# Power, Tokens, Talent: Inside the $7 Trillion AI Compute Race

**Published:** September 17, 2026

![A panel discussion with five speakers on stage at the RAISE Summit, with an audience in the foreground.](https://cdn.sanity.io/images/ult5g8gw/production/052030420759e46f2dd59c6e5ce8c89d93d9c30a-960x540.jpg)

**Watch:** https://fast.wistia.net/embed/iframe/v1o3jte4o9

Industry leaders examine power scarcity, tokenomics, and a massive AI infrastructure buildout over the next few years.

## Speakers

- **Val Bercovici** — WEKA
- **Carmen Li** — Silicon Data, Compute Exchange
- **Parimal Pandya** — Akamai
- **Sam Tabar** — WhiteFiber
- **Piotr Tomasik** — TensorWave
- **Robin Wauters** — Moderator, Tech.eu

## Transcript

### Why $7 trillion will be spent on AI infrastructure by 2030

**Robin Wauters:** Good afternoon everyone and welcome to this panel. Thank you for joining. Also people on stage, of course, thank you for joining. From all of the sessions that I’ve seen so far at [RAISE Summit](https://www.raisesummit.com/), this topic seems to be a very, very timely and relevant one. It’s come up in almost every conversation I’ve had at this event, so I’m looking forward to the discussion, also with this great group of people. They're all participants in this ever-expanding and accelerating AI infrastructure race from different perspectives, which makes it interesting.

We all know that this race is very much ongoing and still in its early stages. There’s a [McKinsey report](https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-7-trillion-dollar-data-center-build-out-how-industrials-can-capture-their-share) that says about $7 trillion is going to be spent on this AI infrastructure race by 2030. That’s less than four years away. It’s a mind-boggling number, and it’s more than just data centers and real estate. It’s about power, it’s about electricity, it’s about chips, it’s about everything really. And these conversations are happening in boardrooms. They’re no longer part of the IT department only. So that makes it very interesting and very lively, I would say, this race.

Rather than everyone introducing themselves, which might take a bit too much time, I’ll just kick off with the first question and maybe you can weave what you do and who you are into some of your answers.

First question is rather simple, really. The title of this session is “Compute as Capital: Redefining AI's New Currency,” which for me begs the question: What was the old currency? What was it before? And do you agree that compute is the new currency in this race, and where do you sit on the whole stack?

### [01:48] Is compute the new currency of the AI economy?

**Sam Tabar:** Hi, I’m Sam Tabar, I’m co-founder and CEO of WhiteFiber. Just taking a step back, I think where we’re at with respect to artificial intelligence is basically that this is the third catalyst of the industrial revolution. The first catalyst is electricity. When electricity was invented, the industrial revolution went off to the races. And the second catalyst was the internet. The internet transformed capitalism. It transformed the industrial revolution. And the third catalyst that we’re experiencing now, this new era, is artificial intelligence.

So compute is indeed the new currency. Just as the internet and electricity were in many ways their own asset classes and their own currencies. I don’t think the financial dynamics and systems and the financial markets have really caught up to how compute is an asset class in itself, and that it’s not behaving in ways that are normal.

One example is that H100s today are in fact higher in value than when they first came out. That’s not normal. There’s normally a three-year depreciation rate on equipment that goes down to almost zero after three years. So we are in a new era. Financial dynamics are not behaving as they usually do. And I think that compute is obviously a new asset class that is yet to be understood.

**Robin:** Great. Anyone else want to chime in?

**Parimal Pandya:** First of all, thanks for having me. I’m with Akamai, and I think many of you might know Akamai’s history over the years, but more recently we’re front and center in the distributed inference computing business, and hence the question is very relevant.

I 100% agree, and I say this based on conversations I’m having with customers, prospects. There is this very keen sense of awareness that compute is a currency, in that it allows you to acquire talent. It allows you to acquire time to market, to acquire speed. And if you have it versus somebody else who doesn’t have it, it can mean a world of difference in terms of your ability to compete and stay competitive in the long term. So absolutely, it is undoubtedly one of the most critical resources today, whether you call it currency or what have you.

