Founder of Speechify: Why every startup should build its own data center

From being rejected by all cloud service providers to building a self-hosted GPU cluster that trained the world's top-ranked speech model — Cliff Weitzman tells a true story about computing power, costs and strategic missteps.

1. A Counterintuitive Decision: Buy GPUs Instead of Renting

In 2022, Cliff Weitzman, founder of Speechify, made a decision that puzzled many people: he spent tens of millions of dollars purchasing NVIDIA GPUs to build his own data center.

At that time, almost all startups were renting cloud services. But he and his brother did the math:

The purchase price of one H100 graphics card is about 30,000 US dollars. If you rent it from AWS or Azure, it costs 3.5 to 5 US dollars per hour. Multiply that by 24 hours and 365 days, and it will cost 35,000 to 50,000 US dollars a year.

"The cost of renting for one year is 1.5 times the purchase price. And the hardware usually has a normal service life of three years, and can be used for another ten years or so after the warranty expires."

From a purely financial point of view, purchasing is obviously more cost-effective.

But what really prompted him to make the decision was the way his engineers behaved.

"We found that the engineers at Speechify were extremely frugal. When using GPUs, they always thought: 'Oh my god, I'm making the company lose tens of thousands of dollars.'"

He drew an analogy: Imagine you are Michael Jordan, you want to enter the NBA, but you have to pay 20 dollars every hour to train at a basketball training center. That's terrible. You want a court that you can use anytime — in fact, you want one in your own home.

So they bought their first batch of GPUs.

2. Training and Inference: Two Completely Different Needs

Many people don't understand, why not just rent cloud services directly?

Weitzman explained that the key lies in distinguishing between training and inference.

During training, you have a hypothesis and want to get the answer as soon as possible. Every minute that passes, you are competing with others. Having GPUs that are orders of magnitude faster is a huge advantage.

During inference, for example, when Speechify converts text to speech, you don't need the top-tier graphics card. An older GPU can also complete the task within 100 milliseconds. "I can use these older GPU models for inference operations at any time."

More importantly, training requires connecting a large number of GPU clusters with massive memory. "It is impossible for me to only rent software from Google and Microsoft and run it at the scale I want, because I need a very large scale."

They need to store data on adjacent memory cards so that all GPUs can access it at the same time. This architecture cannot be met by rental solutions at all.

3. The Reality of Supply Chain: Insurance, Cooling and a Truck Full of GPUs

Building your own data center sounds simple, but the actual operation is full of details.

First of all, you need to rent space in a data center. The data center provides network connections, power, and on-site engineers — they unload the equipment from the truck, install it, and fix any problems that arise.

But the biggest limiting factor is power.

"The data center itself provides network connectivity, it provides power, and that's actually the biggest limiting factor right now."

The second is cooling. Right now, liquid cooling technology has become an inevitable choice, because air cooling can no longer meet the demand. But many data centers have not yet obtained approval to install liquid cooling systems. You need to buy the so-called "side carts", input the liquid cooling information yourself, and then pay the data center staff to install them.

Then there is transportation.

"Somewhere in the United States right now, there's a truck loaded with GPUs. Its value is equal to that of a house. It's supposed to arrive at my data center, right? If the truck gets into an accident, or the humidity is too high, or the GPUs get flipped over, I feel like I've lost a huge amount of assets."

Insurance is very important. The importance of amortization amounts is also very prominent.

He also shared a real negotiation story: they ordered a batch of GPUs from France, but they were delayed for a long time and did not arrive. So he said to the supplier: "Listen, I have a better deal. I'm going to cancel our contract because you didn't deliver on time."

"Remember, I'm paying rent for the space I use in the data center. If the graphics cards are delayed, I still have to pay rent for that space."

That's where the pressure comes from. In the end, the supplier shipped the equipment as promised.

4. The Math of Computing Power Allocation: 80% Self-Hosted, 20% Rented

After having your own GPU cluster, how do you allocate computing power?

Weitzman's model works like this:

Take November as the benchmark (100% usage), October will reach 140% because that's their most important month. December may only be 80%, because everyone is at home, but not necessarily studying or working.

"I can only take up 20% of the normal usage. Let me buy that part, because it's the most cost-effective option."

"I will take up another 25% of the bandwidth usage, and sign long-term contracts with large cloud service providers. For the remaining part, I will rent the so-called 'instance resources' from those large cloud service providers."

In this way, he not only has control over the core computing power, but also retains the flexibility to cope with peak loads.

