Duckbucks: You started your career as a chip designer. One would imagine you had a more industrial view of tech compared to some VCs working today, who came up in an era of software as a service (SaaS), mobile technologies and low-capital models. Has that given you a different perspective now that we’re again in an era of capital-intensive, infrastructure-heavy technology?
Bill Tai: Deep tech is back! That’s a perfect theme [for an interview], because that would kick off my career path in terms of starting in chips and then moving up the stack to communication equipment, networks, SaaS and everything else. I think it was a foundational understanding of how electrons move that led me to Bitcoin, too, and all the stuff I did in crypto and data centers. Most of these SaaS kids couldn’t set up a data center.
DB: And you’ve seen a number of hype cycles in that time, maybe more than you can count.
BT: History doesn’t repeat exactly, but it always rhymes. This is probably my third or fourth mega-bubble, if you want to call it that. Each one is different. Dating back to the ‘80s when I first got into tech, the first company I joined, with Jensen Huang, we were both at a startup together called LSI Logic. That company was founded in 1981 and went public in 1983 with $1.5 million in revenue. It did an IPO through Morgan Stanley and Hambrecht & Quist with, I think, a $120 million print; priced at $21 a share, it traded to $77 on $1.5 million of consulting advisory on how to do chip design.
In that era — I think because Apple had come out of nowhere and gone public —anything ending in “-ology, inc.,” any vendors to Apple or the microcomputer companies, could go public on an idea. We had buildings next to us owned by other companies that were going public with no revenue, just a few years into that wave. Sure enough, that bubble popped, and from 1983 it took about 17 years until the dot-com bubble in ‘99. In that era, anything ending in “dot-com” could go public, just like anything in “-ology, inc.” could go public 17 years before.
What I extrapolated from that was that institutional capital has to lose its memory of the prior bubble to allow another bubble. If you think about what happens in the hedge fund world and the public stock world, the average life of a fund manager might be five to seven years. You start off having a crash, and then after that crash, everybody is a value investor for about one cycle. They’ll all do everything at 10 times earnings. Then towards the end of that window, nobody’s outperforming, so the younger guys will take more risk and they’ll say, “You know what? It’s not 10 times earnings, it’s price/earnings to growth (PEG) ratios. I can pick the companies that have higher growth.” So for about one generation, you have PEG ratios come into vogue. And then after five years, no one’s outperforming, so you get a few bright people who are using products that are exploding in use with no earnings and they say, “You know what? I can tell you which ones will have earnings. And you’re just old, so you don’t get it, so you need to buy price-to-sales.” Then everything decouples until there’s no ground wire — and then the system pops and resets.
‘Institutional capital has to lose its memory of the prior bubble to allow another bubble.’
We’ve had that over and over and over, and I think we’re in one of the ramps now where people see the growth with no earnings — they’re willing to fund Anthropic at $380 billion, they’re willing to fund tens and tens of billions every month at OpenAI with $3 of losses for every $1 in revenue, hoping that eventually the margin structures will tune and get profitable. They may or may not.
I think the first cycle was similar. It was heavy capex for fabs — you had to put up hundreds of millions then, the equivalent of tens of billions today, to put up factories. You were basically shipping molten sand in the form of chips with almost no marginal cost, so it’s an all-gross margin if you can maintain price to cover the load of the capex. Bring the clock forward 15 years, it’s internet data centers. Clock forward today, it’s AI GPU data centers.
With fixed capex, you’ve got to go for high volume and try to get margins to grow. The internet data center market was exactly like that, but it got overbuilt — just as the fabs did in the ‘80s. The industry at the end of the ‘80s was losing money hand over fist. If you took the aggregate earnings of the industry from the start of its history until the late ‘80s, the industry had negative earnings. The internet data center business was like that, and in that case, the revenues were kind of fake because it was ad revenue being traded around. It wasn’t really contributing to margins. I had friends at eBay and AOL that, to make their quarterly earnings, would call each other up at the end of the quarter: “Hey, buy $30 million of ads from me and I’ll buy $30 million and 50 cents of ads from you, and we’ll make our numbers and stock goes up.” Then AOL drove it to $150 billion and bought Time Warner for real revenue, but then eventually those ads went away.
This time, today, what’s a little bit different is that the capital for the AI data centers is — for the most part — coming from companies with enormous cash flow, whether it’s Google, Amazon, Meta. In the dot-com bubble, it was mostly the big buyout firms that moved into tech to fill the gap for funding data centers.
