20VC: How to Build Your Own Data Center & Why Every Startup Should Do It | How ElevenLabs Leapfrogged Us: What I Learned | The AI Talent War: How Your Hiring Process Needs to Change with Cliff Weitzman, Speechify

Hundred percent is on. It’s the biggest strategic mistake I made in the history of Speechify. The best way to lose is not to be in the race. Be in the race. You don’t want to be a fat manager who’s like a general sitting in the back saying, “Take that hill.” You want to be the warrior who runs up with their sword and engages the enemy first.

This is 20VC with me, Harry Stebbings. Now, I am fed up of the simple question, answer, back and forth podcast. Today is a real freaking discussion. Cliff Weitzman, founder and CEO at Speechify, one of the fastest-growing text-to-speech startups in the world, on the show, where we have a real debate about whether it's right to scale into enterprise from a phenomenal consumer business, what it takes to build an amazing go-to-market motion when you've already built this amazing consumer business. And then he also tells us some wild freaking stories about spending tens of millions of dollars on NVIDIA GPUs, and why so many more companies should be doing that over-relying on other providers. This and so much more in the episode today. But before we dive into the show today, today I wanna tell you about how the first AI law firm, Crosby, helped us close a big sponsor. As you know, some of the biggest companies in the world advertise on 20VC. My British dulcet tones clearly convert well. I was working to close this big sponsor, and they wanted to get through legal review quite quickly to close the deal. Crosby turned red lines around in three hours and caught major issues that would've caused us serious problems in the future. Crosby combines AI, some of the best engineers in the world from companies like Ramp and Stripe, and some of the best attorneys in the world from top 10 law firms. Customers get the best of both worlds. An elite human attorney reviews every contract, but they move incredibly quickly, returning red lines in under four hours. They help the fastest-growing companies like Cognition, Ramp, and Clay close deals in hours, not weeks. Learn more at crosby.ai/20VC. If you want to redline NDAs, MSAs, DPAs, and any other procurement contracts faster, go to crosby.ai/20VC. It's speed that you can really trust. While Crosby keeps your numbers sharp, OneMind keeps your customer conversations sharper. Our friends over at OneMind have a hot take. The B2B GTM model we've been using for, ah, the last 20 years, it's collapsing. Predictable revenue isn't so predictable, and buyers are just tired of explaining themselves at every handoff between SDRs, AEs, CSMs, and support. You feel it in your board reporting. Your sellers feel it in their coverage. Your buyers feel it as they wait for answers. Well, enter OneMind and their GTM superhumans. HubSpot, Alphasense on an investment that helped us close an $8 million deal. $8 million, baby, that's a lot of money. That's why I'm genuinely excited to have them as a partner on 20VC. Alphasense combines AI with one of the world's deepest libraries of market intelligence, including expert interviews, broker research, earnings calls, company filings, and real-time news. Every answer is grounded in this incredibly trusted evidence and fully traceable to the original source, which is so important. So you can make really high-conviction decisions with confidence. But the best part, they're building Super Analyst, an always-on AI analyst. So instead of starting your day with another search, you'll start with work you already done, your coverage monitored, the important developments surfaced, and your investment brief already waiting for you. See for yourself. Head to alphasense.com/20VC. That's alpha-sense.com- You have now arrived at your destination. Cliff, it is so good to have you back in the studio, dude. I was looking forward to this one because, when I was writing it up, it was a very different thread of conversation from how I’d normally go. Thank you so much for joining me again today, dude.

My pleasure. Glad to be here, as always.

I wanted to start with this: You’re spending tens of millions of dollars on NVIDIA GPUs, and you’re paying an additional $100,000 per GPU to receive them 4 months early. Why? What do you know that the market doesn’t know?

1. Owning GPUs Beats Renting

In 2022, we bought a huge rack of GPUs from NVIDIA. The reason we bought them is for training. We have a bunch of models. The newest Speechify Simba 3.2 model is ranked number 1 in the world for quality, above all the frontier labs, and is 10× more affordable than stuff like ElevenLabs.

We used to rent GPUs, and we found that engineers at Speechify would be parsimonious with how they used them because they were like, “Oh my God, I’m costing the company tens of thousands of dollars. I don’t want to do that.” The analogy my brother and I came up with is: Imagine you’re Michael Jordan, and you want to be in the NBA. It’s the only thing you care about, and you need to pay $20 an hour just to train in a basketball center.

That sucks. You want one that you can go to whenever you want to. In fact, you want a hoop in your house. Our initial idea was that we wanted a hoop in our house, so we bought a bunch of our own GPUs. That deal ended up being really good for us, and we ended up training really good models. With time, we invested more and more and more.

The second part is actually how the economics work out. If you look at it, the Transformer was invented inside Google in 2017. NVIDIA came out with A100 GPUs in 2019. Shortly after, they came out with H100 GPUs. The original ChatGPT was trained on A100s.

Then they came out with Blackwells, so then B200s and B300s. Now they’ve come out with Rubins, which are the GPUs Elon is sending to space, and they’re liquid-cooled. They’re very, very cool.

Every class of GPU is more affordable per 1 trillion FLOPs. A FLOP is an addition, subtraction, multiplication, or any mathematical operation, and you measure them by how many trillions of operations happen per second in a GPU. The newer GPUs are more affordable in that respect.

If I were to buy an H100 for, let’s say, $30,000—that’s how much a single card would cost—and I wanted to rent an H100 for a 1-hour spot instance from GCP, it could cost me $5. If I rented it from Azure or AWS, maybe it would cost me $3.50 per hour.

If I multiply that by 24 hours and then by 365 days in a year, I’m actually going to end up paying $35,000 to $50,000 to rent that GPU for 1 year. But I could buy it for $30,000. So it’s 1.5× the cost of owning the hardware to rent the hardware for 1 year.

The hardware is typically warrantied for 3 years to work properly, but it’ll keep working after the warranty for—I imagine, I don’t know—10 years. The math just makes way more sense when you buy them.

The other big part is that, if you want to do large-scale training like we do, you need the memory to be colocated with a large cluster of GPUs. I can’t just rent from Google or Microsoft or even Baseten and run the size of training that I want because I need a gigantic memory card next to it, with all of my data that all the GPUs are accessing. That’s why we first started buying them.

The next thing that we found is that, if you run open-source models for coding, you could pay Anthropic. You’re paying for all the tokens and for the branded Fable 1. Or you can run an open-source model, and instead of running it on a spot instance from Azure or anyone else, you run it on your own hardware. Then you’re paying a fraction of a fraction of a cent per token.

For all those reasons, it made a ton of sense, but we can go into all the depth that you want.

I just want to dig in. I completely understand the rationale there, but chips depreciate. You have chip cycles, and they’re accelerating. We’re seeing newer and newer chips being created, and we’re seeing specialization within chips. By buying, you’re locking yourself into one chip architecture, so to speak. How do you think about that?

At Speechify, we still use K80s for a lot of specific operations for inference, and we use older models of GPUs constantly. There’s essentially a difference between when you do inference and when you do training.

For training, I’m like, “Okay, I have this hypothesis. I want to know the answer to this hypothesis as soon as possible.” Every minute that it doesn’t come out, I’m in competition with everybody else. Having a GPU architecture that is faster by orders of magnitude is a huge advantage.

But if you go speech-to-text or text-to-speech with Speechify, I can afford to give you a lower-quality GPU, and it’ll still give you what you need in 100 milliseconds. It’s totally good. I can always use these older GPU models for inference.

