# "If You Dare Borrow, They Dare Lend": Discussing Data Center Financing, Power, Computing, and the...

硅谷101 · 2026-11-05 · 42 min · https://www.youtube.com/watch?v=xzc99JOBMlY

## Transcript

Christina Xu
Wow, we almost got a full house. If you’re here at 9:40 on a Sunday morning, I guess you either really love data centers or you smell money. Either way, you’re in the right room, and we’ve had a great track record with that.

Yixing Xu and I were on the same panel last year talking about data centers. Our conclusion came down to one line. I don’t know if you still remember: data centers are the new oil, right? If you believed in that and were bullish on data centers and new cloud over the past year, you’ve probably already doubled your money in the stock market. You’re welcome.

This year, we’ll make sure the panel is as insightful as last year, and even spicier, as Qian required. So, let’s start with some introductions.

Yixing Xu
Sure. Hi, everyone. My name is Yixing Xu, and I’m currently working at Google, in charge of Google’s data center planning and energy strategy. Previously, I worked at Microsoft and ByteDance, also in charge of similar energy strategies to support the very fast growth of data centers globally.

Before that, I worked at Breakthrough Energy, focusing on the clean energy transition. I have some experience in data centers and energy, and I’ve been supporting these tech companies as they grow very fast. They’re building gigawatt-scale data centers worldwide. Nice to meet everyone.

Bruce Liu
Thank you. Thank you for having us here, and thank you for getting up so early. Hopefully, this panel pays off in the end. We’re going to discuss a lot of interesting stuff. Just enjoy.

Christina Xu
Alex, let’s start with some introductions. I have lots of questions for you.

Alex Yeh
Sure.

Christina Xu
Yeah.

Alex Yeh
Yeah. My name’s Alex, founder and CEO of GMI Cloud. GMI stands for General Machine Intelligence Cloud. I believe we’re moving into the general machine intelligence age. I think this is inevitable.

My mission, and the company’s mission, is to accelerate adoption for everyone—not just frontier labs or hyperscalers, but everyone who wants to use AI. We build foundational GPU infrastructure across the world, from Asia to the US, as well as a Kubernetes management layer for people who don’t want to manage bare metal. We also have a token factory, which is processing over a trillion tokens per day.

Christina Xu
Okay, Yixing, I’ll start with you. Remember last year you mentioned the bottleneck for data centers was energy, especially transmission, then generation. What do you think about the past year? Is it still the same? What changed?

Yixing Xu
Yeah, that’s a great question. Last year, we talked about how the United States didn’t have enough electric power to support the very fast, large-scale growth of data centers. The main bottleneck is the transmission system. You can think of it as the high-voltage transmission lines you see along highways, delivering electricity from generation all the way to the data center.

In the past year, a lot of people realized a very cruel fact: it’s almost impossible to build a large-scale transmission system in the United States. So almost everyone gave up on solving that problem. They pivoted to another one: if we can’t build transmission, which is like a highway, how about building the generation directly at the data center campus?

I think that’s why, in the past 12 months, a lot of people have picked so-called on-site generation, or behind-the-meter generation, as the solution to the data center electricity problem. Some of the data I’ve seen suggests that roughly 30% to 50% of this year’s planned data centers—meaning those that may come online in one or two years—will use behind-the-meter generation.

That means a lot of natural gas turbines: large, medium, and small, especially small turbines. General Electric’s large gas turbines have already been sold out for many years, so the small ones have become hot products in the market. And if you follow the stock market, there’s a company called Bloom Energy that’s become very hot this year, mainly—

Christina Xu
Make sure you take notes.

Yixing Xu
Yeah, mainly because they provide speed to power, meaning they can help you get on-site generation online within several months, sometimes within three to six months. They sometimes claim they can do it within 60 days. That’s become a very hot behind-the-meter solution.

Of course, they have other challenges, like supply chain challenges and whether their technology is mature enough for all data centers. Right now, they’re still small-scale, so there are risks associated with that. However, companies are seriously thinking about this solution, and some of them think it could be one of the big solutions for data centers.

