在20VC的一次激烈讨论中，Harry Stebbings欢迎Speechify的创始人兼首席执行官Cliff Weitzman，坦诚探讨了该公司的战略决策和更广阔的AI格局。 Weitzman首先详细介绍了Speechify的非传统策略，即向NVIDIA GPU投入“数千万美元”，通常每单位额外支付10万美元，以提前四个月获得它们。他解释说，这一大胆举动是为了让工程师能够无限制地访问计算资源，以训练Speechify的尖端AI模型，例如在质量方面全球排名第一的Simba 3.2。他将其比作迈克尔·乔丹为了最佳训练需要自己的篮球筐。经济账很简单：购买一块H100 GPU花费3万美元，比每年租用3.5万至5万美元更具成本效益。拥有GPU还能提供对大规模训练至关重要的共存内存（co-located memory），并能以极低的成本运行开源模型。Weitzman指出，较旧的GPU对推理仍然有价值，而NVIDIA为GPU二手市场提供担保的努力，进一步巩固了它们的长期价值，这类似于埃隆·马斯克（Elon Musk）的SolarCity战略。尽管存在国际运输、保险以及数据中心液冷复杂性等物流挑战，Weitzman认为，获得的控制权、速度以及扩展其AI团队的能力大大超过了这些困难。 话题随后转向了Speechify“最大的战略错误”：未能更早进入B2B市场。Weitzman承认，他最初错误判断了基于API的业务的潜力，认为它们会商品化。他现在认识到，持续创新是关键，并且第一个产品往往是未来产品的“楔子”。他坦言，这一疏忽让Eleven Labs等公司在B2B领域“超越”了Speechify，尽管Speechify在消费者文本到语音应用安装方面占据98%的市场主导份额，并已处理超过7700亿个词。 哈里（Harry）对这一转变提出了质疑，询问进入一个已被Eleven Labs（获得政府大量投资）和Sierra（由Brett Taylor、红杉资本和Greenoaks支持）等强大参与者主导的市场是否明智。Weitzman反驳说，AI领域是寡头垄断，而非垄断，并举例说明了Anthropic和Facebook等“后来者”的成功。他强调了Speechify强大的工程团队，以及他们提供卓越质量、速度和便宜10倍的API价格（Simba 3.2每百万字符收费10美元，而Eleven Labs收费100美元，OpenAI收费196美元）的能力。他强调了“参与竞争”的重要性，并将他们对AI的深刻理解应用于新的B2B问题，即使这意味着最初免费提供产品以进行学习和创新。 在招聘方面，Weitzman不同意哈里（Harry）的说法，即由于OpenAI和Anthropic的诱惑，初创公司现在招聘比以往任何时候都困难。Weitzman承认这些巨头提供丰厚的薪酬待遇，但他认为，对于种子阶段的公司来说，单个创始人/工程师的影响力更大，尤其是在AI代理的帮助下。他强调招聘重点已从手工编写代码转向原始技术能力、智力和强烈的职业道德，更倾向于选择表现出高“斜率”（快速增长潜力）而非高“截距”（当前技能）的候选人。他提倡实用、功能性面试，并观察候选人如何协调AI代理。在公司内部，Speechify衡量成功的标准不是令牌使用排行榜，而是用户实际采用的投入生产的功能，这体现了他们作为一家“应用型AI公司”的身份。 展望未来，Weitzman认为人机界面将主要转向语音，类似于Google搜索或ChatGPT的简洁性，但以口头形式呈现。他还分享了他将AI应用于药理学和生物学的个人热情，这源于他兄弟患有严重的自身免疫性疾病。他详细介绍了他的个人项目：对他兄弟的基因组进行测序，并在GPU集群上进行蛋白质组学和RNA分析，以识别疾病机制；以及他解决传统医学因经济可行性而忽视的“孤儿病”的雄心。他个人与阅读障碍和多动症的斗争，通过技术得以克服，这支撑了他对AI能够显著改善每个人的生活质量的信念。

In a robust discussion on 20VC, Harry Stebbings welcomed Cliff Weitzman, founder and CEO of Speechify, for a candid exploration of the company's strategic decisions and the broader AI landscape. Weitzman opened by detailing Speechify's unconventional strategy of investing "tens of millions of dollars" in NVIDIA GPUs, often paying an additional $100,000 per unit to receive them four months early. This bold move, he explained, was driven by the need for engineers to have unfettered access to computing resources for training Speechify's cutting-edge AI models, such as the Simba 3.2, which is ranked #1 globally for quality. He likened it to Michael Jordan needing his own basketball hoop for optimal training. The economics were simple: buying an H100 GPU for $30,000 was more cost-effective than renting it for $35,000-$50,000 annually. Owning also provides the co-located memory crucial for large-scale training, and allows for running open-source models at a fraction of the cost. Weitzman noted that older GPUs remain valuable for inference, and NVIDIA's efforts to underwrite a secondary market for GPUs further solidifies their long-term value, akin to Elon Musk's SolarCity strategy. Despite logistical challenges like international shipping, insurance, and the complexities of liquid cooling in data centers, Weitzman believes the control, speed, and ability to scale their AI team vastly outweigh the difficulties. The conversation then turned to Speechify's "biggest strategic mistake": not entering the B2B market earlier. Weitzman admitted he initially misjudged the potential of API-based businesses, thinking they would become commoditized. He now recognizes that continuous innovation is key, and a first product often serves as a "wedge" for future offerings. He confessed that this oversight allowed companies like Eleven Labs to "leapfrog" Speechify in the B2B space, despite Speechify's dominant 98% market share in consumer text-to-speech app installs and having served over 770 billion words. Harry challenged this pivot, questioning the wisdom of entering a market already dominated by powerful players like Eleven Labs (with significant government buy-in) and Sierra (backed by Brett Taylor, Sequoia, and Greenoaks). Weitzman countered, asserting that the AI space is an oligopoly, not a monopoly, citing examples of "second movers" like Anthropic and Facebook succeeding. He emphasized Speechify's strong engineering team and their ability to offer superior quality, speed, and 10x cheaper prices for their API (Simba 3.2 costs $10/million characters compared to Eleven Labs' $100 and OpenAI's $196). He stressed the importance of "being in the race" and applying their deep understanding of AI to new B2B problems, even if it means initially offering products for free to learn and innovate. On hiring, Weitzman disagreed with Harry's assertion that it's harder than ever for startups due to the allure of OpenAI and Anthropic. While acknowledging the high compensation packages at these giants, Weitzman argued that for seed-stage companies, the impact of a single founder/engineer is greater, especially with AI agents. He highlighted a shift in hiring focus from handcrafted code to raw technical aptitude, intelligence, and a strong work ethic, preferring candidates who show high "slope" (potential for rapid growth) over "intercept" (current skills). He advocates for practical, functional interviews and seeing how candidates orchestrate AI agents. Internally, Speechify measures success not by token usage leaderboards, but by features shipped to production that users actually adopt, embodying their identity as an "applied AI company." Looking ahead, Weitzman believes the human-computer interface will predominantly shift to voice, mirroring the simplicity of Google Search or ChatGPT but in a verbal format. He also shared his personal passion for applying AI to pharmacology and biology, driven by his brother's severe autoimmune disease. He detailed his personal project of sequencing his brother's genome and running proteomics and RNA analysis on GPU clusters to identify disease mechanisms, and his ambition to solve "orphan diseases" that traditional medicine neglects due to economic viability. His personal history with dyslexia and ADHD, overcome by technology, underpins his conviction that AI can dramatically improve the quality of life for everyone.