Number 2, we have so many experiments that we’re running at every single point in time. Not all of them need to run on the newest hardware. The analogy I always give is: Let’s say you bought an iPhone back in 2011, and it’s an iPhone 3G. Then you bought another iPhone and another iPhone and another iPhone. You could have a drawer in your house with 5 iPhones collecting dust because you can only use 1 iPhone at a time.

But if I own 100,000 GPUs, I’m still going to use all of them at the same time. I’m not losing anything by having more GPUs because not only do I own a bunch, I still rent from the hyperscalers all the time. I rent both dedicated instances that I prepaid for and spot instances.

For example, more people use Speechify in September because everybody goes back to school, so I need to level out the load. The parts of that load that I know for sure I’m always going to use, whether it’s training or inference, I might as well just own them.

On top of that, I have so many other friends who are running training and inference, so I can always rent it out to other people if I have excess capacity, which I don't expect to have. Every once in a while, though, you have an interesting situation. For all those reasons, it makes mathematical and financial sense.

Lastly, if you have excess capital, you either stick it in the bank or buy a bond. The best bond you can buy long-term will yield you around 5%, or you can buy a GPU. Because renting it would cost me 1.5 times as much as buying it for the year, the return is much higher.

So how many GPUs do you buy, then?

Let's talk about Rubin systems, for example. Rubins come in the form of 72 cards in one rack. We'll buy multiple racks of Rubins, and on top of that, we'll buy B300s, which are the newest form of Blackwells because we can get them earlier. When we bought our first instances of DGX H100 GPUs, we bought a bunch of racks of those.

Those get delivered in a truck to the data center. We rent the data center, and the data center provides the networking capability and the energy, which is actually the largest constraint now. It also provides physical engineers who take it off the truck, install it, and fix it if it has an issue. Then it just runs.

Does ElevenLabs do this?

Yeah. ElevenLabs is amazing at this. Piotr at ElevenLabs literally bought a bunch of GPUs early on and set them up in his house, and then they just kept building bigger and bigger and bigger clusters. They do the same thing that we do.

How do you think about forecasting chip buying? It's incredibly difficult to know, A, demand, but also, B, the supply of chips.

Yeah.

How do you think about forecasting chip purchasing?

I want to explain again that it's very different from buying an iPhone or a MacBook. I can only use one MacBook at a time and one iPhone at a time, but I can use all the chips I have at any given point in time. I will still have more demand, especially when I have multiple teammates and 60 million users using inference on my Speechify software, which provides text-to-speech and helps them read and dictate their work. They also use Speechify Work, our newest product, which is agentic, like JARVIS from Iron Man.

Let's imagine I have 100% capacity, which is the average monthly usage I need for GPUs for training my AI models and for inference on my AI models. Inference is when you make a call to Speechify, give me text, and I give you back audio. There's math happening in the background. That's inference.

Training is when I take a gigantic amount of data, all the architecture and software engineering we're doing, and say, "I think this will give me a better model." I'm baking that model in the oven, and I'm going to come out with a new black box. When I give you text, that black box calculates it and gives you back the audio. Those are the 2 usages.

Let's say I have 100%, which is what I would have in a month like November. In October, I'll have 140% because it's a big month for us. In December, everyone's at home; they're not necessarily studying or working, so I might have 80% utilization. I can take 20% of the normal usage and buy it because it's the best deal. I'll take another 25% of the usage and do long-term contracts with hyperscalers. For the rest, I'll rent what's called spot instances from the hyperscalers, and I'm still not even close to overcommitting myself.

That's how we think about the math. We also have 45 engineers, but we want the team to be 150 engineers. Even within my 45-person engineering team, there are a couple of rock stars who have dedicated DGX racks just for them. Twenty-five percent of my team is almost waiting, and I want to double the size of the team. It's like having a football team and needing another field because they don't have enough space to practice. That's how I think about how to allocate.

In terms of depreciation of the asset over time, these are still amazing GPUs. Even A100s can run amazing experiments. It's completely valid to use them as long as they're hooked up and haven't stopped working.

Think about the mileage of a car. If a car gets to 250 miles, you know it's probably going to break at that point. That's not necessarily true for a GPU because it doesn't have as much wear and tear. Yes, it's moving, and yes, all these things happen, but it's in a very clean environment. It's cooled very well and has constant maintenance because it's not moving around. It's very expensive, and NVIDIA just does a really good job.

That asset is going to stay for a very long time. If it gets so outdated that I can no longer run training on it, I'll use it for inference.

One more interesting thing just happened. I believe earlier this month, NVIDIA did a huge deal with Blackstone, BlackRock, Apollo, and Goldman Sachs. They said, "Listen, we want more people to buy more GPUs. We're going to underwrite up to 25% of the value of a GPU for you. If you lend money to someone who buys a GPU—let's say Google or a startup like CoreWeave—and that company goes out of business, and you have that GPU as collateral against the investment, we'll buy back the GPU for up to 25% of its value."

They're succeeding in creating a liquid secondary market for GPUs that they're underwriting, so now the large banks have an incentive to lend money at much better interest rates. This is exactly what Elon did at the beginning of SolarCity. He went to Morgan Stanley and Merrill Lynch and got them to amortize the price of a solar panel over 30 years. The whole invention behind SolarCity was the fact that you could take a loan against the collateral of your solar panel.

NVIDIA has done an amazing job creating a clear floor for the value of the GPU over time.

Do you think the circular-economy fears that people often cast against NVIDIA are justified or not? We saw their CFO push back on them and say, "Enough. Enough of this bullshit." Do you think that's justified or not?

I think a lot of the things that were going on about a year ago between Oracle and OpenAI were way too much. That was ridiculous. I think the NVIDIA stuff is different because you're talking about a real asset.

If you think, for example, about the logic behind the value of Bitcoin—what is the intrinsic value of Bitcoin? I can't really tell you. What's the intrinsic value of gold? Gold can be used for some medical applications because it's an amazing metal, and it's used for jewelry, whatever. But a GPU has intrinsic value. You can actually use that asset for something that's really, really valuable, and it doesn't matter where that GPU is. It could be in Iceland; it's still useful to anybody all over the world as long as it's networked.

It actually has a pretty good store of value, even if new GPUs come online. My one question—and this is the math for everybody to come back to—is: How many teraflops per second can this device do? That's essentially a token. That's the value.

There is intrinsic value. Yes, you can have all these circular things, but at the end of the day, NVIDIA is making a product that's real. It's not complete tulip mania. There's real, real intrinsic value here.

What does no one know about buying chips—

So much.

—that they should know? What's the, "Oh, my God—

I—

—people are so naive about this"?

It's not that people are naive; they just haven't been in the space. I'll give you an example. Imagine you're buying a GPU. You're going to buy an NVIDIA-produced product, but NVIDIA isn't going to waste time talking to Cliff Weitzman. So who do I buy it from?

One of the best-rated vendors is Dell. Everybody thinks Dell is a personal-computer company. No. Dell is a GPU rack supplier at this point.

Then I want to buy it from Dell. Dell has a constraint because there aren't a lot of Blackwells available. It happens that they have some in France. All right, I'll order mine from France.

Then you realize it was supposed to come a month ago, and it's still not here. You have to negotiate to make sure that you get it, which is why we're very willing to pay an extra $100K per month to get them earlier.

So you'll call up Pierre in France and say, "Hey, we'll give you an extra $100K kicker if you get them here in a month"?

Even more than that. In that France situation, which is something that happened to me, I was like, "Pierre, what the heck? We have a contract. You're not delivering on time."

We had a contract with another company beforehand, and they were a few weeks late. I called them and said, "Listen, I've got a better deal. I'm canceling our contract because you didn't deliver. I'm going to go with this other contract, but if you have a better price, we'll go with you. I just need the GPU now."