I started an ISP in ‘95, it built out the first privately owned ISPs in Taiwan and a few other countries. We grew that to become data centers in 10 countries and then went public in 2000 with Goldman Sachs, Morgan Stanley and Salomon Brothers. In the end, the big capital there came from Texas Pacific Group because the buyout firms saw that these are real estate plays, actual physical buildings. They were OK putting equity in and then levering a lot of debt to get them done because there would be cash flow. Ultimately, there was an overbuild so — like with the semiconductor fabs in the ‘80s — marginal prices went down. We’re not there yet; we’re still on the ramp of AI data centers.
DB: Thinking through who the AI winners will be — you have the model providers, you have infrastructure, you have applications. Is it infrastructure where some of the stronger bets are being made?
BT: Under the current architectures, which are pretty much monopolistically driven by NVIDIA, the need for energy is growing at an exponential rate. The data centers that we built for the plain old internet in the late ‘90s would be built to a specification of 15 kilowatts per rack. I would say five years ago, that number was maybe 25 to 45 or 50 kilowatts. There are data centers going up today at 500 kilowatts to a megawatt — one million watts — per server: a few square feet, six feet tall. It’s growing by leaps and bounds.
The demand that is projected for all the GPUs that are supposed to be built out is enormous. Roughly one gigawatt is about one nuclear power plant. If you were to take the announced and funded plans for data centers to be built out by 2027 — a year and a half from now — there needs to be 27 gigawatts. And by 2030, another 63 [gigawatts] on top of that. So if you can imagine a world where the US could even conceive of building a nuclear power plant to completion every single month for the next two years, and then maintain that pace till 2030 — that’s the funded demand. There just isn’t enough electricity generation, and the way the regulatory structures work in America, you can’t get stuff done because of the public utility commissions and all the layers of bureaucracy. We need something that looks like the Telecom Deregulation Act of 1996, which forced AT&T to break up, compete with itself between the parts and drive prices down.
For the moment, I think we’re kind of stuck. You’re seeing huge increases in value in companies that can provide off-grid, behind-the-meter, not-connected-to-the-public-transmission-lines capacity. Take a company that you wouldn’t think of, like Caterpillar Tractor. If you look at their stock chart for the last year and a half, you’ll see the result of them being able to build out facilities and put gas turbines in there.
‘There just isn’t enough electricity generation, and the way the regulatory structures work in America, you can’t get stuff done because of the public utility commissions and all the layers of bureaucracy.’
With Hut 8, we saw this coming. Hut 8 started as a small subsidiary of a Bitcoin mining operation I had funded called Bitfury. From 2014 to 2017, we were arguably the world’s largest Bitcoin mining operation, and we had data centers in Iceland, Finland, Norway, the Republic of Georgia, a few other places in Europe, and then Canada and Texas. We took 100 megawatts, spun that out and took it public as Hut 8. We took another 100 megawatts or so and took that public — today, that’s Cipher Mining. Both of those companies have pivoted to become AI infrastructure providers. We started buying GPUs at Hut 8 in 2019 and then bought a hosting company so that we’d have regular hosting operations and, over time, built up a pipeline now of around 10 gigawatts. The deal we announced recently was for 245 megawatts, about 4% of our footprint — that’s a $7 billion guaranteed-revenue deal with Anthropic, Google and FluidStack. I don’t know that we can 40x $7 billion, but the demand is there. The ability to supply it is hard.
DB: This goes back to where we started, the increasing relevance of the physical world to technology. You can see this in another way with Zoom, where you were the first committed investor — a “sleeping giant” company that seized on the black swan of COVID. Did that experience shape your understanding of market timing and being ready to seize opportunities?
BT: Where I invest is, I look for very large markets that are about to go through a structural change. The thought process through my career — in the very beginning I was a chip designer, so for the beginning of my venture career, which was starting around ‘91, I funded mostly chip companies because that’s what I knew. And then I started funding the next layer above that which were products built out of chips. So I spent a lot of the ‘90s funding communications equipment: ethernet switches, wide-area network equipment kind of chips, then boxes. When the internet hit, it was obvious to me that you would stitch those boxes together and digitize the phone network, so I started an ISP. And that became data centers.
So it was chips, boxes, networks. And then that company went public and I took some time off —I got sponsored as an athlete, and for about three years I was mostly kiteboarding and angel investing. Then in 2002, it was pretty clear to me a new wave was coming because the internet infrastructure had been overbuilt and was largely free. You had kids in dorm rooms, like Mark Zuckerberg, starting companies that would flare up in usage because they had free cost structures. There was a lot of value being created, so I came out of retirement, started funding things in the mobile-first Web 2.0 era. Then the iPhone came out, so that shifted me into mobile applications and cloud.
