Cliff Weitzman on Owning GPUs, Voice AI and Building Speechify

Speechify co-founder and CEO Cliff Weitzman joins Harry Stebbings to discuss why his company buys GPUs, how it manages AI research and what he learned from competitors in voice technology. They also debate hiring, customer support and his hopes for AI in biology.

Stebbings opens with Speechify's investment in GPUs and the cost of obtaining equipment early. Weitzman says the company first bought a rack in 2022 for training: engineers had been reluctant to run expensive experiments on rented machines, so ownership gave them more predictable access to compute.

He compares the situation to an athlete needing a place to practice without paying for every hour. The company continued investing as it developed models. His claims about Speechify's model quality and comparative prices are company claims, not results independently established in the conversation.

Weitzman then offers an illustrative buy-versus-rent calculation for an H100 GPU. He compares a hypothetical $30,000 purchase with hourly rental prices that could add up to more over a full year. Those examples depend on utilization and market prices; the comparison alone does not capture every ownership cost.

He says large training jobs also need GPUs and data storage located together, and he expects company-owned hardware to give the team more control over that setup. He sees another use for the machines in running open-weight models for inference. Actual economics depend on networking, facilities, maintenance, power and workload.

Stebbings objects that chips depreciate and new architectures arrive quickly, locking buyers into old hardware. Weitzman replies that Speechify still uses older GPUs for some inference and experiments. Training experiments may benefit from the newest hardware, but a task with an acceptable response time may not need it.

He says owning a baseline does not mean abandoning the cloud. Speechify also rents dedicated and spot instances to handle fluctuating demand, including seasonal changes as students return to school. In his view, multiple GPUs can remain useful simultaneously, unlike an old phone sitting unused in a drawer.

Weitzman says excess capacity could potentially be rented to others, while acknowledging that he does not expect to have much surplus. He compares the return from useful hardware with holding cash or bonds. That is his company-specific investment judgment, not a guaranteed return or a complete risk comparison.

His financing comparison assumes the hardware is continuously useful to the business. It does not mean the card itself produces a risk-free yield: an owner also faces utilization, obsolescence, repair and resale uncertainty. Stebbings returns throughout the conversation to the value of retaining the flexibility to rent.

He says Speechify buys racks through suppliers and rents space in a data center that provides networking, power and staff to install and maintain equipment. He identifies electricity availability as a major constraint. The company is buying hardware and colocating it in rented facilities, rather than necessarily constructing a new building itself.

To forecast demand, he separates a relatively reliable baseline from peaks. In an illustrative allocation, the business might buy enough equipment for a conservative portion of normal usage, use longer-term contracts for another portion, and rely on spot capacity for the rest. He also considers how future hiring could change demand.

Weitzman uses a hypothetical November baseline, a busier month and a quieter month to show how he would size capacity. Buying only a portion of the baseline leaves room to meet peaks through rented instances. The percentages are examples of his planning method, not a disclosed capacity plan.

He believes older accelerators can continue to support experiments or inference after they are no longer optimal for the most demanding training. That depends on the workload and the equipment's useful life. His expectations about durability are not a warranty, a market-price forecast or a substitute for depreciation and maintenance analysis.

Weitzman describes a reported financing arrangement involving Nvidia and financial institutions that he believes could support a secondary market for GPUs. He argues that collateral and potential resale value can affect lenders' calculations. The interview does not independently establish the arrangement's terms or guarantee a resale floor for any particular chip.

Asked about concerns over circular financing in AI, he argues that a GPU is a usable physical asset and distinguishes that from some other financing arrangements he criticizes. His view does not establish that every GPU retains a particular value or eliminate counterparty, depreciation and financing risk.

Buying chips involves vendors, availability, contract terms and delivery schedules, he says. A late shipment can be costly when a company is already paying for space. He recounts negotiating with suppliers and describes a willingness to pay more for faster access; the discussion does not establish a universal premium per chip.

He describes asking a supplier to deliver equipment under an existing contract and comparing alternative offers when a shipment was late. Each delay also meant paying for reserved data-center capacity that was not yet productive. His account illustrates procurement risk and does not characterize any supplier's practices generally.

He adds transportation insurance, amortization, physical security, networking and cooling to the list of considerations. Newer racks may require liquid cooling, and existing facilities may need additional equipment and installation. He describes one possible setup, not a guarantee that all data centers can support a particular rack.

Stebbings pushes back: renting offers flexibility and leaves the team free to focus on the product rather than freight, insurance and cooling. Weitzman agrees rentals are possible but says a colocated cluster may require expensive long commitments and less operational control. Their disagreement is about priorities as well as headline cost.

Weitzman argues that timely access to compute can itself become an advantage when researchers are competing to test ideas. He expects purchasing to give Speechify earlier access to some hardware. That is his expectation about procurement, not an independently verified guarantee of delivery or market exclusivity.