Remember, I'm paying for the rental space at my data center, so the most expensive part of a GPU delivery is when it's late.

I'm still paying rent for that data center space. So you put pressure on Pierre to send you the thing when he said he was going to send it to you, and then you go to NVIDIA or Dell or whatever and say, “Well, it’s a market. Hey, can I pay more to get it earlier? Skip the queue?” “Yeah, you can.” “Cool.”

Now there’s a truck somewhere in the United States with a GPU whose value is the value of a house coming to my data center. I should have insurance on that, right? Because if that truck gets hit, there’s too much humidity, or the GPU gets flipped, I’ve lost multiple houses’ worth of GPUs. So the insurance is really, really important.

The value of the amortization is also really, really important. There are all these nuances in how to do the math. Then there’s the cooling, right? You’re not only paying for the physical space, the networking, and the energy—and energy is the biggest constraint. We’ll talk about it in a second—but how do you cool that thing?

You have a thing that’s just moving and moving and moving. What’s most new now is liquid cooling, because air is just not enough, and the thermal load of water is much better. There are other liquids that are even better than water. Rubin GPUs are liquid-cooled.

Most of these data centers don’t have liquid-cooling installations already approved. So we had to do a bunch of research and found that we could buy what’s called a sidecar of liquid cooling, put it into the data center, and then pay someone at the data center to install it for us. Now you can have the rack that you want. There’s a big difference between running a purely software company and running a company that includes hardware.

But when I listen to all of this, I’m now more sure than ever that it’s a mistake to price-optimize and spend the money to buy it versus rent it. I get you on the optimization, but you’re not saving 10 times more. It’s 0.5X more per year.

No, no. Per year. Exactly.

Yeah, per year. But you have the flexibility to scale it up and down. You don’t have any of the logistical nightmares of insurance, transportation, security, liquid cooling, or logistics. Then you can build your product around what actually matters most, against ElevenLabs, which is fucking running fast. I don’t want to worry about liquid cooling and insurance for a freight truck.

ElevenLabs worries about the same thing, because for them to train excellent models, they need to have co-located GPUs with a lot of memory available.

You can’t do it if you rent it.

You can. It just becomes, one, ridiculously expensive. Two, you need to commit many years ahead of time, because you need to build a co-located cluster. Then you don’t have as much control because you don’t own it. It becomes difficult. You suddenly need an InfiniBand cable, which allows the memory to flow from one DGX to the other.

The answer then is, well, now my ability to train is so much bigger. I can have a larger AI team. Every person on an AI team is leveraged. I could shoot ahead of everybody so much faster.

Let me make one thing clear. If I want Rubin, which has these much faster GPUs, I’ll get it faster if I buy it than if I wait for Google to buy it and then have other people in front of me in line. I’m going to skip the queue by a lot, and then I’m going to have a year of access to Rubin before everybody else does.

2. Compute Data And Team Win

The way I think about it is the following: How do you build an amazing company in a world where there’s so much competition today? The team is the most important part, but the team is the most important part because the team gets you the other resources.

What are the missing pieces? The missing pieces are data, compute, and architecture. In a world where intelligence is commodified and no one needs to handwrite code anymore, what I’m looking for from our engineers is 10 really good decisions per day. That’s very tiring, rather than optimizing random parts of the code.

Each person has 5 to 18 agents running at any point in time, doing long-horizon tasks on these GPUs, coming up with theses, testing them, and going back and forth. If they don’t have the capacity to train, the team is limited. If they don’t have the data to train, the team is limited.

And, by the way, there’s a lot involved in cleaning the data. You get this raw data in the beginning, and you need to organize it into data sets. You need the GPUs to organize the data sets, too. One of my best engineers right now isn’t even writing models. He’s making synthetic data sets to train models, and so it really becomes an indispensable asset.

Would you ever buy data?

We have, but only small data sets.

I suggest you use Fireworks. Fireworks is amazing. Lin Qiao, the founder, co-founded PyTorch. I had the very obvious realization that every company would have its own specialized models of a certain size, trained on its own data, but it would need supplemental data—

100%.

—like synthetic data or real-world data that you just don’t have. You would buy that from data providers like Mercor—

Yeah.

—which is why I also—

Mercor, micro1, Surge AI—all these companies are amazing, and they shorten the cycle to getting to revenue, by the way.

100%.

If you’re Mercor—shout-out to Brendan Foody—ElevenLabs, OpenAI, or Anthropic is going to make money for the next decade or two on the data they bought from you. They’re willing to pay you a fraction of that 10 years of revenue today to supply them with the data.

Again, it’s all a speed thing. Yes, ElevenLabs, OpenAI, or Anthropic can go and make a team that will get the data, but they don’t want to manage it. The data is key. You need all 3 things: compute, data, and a team that writes great products. Ideally, you also need users who use you a lot and allow you to have a feedback loop about whether the stuff is good or not. So, benchmarking.

When I was doing this as a venture investor, we did outcome-scenario planning, which is the most bullshit exercise to pretend we’re smart by predicting the future. We do it because it makes us feel important.

They predominantly sell to frontier labs today, and that’s where 90% of their revenue comes from. With the rise of specialized models on a per-company basis, with their own data, I believe that you would move that customer base from purely frontier labs to every large-scale enterprise that needs supplemental data. If that’s the case, how big an outcome is the data marketplace?

3. Data Marketplaces Need Operations

The first problem to understand about the data marketplace is that it’s not ARR. It’s not annual recurring revenue. It’s one-time deals every single time. The buyer of the data isn’t required to buy it from you again, so it’s a very risky business.

If you look at the early days of companies like Mercor, they didn’t raise significant venture funding off the bat because investors were very skittish about that fact. Let’s put that aside.

It’s very important that the Ts are crossed and the Is are dotted regarding how you got that data. You need to indemnify the companies that are using you, and that’s part of why they buy it from you as opposed to sourcing it themselves. We’ve seen the lawsuits.

It’s a great business if you can do it well, but you need to be an operations monster. You need to be really, really good at operations. You need to be very fast, and the key is that the company training on your data actually needs to see improvements in its model at the end of the day.

The thing that has always been challenging for Speechify compared to other companies is that B2C customers pay a lot less than B2B customers. ElevenLabs, huge credit to them, leapfrogged us because they sell to B2B, while historically we’ve only sold to B2C.

Our big constraint was that it needed to cost us less than $10 per million characters. ElevenLabs charges $100 per million characters. The OpenAI model on the benchmarks costs $196 per million characters. When we sell it to other B2B companies now, we just launched our API, Simba 3.2, and it costs $10 per million characters.

Dude, I’m too old not to ask the painful questions. I think the joy is when you ask them and are less worried about asking them. You said ElevenLabs leapfrogged you. Is that on you for not doing B2B?

4. Speechify Missed The B2B Wave

Yeah, 100% on me. It was the biggest strategic mistake I made in the history of Speechify.

How do you reflect on that?

I met Piotrek and Mati. I was living in London at the time, in my house there. I think it was 2022, and we were very impressed by them. We wanted to use their model, by the way. It was just too expensive for us to use.

I looked at it and thought to myself, “They’re very smart. They’re going to do well, but I don’t like their strategy because I think that an API for text-to-speech will become commoditized over time.”

You’re going to get to the point where you can run that API on your computer and then on your phone. What are they selling anymore? I don’t want to go into that business, and I made a critical error.

What I didn’t understand is that the point of an AI lab like Speechify or ElevenLabs is to continuously innovate, and the first product that you release is your wedge that gets other people to later use your other technology.