**Robin:** We don’t necessarily have to go one by one, so if anyone else wants to chime in, feel free to jump.

**Val Bercovici:** I’ll jump ahead, and Carmen, I know, has some interesting thoughts here. Val Bercovici from WEKA. I would argue that much like the real world, there’s more than one currency. Some are more liquid than others. Energy for sure — gigawatts, megawatts if you’re finer grain, they are a huge unit of energy in this industry. Compute, very much so.

For me, if you have gems and coins as early currencies in an economy — and the U.S. dollar, for example, is a highly liquid currency today — tokens are the U.S. dollar. I think they’re the ultimate currency in our industry, and they’re probably one of the best-measured currencies because it’s so transparent, so abundant, there’s such a diversity and liquidity of offerings. And it’s something I know Carmen has some pretty interesting things to say about as well.

**Carmen Li:** Way to set up for me. Thanks, Val. Carmen Li, CEO of Silicon Data and Compute Exchange. Silicon Data is an index provider for compute-related indices. We are launching futures and options at [CME Group](https://www.cmegroup.com/) this coming fall, so you can hedge your compute volatility just like you hedge through WTI or Brent (crude oil providers). Separately, Compute Exchange is a compute marketplace. So we love our neocloud partners here; we help them drive a lot of transactions through a liquid and open platform. And then we are doing token forwards as well.

Back to the size of compute. If you think about the scale today — even today, which we can argue is the infancy of the stage — we're talking about $300 billion of capex. Then you think about the financial sector, where derivatives are usually 10x the spot market, so we're looking at $3 trillion of derivative trading. That’s just a starting point.

The way I look at it — I used to be an oil trader — back then we considered the petrodollar system. Everything was backed around the petrodollar. Now we’re looking at the compute dollar. Everything around compute and tokens. And again, tokens, just like different currencies, they have different values. How do you normalize across different GPUs, and tokens can be a very interesting conversation.

**Robin:** Yep. Last but not least.

**Piotr Tomasik:** Thank you. I'll be quick. I completely agree with the gentlemen at the end. By the way, Piotr Tomasik, co-founder and president of TensorWave, an AMD-only neocloud at scale, scaling from, let’s call it 50 megawatts this year to gigawatts by the end of 2027.

But I wanted to double-click on the first gentleman’s comment. I think this is not behaving normally because we don’t know how to comprehend exponential growth as humans, and the demand is unbound. We need to put about $12 billion to work across equity and debt in the next 12 to 18 months. Really bullish, and we have a third industrial revolution on our hands.

### [07:19] What is the real bottleneck now that GPUs are not the constraint?

**Robin:** I want to stay on this point of the price and the value of these H100 chips, that seems to stay or even increase, which brings me to scarcity — the real bottlenecks. I remember not so long ago GPUs were the bottleneck. Now it seems to be everything: power, space, storage, access to capital. So that begs the question: If I would give you 10,000 H100 chips today, what would still keep you awake at night if that was not the bottleneck? Where’s the real scarcity for you in your organization?

**Piotr:** In my case, you’d give me 10,000 MI300Xs. But the next thing that would keep me up at night is access to good electricity, grid power. It’s getting more scarce for many different reasons. In the United States, we’re having protests, people that don’t believe data centers are helpful to their communities. And just simply finding partners that actually understand how to stand up data centers. There are people that think they can see a power line from their land and that means they have powered land. I’m not kidding you, we get calls like this every single day. So getting real access to power, I think, is the thing that I would say.

**Carmen:** So 10,000 — I'm a little greedy here — I don't think 10,000 will make me a happy person. I think a minimum of 50,000, and then B300s to be transparent. So I'm a little greedier than my friend here.