Again, remain cautious. This is still a new technology, but it’s one of the things people are thinking about as a way to solve the energy problem for data centers. I’d say the big shift this year is that—

Bruce Liu
People gave up on solving the transmission problem and are now focusing on solving the on-site, or behind-the-meter, generation problem.

Christina Xu
Okay. On Bloom Energy, I know lots of new clouds use this technology. Is it the same for hyperscalers?

Bruce Liu
Yeah, that’s true.

Christina Xu
Okay. Bruce, I’m sure you invest in different parts of AI data center energy. Where do you think the bottleneck is, and which part worries you the most?

Bruce Liu
Well, as a matter of fact, the whole AI supply chain, including energy, is in significant shortage. Power for data centers is a big one. And it’s not only about the equipment. In the US, I agree with Ethan: the infrastructure wasn’t that good before this AI build-out boom. But it also has its own advantages. Natural gas is one.

I fully agree that behind the meter is the solution, and it just takes time to build out. You need equipment from companies like GE Vernova and Siemens Energy. You also need physical things: construction, labor, permits, all those things. Those are all the bottlenecks we have to deal with and solve in the coming midterms.

States like Texas, Virginia, and Ohio are all facing these challenges. Do we have the permits? Do we have local community support to build out all those ambitious data centers? Those are the things we have to pay attention to going forward.

I remain very optimistic, because in the end, it’s all about trade-offs. When you build something new, it’s going to have both positive and negative impacts on the local communities. All the hyperscalers and new clouds have to prove they can make it a win-win, rather than just, “Oh, I take everything.” I’d say the direction of travel is the right one. To sum up, those are the real challenges we have to face, but people have solutions. That’s where we are, so we stay invested.

Christina Xu
Okay. I can tell we’re all very AI-pilled, positive, and bullish. Just one quick gut check: if you had to pick one thing that’s most short right now, is it GPUs, networking, memory, or power? What would you pick?

Bruce Liu
Okay. Like I said, everything is in huge shortage, just to different extents. GPUs, memory, and even the equipment further down the chain, substrates, and all those things. Even MLCCs—Taiwanese companies are benefiting a lot from those shortages. All those physical things just take time to build out.

If we have to pick one direction, given the urgency, people are now thinking a lot about CPU compute, right? The GPU-to-CPU ratio used to be one-to-one. Now it’s four-to-one. Some people are even bullish on eight-to-one or something like that. That’s where you see huge potential.

But the point I want to make is that we need to be a little creative—

Christina Xu
Right.

Bruce Liu
—and open-minded. This is a very early stage, and this is a growing-pie story. It’s not about who’s going to win or lose; the whole pie is growing. That’s one aspect I want people to pay attention to.

Second, it’s also about technological advantage. You guys are the experts in this aspect, right? All the semiconductor projects you’ve been doing for the hyperscalers are fundamentally different from the stuff you used to be working on.

Yixing Xu
So that’s also going to create winners and losers. For example, power semiconductors used to be very boring, cyclical businesses. But now GPUs and CPUs are using more power and higher voltages, which makes that area very exciting. We’re looking into it. This is an exciting time. That’s what I want to say. Sometimes you don’t need to decide, “Okay, I have to choose GPU versus CPU, memory versus optics.” I truly believe that, in the grand scheme of things, this is a growing pie story.

Christina Xu
Yeah.

Yixing Xu
Yeah.

Christina Xu
That’s great. Okay, great. I don’t know if you guys heard the news that GMI Cloud just raised $668 million. Congratulations, first of all.

Alex Yeh
Thank you, guys.

Christina Xu
I saw that your contracted ARR also grew almost 10 times in a year. What makes GMI so different?

Alex Yeh
I would say three things. Number one, reliable and on-time compute is the most important thing. We have a lot of software layers built on top, but number one is providing on-time delivery and reliable compute. That means having a very strong supply chain relationship in Asia and expanding globally. So that’s number one: supply chain.