So, for example, if you're in text-to-speech, you build the best text-to-speech model in the world for 1 specific voice. Cool. Well, now you can do other voices. Now you can add emotional prosody. Now you can add voice cloning. Now you can add speech-to-text.

Now you can build duplex models that handle ums and ahs, laughter, interruptions, and turn-taking. You add a harness for voice conversations, then you optimize it for sales, customer support, and all these things. What they did is first build an amazing API. They were great at launches, and they built a really great product for creators.

Then they built their best product ever, which was Agents. Agents is amazing because the buyer is no longer a software engineer. The buyer is a CTO, CIO, CEO, or executive in the company. Sierra has this concept called outcome-based pricing. Bret Taylor is amazing, and so you can start fighting on the outcome. Having an AI agent is like having an AI coworker.

But it was my mistake to think that an API product was a bad strategy because I thought it was something that would become commoditized, and I forgot the central thesis about Silicon Valley, which is: constantly innovate, get the user to start using your product—I don't care if it's free—and then sell them other things. That was my big, big, big, big mistake.

How possible do you think it is? I think people underestimate the complexity of building out a B2B GTM.

Super hard.

I think it's a strategic mistake for Speechify to go to B2B.

Hmm. Tell me your position.

You are now competing against ElevenLabs and Sierra, and those two are competing. Whether they like to admit it or not, they absolutely are competing. I'm sure if you ask them off camera, that's Bret Taylor.

Yeah. You don't want to compete against Bret Taylor.

Motherfucker, I don't want to compete against Bret Taylor. That is the tidal wave of ElevenLabs. ElevenLabs is an unstoppable machine at this point, to the point where it has government buy-in across all of the large major Western democracies. Actually, it's insane, the government buy-in they have, and they started 3 months ago.

You just said the key thing: They started 3 months ago.

Yeah. I think they've reached a tipping point where they've just taken the market. I think Sierra is running behind them, chasing, and they're doing a decent job of it, but they've got Bret and they've got Sequoia and Greenoaks and all the royalty of Silicon Valley behind them. They're still running behind, chasing ElevenLabs, with Sequoia kind of pretending to be neutral because they're in both of them, which is incredibly challenging. I just think being third—the Postmates effect—is never a good market to be in when I could be the dominant consumer brand that leads with a really different and compelling story.

So, here are the 2 things to consider. The first one is, if you go to the App Store and search text-to-speech, Speechify has 98% of the installs in text-to-speech for B2C.

Hmm.

Speechify has served more than 770 billion words to users over the last few years, which in terms of time spent listening is like 6,000 years of listening. If you go from today to 0 B.C. and back, you still have thousands of years left. We've completely dominated that market, and it's still a business that's growing really, really fast. We're constantly adding more features into that product.

The thing is, we have a pretty big engineering team, and now everybody's capable of doing 10× what they did before. So I have extra staff. I have a huge AI engineering team with the ability to make amazing models. Where is the highest ROI for that to go? It needs to go both B2C, but it should also go B2B.

One thing that I will never be is a person who doesn't learn. I might as well just freaking learn B2B. Now, to your point about competing against giants like Sierra or ElevenLabs, Anthropic came into the market as a second to OpenAI, and they were second for a very long time, and now they're not second. Facebook came second to Friendster and MySpace, and now they're not second.

The nice part is this space is not a monopolistic space; it's an oligopolistic space. If you look at what happened with ElevenLabs, I'm going to exclude Sierra because the Bret Taylor effect is huge. It's just amazing to see how good of a business that is. It might very well be that for the core offering that they're currently winning on, I will not win. But what did I learn last time? It's fine if I offer my product essentially for free because I'm an AI research lab. As long as people start to use me, with time I'll be embedded in the system, and I'll keep coming out with more and more and more innovations that are useful to them.

There's unbelievable demand from all these companies and governments and everybody else for great tools, whether they be AI agents, APIs, or products. I just want to be on your phone if you're a user or in your stack if you're a company, and supply you with the best front deployment engineer experience and AI orchestration experience, and API experience to give you an amazing experience. There's room for everybody.

I agree there's room for everybody. Value accrues to the top 1 player.

I agree.

I think it's kind of like the inference market, where Fireworks will be a multi-hundred-billion-dollar company, and then I think Baseten will be a 100-billion-dollar company, and then Together and a load of the others will be 50, which is amazing.

It is completely true. Power law, yes.

Huge, hugely powerful, amazing, valuable companies.

But you would then think that OpenAI would be the place where value accrues for voice AI, right? That's what you would have thought 3 years ago.

Mm-hmm.

That's not what ended up happening. You can't not go into the race because there's a big incumbent.

Well, I think, with all candor, that's because of incredibly poor management.

I agree. But that's the thing. Every—

That was theirs to take, and they've fumbled the bag across every spectrum.

And for every company in the world, no matter how exceptional the leadership team is, niches get fumbled, right? Voice AI was a niche for OpenAI. LLMs are the core, and, by the way, they also fumbled AI coding. Now they're trying to catch up because it's such a big space. All respect to Piotr and Mati—I think they're absolutely amazing, and I love working adjacently to them.

I just don't think they're going to fumble the bag.

But they have so much in their net right now, and so much is getting added to the net constantly.

Yeah.

And so you have to go where the football is going.

Yeah.

I think it's too expensive for Speechify not to be playing in B2B as well as playing in B2C. The best way to lose is not to be in the race. Be in the race.

In terms of the products that we build, we were chatting earlier. You said that every startup, say, has to be a compound startup. Can you talk to me about that and how you think about that?

Yeah.

It's not that every startup has to be a compound startup. It's that, at a certain point, you can't afford not to be that.

Do you not think there are a few companies that are just absolutely fucking running rings around everyone else?

Yeah, absolutely. Those are the winners, right?

Yeah.

ElevenLabs is an example. Anthropic is an example. Ramp is an example. Speechify is an example. All the companies that have absolutely maniacal leadership teams and engineering teams. That's why people care about team more than almost anything else, because the right team will iterate fast, get there, and then figure it out.

Now, when everything can be turned into a reinforcement learning problem, where you can have long-horizon agents and orchestrating agents thinking about the problem for 2 weeks at a time, if you set that up, of course you're going to win.

I got into a lot of trouble, as I always do with most of my social posts. I used to be quite a sweet little boy, actually. No, really, I used to be like the Harry Potter of venture capital, and now I'm more like—

Yeah, you lost the glasses.

Lost the glasses, and I kind of became more like Piers Morgan, if you know Piers Morgan in the UK. Highly opinionated. But a question that I have is this: I said if you're a startup, it's never been harder to hire great talent. OpenAI and Anthropic have such a carrot-and-reward mechanism in front of you, especially with impending IPOs, that the best talent just wants to go there, and talent follows talent.

You've seen the founder of Monzo, a multibillion-dollar bank in the UK, go there from YC as a partner. Matt Clifford, the founder of EF, which is a multibillion-dollar company—he should be prime minister, and he's going to join Anthropic. Am I wrong that this is the hardest time ever for startups to hire because the prizes of Anthropic and OpenAI are so great?

5. AI Changes Hiring And Execution

My favorite type of person to hire is a CTO of another company. When we were 21 people at Speechify, 18 of the folks at the company were previously either CEO, CTO, or VP of engineering at their last company.

Anthropic—I have never seen a company like this—hires so many CTOs of publicly traded companies and other successful startups.

Workday was one of them.

The reason is that they build the best, most beloved product for engineers in the history of the world, so it’s easy to hire CTOs. By the way, they hire many more CTOs than CEOs because CTOs are the ones who get the most excited about this product.