And then even with that, you can say, OK, so now I have the assets, which is amazing. How do I think about the depreciation curve? Obviously we know H100 resale value has been going up 20% year to date, which is great. But how do I know the future? I don't know the future. How do I protect myself against that? Insurance, residual value calculation, depreciation — those things. That’ll be very interesting.

**Val:** I love the premise of the question. I can’t call it easy, but it’s straightforward to be able to get funding for 10,000 H100 equivalents. It doesn’t have to be H100s. However, what would keep me up at night is that’s becoming a commoditized market. So yes, backers will see that I'm doing multi-year contracts for those GPUs and will back me on that.

But I take a look at some really good market observations. There’s a Stanford course now, and there was a session by Altimeter Capital, that showed that for a gigawatt of energy annually, you can rent that gigawatt for $15 million of GPU hours, or $30 million for tokens — same gigawatt. What would keep me up at night is making sure that there are continued guaranteed contracts for that token consumption, because the perception, especially amongst investors, is that that’s a spot market. [OpenRouter](https://openrouter.ai/), for all of its benefits, also kind of gives that perception. A lot of leading developers on X are talking about flipping models all the time. But it’s not.

And futures contracts for tokens are a great way to show that you can collateralize actual reserve token revenue, convince backers there’s an even better business than renting GPUs, believe it or not. And what would keep me up at night is convincing backers that that’s a real thing. And you’re going to help a lot with that.

**Parimal:** Short answer: power. I think it is the linchpin. Without power, you can have everything and can't do anything with it. So that’s the starting point. And then I know people who have access to power but don’t have the talent or the know-how to actually build and deploy, and access to the supply chain and the sophistication to build large-scale infrastructure with reliability, with stability. And I think that know-how, you can’t just get it overnight. It comes over decades. Those are the things that I think people should worry about.

### [11:28] The three ingredients of a working AI data center business

**Sam:** So the question was, if you received thousands of GPUs, what would keep you up at night? What’s the challenge?

Well, there are three ingredients to a successful business model when it comes to this business. It's GPUs — that’s the first ingredient. There’s power, which a lot of people talked about. And the gentleman to my left also mentioned talent. Those are the three ingredients. If you’re missing one of those ingredients, you have a hole in your business model.

So we solve the first one, the first ingredient, by you giving me thousands of GPUs. Thank you. The second ingredient is power. And so what’s very important is, if you don’t have access to power, you can’t use your GPUs. What we’ve done is — we have four data centers across North America — we’ve secured power in a very, very special way. We’ve taken over old factories that were left for dead, left abandoned, and we converted them into Tier 3 data centers.

So for example, in Canada, we took over what was Canada’s largest mattress factory. We turned that into an AI campus, and we did that for Cerebras. In North Carolina, we took a one-million-square-foot textile factory. We turned that into a Tier 3 data center. That was for Nscale, an $865 million contract. I can mention all this because we’re a public company and this is all public information.

The comment about talent is very important. It’s true that it takes decades to make this work. We were once upon a time a Bitcoin miner, and when you build a Bitcoin mining facility, that’s like building a Hyundai. But when you’re building an AI campus, that’s building a Lamborghini. Two very different cars. It’s a car, but it’s a very different skill set.

So we knew what our limitations were, and we acquired for $48 million a few years ago a team in Canada that’s been building data centers for hyperscalers like Amazon and Microsoft. And they’ve been doing this for 20 years, well before the space got very sexy. We also acquired in that transaction a data center, but we actually didn’t care about that data center. We cared about their skill set, because it does take decades to do this. Otherwise, you'll be making really expensive mistakes.

And so our moat is speed to market. We can build a data center in six months, unlike most other people, which takes about two years. Why? Because we have this retrofit approach of taking over facilities and leveraging the skill set that we acquired a couple of years ago with this particular team.

But one thing that's really important with respect to power that we were able to solve today, and it’s about to be announced, is we invented a technology that will be announced in a few minutes, where we can actually use two disparate locations and turn that into a virtual supercluster. Normally the constraint was bandwidth of one particular site.

