are either hyperscalers or people who are planning to lease capacity to hyperscalers. I assume that’s a correct assumption.
Harqs Singh: The client space has grown quite significantly. To your point, there’s the hyperscalers — U.S. or Chinese hyperscalers. Then there’s the Tier 2 colos.
This neo-cloud — and that’s been created: a lot of Bitcoin mining power has been migrated to building AI. Fourth segment is enterprises. We’re starting to see enterprises — certain sectors: financial services, healthcare, insurance — look to build small AI factories for their private data. And then the last segment is the sovereign AI space — where we’re going to see governments, for national security but also for public services for their citizens.
So what we’ve seen is it’s actually expanded significantly from the hyperscaler in the past to a few new segments in the market.
James Kaplan: To what extent do different segments or different customers within a segment require a slightly different product — or to what extent can they use a standard product? How much variation, how much tailoring do you have to do by segment or by customer?
Harqs Singh: Our starting position is always — our gold standard is the NVIDIA standard, the DGX standard. We start from there, which is a great starting point. Typically what happens is the modifications that take place are either for real requirements like security in the sovereign space, or some enterprise requirements — or it’s because the technical teams on the other side want to add their value, if we could put it that way. They tend to put in a few things that they would see as imparting their value to the firm — not always necessarily needed. We see that the hyperscalers — and the people that have these technical teams — tend to be slower than the others.
So it depends on whether there are real requirements like security or enterprise — or whether it’s technical teams adding their two cents.
James Kaplan: A head of infrastructure — one of the guys used to describe this as the blue cable problem. An application development team would say: build me a server that looks like this, and the cable should be blue. The cables cannot be red — I insist on blue cables, right?
Harqs Singh: James, your analogy around using color — we’ve seen requirements around using specific colors — but I think you have to go back to: does this fundamentally change the product? A lot of the time it doesn’t. It’s preference.
Some people want greater ceiling height; others want it to be within a shell of a building — all things that are doable. I think the most important things are speed and revenue. A lot of what we’ve done in technology comes back to revenue growth, helping the business expand to where it needs to be.
That’s the bit we should never lose: we’re building infrastructure, but it’s to help drive business value. That’s where you and your teams have been focused fantastically — what’s the intersection between technology, data, AI, and revenue, and how does it create real value? That’s the part I try to focus on in my day-to-day as CTO of the organization. We always need to come back to: how do we create value?
On that basis, one of the most common questions I get asked is: what is your standard reference design optimized for? People say — oh, is it the size of the generator? Is it the size of a panel? I say: look, if you look at the TCO of a data center, most of the cost is in the chips. Seventy-five percent of the cost is in the chips. So we’re optimized for that. We’re optimized to get the most out of the chips as possible, because it’s the most expensive resource.
You want to make sure it’s idling for the least amount of time — to the point about making sure your network connectivity is right — and that we deploy SuperPOD after SuperPOD, because that’s how you get the most efficiency out of the most expensive resource and the most value. We’re laser-focused on making sure that intersection between technology, physical, IT, data, and revenue — we’re always focused on optimizing for revenue and business value.
James Kaplan: When you have discussions with customers around customization — to what extent are you involved in dialogues like: if you want the off-the-rack version you can have it in these months, but if you want the blue cable — to use a metaphor — that’s these months? To what extent are you involved in dialogue around that tension between speed and bespoke configuration?
Harqs Singh: That’s exactly why we have a standard to start with.
The standard off-the-shelf is aligned with the latest chip technology and can be deployed much more quickly. It’ll save you three months by taking off-the-shelf. If you want the yellow cable or the red cable — whichever specification you have — that will add a number of months. If you want to do that, we can do that too. But it’s the tradeoff of time.
What we’ve found is: those that don’t have a large technical team will go off-the-shelf — it makes sense. Those with the technical team will say: no, we have standards here; we need to align with them — blue cable, please. And that creates that time.
A lot of the opportunity is the gap between those organizations — those that can run really quickly, especially when you think about there is this supply-demand imbalance in early 2027. Those are the people that will take advantage of that supply-demand imbalance.
James Kaplan: Let me ask one final question — I realize we’re coming up on time. At one point when we were chatting previously, you talked about, for want of a term, a generational waterfall for data centers — that some things for a year or two, or even less, might be on the hardest training problems and then would get repurposed to inferencing problems.
I’m sure I’m not describing this correctly — but I was wondering if you could talk a little bit about how you think about that.
Harqs Singh: Absolutely. If you think about the latest chips that are coming out, it’s really interesting to see how they’re doing in some of the benchmarks now. When we spoke last time, James, there was this piece where you put the n chip on training and then n minus one you look to reuse for other capabilities that are generating some revenue.
I can see right now, from some of the demand-supply imbalances in the whole market, even some of the older chips that have been in the market for four or five years are still quite valuable in terms of the GPU cloud prices they’re getting. So there is this sort of short-term skew right now because capacity is not coming online quick enough, essentially.
You need to think: are we building an upgradable design? If you deploy GPUs today and you get three, four, five years’ worth of value on them on a contract with an offtaker, then in five years’ time, how do you change the design such that you can deploy the latest chip from the industry — from NVIDIA, essentially?
What we’re seeing is: you need to update some of the power plant from AC power to DC power. You need to upgrade some of the densities they’re going to be operating at with liquid cooling. So we’ve designed that in from the start.
What that enables you to do is have flexibility — and also be able to create this into a long-term sustainable asset. Flexibility is really important because if you get out to five years, you’re then trying to guess for that chip you deployed five years ago: what’s the market value of that, and what’s the revenue potential of it?
I think at that point everyone is optimizing for flexibility because nobody knows what’s going to happen — as many options as possible. To your point, part of that option could be: you relocate the chip to inference. It could be: wait a minute — I heard a story just two weeks ago where a client signed up to a five-year deal; after five years, for the same chips, they signed up to a 20% price increase because the demand-supply imbalance was so bad.
James Kaplan: That’s a new thing — at least in my recollection.
Harqs Singh: Absolutely. What we’re trying to do is give you as many flexible options as possible — whether it’s relocating it; whether it’s saying that chip is no longer relevant because there is a demand-supply imbalance and you need to upgrade to the next chip; or whether there is real revenue and market value to that chip even if it’s five years old and fully depreciated. Even if it’s fully depreciated, James — obviously from an economic perspective, you don’t need to get the return you had before because it’s a fully depreciated asset, and that gives you some flexibility as well.
We’re fully focused on giving you full flexibility and options so you can maximize revenue at the time based on what the market is doing. And that’s the bit we don’t know: what the market is going to do.
James Kaplan: Fantastic. Thank you so much. I hope you found this conversation as enjoyable and as interesting as I have.
Harqs Singh: I have. Thank you for having me, James.
James Kaplan: Thank you so much. It was great.