Number two is East and West. As we face a lot of political headwinds in the U.S., even on megabuilds and smaller capacities, we have data center footprints across the U.S., East Asia, South Asia, and Southeast Asia. That provides different locations for companies that want to expand rapidly.

And number three, of course, is financing. We’re able to tap into Asian capital. It used to be yen trade, but we have an NTD trade, which is Taiwanese dollar trade. We’re able to borrow at a much lower cost than our peers in the U.S., who have to borrow at close to 10% or more.

Christina Xu
That’s great. With all the capital raised, what is still the hardest thing to buy? Early this year, you mentioned the link, right? What about right now?

Alex Yeh
Right now, it’s still power. I think power is the most critical issue. If you follow Jensen’s Q2 earnings comments and all the Wall Street discussion, it’s all about power scarcity. This is a global phenomenon, not just in the U.S. Compute with power that’s ready today, or even in 2027, is extremely rare.

Christina Xu
Okay, great. Let’s dive a little deeper into this question. If a company signs up for a megawatt today, this might be for you, Yixing and Alex. How long until that turns into usable compute? Just give us a sense.

Bruce Liu
In the United States, the timeline is normally about 2 years, maybe 2 and a half years for very large-scale data centers. We’re talking about data centers of 500 megawatts or more, sometimes at the gigawatt level. For smaller ones, it could be faster, maybe 1 and a half years. But for large ones, it’s roughly 2 years.

One thing I want to share with the audience is that, based on what I’ve observed over the past half year, there’s a potential huge risk to the whole data center and AI industry right now. I think—

Christina Xu
You mean delays?

Bruce Liu
Yes.

Christina Xu
Okay.

Bruce Liu
I think there are delays in data center construction. A lot of data centers were planned last year, so they weren’t going to be delivered this year; they’re expected to be delivered in 2027 or 2028. A lot of problems haven’t surfaced yet. We’ve noticed that the Wall Street Journal and investment bank reports have also mentioned signs of delays across the board in the United States.

I think this could be a huge risk, and there are a lot of reasons behind it. One is what we discussed: electric power could be an issue, and all types of electrical equipment could be another. There are also all types of regulations, including new regulations that limit the progress of data center construction. One example is the FAA, the U.S. aviation regulator. It has nothing to do with data centers, but somehow it has new regulations limiting data centers built near airports because they think there could be safety issues.

Another thing is what Bruce mentioned: local community opposition is becoming a huge trend in the United States. If you read mainstream media in the United States, like the Wall Street Journal, you’ll notice there are a lot of stories about data centers. They all have the same narrative: that local communities in the United States are against the AI industry or AI data centers. It’s unfortunate that all these stories are framed that way, because I personally don’t believe that’s an objective view. However, that’s the kind of story people click on and read, and it’s not necessarily the truth.

What’s happening now is that, before the November midterm election, almost all politicians in the United States are expressing views against data centers so they can get more votes and be elected. That’s become a trend, even in Texas. Early this year, the Texas governor was fully supporting data centers. But a few months ago, his view changed 180 degrees; he became completely against data centers because of pressure from local communities on those politicians. That could be another huge force causing data center delays.

I think this could be a huge risk. As an example, SpaceX rented out some of its data centers to Anthropic. I think the number is around 300 megawatts of data centers for almost $1 billion a month.

Christina Xu
Yeah, it’s like—

Bruce Liu
Right?

Christina Xu
—three times.

Bruce Liu
Right.

Christina Xu
Triple the price compared to last year.

Bruce Liu
Yeah, that’s a huge price. You can use this number as a reference: 300 megawatts of data center is worth $1 billion per month. What if several gigawatts of data centers are delayed for a year? How much revenue is going to be lost or delayed because of those construction delays? That number could be $100 billion for the whole industry.

We were also talking about financing on this panel. All the money is being put into data center construction, but you won’t get the revenue until maybe a year later, causing a $100 billion revenue delay. That’s a huge blow to the industry. I think this could be a huge risk if it’s not managed well.

Christina Xu
Okay. Alex, you mentioned that it takes about 2 years from when you sign up for power to when you get compute. Is it the same for new clouds, or is that actually your opportunity?