The technology that we’re announcing today is that we’re able to do a virtual supercluster amongst two different sites using dark fiber that’s within 83 kilometers of site A and site B. I think that will transform the industry. We’ll be talking about it a lot next year, and that’s an invention that solves one of the bandwidth constraints of basically relying on just one particular site.

### [15:07] Are we in an AI bubble or a structural supply shortage?

**Robin:** Great insights. Power, talent, of course people complaining, neighbors complaining when data centers get built. Great insights, really appreciate it.

If you come to RAISE Summit three years from now, what will we be laughing at? Meaning, what is an assumption that’s really alive in the industry today that you know not to be true?

**Sam:** I’ll be very quick. Everybody thinks that we’re in an AI bubble, which — I’m sure everybody else here is going to agree — we are not. This is not a cyclical situation. This is a structural situation that's going to take a very long time to solve. And so the demand will far outstrip supply for many years to come. We are not in an AI bubble. Anybody who is saying that today, you’ll be laughing about it three years from now.

**Robin:** Does anyone believe we’re in an AI bubble? Please go first.

**Parimal:** That’s a very long conversation, which could consume the 20 minutes we have left. But I’ll say, I don’t know what we’re going to be laughing at. I don’t know. What I know is what I see today, what I see down the road in the near future. And I take the world as it comes, and I try to optimize for what I can and how much we can in the future.

And I’ll just say this: Every assumption that I’ve made, or everything that I’ve said I learned today, has proven to be untrue three to six months later. I’ve had to learn new things all over again. The dynamics change, the game changes, the equations and inputs change. So I just can’t be presumptuous about anything. I just have to make sure I’m making the best decisions in the moment with what I have in front of me.

In so far as, is there a demand bubble or a bubble of any sorts? I will say this: Maybe there’s a little bit of overconsumption going on right now. Maybe that’ll settle down. But I think over time, once the ROI and the business cases get proven and wider adoption takes place, I think it's here to stay.

**Robin:** All right, Carmen. What do you know or realize that nobody else does?

**Carmen:** Oh boy. Where — can you give me an easier question? Obviously as the index provider I cannot give any forward guidance, but what we see in the data, which we have 10 structured products on, is you look at longer-term contracts, the 36-month contract versus on demand. If you think about this, the longer the term you sign, the cheaper it should be per GPU per hour. If the curve is backward-dated, meaning you sign a long-term contract but you’re paying a premium to the shorter-term contract, you can argue the supply-demand curve is distorted. So people are paying a premium to have the security of GPUs versus saying, “Hey, I’m entitled to a discount.”

We were in that backward-dated situation for the last, I would say, three months, starting January all the way to April. Now our longer tail is flattened, meaning people are paying the same prices for a three-year contract as for one year. It tells you the market is coming to a better equilibrium, at least for the time being.

And then back to the bubble conversation. I think it’s healthy for any market to have people shorting the market, going long in the market. That’s what keeps prices at a reasonable range. If everyone’s long, then we will enter a bubble phase. I think it’s very healthy that people are challenging each other’s valuation, each other’s expectations and any sort of assumptions you build in. It’s very healthy. There’s nothing wrong with that. It doesn’t signal that a bubble has formed or that we have a bubble. Obviously overvaluation happens in every other sector, every other stock. That’s completely normal.

The second thing I want to point out is our LLM token indices are very widely cited everywhere at this point. So one of the things I want to point out is our token indices are a per-unit token cost. Just by it going down doesn’t translate to people consuming less. It just translates to per-unit cost coming down. In terms of the market’s overall consumption, everyone has their own opinions. So that’s one thing I want to point out: I don’t want my indices to become this bull-versus-bear sort of calling. It’s really meant to be a per-unit cost measure.

