finance
The Twenty Minute VC (20VC) interviews the world's greatest venture capitalists with prior guests including Sequoia's Doug Leone and Benchmark's Bill Gurley. Once per week, 20VC Host, Harry Stebbings is also joined by one of the great founders of our time with prior founder episodes from Spotify's Daniel Ek, Linkedin's Reid Hoffman, and Snowflake's Frank Slootman. If you would like to see more of The Twenty Minute VC (20VC), head to www.20vc.com for more information on the podcast, show notes, resources and more.
Erik Bernhardsson, co-founder and CEO of Modal, discusses the company's recent milestone of reaching $1 billion in annual revenue and its recent $750 million Series D funding round led by Accel at a valuation of $15.75 billion.
Modal CEO Erik Bernhardsson reveals that the company's sandbox product, launched in 2023, has grown rapidly and now represents a $300 million run rate business, accounting for about a third of the company's total revenue.
Modal CEO Erik Bernhardsson highlights the extreme competition for talent in the AI industry, noting that exceptional new graduates are receiving compensation packages exceeding $20 million from major AI labs.
Microsoft CEO Satya Nadella shared that his biggest change in perspective over recent years is realizing that alignment will not automatically result from scaling laws alone, pointing to deeper science challenges that need to be addressed.
Discussing the security risks of increasingly capable models, Satya Nadella explained that Microsoft builds zero-trust environments with agent identities, logged model calls, and auditable English chain-of-thought outputs to contain potential malfunctions.
Satya Nadella discussed the 'reverse information paradox,' arguing that firms must retain the economic surplus generated by AI usage rather than leaking value to model providers, suggesting that every company will develop private evaluation benchmarks as core IP.
Reflecting on industry shifts, Satya Nadella mentioned that the depth of each layer and the infinite demand for inference were things he wished he had known twelve months ago, while expressing hope for horizontal specialization with multiple winners across the stack.
AMD has acquired Fei-Fei Li's World Labs for $8.2 billion in stock, marking a significant exit for the AI research lab. This acquisition is seen as a strategic move by AMD to bolster its AI capabilities and compete more effectively in the rapidly evolving AI landscape. The deal is particularly noteworthy as it represents a substantial return for the investors and highlights the increasing value of specialized AI research entities.
MongoDB's CEO has reportedly been poached by Meta to lead its enterprise division. This move comes shortly after the CEO's return to MongoDB following a previous departure. The significant compensation package offered by Meta is cited as the primary driver for this transition, highlighting the intense competition for top talent in the tech industry, particularly in AI-focused companies.
AI startup Instinct has secured $1 billion in funding at a $10 billion valuation, signaling strong investor confidence in its agent-based AI assistant. Benchmark led the investment round, highlighting a significant bet on the company's approach to personal AI. This valuation comes despite the company being pre-revenue, underscoring the market's current focus on potential and paradigm shifts in AI.
Anthropic's leaked draft S-1 filing projects $4.6 billion in revenue for 2025, with an operating loss of $8 billion. The filing also noted that two customers accounted for 25% of revenue, indicating significant reliance on a few major clients. While the 2025 numbers are seen as largely historical, the Q3 revenue figures are considered crucial for assessing the company's current trajectory amidst intense competition with OpenAI.
Oura, the company behind the smart ring that tracks sleep and activity, has decided to pull its planned IPO. The company had been aiming for a valuation of $16 billion, but has opted to delay its public offering due to current market conditions. This decision highlights the challenges some companies face in accessing public markets, even with strong consumer recognition and product appeal.
Nubank is reportedly considering an acquisition of the UK-based digital bank Monzo, with a potential price range of $8 billion to $12 billion. This move would signal Nubank's expansion into the UK market, a region where Monzo has a strong presence. The valuation range suggests a significant strategic interest in Monzo's customer base and technology, despite potential challenges in integrating across different regulatory environments.
Bessemer Venture Partners has closed a new fund totaling $5.75 billion, with a significant portion allocated to growth-stage investments. This move aligns with a broader industry trend of firms raising larger funds to invest more aggressively in growth-stage companies. The fund's structure, with a substantial seed component, indicates a strategy to capture value across different stages of a company's lifecycle, particularly in rapidly expanding sectors like AI.
Alex Mashrabov, co-founder of Hixfield, revealed that the company achieved $1 billion in annual recurring revenue in just 18 months, a faster pace than Cursor. The company, with 300 employees originating from Kazakhstan, is rumored to be raising at an $8 billion valuation.
Alex Mashrabov disclosed that Hixfield's internal usage of AI models costs over $4 million per month, with individual team members spending an average of over $10,000 monthly on various models. Some employees have reportedly spent over $30,000 in a single week on specific models like Astra.