Alex Yeh
I think it depends on the scale you’re talking about. Gigawatt-scale compute is a huge issue, and there’s almost no way the grid can support that scale. It typically requires a microgrid or a behind-the-meter setup. As you mentioned, behind the meter basically means using Bloom, some sort of reciprocating engine, or some sort of turbines. That means burning natural gas, so the most critical issue is actually air permits with the ESA for carbon dioxide emissions.

For anything that isn’t permitted yet and where construction hasn’t started, it’s T plus 14 months, I would say. Even for a smaller-scale modular build, skids probably add another 2 months, more or less. Then add another 2 to 3 months for tokens. So from white space ready to token ready, you’re talking about 16 or 17 months if you haven’t started construction.

I think the situation is very dire. In the past, our customers, including hyperscalers, didn’t want to talk about anything below 100 megawatts. Now I’ve seen people asking for even 20 or 30 megawatts. If you can deliver anything before 2027, that’s awesome. They’ll take everything.

That’s the situation right now in the U.S. In Asia, I think you can get it sooner. There’s just more abundant energy, and it also depends on which locations have more supportive governments and a top-down approach. You can have these hundreds-of-megawatts campuses built relatively quickly, but you still have to find the right pockets.

Christina Xu

Yeah. Yeah, that could be a great strategy. Think about how Nuvias and those companies also go global. I want to follow up on the point you mentioned about delays. How much delay are we expecting?

Yixing Xu

That’s a good question.

Christina Xu

For the general—

Yixing Xu

Yeah.

Christina Xu

—hyperscalers.

Yixing Xu

Yeah. I think there are already some reports and numbers out there, maybe from SemiAnalysis or different mainstream media. Right now, no one really knows the scale of the delays, but based on the data, we can make a rough estimate. Maybe 10% to 20% of data centers could be delayed by more than a year.

Bruce Liu
Can I add one question?

Christina Xu

Yes.

Bruce Liu
Is it a matter of delay or cancellation? That’s what I have trouble understanding. If it’s delayed, maybe it’s fine because the whole supply chain is so tight. But if it’s canceled, that might be a bigger problem. I just want to get some sense from you guys.

Alex Yeh
Yeah. You have to understand that cloud and data centers are a regionalized business. Governor Abbott put a pause on the Texas site, and suddenly our sites in Denver and Atlanta become really hot, sought-after assets. It’s a regional event. This hits pause on the Southwest, and the Northeast becomes popular. Then everyone starts scrambling. It literally moves quarter by quarter.

Yeah.

Alex Yeh
Our house view is to start speculative builds around the world, in East Asia and Southeast Asia, as well as in parts of the U.S. with less data center concentration. Then you can find pockets of opportunity. When the time comes and something happens, you can pick up the projects because we have contacts with hyperscalers and frontier labs. When projects get paused or canceled, opportunities come along, but you have to be ready.

Yixing Xu

Yeah. Just to follow up on Bruce’s question, I think most of them are delayed, not canceled. A small percentage could be canceled, but most are just delayed.

Bruce Liu
Okay. Thank you, guys.

Christina Xu

Nobody wants to cancel right now.

Bruce Liu
Well, I actually have another question.

Yeah. Sure.

Bruce Liu
Following Alex’s answers, from your perspective, is financing the major problem, or is it the supply chain shortage and those kinds of things?

Alex Yeh
I think it’s definitely a huge issue. That’s why you see Broadcom, Nvidia, and Google all stepping up on financing. There’s a huge mismatch in financing, and whoever can solve this can make a lot of money. There are a lot of opportunities.

Data centers typically require leases of 12 to 15 years, but compute contracts are typically 4 to 6 years. There’s clearly a mismatch in contract durations. Data center operators want certainty about the lease, so Nvidia, Broadcom, or Google—the largest investment-grade companies—can come in. The most common question we get from data center operators or power companies is, “Okay, who’s your investment-grade counterparty?” They want the project to be investment-grade wrapped. That’s where the mismatch comes in.