He identifies three ingredients for his team's work: data, compute and engineering decisions. He says engineers coordinate multiple agents, test hypotheses and spend time cleaning or generating training data. In his account, a shortage of either usable data or compute can limit what an otherwise capable team can do.

Both speakers expect demand for specialized data as more companies adapt models for their own needs. Weitzman says Speechify has purchased small data sets and that external providers can save time. He adds that a company buying data needs confidence about provenance and rights, especially given ongoing disputes over training material.

Stebbings suggests that companies may buy supplemental data even when they train on their own proprietary information. Weitzman thinks specialist providers can shorten development cycles because a lab does not have to organize every collection effort itself. They also note that the legal provenance of training data is important.

He cautions that data sales are not necessarily annual recurring revenue: a buyer may not return for another purchase. A provider also needs operational speed and data that actually improves a customer's model. Stebbings sees a potentially larger future market if enterprise buyers join frontier labs, but that is a scenario, not a measured forecast.

Weitzman says consumer customers historically paid less than business customers, creating pressure to run text-to-speech cheaply. He praises ElevenLabs for building a business-facing offering earlier and acknowledges that his own company focused on consumers. His quoted comparative prices and market positions are his account, not independently verified benchmarks.

He calls his decision not to pursue a business API earlier Speechify's biggest strategic mistake. He expected text-to-speech APIs to become commoditized and underestimated how an initial product could become an entry point for later innovations, including additional voices, speech recognition and conversational tools.

In his account, ElevenLabs built a developer API, creator tools and later agents for business users. He says an AI company must keep innovating after the first product rather than assuming a single feature will remain unique forever. He sees customer access and feedback as valuable foundations for expanding the product.

The business API, in his view, can be a starting point for relationships with engineers and later with executives who buy more complete agent products. He mentions outcome-based pricing as another possible commercial model. These are strategic interpretations, not proof that any given API customer will buy subsequent products.

Stebbings challenges Speechify's move into business sales, arguing that competing with well-funded incumbents requires a difficult go-to-market capability. He questions whether a third player can capture enough value when existing rivals have strong distribution and financing. This is his strategic objection, not an established outcome.

Weitzman replies that Speechify's consumer presence and engineering resources can support both consumer and business products. He cites company usage figures and believes the market can accommodate more than one provider. Those figures and competitive comparisons are self-reported, and the business strategy remains unproven in the interview.

He points to companies that entered markets after an early leader and later found success. Even if Speechify does not win every existing use case, he wants its API or agents to be available to customers so the company can learn and add features over time. Stebbings remains skeptical that market value will be evenly distributed.

Weitzman acknowledges that competitors may have strong positions in their existing products. He argues that large firms can still leave gaps as markets evolve, and that a challenger needs to learn from customers and keep improving. Stebbings agrees there is room for multiple companies but doubts their returns will necessarily be comparable.

On 'compound startups,' Weitzman clarifies that not every new company must immediately build many products. At a certain stage, he thinks a capable team may need to expand its offering. He emphasizes iteration and execution, though his prediction that agents will make it easier is a view about future workflows.

The discussion turns to whether startups can hire when large AI labs offer extraordinary compensation. Weitzman says established growth-stage companies face strong competition for senior leaders, while early teams may have a broader pool if AI tools help capable people learn new skills quickly.

He says Speechify increasingly values technical aptitude, problem-solving and the ability to learn, including candidates from mathematics, physics or competitive programming. He still values engineering knowledge, but is less focused on a candidate's ability to handcraft every line of code. This is his hiring preference, not a rule for all engineering jobs.

Stebbings counters that some unusually large 'seed' rounds can afford substantial compensation and compete directly for experienced AI talent. Weitzman distinguishes those well-funded teams from founders raising much smaller first rounds. They agree that the label 'seed' can cover very different company sizes and hiring situations.

They also debate whether employees now prefer liquid, more certain compensation over speculative startup upside. Weitzman believes each candidate weighs risk and certainty differently. He notes that public listings can affect recruiting by making the value of equity more tangible, alongside other reasons a company might list.

Inside Speechify, he says AI researchers already coordinate agents, while some experienced product engineers initially preferred familiar hand-coding workflows. He tries to encourage adoption by having strong practitioners demonstrate their methods and share resources. He names Claude Code and Cursor as frequently used tools at the company.

He says one way to encourage a team to adopt agent workflows is to screen-share a concrete example from a skilled engineer, then provide resources for learning. Speechify also uses project-management tickets to organize work. In his account, the person remains responsible for inspecting the result and handling edge cases.

Weitzman describes an engineer's role as setting direction, checking generated features, identifying edge cases and making architecture decisions. He wants agents to handle work while people evaluate whether the output meets the intended goal. His description is a management approach, not proof that automated code is reliable without review.

Rather than rewarding a token leaderboard, he asks for demonstrations and shipped features. A promising model experiment does not earn the same credit until it reaches production and users can provide feedback, he says. His example is a team member who quickly addressed issues after watching a recorded product review.

