How to Build Your Own Data Center & Why Every Startup Should Do It

Why does it matter? Because renting compute from AWS doesn't just cost more — it psychologically caps your engineers at one-seventh their speed

Speechify's biggest breakthrough wasn't a new model or a product launch. It was the moment Cliff Weitzmann watched his engineers deliberately slow themselves down because every GPU run felt like a financial sin. Buying hardware instead of renting it didn't just improve unit economics — it changed what kind of company Speechify became.

• Renting an H100 for one year costs $35,000–$50,000; buying it outright costs $30,000 — ownership pays back before the warranty does • Rented compute creates a psychological tax that kills experimental velocity before a single experiment starts • The first product you ship is never the product — it's the wedge for everything that comes next • Nvidia has now turned GPUs into a financeable asset class, with Goldman Sachs, Blackstone, BlackRock, and Apollo underwriting 25% of residual value

Renting a GPU for one year costs 1.5x its purchase price — and the hardware keeps working for a decade

$30,000 buys an H100 outright. Renting that same card from GCP costs $5 per hour; Azure and AWS run closer to $3.50. Multiply by 24 hours and 365 days: "I'm actually going to end up paying $35,000 to $50,000 to rent that GPU for one year, but I could buy it for $30,000." Cliff's conclusion: "It's 1.5x the cost of owning the hardware to rent the hardware for a year."

The depreciation concern doesn't survive contact with how GPUs actually get used. Older cards don't retire — they get reassigned. Training runs on the latest Blackwell generation; inference cycles down to H100s and A100s; even K80s still serve specific operations at Speechify. With 60 million users generating continuous inference load and a team built to triple in headcount, idle racks don't exist.

There's also a structural training requirement that hyperscalers can't solve competitively: large-scale model training requires GPUs co-located with massive memory cards. Renting a cluster large enough demands multi-year prepayment and still leaves you without control over the InfiniBand cables connecting DGX nodes. Owning means skipping the cloud-provider queue for new GPU generations by roughly a year — Speechify is receiving Rubin access well before most hyperscaler customers will.

"The best bond you can buy long tail will yield you like 5%. Or you can buy a GPU" — at a 1.5x annual return by displacement of rental cost. That's the math.

Rented compute trains your engineers to think like accountants

The performance tax doesn't show up in your AWS invoice. It shows up in the experiments your team didn't run.

"Engineers at Speechify would be parsimonious with how they use the GPUs because they were like, oh my god, I'm costing the company tens of thousands of dollars." Cliff and his brother named the problem: Michael Jordan, wanting only to make the NBA, paying $20 an hour to access a basketball court. You ration your reps. You skip the shot that might not pan out. The solution isn't better coaching — it's a hoop in the house.

Once Speechify moved to owned compute, the gap snapped into focus: "We had really talented engineers who were essentially moving at 1/7th of the speed they could have if they had the compute one-to-one with their creativity and ideas." A 19-year-old on the team now resolves fourteen product notes in a single morning while waiting on training runs to finish — not because he became more capable overnight, but because the GPU stopped feeling expensive.

Before attributing slow AI development to team quality, audit whether your engineers are rationing experiments because compute feels like a budget line. The fix isn't a performance review.

Speechify's biggest strategic mistake: the first product is never the product

"100% it's on me. It's the biggest strategic mistake I made in the history of Speechify."

In 2022, Cliff met the 11 Labs founders in London and dismissed their API-first approach. Text-to-speech APIs would commoditize — inference gets cheaper, models move to devices, the API becomes a rounding error. Smart founders, wrong strategy — that was his read.

What he missed was structural: "The point of an AI lab like Speechify or like 11 Labs is to continuously innovate, and the first product that you release is your wedge that gets other people to then later use your other technology." 11 Labs used a TTS API as the entry point, then stacked emotional prosody, voice cloning, duplex turn-taking, speech-to-text, and eventually AI agents priced on outcomes for CTO and CIO buyers. The API was never the destination. It was the foothold.

"It was my mistake to think that an API product was a bad strategy because I thought it was something that would become commoditizable. And I forgot the central thesis about Silicon Valley, which is constantly innovate." Get users into your product — even for free — then sell them what comes next. 11 Labs now has government contracts across major Western democracies, a position that started with a single excellent API Cliff wrote off as a dead end.

Hire the math olympiad who's never coded — skills take six months, intelligence takes decades

Drop the years-of-experience filter.

Speechify's current hiring profile: math olympiads, competitive programmers, Kaggle winners, physics and math graduates — many of whom have never shipped production code. "The thing I care about the most today is technical aptitude and just like raw technical intelligence because I know that we could teach you everything else and in 6 months you could be a machine."