There are solutions. One is a lockbox: you put 5 or 6 years of cash flow into an SPV, and then that gets—

Bruce Liu
—paid to the data center operator first. There are precedents for that. Another option is some sort of residual value guarantee on the data center side. Everyone is trying to be creative right now.

We had that video talking about the AI data center buildout and debt. I’m sure lots of data centers have started tapping public credit. That’s a trillion-dollar market, and AI data centers are only 5% of it, so there’s lots of room for capital to shift from traditional software credit to data centers. As an investor, what worries you about this trillion-dollar credit and debt market?

Bruce Liu
I’ll give the audience some perspective. We need to step back a little and understand the big picture. Ray Dalio always talks about how we’re at the peak of leverage in the public sector globally, which is true. We have to deal with this problem. The entire developed world is so leveraged that it has to deleverage.

There are only a few proven ways to deal with this. Today, we’ll focus on one: growth. As I’ve mentioned many times, that’s one of Scott Bessent’s pitches. We need to grow out of our debt problems, and that’s true. Looking at history, it’s a major way to do it. In the grand scheme of things, this is a historical mission for AI. If it weren’t AI, it would be something else, but fortunately, this has fallen on AI’s shoulders. The whole world needs AI to grow, so all the money is going to come in. That’s something you need to keep in mind.

Second, this cycle is very different. Think about what happened after 2008. The public sector has been levering up. Governments have issued debt and printed money, and they’ve continued to leverage up until today. They can’t do it anymore. But the private sector has been deleveraging since 2008. Its balance sheets—both corporate and household, especially in the U.S.—are relatively clean and healthy.

In the new cycle, that’s going to reverse. The private sector is going to leverage up, anchored by AI. AI gives it an opportunity and an excuse to do so. The public sector has to deleverage. That’s the big picture we have to put ourselves in. For so many years, we’ve been used to asset-light business models and free cash flow. We don’t like companies to leverage up, but that time has passed. We have to get used to it.

When you hear people say that Oracle’s credit spreads are widening, that’s a little crazy, to be honest. But it’s not a collapse. It’s a repricing. Oracle used to be a cash cow and very safe from a credit perspective, so its credit spread was really tight. Now it has to change its business model to stay relevant. That’s a cost it has to pay, which is why its credit spreads are widening. It’s normal. That’s all I’m trying to say. We’re in this process, and this is the first year the private sector has seriously leveraged up. It’s going to continue for the foreseeable future, so we have to get used to it.

Next year, Amazon and even Meta are going to use all their free cash flow, so they’ll have to borrow. But their balance sheets are very healthy, and they’re all investment-grade. It’s still been a fantastic year. If I remember correctly, new issuance year to date is around half a trillion. But there’s still plenty of room. It’s only about 7% of the whole IT market. We haven’t really used the high-yield market yet, not to mention private equity and private credit.

Financials used to be such a dirty word, but for this cycle, you have to research the banking sector. You really need to see whether the banking system is healthy enough and has the capacity to support this kind of leverage. We have limited time today, and it’s a really exciting topic, but the short answer is yes. The U.S. banking system is healthy and has the capacity to support more leverage. It really depends on whether you’re able and willing to borrow. If you dare to borrow, they dare to lend. It’s as simple as that.

Christina Xu

Alex, I’m sure that’s great for you to hear.

Bruce Liu

I’m not saying there are no problems. There will be. Even today, hyperscalers and semis are healthy, but newer cloud companies and some power buildouts aren’t. Their credit spreads are crazy, so they might be the problem. But this is still very early. It really depends on how AI is growing. If revenue comes in better than people expected, you don’t need this kind of leverage.

Bruce Liu

Leverage works as a buffer, right? In a perfect world, if you generated revenue right away, it would cover your CapEx, but that’s not the case. That’s how capitalism works. During the early stage of the whole thing, you need this buffer. The buffer is leverage. That’s all.

Speaker 1

Yeah.

Bruce Liu

That’s all I’m trying to say.

Christina Xu

That’s so great to hear. We’re still very early in this AI era, so there’s no bubble. Well, I promised to make it spicy, so sorry.