Alex Mashrabov, founder of Hixfield, shared his early life in Kazakhstan, where his parents encouraged his pursuit of technology. He excelled in competitive programming, reaching top three globally by age 19, before pivoting to startups. His previous venture, AI Factory, developed state-of-the-art language translation systems and was acquired by Snap for over $160 million.
Alex Mashrabov explained that Hixfield's growth is driven by addressing the unmet needs of companies struggling to create relevant content for social media. He noted that while direct response advertising is achievable, keeping up with the daily trends on social platforms is a significant challenge for most businesses.
Alex Mashrabov believes that advancements in AI, particularly in scaling laws, will revolutionize video content creation, similar to their impact on LLMs and coding. Hixfield is betting on this trend, focusing on tools that can transform existing assets into short, vertically oriented videos suitable for social media, addressing a gap left by simpler upload tools or manual video editing.
Alex Mashrabov reflected on the early days of AI companies, noting that capital was scarce and multiples were significantly lower compared to today. He recalled that in the past, securing $1-2 million in investment was considered a major success, contrasting with the current market where AI company valuations can reach 200x revenue.
Pagg Agrawal, founder of Parallel, describes his company as the "Google for agents," aiming to build the technology and business models for how agents search the web. He posits that agents will use the web 1000X more than humans, necessitating new, more efficient search technologies and business models.
Pagg Agrawal explains that the scale of agentic web usage necessitates a 10-100X improvement in search efficiency compared to current human-centric models. This is due to agents potentially being more impatient (requiring sub-second responses) or operating in the background with different output needs (tokens or files, not just blue links).
Pagg Agrawal details how web search can be optimized to save compute for AI models. The process involves narrowing down trillions of documents to a few thousand tokens for the model's context window, with the amount of compute allocated to web search depending on the cost and efficiency of the AI model itself.
Parallel's API allows programmers or agents to specify their needs, such as prioritizing accuracy over latency or vice versa. Agrawal highlights different modes, including a very fast API for voice agents that need immediate responses, and another for background agents that can afford more time for the best possible answer.
Pagg Agrawal predicts a dual trend in AI models: frontier models will continue to grow larger, while smaller models will achieve fixed performance levels more efficiently over time. He notes that the gap between the value frontier models can provide in specific use cases will drive their continued development, with no apparent end to scaling laws.
Pagg Agrawal challenges the consensus view on the split between open and frontier models, suggesting the 90/90 split might be inaccurate. He argues that a useful small model can be 1000X cheaper than a frontier model, and it's difficult to predict which use cases will align with which model scale over time.
Pagg Agrawal explains that making AI models more efficient (smaller) while maintaining performance can lead to a loss of parametric memory. He contrasts this with the recall of 'head facts' that models can memorize, noting that finding patterns rather than memorizing is key for more complex information, especially for individuals not in the public eye.
The founder of Muse AI claims their product is the first genuine competitor to ChatGPT. They believe that while LLMs are generally good and trained on vast datasets, Muse is designed as a Trojan horse to specifically challenge ChatGPT's dominance in the AI landscape.
Thomas Sohmers believes that the growing opposition to data centers is primarily a Chinese psyop, arguing that individual water usage by single in-and-out users can exceed that of large data centers. He expressed overall opposition to the current direction and pace of the frontiers of AI development.
Thomas Sohmers, co-founder of Positron AI, shared that his company recently raised an $875 million Series C funding round at a $5 billion valuation. The funding round included investments from notable investors such as a16z. Positron AI focuses on building hardware, software, and full systems for generative AI inference.
Thomas Sohmers explained the fundamental differences between training and inference in AI hardware. Training is primarily compute-bound, requiring more FLOPS, while inference is heavily memory-bound due to the need to read parameters for each token generated.
Thomas Sohmers, co-founder of Positron AI, described the company's focus on building the hardware stack for generative AI inference. This includes chips, software, and full systems, aiming to power the applications driving current excitement in AI.
A discussion erupted regarding the call to "pace the frontier" of AI development, initiated by Dario Amodei and echoed by Sam Altman and Elon Musk. The speakers largely panned the idea, with one arguing that the cyber and economic risks are real, but the existential risk of losing control of AI is difficult to assess and that proposed solutions like third-party monitoring are impractical.
One speaker posited that Anthropic's CEO, Dario Amodei, is strategically highlighting AI risks, including a supposed 10% chance of humanity's destruction, as a preemptive measure for an upcoming S-1 filing for their IPO. The goal is to de-risk the IPO by discussing these concerns publicly beforehand, making them a non-issue by the time the company goes public.