They’re the fastest-growing company ever, especially at the scale that they are, so they’re going to keep growing. You had this very condensed period—like fireworks of growth—in both of those companies. It’s very hard to hire.

But remember, they’re hiring people whose annual compensation needs to be $15 million a year minimum. What startup is hiring someone and paying them $15 million a year? You’re not. You’re a seed founder. That’s not someone you’re going to hire.

And so I will push back against it. The competition for growth-stage companies hiring exceptional leadership talent is more difficult. For seed companies, I would say it’s the easiest time ever because the impact of even just the founder on their own is bigger because they can orchestrate agents. The same thing is true for hiring.

One thing that we have changed about our hiring in the last 6 months is that we really cared that you read a ton of textbooks about software engineering and that your handcrafted code was amazing. I still care that you read a lot of textbooks about software engineering and understand it, but the thing I care about the most today is technical aptitude and raw technical intelligence because I know that we could teach you everything else in 6 months. You could be a machine.

And so we hire a lot of math Olympiads and LeetCoders, Kaggle award winners, and people who studied physics and math. They might not have coded before. I just need the hunger, the work ethic, and the intelligence. Anyone can become so good so fast now.

The pool for hiring exceptional talent is bigger than ever before. Duolingo did this really well. They love hiring college grads and then coaching them. I wouldn’t say that it’s harder to hire than ever before for seed companies. Seed companies now—almost anyone can be someone that you hire if they’re smart and hardworking because you can teach them very fast.

What is more challenging is hiring for growth companies because you’re fighting with absolute juggernauts.

And you know, the growth company…

So it’s challenging for us.

Yeah.

Why do you think it’s hard to hire a really good salesperson?

I totally get that. I will see comp packages in the $50 million-plus range, by the way.

Yeah, exactly.

$15 million is child’s play.

With the greatest respect, I will even see $15 million on the table for comp packages for seed companies today. That is the dislocation that I think—

Wait, wait, wait. Sorry. This is a seed company that has raised how much money, at what valuation?

You’ve got to understand, a seed round today will be a $150 million or $200 million raise. There are several of them. There are 30 or 40 companies that, at seed, have raised $100 million to $300 million.

And this is a company with a guy who’s 1 year out of university?

No, no, no. This is a guy who’s probably spent 4 years at OpenAI or spent 4 years at DeepMind.

All right.

No, no, no. This is a guy who’s probably spent 4 years at OpenAI or spent 4 years at DeepMind.

So then what about the company with the guy who’s been at university for 2, 3, or 4 years and is now starting a company? Or do you think those people are out of the water now?

No, no, no. I think that’s just a very different world. They’ll raise $10 million seed rounds.

Yeah.

Yeah.

So for the company that you just described that raised a seed round at a $150 million valuation and raised, I don’t know, $20 million—

No, I said it was a $150 million raise.

Oh, I wouldn’t call that a seed round. Maybe that’s the name.

But my point is that the talent is concentrated. The people who really get AI and systems, and have seen the magic inside OpenAI, Anthropic, and DeepMind—

Yeah. I agree with you that if you have a company that’s raised $150 million—

But there are a lot of them.

—at a $500 million to $2 billion valuation, definitely that company should give a $15 million comp package. No question.

And there are a lot of them.

Yeah, that makes perfect sense.

And there are a lot of them. There are 30.

And those 30 take 30 people, and there are 1,000 people now. That’s fucking hard.

What you just described is exactly what used to happen with Google and Meta, let’s call it 6 years ago. If you were really cracked, there was essentially a maximum amount that you could get paid at a company like Google or Meta, and the best way for you to make a life-changing amount of money was to go to a company that was small, be there right from the beginning all the way through, and be a really solid founding engineer at that company.

I think people want more certainty of cash today than upside, which sounds—

No, I think the equation is the same as always, which is that each person has their own equation in their head of how much certainty and how much risk they’re willing to take.

But I think people would rather know that the certainty of $10 million from Anthropic is better than $60 million from that quirky startup they could make.

This is the reason why companies IPO. There are 2 reasons: either you want a ton of money, or you want a lot of credibility in B2B, like Zoom did, or you’re hiring, and the value of the package that you offer is so much better when your stock is liquid.

When we look at that dev team for you today, you said, “I wanted to go in. I want to see how we’re orchestrating agents.” What did you find? What did you learn in that discovery process around agent orchestration internally?

6. AI Reshapes Engineering Work

Inside our AI research team, everybody’s orchestrating agents. When you go lower, into the product-facing things that we build—for example, the platform team, the iOS team, the Mac team, the Chrome team, the web team, or the Android team—these are super-smart folks who have been working in those domains for about 10 years, and they know iOS like the back of their hand. They know Kotlin and JetBrains like the back of their hand, so it’s very easy for them to hand-code things because you’re not dealing with something that’s super new. So why change? It’s hard to change, right?

You just need to force them to change. The best thing is to inspire them. You do a Zoom screen share and show them how the best engineer on the team is orchestrating agents, and they’re like, “Oh, wow, I didn’t know you could even do that.” Then you say, “Yeah, please do it.” You recommend blog posts, books, and Twitter threads for them to read.

What’s the team using: Claude Code, Cursor, or Codex?

Cursor and Claude Code. Those are the 2 most popular. There’s a little bit of Codex usage, but it’s not that big. I would say Claude Code is number 1, then Cursor, and then Codex.

We want you to use as many tokens as possible, in whatever harness is best for you. You mentioned Linear. Linear is amazing. Automatically cutting tickets from Linear is fantastic, and just being able to go into your agents and say, “Okay, I have these 6 Linear tickets. Start on them.”

A good engineer today is an exceptional QA. The AI will make the feature. You will test the feature, see if it’s good, figure out where the edge cases are, prompt it to fix them, and then try to make it as efficient as possible, which is hard to do. Then you need to make roughly 10 really good product and engineering architecture decisions a day.

How do you think about token allocation internally? We’ve seen leaderboards be used, which I think is the most fucked-up form of incentive play.

There are a lot of people who are a lot of talk. I’ll ask for examples, and I’ll read the examples, and they’re like, “You know, I’m doing this, I’m doing this, I’m doing this, I’m doing this.” Then you look, and I’m like, “Eh.”

I think about it in terms of demos. Can we hop on a Zoom call, and will you show me what you built? Then I use it myself, and I’m like, “Wow, that’s amazing.” Or you send me a screen recording of a feature or technology that you built, and I’m like, “Wow, that’s so good.”

We give credit when things get shipped to production for users. Even inside the AI team, if you build something really amazing, we give credit when it gets shipped to production for users. This is part of why Speechify ended up winning. You asked, “How did you build Bigger Labs?” The answer is that we ship to production all the time. That’s how we won. We are not in the theory space.

We are an applied AI company. That’s why we win. And so, if you’re an engineer at Speechify, the analogy I always give people is: imagine that you’re in the milk delivery business, and you make me a beautiful bottle of milk and leave it down the road. The milk will spoil. You have to get it to my door and knock. If you didn’t do that, you get no credit.

If you carry the football all the way to the line but don’t cross over to the end zone, if you don’t kick it into the goal, you get no credit. If you bring the ball just to the rim and don’t put it in the rim, you get no credit. Putting it in the rim means pushing it to production with no bugs, with users actually using it, and then getting feedback.

How many companies do that iteration cycle fast? Almost no one, definitely not with the user base that Speechify has. And so, in the AI team at Speechify, you make some amazing discovery, and we’re like, “Great. Push it to production.” Then you go, “Oh, wait, there’s this QA problem and this QA problem, and if you have this many people using it on the AI serving layer, then you have this other issue.” Cool, you get no credit from me. It’s not in production. I can’t use it on my phone. When I can use it on my phone, I will give you credit.