**Piotr:** I think an easy one for me that we’re all going to be laughing about is financiers believing that there’s no residual value to GPUs, and that we have to underwrite everything to these short time periods that are coterminous with our long-term contracts. I think all of us who are able to deploy GPUs in this time period are going to have an incredible money-printing machine on the other end of these contracts that are fully paid for and have plenty of residual value left.

**Carmen:** 100%. Can I add on to that? Because we do refurbished sales too, and the price for our refurbished units has been going up. You can argue, OK, if you have a car this year, then at the end of the year it depreciates in value. But H100s — sorry for the examples — actually appreciated by 20% this past year, which is very remarkable.

### [20:28] Why tokenmaxxing went from badge of honor to badge of shame

**Robin:** But these economics that are being changed, is that a temporary thing? Is it long-term? Is it just because we're in a massive boom that's accelerating and it's going to flatten out? Or do you think this is just going to continue like this?

**Val:** To tie it to the previous question, we’re still at a low phase of maturity as an industry. Meaning that “[tokenmaxxing](https://www.weka.io/video/how-to-get-signalmaxxing-out-of-tokenmaxxing),” if you remember that term, which was very popular just a few months ago, we were talking about it a lot. It was a badge of honor to be at the top of a tokenmaxxing leaderboard a few months ago. Today it’s a badge of shame, because it shows you’re being very inefficient with your token spend. A lot of maturity is beginning to come into the industry, where you have to be much more efficient, and that drives down the rest of the capital requirements and economics, or right-sizes them.

What we’ll be laughing about, I think, two or three years from now, is that today there are a lot of toy makers that release teddy bears and want to make them intelligent, interactive, so they just literally put a link to Gemini or ChatGPT in the teddy bear, as long as it’s Wi-Fi connected, I guess. And your children can talk to the teddy bear about anything they want. Not exactly appropriate.

So we need much more fine-grained models. We need much more domain-specific models in the industry, and they will exist, certainly, just out of economic necessity, out of practical safety necessity, within two to three years. And we’re going to laugh at the fact that in 2026 we were tokenmaxxing and releasing these general-purpose models everywhere irresponsibly.

### [22:03] Who captures the value in the AI infrastructure race?

**Robin:** So maybe we can call it “computemaxxing.” That’d be funny. Or “powermaxxing.”

Maybe on the back of that, I would ask the question, “Who's going to capture the value the fastest in this race?” Because when we’ve seen infrastructure buildouts in the past, whether it’s railways or the internet, it wasn’t always the most obvious companies that captured most of the value. You can only see that play out over the longer term, looking back. But now that we’re in the middle of this race, how do you see this playing out? Who’s going to capture the value? Is it the model makers? Is it the infrastructure? Is it the cloud providers? Is it the power grids?

**Val:** I mean, it’s rapidly shifting over time. Just a few years ago the value was all in the cloud and in the SaaS layer of the apps, and it’s pretty much inverted completely right now. The value is entirely in the scarcity of the supply chain. That’s not sustainable, but it’s going to last a few years. And probably, again, as the industry matures, right around the turn of the decade — 2030, 2031 — we should see the value go up the stack again. In theory, some of these fabrication providers will actually capitalize and build fabs and give us more DRAM and more storage and more GPUs, and a lot more CPUs as well. But the opportunity is in the near term. Where is the value today, not three years from now? And it’s inverted. It’s in the supply chain and it’s in the infrastructure.

**Robin:** Sam, would love to get your take on this as well.

**Sam:** I remember when the internet was invented, nobody could ever imagine the permutations of what would come out of that, like Uber. In 1995, when I remember I was using email and there were very rinkydink websites, I would never have imagined Uber, or dating apps like Tinder, or Instagram. Who would have imagined all of that stuff?

I think it’s extremely difficult to imagine who’s going to really capture the value in the medium term and long term, because we just literally started this. And frankly, I would say the day we started this — although AI and machine learning have been going on for many years — it’s really when everybody in this room used ChatGPT for the first time. That was the “eureka” moment. I am sure all of you had that “what the heck” moment when you were using that. So I would say that’s day one of when this era happened, and what day 10,000 looks like in this era is very difficult for me to predict.

