Sep 5, 2026 · 1h 4m · 20vc
How to Build Your Own Data Center & Why Every Startup Should Do It
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this 20VC episode, Speechify founder and CEO Cliff Weitzman breaks down the contrarian economics of purchasing physical GPU clusters, reflects on his strategic B2B missteps against rivals like ElevenLabs, and explains how autonomous coding agents are revolutionizing startup engineering culture.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 20.5% of the talking time here. How this is scored →
Harry as informed peer 5.3 Guest teaching 5.0 Guest disagreement 2.7 Harry pushing back 3.9
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 59:03 Weitzman rejects Stebbings' thesis that Meta is better off without Zuckerberg
Weitzman completely dismisses Stebbings' premise that Zuckerberg's heavy CapEx hurts Meta, flatly refusing the notion that replacing him would increase market value.
Hardest push from Harry ▶ 17:13 Stebbings attacks Weitzman's GPU procurement overheadStebbings aggressively challenges Weitzman's operational discipline, insisting that water cooling and freight logistics are a wasteful distraction while 11Labs executes faster.
Biggest teaching moment ▶ 10:48 Weitzman explains NVIDIA's bank-underwritten GPU secondary marketWeitzman exposes a sophisticated financial mechanism where NVIDIA, Goldman, and Blackstone guarantee GPU collateral floors, catching Stebbings completely off-guard.
Harry holds his own ▶ 25:28 Stebbings presses Weitzman on entering B2B late against Sierra and 11LabsStebbings leverages his deep market knowledge of venture backing and government contracts to confront Weitzman with the harsh reality of the 'Postmates effect'.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| 20VC Title Sequence and Guest Montage | 4 | 6 | 1 | 2 | Stebbings opens by probing why Speechify is paying massive premiums to buy GPUs months early. Weitzman delivers an extended, highly technical masterclass comparing the cost of renting versus owning compute, InfiniBand co-location, and token economics. | |
| GPU Obsolescence, Secondary Markets, and Depreciation | 5 | 5 | 2 | 4 | Stebbings pushes back on hardware lock-in and rapid chip obsolescence. Weitzman reframes by explaining how older cards are repurposed for lower-latency inference while newer chips handle cutting-edge training runs. | |
| GPU Procurement Logistics and Colocation Infrastructure | 3 | 7 | 1 | 2 | Stebbings asks how Weitzman models forecasting for GPU buying. Weitzman details seasonal demand curves, colocation facilities, and reveals NVIDIA's secondary market underwriting deal with major financial institutions. | |
| Debating the NVIDIA Bubble and Intrinsic Compute Value | 5 | 5 | 2 | 3 | Stebbings queries whether NVIDIA's business is circular financing bubble behavior. Weitzman reframes compute as having hard intrinsic value measured in FLOPS per second, unlike speculative crypto, before walking through real procurement logistics. | |
| The Compute Advantage: Why Ownership Fuels AI Labs | 6 | 6 | 5 | 7 | Stebbings directly attacks Weitzman's strategy, arguing that price optimization is foolish compared to the operational headache of freight insurance and water cooling when 11Labs is running fast. Weitzman counters firmly that co-located hardware and jumping queue for Rubens provides unmatched model-building speed. | |
| Sourcing Training Data and Model Unit Economics | 6 | 4 | 2 | 3 | Stebbings brings industry knowledge regarding specialized models and synthetic data providers like Mercor and Fireworks. Weitzman agrees and breaks down data contract dynamics, legal indemnification, and Speechify's low-cost API advantage. | |
| The B2B Blindspot: Cliff's Biggest Strategic Mistake | 4 | 4 | 1 | 3 | Stebbings asks a pointed question about whether 11Labs leapfrogging Speechify was Cliff's fault. Weitzman takes full accountability, candidly unpacking his false assumption that APIs would quickly commoditize. | |
| Speechify's B2B Expansion and Market Structure | 7 | 5 | 5 | 7 | Stebbings forcefully argues that entering B2B now is a mistake due to incumbents like 11Labs and Sierra capturing the market. Weitzman rejects the winner-take-all premise, citing historical precedents like Anthropic vs OpenAI and Speechify's massive B2C scale. | |