5. The Biggest Strategic Mistake: Missing Out on B2B

During the interview, Weitzman repeatedly mentioned a decision that he regretted.

In 2022, he met the founders of 11 Labs. At that time he lived in London and was deeply impressed by the two founders. But he made a mistake.

"I didn't like their strategy. Because I thought that over time, the API for text-to-speech functionality would gradually be standardized. You will soon reach the point where you can call that API. Check it on your computer, then on your phone, so what else are they selling now?"

He thought at the time that APIs would become a commodity that was not worth investing in.

"So I didn't want to get into that business, but I made a serious mistake."

What he didn't realize was that the significance of AI labs like 11 Labs lies in continuous innovation. The first product you launch is your wedge, and then you can keep adding new things:

"If you are using text-to-speech functionality and you have created the best model in the world. Now you can try to use other voices. Now you can add emotional factors. Now you can add voice cloning features. Now you can convert speech to text."

"That was my mistake. I thought having an API, this product was actually a bad strategy because I thought it might become something that could be commoditized. I also forgot the core argument about Silicon Valley, which is that Silicon Valley is always innovating nonstop."

He concluded: "The best way to fail is to not compete."

6. The Logic of the Data Market: One-Time Transactions, Not Recurring Revenue

When asked about the data market, Weitzman gave a sober judgment.

"The first thing to understand about the data market is that it is not an orderly system. This is not annual recurring revenue. Every time it is just a one-time transaction."

Therefore, data buyers do not need to buy the same data from you again. "So this is a very risky business."

He took the early development of companies like Mercor as an example: they failed to get a lot of capital immediately because investors were very worried about this situation.

But data also has its unique value. "You need to provide compensation for these losses. That's one of the reasons why they buy from you instead of sourcing it themselves."

"This is a very good business if you can do it well, but you need to be a very excellent operations expert. You have to be extremely fast."

7. The Most Important Personal Story: Curing Rare Diseases with AI

At the end of the interview, Weitzman shared a field that really excites him — applying AI to pharmacology and biology.

One of his family members has suffered from severe autoimmune neuroinflammation for six years. He took blood samples from him every week for 15 consecutive weeks, sent them to the lab for genome sequencing, proteomics research, and RNA analysis, then compared the results with self-reported quality of life data and mood changes.

"Every day I have about six years of data about him. I run it on the GPU cluster and I find out many things that no doctor could tell me."

This disease is called an "orphan disease" because not many people suffer from it. There is a Facebook group for people with this disease.

"So I'm organizing some gatherings. All people with this disease need to have their genomes sequenced. Then we run comparative analysis on all this data on a huge GPU cluster to find out the epigenetic factors. There is a common thread between all of them."

"I know I will cure this disease. In the past, I never had the chance to do something like that."

Then he told his own story: at the age of 8, he couldn't read. His father had to open a book and read it aloud to him. At the age of 13, he moved to the United States and couldn't speak English. He listened to all 22 Harry Potter audiobooks, over and over again, until he could memorize the content of the first chapter.

"I went to a mediocre high school, and I didn't choose to enter the AP program, because I modified a lot of spelling mistakes by myself. I also had to train myself to read the English section of the SAT. I couldn't read the text directly. I would read all the answers first, then go find the real correct answer."

Later, he was admitted to Brown University, majoring in renewable energy engineering — because he was not suited for a career in literature.

Technology solved my dyslexia problem, and also alleviated my ADHD symptoms. It will cure my younger brother's disease. Moreover, this method has already successfully cured my father's prostate cancer."

"That's what excites me — the quality of life of everyone has actually been improved. Because you have this amazing machine that can process a trillion operations. You can run it with as many GPUs as you want, and it can solve problems that we can hardly solve."

"It's shocking that we still have some orphan diseases today. That means the number of people affected is too small to be economically viable. We have the ability to try to solve this problem."

Conclusion

Cliff Weitzman's story is not a speech about technology, but a story about how to use computing power to solve real problems.

He built his own data center not to show off his skills, but because renting was too expensive, too slow, and too restricted. He missed out on B2B not because he wasn't smart enough, but because he underestimated the power of continuous innovation. He devoted himself to pharmacology not because of business opportunities, but because his family members were suffering.

"You need data, you need computing power, and you need to ask good questions. Like I said, make 10 right decisions, once a day."

"Or decision-making problems about actual products, you can solve these difficult problems. It's absolutely amazing."

This article is from the WeChat official account "Postwave Evolution Planet", written by Mark, and published with authorization from 36Kr.