But I will definitely venture to say that power will be the most sought-after commodity in the medium and long term as this era continues.

**Parimal:** Just a couple of quick comments on that. Two things. One is there’s a huge opportunity to find efficiencies, and I think there is value to be unlocked there: efficiencies in how we use the infrastructure, the scarce resources up and down the stack, in the use of the algorithms. Any which way we can find efficiency, I think there’s value to be unlocked.

And two, if you look at the majority of the consumption today, it’s very cornered with a handful of large organizations. The large enterprise user is still testing the waters. They’re still looking to find that use case that’s part of their business process that’ll make some business processes more efficient or create a new revenue stream. And I think once we unlock that, then you see the doubling down, the tripling down by the CFOs into those workloads, into those agents, into those use cases. And whoever can figure out how to get the most boring company in the world to take advantage of these technologies to deliver tangible ROI, there is value to be unlocked there.

So, two broad themes. I can’t give you a simple answer, but I do agree — we don’t know what the Uber is going to be of this wave.

**Robin:** That’s what makes it exciting really. And it’s going to be all of the above. Obviously there’s not one part of the market that’s going to capture all the value. It’s all of the above. But to what degree and how fast, nobody knows. Unless you know, Piotr.

**Piotr:** In the short term, I think the infrastructure layer is going to capture the most. But like everyone else is saying, it’s very hard to predict where everything will end up.

**Carmen:** I think that’s all hedge funds do for a living, figure out alpha for the next five years. So if you believe in capital market efficiency, the fewer competitors you have, the more pricing power you're supposed to have, and the higher margin you can capture.

So think about it: There are many layers of cake, as (NVIDIA CEO) Jensen Huang put it. At the bottom there is only one provider. By theory they can capture everything. But they’re not, and there’s a reason why they’re not, and we can talk about that. Then the second layer, there are a few design houses for chips. By theory they can capture whatever they want to capture. In the model layer, if you’re the best and only provider for a certain modality, and everyone else is 30% to 50% worse than yours, by definition you can capture a lot more value.

So it really depends on how competitive the market is. The way I look at it is, it’s very transparent in terms of pricing. If your pricing is not associated with your cost, then by translation you have pricing power. And we know who has pricing power. It’s pretty transparent in the current world. And if you model it out, it’s pretty transparent for next year as well.

### [27:54] Can compute be traded like oil on a regulated exchange?

**Robin:** I want to stay with you for a bit, because I want to talk about the first part of the session title: “Compute as Capital.” Are we witnessing the emergence of computing as a new financial asset class, something that can be priced and traded, as you mentioned earlier? And how is that going to affect the market?

**Carmen:** One of the things I want to address is that people assume, “Hey, now Wall Street, CME, all the capital markets are in this game, it will for some reason make things worse.” I want to address that.

Number one, you want the banks to be heavily involved. The banks have the cheapest cost of capital. If they can manage the exposure and join the ecosystem, it will only lower the cost of capital for everybody in this room and every client down the street. So it’s good. We want our banks to be involved, because they’re regulated, they need transparency, everything’s centrally cleared.

If you and I do a deal tomorrow OTC (“over the counter”), you'll be worried: “What if Carmen goes out of business tomorrow?” But if you and I do the trade at CME, you don’t care. It’s centrally cleared. There’s no longer counterparty risk. You want things to be regulated in a transparent, clear center. With CME’s involvement, with all the banks’ involvement, it’s good for the industry and for global society.

It's a lot more transparent than anything else, with benchmarking. You understand depreciation value. You understand, “Hey, it’s actually benchmarked.” Your machine is worth this, versus zero, which is everyone’s assumption. So you want that kind of clarity, that industry standard.

**Val:** I would actually double down on what my colleague from Akamai was saying earlier on. Whether it’s compute efficiency or capital efficiency, efficiency is really vital in this particular era of the industry.