The discussion highlights that public awareness and fear around AI's potential to cause harm are growing, evidenced by concerned texts from non-tech individuals. Speakers predict that this public sentiment will inevitably lead to congressional committees and hearings, impacting policy and industry direction.
The podcast acknowledges Meta's launch of its AI assistant product, Muse. This development is placed within the context of an ongoing "AI Assistant War," with mention of Instinct and Town on the startup side and Meta (Zuck) on the incumbent side.
The acquisition of Miro by Bending (or Spritz, as mentioned) for $1.35 billion is highlighted as a significant event in the tech industry. This acquisition is presented as part of a broader trend of European companies making substantial moves in the market.
French AI company Mistral has reportedly raised $3 billion, marking the largest tech funding round ever in Europe. This development is mentioned in the context of European companies achieving significant milestones.
David Morehead, CIO of Baylor University's endowment, explains that the primary goal for private investments is generating money, emphasizing the velocity of capital rather than just returns. He notes that a key rule in their office is to always discuss returns in conjunction with the time it takes to achieve them.
David Morehead, CIO at Baylor University, expresses perplexity over the extended duration of some venture funds, questioning whether General Partner (GP) incentives truly align with the financial objectives of limited partners (LPs) like endowments. He highlights that the endowment's allocation to Anthropic is approximately 2.5%.
David Morehead, CIO of Baylor University's Office of Investments, outlines a strategic shift towards improving upside performance while maintaining historical strength in downside protection. He explains that this strategic pivot, initiated about five years ago, is a response to anticipated challenges in higher education finance, particularly declining domestic student enrollment.
David Morehead, CIO of Baylor University, discusses strategies to improve investment performance, particularly addressing the challenge of trailing in growth-led markets. He explains that over the last three to five years, they have focused on incorporating convexity into their strategy without incurring a significant theta (time decay) cost.
David Morehead, CIO of Baylor University, explains their strategic move away from commingled funds towards direct relationships with General Partners (GPs). This approach allows them to have greater control over their risk-return profile and portfolio composition, avoiding the 'average' risk-return profile that is typical in commingled funds.
David Morehead, CIO at Baylor University, details their approach to portfolio construction, prioritizing the private allocation first due to its liquidity constraints. He states that their allocation range for private investments is between 35% and 55%.
David Morehead, CIO at Baylor University, discusses how endowment distributions are becoming increasingly crucial for higher education institutions facing financial pressures, such as declining student enrollment. He notes that their office has historically performed well in down markets and is now focusing on improving upside performance.
Nvidia CEO Jensen Huang has boldly declared that Artificial General Intelligence (AGI) has now arrived, marking a significant shift in the perception and development of AI. This pronouncement suggests a potential acceleration in the AI model race.
OpenAI has announced the release of its latest AI model, GPT Astra. This development is seen as a direct contribution to the intensifying race among major players in the AI model space, with products like GPT Astra potentially setting new benchmarks.
Index Funds have reportedly ceased investing in a company referred to as 'Town' following its involvement with 'Instinct'. The exact nature of this involvement and the implications for Index Funds' investment strategy remain unclear.
The competitive landscape for AI assistants continues to evolve with Matter announcing the release of its product. This move occurs amidst a broader, accelerated race in the development and deployment of AI assistant technologies.
Wonderful has achieved a significant financial milestone, doubling its valuation to $5 billion in just six months. The company also successfully raised $170 million through secondary transactions.
A debate is emerging regarding the ethics and sustainability of AI startups that operate by 'breaking rules,' whether legal or terms of service. While some see this as a necessary path to innovation, others caution against the long-term viability of such models.
The discussion highlights the critical role of form factor and user integration in the success of AI products. It's suggested that products seamlessly embedded into existing platforms, like WhatsApp, have a significant advantage.
Venture capitalist Jason Calacanis voiced strong doubts about the business model and long-term prospects of the company 'Instinct'. He predicts a potential downfall due to uninteresting technology, weak moats, and the impending wave of open-source competitors.
Jean Denis (JD), founder of Town, discussed the company's pivot from AI tax preparation to an AI assistant that integrates with email and calendar. The initial venture into tax prep did not achieve sufficient product-market fit, leading to a strategic reset and focus on areas with high user engagement like email and calendar.
JD, founder of Town, believes that AI assistants are a top-three priority for major tech companies like Google and Apple within the next 12 months. He acknowledges the competitive landscape but feels confident in his company's direction.
JD argues that the winning AI assistant product will likely have a network effect at the agent level, referencing Town's 'agent to agent' feature. He also highlights distribution as a critical factor, with Meta and WhatsApp being significant potential competitors due to their existing user base.
JD posits that AI is fundamentally changing the software landscape, predicting that AI assistants will become the primary entry point for digital interactions, eventually replacing traditional websites and mobile apps. He believes owning the user's primary AI interface will be crucial for market dominance.