When that transition occurred, it was so obvious to me that the cost structure was going to radically change for certain kinds of applications. You could look at a string of companies I backed then — whether it was Zoom in cloud video, Treasure Data in cloud infrastructure, Canva in cloud design, Wish in cloud shopping — there was a different approach to an architecture that was lower-cost, easier to distribute.
‘Where I invest is, I look for very large markets that are about to go through a structural change.’
Zoom is a perfect example. In 2012, the company was just a year or two old, running out of money, and no one wanted to fund it. Every time I sent Zoom to a regular venture person, the typical response I’d get is, “Bill, you’re nuts. There’s like 20 video conferencing companies out there and no one’s making any real money. You’re going to be competing with free products from Google and Microsoft —Hangouts and Skype — and they have infinite money and there’s no revenue. Why would you do this?” Everybody turned it down. So I sent a note to [former Yahoo! co-founder] Jerry Yang, who had backed me in Treasure Data, and I said, “Jerry, you need to look at this.” He didn’t even want to look at it, he sent it to an associate.
We had just launched a product; we got 60,000 downloads, only 20,000 people opened it and 50% only used it once. Nick Adams from Jerry’s shop sends me a note back saying, “Bill, the engagement numbers are horrible. Why should we even look at this?” I wrote a three-page email — which I never do — where I laid out everything that would happen: “Look, there’s a huge opening because of the cloud transition, tablets and phones are going to change the way people do this. There’s enough revenue flowing through WebEx and GoToMeeting that you can build a substantial revenue line here, and Skype is going to die because there’s an architectural limitation.” Everything I predicted actually happened.
What people don’t know about Zoom — they think of it as a COVID company, but Zoom went public a full year before COVID even presented itself. It wasn’t a giant business, but it was a $300 million trailing revenue business, 83% gross margins, 27% net income. It was very profitable because they could carry bits over the cloud at very low cost in a way that was a thousand times more scalable than the other architectures, which were more peer-to-peer. In the old days, using WebEx or Skype, if you wanted to have 10 people on a call, your browser or your computer would have to generate 10 separate streams and maintain 10 separate links. In the Zoom case, you only need to make one into the cloud and then you can spin up 50,000 servers if you want and have 50,000 people on your call. It was just inherently more scalable and there’s just many more things you could do. I knew that that was going be the case because I had run a data center before.
That era was just transformative in cost structures. Things like Canva — instead of Adobe selling $900 CD-ROMs in boxes, which had to pack all the complexity into one loading onto your computer, Canva could send you bits and pieces and have a very scalable backend and a marketplace for users to put content in to lower your cost because it connected you to everybody that had already done work templates that you might need. That’s far more scalable in use case and architecture and cost structure.
DB: You haven’t mentioned a single founder. Is how you assess opportunities more a matter of architecture and infrastructure?
BT: Well, that just tells me what may change in the landscape. But when I fund something, it’s when the right person meets that structural change.
Many, many years ago, at the startup I was in with Jensen Huang — our backers were Fund II of Sequoia Capital, Fund II of Kleiner Perkins and IVP. Don Valentine was on our board, the founder of Sequoia, and Tom Perkins was on our board. When I went into the venture business in the ‘90s, I would kind of tag along with Don. I learned a lot from him in terms of how he viewed the world; he was a great investor. We were at an airport once, stuck after a board meeting having dinner, and I said, “Don, what do I have to do to be good at this?” This is when I was really young. And he said, “You only have to get three things right: One, the market. Two, you swap the team in and out until it works. And then three, don’t overpay. If you do those three things right, you’ll make money every time. But if you get the market wrong, you can have the best team in the world and infinite money and you will lose money every time. You got to get the market right.”
Times have changed because in that era, businesses were asset-heavy, and Don used to go around saying, “We own the desks.” So he could swap people in and out. But over time, when things went capital-light, it became much more founder-centric. I’m now kind of a hybrid where I’ll look for that big market with a structural change — and then I have to pick the right person that has leadership charisma, the ability to put a team together, the right level of grit. Somebody like [Zoom founder and CEO] Eric Yuan or [Canva co-founder and CEO] Melanie Perkins or Hironobu Yoshikawa, who founded Treasure Data, or those at many of the companies I work with today.