| Category Winners, Incumbent Fumbles, and Staying in the Race | 6 | 4 | 4 | 5 | Stebbings contends that top players capture all the value and doubts 11Labs will fumble the bag. Weitzman pushes back, arguing large incumbents inevitably drop balls across expanding product surfaces and that staying in the race is essential. | |
| Talent Wars: Hiring Technical Athletes in the AI Era | 6 | 5 | 4 | 5 | Stebbings argues it has never been harder for startups to hire given compensation packages at frontier labs. Weitzman pushes back by distinguishing seed from growth companies, explaining how early-stage teams can leverage raw technical athletes and agentic workflows. | |
| Compensation Dislocation and Risk Appetite in Tech | 7 | 3 | 5 | 6 | Stebbings challenges Weitzman on seed-stage reality, citing $15M comp packages and mega seed rounds. The two clash on the definition of a seed company and whether employee risk appetite has fundamentally shifted away from equity. | |
| Driving Agent Orchestration and a Production-First Culture | 5 | 5 | 2 | 3 | Stebbings inquires about internal dev agent adoption. Weitzman delivers an operational breakdown of forcing legacy engineers toward Claude Code and Cursor, contrasting shipping to production against useless vanity metrics. | |
| Balancing Token Budgets, Efficiency, and Agent Loops | 4 | 5 | 1 | 2 | Stebbings questions whether companies are actually token-pilled and how token budgeting is handled. Weitzman clarifies the engineering discipline needed to avoid brute-forcing token burn and emphasizes iterative task loops. | |
| Managing AI Dev Teams, Babysitting Agents, and Hiring for Slope | 3 | 6 | 1 | 2 | Stebbings asks how founders must adapt hiring in the agentic era. Weitzman explains the necessity of hiring high-slope talent who relentlessly babysit autonomous agents around the clock. | |
| Stealth Execution vs. Loud PR and Challenging Legacy Giants | 6 | 5 | 3 | 4 | Stebbings questions whether niche voice tools like Whisperflow have commoditized and questions Speechify's quiet PR posture. Weitzman argues Whisperflow's loud PR invited intense competition, whereas Speechify quietly owns consumer market share. | |
| Customer Support AI Markets and Speechify's API Strategy | 6 | 5 | 2 | 4 | Stebbings critiques the AI customer support market as overcrowded and notes top tech companies build their own solutions. Weitzman agrees on support agents but clarifies that Speechify's play is selling the underlying low-cost API infrastructure. | |
| Comparing Sierra and ElevenLabs in the AI Landscape | 7 | 3 | 3 | 5 | Stebbings asks Weitzman to compare Sierra and 11Labs over a five-year horizon and argues Sierra is recreating Salesforce through tool-calling rather than mere voice. Weitzman concedes the point and praises Bret Taylor's product direction. | |
| Quickfire Debate: Investing in Meta vs. SpaceX and Founder Dynamics | 7 | 4 | 6 | 6 | A lively debate erupts over investing in Meta versus SpaceX. Stebbings claims Meta's stock would rise if Zuckerberg left due to reduced CapEx discipline, which Weitzman emphatically rejects using Benjamin Graham valuation principles. | |
| AI Breakthroughs in Biology, Orphan Diseases, and Personal Health | 3 | 7 | 1 | 1 | Stebbings asks what Cliff is most excited about. Weitzman delivers an emotional monologue on using personal GPU clusters to sequence genomes, analyze proteomics, and discover treatments for his brother's orphan disease. |
Statements from this episode (48)
Weitzman: Speechify Simba 3.2 ranks number one globally for quality
“The newest Speechify Simba 3.2 model is ranked number one in the world for quality. Above all the frontier labs, 10 x more affordable and stuff like 11 Labs.”
Weitzman: Renting an H100 GPU costs 1.5 times more than buying it
“If I was to buy an H 100 for, let's say, 30,000 dollars, that's how much the, kind of, a single card would cost. If I wanted to rent an H 100 for one hour spot instance from GCP, it could cost me five dollars. If I rented it from, like, You know, Azure or AWS,…”
Weitzman: Large-scale AI training requires co-located memory and GPU clusters
“If you want to do large scale training like we do, You need the memory to be co-located with a large cluster of GPUs. I can't just rent from Google or Microsoft or even base 10 and run the size of training that I want because I need a gigantic memory card next…”
Weitzman: Speechify still runs inference operations on older NVIDIA K80 GPUs
“At Speechify, we still use K-eighty's for a lot of specific operations for inference, and we use older models of GPUs constantly.”