And if we just take a look at what we generalize as compute, it’s really GPUs, CPUs, networking, memory, and storage. That is actually all of compute. If you’re able to use first principles and look at it, the majority of the dollars in the industry today are going toward inference, about two-thirds. And inference is actually a memory-bound problem more than anything else. It’s not a GPU or compute-bound problem.

Any efficiency you can introduce into that bottleneck, into the [“memory wall”](https://www.weka.io/learn/ai-ml/ai-memory-wall/) as it's known in the industry, yields tremendous returns. It literally takes a dollar of compute capital and can generate as much as $10 worth of value out of that, by engineering just that portion correctly. So engineering, and the talent pool that has the ability to engineer that, is really vital.

### [30:20] What each company needs for the AI buildout to pay off

**Robin:** Great. Predictably, we’re running out of time very fast, but I want to go one by one and ask you, from your organization’s vantage point, what needs to happen in this race, in this infrastructure buildout, for you to be Uber-successful?

**Piotr:** What needs to happen for us is the lowest possible cost of capital that we can get. Thankfully, banks are stepping up in our particular area and are offering very small spreads over SOFR (Secured Overnight Financing Rate), as long as we get the right people to the table. The continued performance of the entire capital markets in that way is the most important thing for TensorWave to succeed.

**Carmen:** I already talked a lot, so this will be short. I think for us to be successful is for everyone else to be successful. We’re an index provider, we are a marketplace, and we want a very vibrant ecosystem with a lot of different providers providing amazing but different solutions for different use cases. So in our way, we want to make sure things are transparent, things are efficient — from our Compute Exchange point of view and from our Silicon Data point of view. We want things to run very efficiently, and then there are fewer bottlenecks we can get.

**Val:** I would say what we need to be successful is already beginning to happen, which is just the transparency and the liquidity of tokens. When you can go on markets like OpenRouter and very transparently see, for any model in the world, dozens of providers each with different price points, different performance points, different locations of availability, different uptime — all that transparency is wonderful for us, because it creates this very open, liquid marketplace. And again, the best-engineered solution ultimately wins in an open, transparent, capital- and token-efficient marketplace.

So I just hope that this trend continues. I don’t see why it wouldn’t. This genie’s out of the bottle, and I think it’ll drive this era of the industry really, really well until it gets boring, and then it’s going to get even more lucrative at that point.

**Parimal:** My take is simple. When our customers succeed, I think we’ll automatically succeed. And we’ve seen that over Akamai’s 28 years of history. What I mean by that is, we have deep, long relationships with every major organization in every vertical in every country across the globe. And when those large enterprises can find viable use cases — which we’re beginning to see happen — where the ROI can be proven repeatably, scalably, I think that’s a sure-shot sign that we will all succeed.

For Akamai, definitely, we see a big opportunity with that. In the meanwhile, of course, the trend that’s driving the consumption right now will continue to perpetuate. But I think when that happens, I know we’re in the right space, in the right spot. We have the relationships, the ability to deliver at global scale, which I think is truly unique.

**Robin:** Sam, last but not least.

**Sam:** The two things I’d like to see that are markers of success: That our reputation for speed to market, and building data centers in six months instead of two years, is beginning to grow, which is why we’re getting clients who have booths here at this conference. And this invention that has just been announced on the internet a few minutes ago, the commercialization of that product is something that I really want to see happen. That'll be a big marker of success.

But with respect to the industry’s success, I think that’s already happening. The industry is a successful industry. It’ll continue to be successful. There will be gyrations, cyclical gyrations. To your point, there are going to be companies that are overvalued and some that are undervalued. There’s going to be some of that gyration going on, but structurally I do see upside for a long time.

### [34:26] What could actually stop the AI infrastructure train?

**Robin:** We actually do have some time left, so I’ll flip the question. Is there anything that can stop this train? Is there something that could really happen that takes the wind completely out of your sails at your organization?