Bruce Liu

No, it’s actually a good question, because a lot of people like challenges. At least in the equity space overall, do you know how much—what’s the multiple for QQQ these days? I don’t think you pay attention to that. It’s 20 or 21 times next year’s earnings. It’s cheap relative to the financial conditions, even after this run-up in real yields. It’s just cheap.

Why is it cheap? Because QQQ right now is really a proxy. We did that research, actually. You can literally calculate the secular growth premium, which is all the risk premium excluding the cyclical premium. What’s left is your secular growth premium. QQQ carries around 26%. The S&P carries around 14%. Small caps—that’s the trouble. They don’t benefit as much. It’s only 5%.

So, actually, people, you’ve got to trust the market. The market is very efficient. It’s not dumb at all; it prices everything very rationally. That’s what I want to say.

Christina Xu

Okay, great. Let’s switch gears a little bit and go back to Alex. I’m sure you’re happy to hear that there’s still lots of debt to borrow. In that case, do you just try to borrow as much as possible and expand as soon as possible?

### Cloud Economics and Tokenization

Alex Yeh

Well, yeah. You see 2 companies in the market: Nebius and CoreWeave. CoreWeave has historically emphasized long-term contracts—5 or 6 years—and they’ve done well. Nebius has, I think, around 50% long-term compute and 50% spot, though I don’t know the exact ratio. That’s a more speculative build.

When supply is very constrained, CoreWeave is unable to take advantage of the price increases across compute, while Nebius can take advantage of the situation and charge a much higher premium: I think $25 to $30, or even $50 per watt, while CoreWeave is roughly $15.

We’re taking the same approach. We’re building speculative projects as well as long-term projects so that we remain healthy. On the long-term projects, we can borrow at a much lower cost of capital from the U.S. and Asia. Obviously, we’re not leveraging all the way up; it has to be healthy. We ring-fence all the borrowing within each SPV.

Speaker 1

Mm-hmm.

Alex Yeh

That way, the company can grow at a healthy pace.

Speaker 1

I notice you’ve also started to sell tokens, right?

Alex Yeh

Yeah.

Speaker 1

So GPU hours aren’t enough?

Alex Yeh

Yeah.

Speaker 1

Is it just for the economics, or is there a strategic reason?

Alex Yeh

Look, you’ve got to look at the cloud business in general. What’s the fundamental business of a cloud company? It’s all about utilization rates. Think of a cloud as a massive laptop. If I lend it to you and you use it for 12 hours, you’re not going to use it more than that. You have 12 hours left, so I’m going to rinse and repeat and lend it to another person while you’re asleep. That’s utilization. If I lend it to only one person, utilization might be 50%.

The cloud business is about increasing utilization. Take an empty cup. The long-term compute contracts—the 5- or 6-year contracts—are the pebbles. You put them in a glass and fill it to 50%. Great, you’re cash-flow positive, but you’re not making much margin. Then you start throwing in sand. The sand is short-term contracts: 6 months, 1 year, 2 years, 3 years, to fill that gap.

Then you pour in water to fill up utilization, because you’re going to have hot swaps. We have to fulfill our SLAs, so you might have 5 or 10 spare nodes. What are we going to do? No one’s going to rent those, so we put them on a marketplace. Not many people will absorb all these one-node, two-node, or three-node blocks. We have that platform, by the way. You can also turn them into tokens, because those are smaller increments and chunks.

I think it’s inevitable for neoclouds and inference platforms to consolidate into a vertical stack. The only way for cloud to increase its margin is by cutting GPU time into smaller and smaller chunks. That’s what the token business is about.

The token business is essential for us, because all our compute is already sold out through the end of 2027. But we have these spare hot swaps and compute that we buy externally, and we turn that into tokens. To us, that’s pure margin.

### Predictions for the Year Ahead

Speaker 1

Okay, great. We don’t have much time left, but I want each panelist to make a prediction for next year. Give a prediction and a time. Who wants to start?

Speaker 3

I can start, because I’m a market guy.