Weitzman: Speechify will purchase multiple racks of NVIDIA Rubin and B300 GPUs
“So we'll buy multiple racks of Rubens. And on top of that, we'll buy B 300, which are like the newest form of Blackwells because we can get them earlier.”
Weitzman: ElevenLabs builds and operates its own physical GPU clusters
“11 Labs, I think Piotrek at 11 Labs literally bought a bunch of GPUs early, 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.”
Weitzman: NVIDIA guarantees 25% GPU buybacks with Wall Street lenders
“So I believe earlier this month, NVIDIA did a huge deal with Blackstone, BlackRock, Apollo, and Goldman Sachs. And they said, listen, we want more people to buy more GPUs. We're going to underwrite for you up to 25% the value of a GPU. That if you lend money t…”
Weitzman: Networked GPUs have intrinsic utility and store value globally
“But a GPU, it has intrinsic value. Like, 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. …”
Weitzman: Speechify will pay $100k extra monthly for faster GPU delivery
“We're very willing to pay a hundred K per month extra to get them earlier.”
Weitzman: The biggest cost of delayed GPUs is unutilized data center rent
“The most expensive part of a delivery of a GPU is if it's late, I'm still paying rent for that data center space.”
Weitzman: Most data centers lack approved liquid cooling for NVIDIA Rubin chips
“Rubens are liquid cooled. But most of these data centers don't have liquid cooling installations already approved”
Stebbings argues buying GPUs is a mistake due to marginal cost savings
“It is a mistake to price optimize and to spend the money to buy it versus to rent it because I get you on the optimization, but you're not saving 10 times more. It's .5 X more per year.”
Weitzman: Buying Rubin GPUs yields one year advantage over cloud renters
“If I want a Reuben, which is like these much faster GPUs, I'll get it faster if I buy it than if I wait for Google to buy it, and then there's other people in front of me in line. So I'm gonna skip the queue by like a lot, and then I'm gonna have like a year o…”
Weitzman: Speechify engineers concurrently run 5 to 18 autonomous coding agents
“Our engineers, really what I'm looking for is 10 really good decisions per day, which is very tiring, not like optimizing the random parts of the code. And each one has like, you know, five to 18 agents running at any point in time, doing long horizon tasks on…”
Weitzman: A top Speechify engineer focuses entirely on generating synthetic data
“Like I have one of my best engineers right now is not even writing models. He's making synthetic data sets to train models.”
Weitzman: AI data marketplaces are high-risk because revenue is not recurring
“The first problem to understand about the data marketplace is it's not ARR, right? It's not annual recurring revenue. It's one-time deals every single time. So the buyer of the data is not required to buy it from you again. So it's a very risky business.”
Weitzman: ElevenLabs leapfrogged Speechify by selling B2B while Speechify focused B2C
“11 Labs, huge credit to them, leapfrogged us because they sell to B to B. Well, historically we've only sold to B to C.”
Weitzman: Speechify's Simba 3.2 API costs $10 per million characters
“11 charges 10, a hundred dollars per million characters. The OpenAI model on the benchmarks cost a 196 dollars per million characters. So ours, when we sell it to other B to B companies now, we just launched our API, Simba, 3.2, it costs 10 dollars per million…”
Weitzman: Agents is ElevenLabs' best product because buyers are C-suite executives
“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, executive in the company.”
Stebbings: Speechify entering B2B is a strategic mistake
“I think it's a strategic mistake for Speechify to go to B to B.”
Stebbings: ElevenLabs has secured government buy-in across all major Western democracies
“Now, 11 Labs is an unstoppable machine at this point, to the point where it has government buy-in across All of the large major Western democracies.”
Weitzman: Speechify holds 98% of B2C text-to-speech App Store installs
“The first one is if you go to the App Store and you search text-to-speech, Speechify has 98% of the installs in text-to-speech for B to C.”