Yesterday, I had a call with our AI engineering team, and I said, “Listen, the product that we have running for duplex models and AI conversational harnesses is something I’m really excited about, and it’s been moving fast. I want it to move faster. Here are 14 notes that I want.” Then what I always do is get on a Zoom call, flip my computer around to face my phone, and use the product in front of them. We record it, so then they see all the bugs, and I send the recording in the chat.

Someone on our team—he’s 19 years old—sent me a demo this morning from that conversation that solved all of my problems. He was like, “Hey, I was waiting for 3 training runs to finish, so I had a little bit of time while I was waiting. I implemented everything that you asked.” It blew my mind. It was so good. That’s using AI correctly. It’s not a token leaderboard; it’s what you showed in production that was good.

How many companies do you think are actually as token-pilled and AI-centric as we think, in terms of developers?

I think there’s a guy, Jason Jaeger, who used to work at Speechify, and now he has My Tech CEO on Instagram. He’s super funny, and so he makes a lot of videos about crazy CEOs who are all saying, “Use tokens, use tokens.” I think all founders, in some way, have that animal inside of them because they know that it’s the right path. But there is a difference between reality and theory, and you need to make sure that you don’t overdo it.

Do you have any price sensitivity on tokens?

Yeah, of course.

Yeah.

Absolutely. I’ll lose my mind if, to implement a tiny feature, you use 50,000 tokens. Why did you do that? We will let people go if they just go bananas with something for no reason.

Are you able to accurately budget tokens on a model?

Not accurately, but within bounds.

Yeah.

The other thing is that a lot of engineers are—look, you go into engineering because you like optimization. Most engineers are not blind, and it physically hurts them to overspend tokens. Again, I always think that the best way to interact with AI is that you are chatting in the chat, or actually doing it verbally, and you’re essentially pseudo-coding with your words constantly and explaining architecture.

A great example would be someone I know who has no engineering background and wanted to build an app. They built exactly what they wanted. It took them 2 hours, but they needed an API call, and they needed to scrape this website. They basically scraped every single page and every single part of the website, so the bill they got for the scraping was gigantic.

Then I was like, “Why are you doing it like that? Why aren’t you going into the database to this exact URL and scraping that from the URL?” The number of nodes they needed to hit became about 20 instead of 25,000. An engineer will spend their time making sure that the thing is optimized like that. That’s how you build a good database or a good architecture system, or whatever. You do the same thing when you’re interfacing with the agent: you want the agent to take the path of least resistance, not the path of most resistance.

Totally get that, and I agree with you. I think one of the biggest problems is that agents are goal-seeking, and so they—

Yeah.

—they’re like—

It’s all about the target. You need to be good at picking the right target. I think Anthropic published this paper when Fable 1 came out about long-horizon tasks with Fable. The first thing was that it was much better at running a 2-week task, and it could burn $12,500 worth of tokens in 2 weeks and basically make a better model with that. That’s a perfect, amazing way of using tokens. That’s exactly what you want.

What you don’t want is burning 12,000 tokens in the span of 5 hours doing something that’s totally unnecessary and doesn’t make any sense. You need the loops to happen, and then you need to check the result. What you want to build—and Boris, who’s the inventor of Claude Code, talks about this all the time—is all about the loop. You say, “Here is the target. Here’s how you measure the target. Now iterate against the target over and over and over again until you get it.”

What did you not know about building an AI-centric dev team that you wish you’d known?

How useful is it to own your own GPUs?

What was that realization moment? Did you see—

Yeah.

—a bill one day?

The realization moment was when we realized that we had really talented engineers who were essentially moving at 1/7 of the speed they could have if they had compute 1-to-1 with their creativity and ideas.

If you’re a founder listening to this, how should I change my hiring process in a new AI world?

Number 1: functional interviews. Build this, and then you see if they can build the thing, and then you run it through unit tests. The second one is to give them a large code base, even an open-source repository, and have them understand the code base, make changes, and then check what they broke.

They have to be able to orchestrate agents well. If they’re not doing that, it’s kind of not worth having the person. The next thing I’ll say is that it’s more fun to have a smaller team. Having a big team is great, as long as everyone’s carrying their weight.

The way I think about it is, yes, I can have multiple agents running on my computer, or I can have several Slack chats with really smart people who are much bigger domain experts than I am. Basically, that human being is the outcome owner for that task, and they have the agents. As a founder, I can run multiple projects at the same time to a really amazing level of granularity.

I think about moments earlier in the year when my brother Tyler would literally have an alarm to wake up at 3:00 in the morning because he needed to check what the agent was doing at 3:00 in the morning. He’d wake up, make sure it was good, and go back to sleep. You want to babysit your agent basically every 3 hours.

The beautiful thing now is that you can go work out, and the agent will tell you the answer. Then you voice-note back with Speechify what you want it to do next, and that’ll happen. You want people who are essentially that level of addicted. Obviously, that creates massive AI fatigue, so make sure your teams don’t burn out. But you want someone who is that level of excited.

I think hiring for slope more than intercept is more important today than ever before. Said another way, I look for the potential the person has more than I look for where they are today.

When I look at Wispr Flow and Willow, I did this tweet, and I deleted it because I don’t ever want to be sulky and miserable. It’s an amazing thing to build a company. You should be incredibly credited for doing so as an entrepreneur.

But I found WhisperFlow’s product was just getting worse. I said it on Twitter because I honestly just wanted alternatives. I really need this product, and I wanted alternatives. I got 500 different alternatives, and I was like, “Well, talk about the commoditization of a market. That is not one that I want to be in.”

Can you help me understand? Are we seeing the complete commoditization of the WhisperFlow and Willow speech-to-text market for productivity?

7. Speech Products Become Compound

What they came to the market with first was not necessarily their own model. Part of the reason they got worse is that they switched to their own model because it’s a lot more affordable. They had a harness that tied together a bunch of other things. Probably it was DeepL under the, or Deepgram under the hood, with a bunch of optimizations and more products.

Now they’re trying to do notes, and they’re trying to move more into productivity.

And I think they’ll be successful with it, yeah.

100%.

Yeah. So that’s to your point about the compound startup. One of my biggest mentors—

When you look at them, do you not reflect on what we said before—not announcing fundraises, not announcing anything?

They’re the opposite of me.

They've announced everything.

They're the opposite of me.

They announce going to the bathroom.

Correct.

And hence they have, I would say, a bigger brand.

Not in terms of users. If you walk down the street in New York City, way more people will know Speechify than know WhisperFlow, simply by virtue of the fact that we have way more users. But in the tech world, WhisperFlow has a way bigger brand, right? Investors know who Wispr Flow is because they announce. We intentionally don't announce.

But we don't have any competitors. Who are you going to use instead of Speechify to do text-to-speech for your models? The closest thing is ElevenLabs, and we're so much bigger than ElevenLabs. We are unique in our market.

So because WhisperFlow was so public about it, they now have a lot of competition. Peter Thiel says, "Only losers compete." Try not to compete.

What happens to that market? Does WhisperFlow take the majority, and then are there thousands of ankle-biters?

I don't know. I want them in that market, right? I think that market becomes oligopolistic as well. And this, by the way, is another mistake that I made.

I built my own speech-to-text experience that I've been using on my computer for the last 7 years, sideloaded on my iPhone and on my computer. But I figured it's a commoditized product, right? Apple's going to release it instead of the button. It'll be great, and there you go. But Apple keeps not doing it.

