Buying vs Renting Nvidia GPUs for AI Startups

Cliff Weitzman, founder and CEO of the text-to-speech company Speechify, pays Nvidia's hardware partners an extra $100,000 per month to get the chips four months earlier than everyone else (03:46). A GPU, short for graphics processing unit, is the chip that trains and runs AI models. Paying a premium to jump the line sounds like the kind of decision a company makes when it has more money than sense. Weitzman's explanation is a basketball analogy about ownership and access.

"Well, that sucks. You want one that you can go to whenever you want to. In fact, you want a hoop in your house." - Cliff Weitzman [04:27]

The point is about friction. Speechify used to rent GPUs, and Weitzman noticed his engineers rationing their own experiments because every hour on a rented chip showed up as a cost the company would notice. Renting turns compute into something an engineer has to justify. Owning turns it into something an engineer just uses.

The math behind that decision is more concrete than the metaphor suggests. An Nvidia H100 GPU costs about $30,000 to buy outright. Renting the same chip from a major cloud provider, at roughly $3.50 to $5 an hour, adds up to $35,000 to $50,000 over a year, which works out to about 1.5 times the purchase price (06:07). The hardware carries a three-year warranty, but Weitzman expects it to keep running well past that, since a GPU sitting in a climate-controlled data center takes none of the wear a car engine takes on the road (12:39). Large-scale model training adds a second reason to own rather than rent: training needs a big memory pool sitting right next to a large cluster of chips, and cloud providers will not build that configuration for a single tenant the way an owned cluster can (06:32).

Nvidia has also been busy turning GPUs into something closer to a financial asset than a depreciating expense. Weitzman describes a deal, done this year with Blackstone, BlackRock, Apollo, and Goldman Sachs, in which Nvidia agrees to underwrite up to 25% of a GPU's value if a bank lends money against that chip as collateral and the borrower defaults (13:38). That guarantee gives banks a reason to lend against hardware at better rates, which is the same trick Elon Musk used at SolarCity, where he got Morgan Stanley and Merrill Lynch to amortize solar panel costs over 30 years by treating the panel itself as loan collateral (14:05). The effect, in both cases, is a price floor. If a GPU has a guaranteed resale value, it behaves less like a wasting asset and more like a bond.

None of this is free of headaches, and Weitzman is candid about the logistics that come with owning physical hardware. Racks get delayed in transit, and delay is the most expensive part of the whole process, because a company keeps paying data center rent whether or not the chips have arrived (16:50). Weitzman describes calling a vendor in France to renegotiate a late shipment, and paying extra to skip the delivery queue entirely once a better offer came along (16:36). A truck carrying an unreleased Nvidia rack can be worth as much as a house, which means it needs its own insurance policy (17:09). And the newest chips, like Nvidia's Rubin line, are liquid-cooled, which most data centers are not set up to handle, so Weitzman's team had to source and install a separate cooling unit before the racks could even be turned on (18:06).

The Mistake That Let a Rival Win

The hardware conversation sets up a more personal one. Weitzman met the founders of ElevenLabs, a rival AI voice company, in 2022, and decided against building an API business because he expected text-to-speech APIs to become commoditized, cheap and interchangeable, over time (24:34). ElevenLabs made the opposite bet, and it paid off: an API became their wedge product, which led to voice agents, which led to partnerships with Western governments (25:29). Weitzman now calls the decision the worst one he has made as a founder.

"It was the biggest strategic mistake I made in the history of Speechify." - Cliff Weitzman [24:17]

What he missed, in his own telling, is that the first product an AI lab ships is rarely the business. It is the entry point that earns the right to sell everything that comes after it (24:50). Speechify still dominates its original market. It holds 98% of installs in consumer text-to-speech apps and has served more than 770 billion words to users (27:45). But dominance in one lane did not stop a competitor from building an entirely new lane next to it. Speechify has since launched its own API, called Simba 3.2, priced at $10 per million characters, compared with $100 from ElevenLabs and $196 from OpenAI's benchmark model (23:58). Weitzman frames the decision to enter B2B late as a cost of staying out of a race rather than a cost of losing one.

"The best way to lose is not to be in the race. Be in the race." - Cliff Weitzman [28:55]

Hiring in the Age of Agents

The same competitive pressure shows up in hiring. Weitzman says compensation for elite researchers at frontier labs like OpenAI and Anthropic now starts around $15 million a year, a number that Harry Stebbings suggested only a small number of heavily funded seed companies, some raising $150 million to $300 million rounds, could realistically match, though Weitzman was skeptical that such rounds should even be called seed rounds (33:21). For most seed-stage founders, he argues, that competition barely matters, because the pool of people worth hiring has actually gotten larger. A founder no longer needs someone who has memorized a decade of software engineering textbooks. Raw technical aptitude and a willingness to learn fast now matter more, since AI tools can close the skills gap in months (33:45). Weitzman says 18 of Speechify's first 21 employees had already been a CEO, CTO, or VP of engineering somewhere else, and that Anthropic itself now recruits more CTOs than CEOs, because CTOs are the ones who get most excited about a product built for engineers (32:40).

What has changed inside Speechify's own engineering team is the definition of the job itself. Coding tools like Claude Code and Cursor now write the feature. The engineer's job is to catch what is wrong with it.

"And then really a good engineer today is just an exceptional QA. The AI will make the feature. You will test the feature, see if it's good." - Cliff Weitzman [38:59]

QA here means quality assurance, the discipline of testing a feature for bugs and edge cases before it reaches a user. Weitzman says his AI research team runs five to 18 agents at once on long-running coding tasks, and a good engineer's real job has become making about ten sound product and architecture decisions a day rather than typing code line by line (20:12). Progress is not judged by how many tokens a team burns or how a leaderboard ranks them. It is judged by whether a feature actually reaches production with no bugs and real users touch it. Weitzman compares a feature that almost ships to a football carried to the goal line but never crossed over: no credit (40:15).

The clearest illustration of what this workflow makes possible is not about a company at all. Weitzman has a family member with a rare autoimmune neural inflammation disorder. Over 15 weeks, he collected blood samples, sequenced the genome, ran proteomics and RNA analysis, and compared the results against daily self-reported mood and quality-of-life data, then ran all of it through a GPU cluster looking for patterns no doctor had found (58:52). He describes using the same approach, with a GPU cluster, to help locate a cancer lesion in his father's body that doctors had missed (61:11). Weitzman was a child who could not learn to read until his father read him Harry Potter aloud, and later built his own version of a text-to-speech tool just to get through college. The line he draws from that childhood to a GPU cluster searching for a cure is the clearest argument in the whole conversation for why he thinks owning compute is not a luxury.

"Technology solved my dyslexia and it solved my ADHD, and it's gonna solve my brother's disease." - Cliff Weitzman [61:11]

Weitzman closes with a prediction that ties the hardware argument back to product design: within a few years, he expects most human-computer interaction to move from screens to voice, the same way ChatGPT's plain chat box beat GPT-3's raw interface by removing friction (54:11). Whether or not that timeline holds, the throughline of the conversation is consistent. Every advantage Weitzman describes, cheaper compute, a faster hiring pool, a workflow built around shipped features rather than clever demos, comes down to removing the friction between having an idea and testing it. Owning the hoop, in other words, is not really about the hoop.