Stebbings: OpenAI fumbled voice AI due to poor management and hiring
“Well, I think with all candor, that's because of incredibly poor management. And hiring, and that was theirs to take, and they fumbled the bag across every spectrum.”
Weitzman: OpenAI fumbled the AI coding space alongside voice
“LLMs are the core, and by the way, they also fumbled AI coding. Now they're trying to cache because it's such a big space.”
Weitzman: 18 of Speechify's first 21 employees were former executive leaders
“When we were 21 people at Speechify, 18 of the folks at the company were previously either CEO, CTO, or VP of engineering at the last company.”
Weitzman: Seed-stage hiring is easier than ever, but growth-stage hiring is harder
“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…”
Weitzman: Speechify now prioritizes raw intelligence over handcrafted coding ability
“So one thing that we have changed about our hiring in the last even six months Is 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 softwa…”
Stebbings says Chief Revenue Officer compensation packages can exceed $50 million
“I will see CRO packages in the fifty million plus range, by the way.”
Stebbings sees compensation packages up to $15 million at seed-stage companies
“I will even see fifteen million dollars on the table for comp packages for seed companies today.”
Stebbings: 30 to 40 startups have raised $100M to $300M seed rounds
“I mean, there's 30, 40 companies that at seed have raised a hundred to three hundred million.”
Stebbings: Tech talent prefers guaranteed cash over startup equity upside
“I think people want more certainty of cash today than upside”
Weitzman: A good engineer today is essentially an exceptional QA tester
“Really a good engineer today is just an exceptional QA, right? The AI will make them feature. You will test the feature, see if it's good, you'll figure out where the edge cases are, you'll prompt it to fix it, and then you try to make it as efficient as possi…”
Weitzman: Speechify will fire employees who waste tokens unnecessarily
“We will let people go if they just go
Bananas with something for no reason.”
Weitzman: The best way to use AI is verbal pseudocoding and architecture explanation
“I always think that the best way to interact with AI is you are
Chatting in the chat, or actually doing it verbally, and you're essentially pseudocoding with your words constantly, and you're explaining architecture”
Weitzman: Engineers moved at one-seventh potential speed without sufficient compute
“The realization moment was when we realized that we had really talented engineers who were essentially moving at one seventh of the speed they could have if they had to compute one to one with their creativity and ideas.”
Weitzman: Engineers who cannot orchestrate agents are not worth hiring
“They have to be able to orchestrate agents well, and if they're not doing that, It's kind of not worth to have the person.”
Weitzman: Whisperflow degraded product quality by switching to their own cheaper model
“Part of the reason they got worse is they switched their own model because it's a lot more affordable.”
Stebbings: Tech companies like Klarna build their own customer support AI
“And then 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.”
Weitzman: Sierra relies on third-party models and Bret Taylor's go-to-market
“Sierra doesn't have their own model team. They use other people's models, right? Because the value of Sierra is the go to market. It's Brett Taylor.”
Stebbings: Bret Taylor is building Sierra to recreate the next Salesforce
“I think Brett Taylor's actually trying to recreate the next generation of Salesforce. He is absolutely not playing the customer support game. He's moving into pre-sales.”
Weitzman: Sierra's core value is tool calling, unlike ElevenLabs
“If you use a tool like Sierra, the wedge right now is voice, but the important part is tool calling. 11 Labs lets you do some tool calling, but that's not the bread and butter.”
Weitzman predicts voice will overtake screens as primary interface within five years
“Human computer interface is going to become primarily voice, as opposed to a screen.”
Weitzman: ChatGPT voice AI is too slow and uses a dumber model
“If you use voice AI from ChatGPT right now, It sucks. It's too slow. The LLM is much dumber than the core LLM. The escalation to the higher quality LLM is pretty weak.”
Weitzman: Meta would be ripping without GDPR and US data regulations
“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.”
Stebbings predicts Meta's stock price would rise if Mark Zuckerberg departed
“If Zuck expired, Meta's stock price would increase.”
Weitzman: Alex Karp inflates Palantir's P/E ratio while Zuckerberg depresses Meta's
“What Meta doesn't have is what Palantir has, which Palantir has the Alex Karp effect. Alex is really good at pumping up the PE ratio of the stock, and Zuck, I agree, is the opposite.”