Speaker 1

Okay.

Bruce Liu

Here’s the situation in the U.S. equity market. We’re kind of at the bottom of the valuation range for a reason: financial conditions have been tightening. It has nothing to do with cloud or AI. The U.S. Treasury market has been making some dramatic moves. That’s a good thing, because you need that to deal with inflation. Hopefully, going forward, we can really deal with inflation, and that will give us room.

All I’m saying is that financial conditions have kind of peaked here. If that works out, it won’t be enough to really damage the growth we have right now. Imagine QQQ is growing 50% year over year right now. Even if that slows down a little, it’s still going to be in the 40% range. From a growth perspective, from a growth-cycle perspective, we’re really in the middle of something.

I also encourage everybody to be open-minded. This thing can last longer than you imagine. I’m very positive, but I’m not betting on valuation. I’m looking for places with solid, resilient growth. Unfortunately, so far, that’s AI, because all the other cyclical areas—homebuilders, retailers, even traditional industrials—are having trouble because they’re very sensitive to higher interest rates.

Given this background, that’s where I’m going to place my bet toward the end of this year, even in the second half.

Christina Xu

To make it more concrete.

Bruce Liu

The long AI—

Christina Xu

Very short. One sentence.

Bruce Liu

Long AI infrastructure, by the way.

Christina Xu

Own AI infrastructure.

Bruce Liu

I don’t want to pick winners or losers, given where we are already, but you should have broader exposure. I think that’s the right play.

Christina Xu

Okay. I might come back to you. Okay, Ethan.

Yixing Xu

Great. In doing—

Christina Xu

It doesn’t have to be a stock. It can be anything. A prediction.

Speaker 2

During breakfast I talked with Bruce, and I think we’re pretty much aligned. I do see some risks, as I mentioned, like delays. There are going to be some pains in the process, for sure, meaning there will be some ups and downs. But in the mid to long term, I’m very positive and optimistic about AI infrastructure. Looking at the demand for compute and the supply—

Yixing Xu

Yes. There’s a mismatch between demand and supply, with a gap. I think the power sector for AI infrastructure is going to do really well by the end of next year, even with some ups and downs.

Christina Xu

And even more specifically, as you mentioned, that might be behind-the-meter power generation.

Speaker 2

I would say behind-the-meter generation and technologies. Similar to what Bruce said, I think a lot of electrical equipment is going to benefit from it.

Christina Xu

Okay, Alex.

Alex Yeh

One sentence, right?

Speaker 1

Great.

Alex Yeh

I think we will have a permanent compute shortage, at least until 2030, as far as I can see. I’ll explain why. Right now, the market talks about pre-training and post-training, but what’s actually using compute? That’s inference. It’s mostly coding—things from Codex to Claude Code. I think that’s priced in. You also have multimodal, which is image generation and video generation. You probably see that on Reels, on TikTok, and elsewhere.

Alex Yeh

Those people already see that. But you have to understand: talking to researchers, recursive self-improvement, or RSI, is coming very soon. That means compute—or AI—is going to self-train and improve without a human in the loop. So compute demand is going to skyrocket. Talking to a lot of prominent researchers, their prediction is that it could be completed in as little as 6 months to 2 years.

The second thing is robotics. It’s the first time we’ve seen Astra able to utilize robots in a phenomenal way. Traditionally, we didn’t even know whether robots could generalize. You had to train them on a specific task, like picking up the mic, for example. What Astra has shown is that it stretches first. It wiggles before moving things. So now you see the light at the end of the tunnel: brute force, continuing to layer in compute and egocentric data, and we can get to true AGI with robotics.

What’s the market now? It’s basically only coding. And not to mention Muse and Instinct, which are driven by consumer demand. That hasn’t picked up. So we’ve only priced in the business side. We haven’t talked about robotics. We haven’t talked about consumers. And then there’s RSI. Incredible. We’re just at the beginning.

Christina Xu

Okay. All right, everyone. Memorize all those predictions and come back next year to see who was right.

Bruce Liu

Thank you.

Christina Xu

Thank you.