The interview has changed accordingly. Functional build-and-break assessments first: give a task, run unit tests on the output. Then drop the candidate into a large unfamiliar codebase, have them make changes, and identify what broke. The filter isn't syntax — it's reasoning speed and instinct for agent orchestration.

"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." A brilliant 22-year-old with six months of GPU cluster exposure beats a senior engineer who spent a decade writing optimized Kotlin for a framework that agents now render irrelevant. The talent pool for AI-era companies is larger than it has ever been. Founders are screening it out with requirements built for a different decade.

Nvidia turned GPUs into a financeable asset class — and Goldman, Blackstone, BlackRock, and Apollo are co-signing the floor

The residual value risk on owning GPUs just got underwritten by four of the world's largest capital pools.

Nvidia struck a deal with Blackstone, BlackRock, Apollo, and Goldman Sachs to guarantee up to 25% of a GPU's value as collateral. If a borrower defaults and a lender seizes the hardware, Nvidia will buy it back. "They're succeeding in creating a liquid secondary market for GPUs that they're underwriting. So now the large banks have an incentive to loan money at much better interest rates."

Cliff maps it directly onto Solar City's origin. Elon Musk went to Morgan Stanley and Merrill Lynch to underwrite solar panels over 30-year terms — the innovation wasn't the panel, it was making the asset lendable. Nvidia executed the same structural move: GPU racks are now collateralizable infrastructure with a defined downside floor, and banks will finance against it at rates that didn't previously exist.

For any startup modeling the buy-vs-rent decision, the residual risk that previously justified renting is now priced and backstopped. Debt acquisition against that floor is a live alternative to full upfront capital deployment — and the math only gets more favorable the longer you hold.

A great engineer's only job is ten good decisions per day — not code

"A really good engineer today is just an exceptional QA, right? The AI will make them the feature. You will test the feature, see if it's good. You'll figure out where the edge cases are."

Each Speechify engineer runs five to eighteen agents in parallel on long-horizon tasks — proposing hypotheses, testing them, iterating continuously on the GPU cluster. The human bottleneck isn't implementation speed. It's the quality of roughly ten product and architecture decisions made per day.

Performance gets measured by one criterion: did it ship to production, and are users touching it? "If you carry the football all the way to the line but you don't cross over to the end zone, you get no credit. And in the rim means push to production with no bugs. And users are actually using it." Token volume, test coverage, lines of code — none of it registers. The 19-year-old who cleared fourteen product notes in a morning while waiting on training runs is the benchmark. Not what he wrote. That it shipped and worked.

Engineering reviews built around code volume and ticket velocity are now measuring the wrong thing entirely. Shipped-to-production outcomes and decision quality — that's the whole framework.

One person, a GPU cluster, and a portable sequencer can now solve orphan diseases that pharma won't fund

Fifteen weeks. Weekly blood draws, genome sequencing, proteomics analysis, RNA readings — all compared against six years of daily self-reported quality-of-life data, run on a GPU cluster in Scottsdale, Arizona.

Cliff is doing this for a family member with severe autoimmune neural inflammation. The condition is rare enough to qualify as an orphan disease: too few patients to justify pharmaceutical R&D investment. "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."

The toolkit is already in circulation. A $5,000 portable device sequences a full genome from hair, saliva, or blood. AlphaFold predicts protein structures. CRISPR edits them. Twist Bioscience synthesizes custom RNA or DNA sequences and ships them to your door or lab. "I can design not just the protein that is creating these issues but I can design the molecule that needs to bind to that protein to either turn it on or off."

Cliff is now organizing meetups with every member of the condition's Facebook group to sequence all their genomes and compare epigenetic threads across the entire patient population on a shared compute cluster. The constraint was never tools or compute. It was organized data collection and the willingness to attempt something pharma's spreadsheets said wasn't worth it.

The economic logic that made renting compute 'sensible' is the same logic that made orphan diseases 'unfundable' — and both expired at the same time

The thread connecting GPU ownership, hiring olympiads, and sequencing rare genomes isn't any particular technology. It's the collapse in the cost of attempting ambitious things. The assumptions that justified renting compute, filtering for senior experience, and leaving orphan diseases to Facebook support groups were all calculated against a world that no longer exists. Institutions built on those assumptions are the last to revise them.

The frontier belongs to whoever updates their priors first.

---

Topics: GPU infrastructure, AI compute, data centers, hardware ownership, Speechify, 11 Labs, AI hiring, agent orchestration, voice AI, drug discovery, AI biology, startup strategy, B2B GTM, compound startups