For Weitzman, a useful demonstration includes the product working in a user's hands, not just an impressive research result. He records himself using features so engineers can see failures and reproduce them. That emphasis on production feedback is his management philosophy and does not imply that every defect can be eliminated.

He says he is sensitive to unjustified token spending even though he wants teams to use AI extensively. The right amount depends on the task. His example contrasts a broad scrape that touches thousands of pages with targeting the needed information more directly, reducing unnecessary work.

He illustrates waste with a person who built an app but retrieved far more web data than the task required, producing a large bill. An experienced engineer, he says, can narrow the request and define a more efficient path. His point is to monitor work and cost rather than simply maximize the number of calls.

For long-running agents, he emphasizes choosing a clear target, defining how to measure it and checking the results of repeated iterations. A large spend can make sense when a difficult task produces useful progress, but wasteful loops can consume resources without creating value. The output matters more than the token count alone.

His suggested hiring process includes practical exercises: build a feature, test it, and inspect an existing codebase for the effects of a change. He wants to see how candidates use agents as well as how they reason about code. He also warns that intense AI-assisted work can exhaust teams.

Stebbings raises concerns that speech-to-text products are becoming commoditized as alternatives proliferate. Weitzman discusses his impressions of competitors and says some products combine models with a broader interface. The interview does not independently establish the underlying models or the reasons for any particular product's perceived performance.

Weitzman says he looks for a person's potential to improve, not only current skill. He describes working with human colleagues who own outcomes while using agents underneath. Although he personally values intense engagement with work, he explicitly acknowledges the need to avoid burnout.

Weitzman says he previously built a personal dictation experience but did not launch it because he expected major platform providers to make the feature commonplace. He now regards waiting as another mistake and says Speechify has launched competing offerings. The lesson he draws is to test an opportunity rather than assuming an incumbent will take it.

Stebbings questions the crowded customer-support market, where established vendors, new AI companies and in-house teams compete. Weitzman distinguishes Speechify's core business API from a complete customer-support application. He says the API aims to supply voice capability to customers building a range of products.

Weitzman says an API supplier and a company selling an entire customer-support workflow may occupy different layers of the market, although products can overlap. He describes business agents as a way to understand customer needs, not merely an attempt to copy another company's complete offering.

At the same time, Weitzman wants to work directly with business customers on agents so his team can learn their problems and identify new product opportunities. He presents claims about relative price and quality as part of Speechify's pitch; those comparisons are not independently evaluated in the interview.

Asked whether Sierra or ElevenLabs will be larger, he predicts both have significant potential and describes AI agents as a broad market extending beyond customer support. Stebbings sees differences in their product ambitions. Neither speaker establishes future valuations or guarantees that the companies' strategies will succeed.

In a quick-fire prediction, Weitzman expects voice to become a much more common interface for computers and wearables. He links that expectation to simpler interaction rather than assuming screens will immediately disappear. His assessment of current voice products and their future adoption is a personal view.

The speakers compare a hypothetical investment in Meta and SpaceX. Weitzman favors Meta, citing its data, founder leadership and valuation as he sees them; Stebbings offers a different view of the opportunities around Musk's businesses. Their comments are personal, speculative opinions and not a recommendation to buy either company.

In the investing discussion, Weitzman says he prefers founder-led companies and believes Meta has valuable assets. Stebbings argues that other technology businesses have a broader set of potential markets. The exchange is a hypothetical comparison based on their beliefs, with no verified forecast of relative valuations.

They debate the role of founders, capital spending and valuation in public and private technology companies. Weitzman thinks underlying technology and business assets matter, while Stebbings questions how markets might react to management changes. Neither can establish a future stock price or investment return from the discussion.

Asked what excites him most, Weitzman turns to AI applications in biology and discusses a family member with a serious rare condition. He describes collecting biological and self-reported data and analyzing it with computing tools. These are personal research experiences and hopes; the interview provides no independent clinical validation.

He speculates about using computational tools to investigate biological mechanisms and possible interventions. His confidence that he can solve a disease or identify a treatment is aspirational, not evidence of a diagnosis, validated therapy or cure. Such research would require appropriate scientific and clinical evaluation.

He describes comparing measurements from repeated samples with self-reported quality-of-life data. He hopes computational analysis could identify useful patterns, but a correlation in an individual account does not establish a disease mechanism or effective intervention. The interview does not report independent peer review or clinical results.

Weitzman connects his interest in accessible technology to his own experience growing up with dyslexia, learning English through audiobooks and building a text-to-speech tool at Brown. He says technology helped him learn and work. His personal account should not be generalized into a claim that the software treats or cures dyslexia or ADHD.

He shares another family health experience and hopes better tools can improve research and quality of life. The medical claims are anecdotal and not independently confirmed in the interview. Stebbings closes by thanking him for an open discussion spanning hardware, company strategy and the potential of AI.