If you remember, 2 years ago Apple announced a partnership with ChatGPT that would improve Siri. Nothing happened. That's also the reason why I never went after Siri.

And so now we've launched a product to compete with Siri, we've launched a product to compete with WhisperFlow, and we've launched a product to compete with ElevenLabs because I learned a lesson that I should have learned before, which is the same lesson from ElevenLabs: the way that you win is you offer an excellent product for free, and then you have a wedge, and then you add more and more and more things.

So I don't know what happens with the Wispr Flow space. I just know that if you're a founder, you should always try.

The final one—another one that I get in trouble for, but I stand by strongly—is that I just think the customer support market is a challenging market to really get behind.

Yeah.

You have Sierra and Decagon out in front with the majority of funding and attention. But to say that there are 18 companies that have now raised over $100 million in the last 18 months, there is what I would call the mid-tier, which is your Intercoms, your Talkdesks, and your Crescendos—all these companies that aren't old, but are old enough.

Yeah.

They're 8 to 10 years old, and they're pretty good.

Yeah.

And then you've got Salesforce, Atlassian, and the much older ones. The worst thing about this market is that, for any sophisticated buyer—an Airwallex, a Klarna, a Navan, a technology-facing company—everyone has built their own system.

Yes.

Because they need something sophisticated.

Of course. But why would you pay a tax for it?

What am I missing?

8. AI Agents Need Forward Deployment

The first thing you're missing is that the core product we're offering B2B is the API, not the agents, right? Sierra doesn't have its own model team. They use other people's models because the value of Sierra is the go-to-market. It's Bret Taylor.

And so that's why, if you talk to Monty and Piotr, they'll tell you we're not competitive with Sierra because its main business historically has been the API. That's the first thing.

In the API business, you have Speechify, ElevenLabs, Gemini, Groq, and SpaceX is now in the race. That's kind of it. So that's not that competitive a space compared to the B2B customer support space. Everybody's in that space—Fin, everybody.

I'm not building that product. What can I offer you that's 10 times better than the next person? Not much. On the core API side, I can offer you better quality, faster speed, and 10 times cheaper. Good offering.

But then I also have to offer agents because there are so many pockets of value that have not been unlocked. And unless I am—again, I have this model for leadership—you don't want to be a fat manager who's like a general sitting in the back saying, "Take that hill." You want to be the warrior who runs up with their sword and engages the enemy first.

You need to be the same thing with your product. You need to be the number-one user of your B2C product, and you need to help your customers use your product better. If you do that, you will learn their problems, and then you will figure out what the next product is that you need to offer them.

So unless I have forward-deployed engineers working with my B2B customers, building agents for them using our technology, I will not figure out what the really amazing next innovation across the hill is.

And so you mentioned the right thing, which is that ElevenLabs now has all these partnerships with governments. Government is not exactly customer support. They would have never gotten to governments had they not done a great job in the private sector first.

I agree. ElevenLabs, in addition to OpenAI, is the most integrated AI company with governments right now. That means they figured something out, but you've got to start in something like customer support.

Again, if you're a founder, you need to try. You cannot not try. You cannot give up before you're even in the race.

What will be a bigger company in 5 years, Sierra or ElevenLabs?

I think they're both going to be massive.

Give me 1 name.

Bret Taylor has the best résumé, I think, of anyone in the world, right? I think he started Google Maps, then he was CTO of Facebook, then he was co-CEO of Salesforce. He's on the board of OpenAI, and now he founded Sierra.

I would never try to fight Bret Taylor, and I think the field is so large. We don't understand how big the space for AI agents is—not even close. In the same way that people didn't understand how big the field was for LLMs in 2019, and the same way people didn't understand how big the space was for AI coding agents in 2021, this is the next huge space.

Both those companies are going to be massive.

I think they're playing very different games.

All right.

I think Bret Taylor is actually trying to recreate a next generation of Salesforce. He is absolutely not playing the customer support game. He's moving to pre-sales.

Everything.

He's moving to post-sales.

But neither is ElevenLabs. ElevenLabs has a product that does that, too. That's why I call it AI agents, not customer support.

But I think—

ElevenLabs is not Fin.

Monty's building a very opinionated, voice-centric company. It's voice-centric.

Correct.

Oh, you think Bret Taylor is doing all of it?

I think Bret Taylor is doing all of it.

Put another way, if you use a tool like Sierra, the wedge right now is voice, but the important part is tool calling. ElevenLabs lets you do some tool calling, but that's not the bread and butter.

There was a really good presentation that Bret Taylor did—a screen share of him building a guitar store on Shopify and how he uses Sierra to do customer support and sales and everything else. It was extremely impressive. If you haven't searched for this, you should. Bret Taylor is a big guitar guy.

That is a very different product from what ElevenLabs is doing, and so they're both going to crush. I agree with you on the Sierra conclusion.

What today is a no and in 5 years' time will be like, "Yeah, of course"?

9. Voice Becomes The Interface

The human-computer interface is going to become primarily voice as opposed to a screen. Part of the reason why Google succeeded is that it has a very simple interface. There's a text box and a button. That's it. Anyone can learn how to use it.

The reason why ChatGPT worked as opposed to GPT-3 is because it was also a very simple interface: just chat. There's a text box and a button. You get a response. That's it. The simpler version of that is just having a conversation. I say something, and I hear something in response.

If you use voice AI from ChatGPT right now, it sucks. It's too slow, the LLM is much dumber than the core LLM, and the escalation to the higher-quality LLM is pretty weak.

I think what will happen—and Meta has the right idea, by the way, so go, Chris Cox—is that people are going to be talking to their computer and phone and some wearable constantly throughout the day, and using screens a lot less.

You can buy one: xAI or Meta. Which do you buy?

Meta.

Why?

Elon’s distracted.

Is he distracted, or is he building full-stack? Because, actually, I think he's never been more strategically positioned, and he has an outlet for each of the different products that he's built, and each one feeds the next.

When you look at Zuck and Meta, bluntly, the compute spend that he's producing, the outlet is increased conversion on an ads business, which is the biggest ads business in the world. Seven percent on $240 billion is a lot of fucking money.

Yeah.

But it's actually not in the same quantum league as doing space data centers.

Yeah. So let's take the space data centers out for a second. I think space data centers are a very interesting idea, and what they do really well is let me underwrite a gigantic TAM for my expectation for SpaceX.

It ruins all estimates.

Right? And so, that was a great rabbit out of the hat by Elon in order to pitch investors really well. Let's take that out for a second, and I'm going to talk to you about SpaceX and Tesla like they're one company, because really I'm assessing Elon; I'm not assessing SpaceX as an individual stock.

For data centers, the biggest constraint right now is memory chips, and then very soon it's going to be energy. It's energy a lot of the time. So what do you need for energy? You need energy supply and energy storage. The best energy storage right now actually comes from Tesla. Tesla also has a chip-manufacturing operation that they're doing, basically competing with everyone else. That's going to do really well.

And if you saw that Joe Rogan interview with Elon maybe 2 years ago, he was explaining that the hard part is not building the product; the hard part is building the manufacturing for the physical product. So Elon is number 1 in the world for manufacturing complex items like that. That's very exciting. And the TAM for Elon's companies is bigger.

However, I think that Meta trades—what does Meta trade at right now?—less than SpaceX.

It's less than SpaceX. It's fucking dumb.

And so, I think Meta has more data than anybody else in the world. I think Meta is actually super-hampered by laws like GDPR. If GDPR didn't exist and the other laws in the US didn't exist, Meta would be ripping. They just can't train on their data properly.

And so they'll figure that out at some point in some way. I don't know how, but I believe in Zuck. At the end of the day, I'm a huge believer in founder-led companies. We're talking about 2 of the best founders in the world.