**Val:** I think there are two important headlines impacting the industry right now. One was obviously the Mythos/Fable moment in governments — U.S. government at least — interjecting very forcibly around what is a cap in terms of unregulated models. And we’ve hit it. So it looks like we’re in this other era, this new era right now, of all the powerful, leading-edge models, which will be open source in about three months is my estimate — and I’m usually wrong about this, exponential progress being what it is. So we’re seeing a new ceiling in terms of what open free models — well, open available models, I should say — are, whether closed or open.

And so we’re going to figure out, is this just a one-way street? There are headlines that the Chinese government now wants to restrict access to their powerful models as export controls outside of China. That’s going to dramatically impact a lot of us on stage here. It won’t fundamentally alter our trajectories, but there will be a readjustment period in terms of what’s going to fill that void of open models, for example. So just really interesting headlines in the industry right now.

**Sam:** There’s one headline that will affect everybody here in the industry if this trend continues, which is that data centers have now become the new boogeyman. There are some moratoriums now that are being announced on data center development. There are protests and so on.

I’ll even say, at WhiteFiber, when we were developing our North Carolina site, we had to have a community day where we invited the community and we explained that we’re actually using 85% less water than our previous tenant, that there’s going to be less noise than our previous tenant. That’s one of the benefits of a retrofit model; we're using a site that already exists.

But people who are building data centers from scratch, there’s going to be a lot of community resistance when people see this massive facility being built where there was just greenery initially, and the optics of that are very negative. And if this movement against data centers ends up snowballing, that means it’s going to hurt the power availability for AI, and that will hurt the entire industry. So I think this trend against data centers can be quite dangerous for the industry, and I hope we’re able to distinguish good actors from bad actors.

**Robin:** Piotr, you’ve actually alluded to this as well earlier, so I’m guessing you concur.

**Piotr:** We had to get through a moratorium in a place in the United States a couple of weeks ago. It was like a movie scene, packed door-to-door with people in city hall. Thankfully, we were able to overcome that, but it’s going to get harder and harder.

Data centers are the physical manifestation of people’s fears around AI. And the frontier model companies did not do a good job, from the ChatGPT moment to now, saying that there’s a place for people in the future. Now they’re wanting to go public, so they’re changing their tune — and I think rightfully so, because there are going to be jobs in the future, just like every other revolution changed jobs and industry and society.

**Carmen:** I think it’s interesting, we have different pushbacks. Again, that’s against the proliferation of the industry, from the GPU level to the data center level to the LLM model level. So think about it: If we all start building in the United States and people scale up on my platform, all I’m going to route them to is European data centers, or Southeast Asian data centers, or South American data centers — which is fine by me, compute-wise. But at the same time, are we really doing ourselves a service or a disservice?

On the model layer as well, we want to see proliferation. Within tokens, the price will be the ultimate indicator of how much you want to pay. If you stop the models — a sector of models — will that really help your economy, or hurt other sectors that you’re trying to serve? And at the end, AI has no rights, humans do. How do we figure out enabling humans with AI? That’s the most important topic.

**Parimal:** Short answer: I think we’ve talked a lot about this scarcity of all the different resources, and that’s definitely in the near term a big cause of concern, and one that could trip the whole thing. And then long term, I think just how the industry evolves, how adoption evolves, and our ability to execute.

**Robin:** Great. Well, that’s all the time we have. Perfectly timed. As you could tell, five experts on stage and then me — but I learned a bunch, different perspectives as well. So I really appreciate that. Thank you for sharing your insights.

## Related Videos

- [Why AI Inference Margins Are Suddenly Getting Better](/video/why-ai-inference-margins-are-suddenly-getting-better)
- [Inference Is Eating Memory, and Tokens Now Run AI Economics](/video/inference-is-eating-memory-and-tokens-now-run-ai-economics)
- [Build for What's Next in AI with NeuralMesh](/video/build-for-what-s-next-in-ai-with-neuralmesh)