The last thing I'll say: look at Zuck's age and look at Elon's age. Zuck's not going to stop and Elon's not going to stop, but at a certain point, one of them will expire. And so Zuck has 20 extra years. Depending on how long you're investing, I'm younger than Zuck. Let's see what happens.

I think if Zuck expired—

Meta's dead.

No, because you'd have a CEO who comes in and understands. And this may be short-term,

Yeah.

but that's saying we're going to invest more and more in CapEx when we don't have an outlook for it.

Yeah.

You'd actually see stock-price appreciation in the short term. Every time Zuck steps out on the podium and says, "CapEx, CapEx, CapEx," he's hammered for it. Say I'm going to—

But that's why Meta is a good investment right now, because what Meta doesn't have is what Palantir has, which is the Alex Karp effect. Alex is really good at pumping up the P/E ratio of the stock. And Zuck, I agree, is the opposite.

It's the same as Elon. It's the Elon prism.

Same as Elon. Exactly. And so—

Like, if Elon were to be removed—

The intrinsic value of Meta—

he loses 70% of that value.

Exactly.

If Zuck's removed, you definitely don't lose 70%. You maybe lose—I don't think you lose anything. I think you get an experienced exec in who says we're an ads business.

Charlie Munger and Warren Buffett—actually, no, it's Benjamin Graham—have this concept of the cigar butt. What's the intrinsic value of a company? They approach it from an accounting perspective. I think about it from an underlying technology and business perspective.

The underlying asset, the intrinsic value of Meta, is so large in relation to how it's valued in the market today. You're correct: what's the P/E ratio of Meta? 32? Something like that. SpaceX is insane. Tesla is also in the multiple hundreds.

I think that there has to be a correction that happens, unless Elon succeeds with a big vision, in which case he wins.

What are you most excited by?

10. AI Can Solve Orphan Diseases

I'm most excited by applications of AI to pharmacology and biology. I have a family member who has very severe autoimmune neuroinflammation. He's had it for 6 years. I took a blood sample from him every week for 15 weeks, sent it to a lab, sequenced his genome, did proteomics on it to figure out how the proteins are expressing in his body, and ran an RNA analysis each week. Then I compared that to self-reported data on his quality of life and mood every day.

I have 6 years' worth of data on him. I ran it on a GPU cluster, and I found so many things that no doctor could ever tell me. He has a very rare disease. It was an orphan disease because there aren't that many people with it. There's a Facebook group for this disease.

I'm buying basically a $5,000 device you can fit in your pocket, but if you put a piece of hair, saliva, or blood into it, it can sequence your entire genome. I'm organizing meetups with all the people who have this disease to sequence all of their genomes and then compare them all on a gigantic GPU cluster to figure out what epigenetic common thread there is between them. I know I'm going to solve this disease.

It gets even more beautiful because I can then take all the conclusions that I have about it and put them into AlphaFold from Isomorphic Labs. I can design not just the protein that is creating these issues, but the molecule that needs to bind to that protein to either turn it on or off. I can use CRISPR to do the same thing.

I can use a lab like Twist where I can tell it, "I want you to make me this RNA sequence or this DNA sequence," and it can make it for me and ship it to my lab or my house. I can create amazing outcomes with it, and I can simulate all of it on my computer that's SSHed into my GPU cluster in Scottsdale, Arizona. I could cure my brother.

My experience is that, when I was 8 years old, I couldn't learn how to read. My dad had to open a book and read Harry Potter to me, and that's how I learned how to read. When I was 13, I moved to the United States of America, and I didn't speak English. I listened to Harry Potter audiobooks 22 times in a row, and I still have the first chapter memorized.

Then I couldn't get into the private high school that my brother went to and that my sister went to, and I was really bummed. I went to a lower-quality high school. I didn't get into AP US History because I made a bunch of spelling mistakes in my essay, and I couldn't read the passage in time.

I needed to train myself to read the SAT English portion. I wouldn't read the passage; I would read the answers, and then I'd go and hunt for the answer. When I got to college, somehow, by the grace of God, I ended up going to Brown and starting a major in renewable energy engineering because I couldn't do literature. I built a text-to-speech tool that would read all my books to me, and that's why I graduated.

Technology solved my dyslexia, and it solved my ADHD, and it's going to solve my brother's disease. It's already solved my dad's prostate cancer because I figured out, with a bunch of help from other people, how to use GPUs to identify where in his body the lesion was.

That's what I'm excited for: this better quality of life for literally everybody because you have this magical machine that can run a trillion operations per second on as many GPUs as you want, and it can solve problems that we can't.

I find it staggering that still today we have orphan diseases, which is like, "Oh, there's too few people to make it economically viable for us to try and solve." There are hundreds, thousands, low thousands, but low thousands of—

And again, it's the same thing.

Wow.

You just need data, you need compute, and you need to ask good questions. Like I said, 10 good decisions per day, either hypotheses or actual product decisions, and you can solve these problems. Freaking amazing.

Cliff, it's been so great to have you on the show. I much prefer it when it's a discussion.

Talk to you soon.

But before we leave you today, today I wanna tell you about how the first AI law firm, Crosby, helped us close a big sponsor. As you know, some of the biggest companies in the world advertise on 20VC. My British dulcet tones clearly convert well. I was working to close this big sponsor, and they wanted to get through legal review quite quickly to close the deal. Crosby turned redlines around in three hours and caught major issues that would've caused us serious problems in the future. Crosby combines AI, some of the best engineers in the world from companies like Ramp and Stripe, and some of the best attorneys in the world from top 10 law firms. Customers get the best of both worlds. An elite human attorney reviews every contract, but they move incredibly quickly, returning redlines in under four hours. They help the fastest growing companies like Cognition, Ramp, and Clay close deals in hours, not weeks. Learn more at crosby.ai/20VC. If you want to redline NDAs, MSAs, DPAs, and any other procurement contracts faster, go to crosby.ai/20VC. It's speed that you can really trust. While Crosby keeps your numbers sharp, OneMind keeps your customer conversations sharper. Our friends over at OneMind have a hot take. The B2B GTM model we've been using for, oh, the last 20 years, it's collapsing. Predictable revenue isn't so predictable, and buyers are just tired of explaining themselves at every handoff between SDRs, AEs, CSMs, and support. You feel it in your board reporting. Your sellers feel it in their coverage. Your buyers feel it as they wait for answers. Well, enter OneMind and their GTM superhumans. HubSpot, Alteryx, and ZoomInfo are just three of the tech companies using OneMind superhumans to qualify buyers, ride along with sales reps, and coach customers. Whether you're ahead of target or under pressure to increase bookings ahead of hiring or behind on target even and need every rep to hit quota, OneMind is the solution. See for yourself at onemind.com. That's onemind, M-A-N-D, .com. While OneMind helps convert more buyers, AlphaSense helps you spot what matters next. We used AlphaSense on an investment that helped us close an $8 million deal. $8 million, baby. That's a lot of money. That's why I'm genuinely excited to have them as a partner on 20VC. AlphaSense combines AI with one of the world's deepest libraries of market intelligence, including expert interviews, broker research, earnings calls, company filings, and real-time news. Every answer is grounded in this incredibly trusted evidence and fully traceable to the original source, which is so important. So you can make really high conviction decisions with confidence. But the best part, they're building super analyst, an always on AI analyst. So instead of starting your day with another search, you'll start with work you already done, your coverage monitored, the important developments surfaced, and your investment brief already waiting for you. See for yourself. Head to alphasense.com/20VC. That's alpha-hyphen-sen-