Town's AI assistant features an 'agent to agent' functionality where one assistant can query another if it doesn't have the answer. This feature is highlighted as a key aspect of Town's strategy to build a network effect, making it difficult for users to switch to competing products once integrated within a team.
JD discusses the perception of the AI assistant market, noting that while some see it as commoditized, he views talking about 'moats' as a luxury. He believes success will come from deep product-market fit and attracting a large user base, rather than solely relying on defensibility at this stage.
Cliff Weitzman, CEO of Speechify, discussed the company's strategy of purchasing Nvidia GPUs, even at a premium of $100,000 per unit for early access. He explained that this investment is driven by the economics of training large language models and the need for dedicated, high-memory infrastructure that is not readily available through cloud providers.
Cliff Weitzman stated that Speechify's latest model, Speechify 3.2, is ranked number one globally for quality, surpassing even frontier labs. He also highlighted its cost-effectiveness, being 10 times more affordable than solutions like ElevenLabs.
Cliff Weitzman presented a financial argument for purchasing GPUs over renting them. He calculated that renting an H100 GPU for a year could cost between $35,000 to $50,000, whereas buying one for $30,000 and using it for its warranted three-year lifespan (or longer) is more economical.
Cliff Weitzman elaborated on why Speechify chooses to own its GPU infrastructure, emphasizing the necessity of co-located memory with a large cluster of GPUs for effective large-scale training. He stated that renting from cloud providers doesn't offer the required memory accessibility for their specific training needs.
Cliff Weitzman argued that running open-source models on self-owned hardware is significantly more cost-effective than using third-party token-based services. He noted that this approach allows for a cost per token that is a 'fraction of a fraction of a cent'.
Cliff Weitzman shared an observation that when Speechify engineers used rented GPUs, they tended to be 'parsimonious' due to the high cost. This led to the idea of owning GPUs to encourage more liberal usage for model training, comparing it to Michael Jordan needing affordable access to a basketball court.
NVIDIA has announced a record-breaking quarter with $96.2 billion in revenue, exceeding analyst expectations. The company is also nearing a $12.9 billion deal to acquire Hugging Face, a move that could significantly impact the AI landscape.
NVIDIA has provided a strong outlook, projecting 70% revenue growth for the next fiscal year, significantly higher than the 44% expected by analysts. This projection is driven by sustained, intense demand for compute power.
The acquisition of Hugging Face by NVIDIA for $12.9 billion is seen as a strategic move to win across all market segments, including open-source AI. This positions NVIDIA to capture a significant portion of the AI dollar, even if it means subsidizing open-weight models.
OpenAI has reportedly cut off access for Cursor, an AI-powered code editor. This action drew a notable response from Mike Trout, who expressed concern over the impact on traffic, and a brief comment from Elon Musk regarding Sam Altman.
Discussions arose regarding Cursor's flexibility in using custom codecs and the perceived inconsistency in data related to its usage over the past 60 days. The reliability of the '5% traffic' figure for OpenAI was questioned.
Instinct, a new AI assistant from Silicon Valley, has reportedly raised capital at a $2.5 billion valuation. The company is generating significant excitement within the tech community.
Aaron Katz, founder of ClickHouse, notes that the current AI growth cycle is accelerating at an unprecedented pace compared to previous internet and mobile cycles. He highlights that companies are growing rapidly and the demands on their systems are unlike anything seen before.
Aaron Katz suggests that while many AI companies operate with lower gross margins than traditional SaaS companies, investors may be satisfied as long as a path to margin expansion is demonstrated. He contrasts this with concerns about gross margins five years ago in enterprise software.
Aaron Katz identifies the durability of revenue as the single biggest risk for AI companies, particularly generative applications. He notes that switching costs can be very low for these applications as model providers rapidly leapfrog each other.
Aaron Katz describes a moment of 'AI awakening' at ClickHouse, where they realized they were not adopting AI coding applications fast enough. After sending an email to Anthropic, the company saw explosive growth in their usage of these applications.
Aaron Katz reflects that if he could change one thing over the past two years at ClickHouse, it would be increasing sales capacity earlier. He contrasts their 100 sales people with competitors who have thousands, covering fewer use cases.
Aaron Katz shares his biggest takeaway from his 12 years at Salesforce, working with Mark Benioff: 'That you can overestimate what you can achieve in one year and underestimate what you can achieve in five.' He notes how Benioff's vision and marketing created a perception that allowed Salesforce to grow into a major competitor.
Aaron Katz states that AI tooling has enabled ClickHouse to ship products faster and enter new product categories two years ahead of schedule. He explains their acquisition strategy: building organically when internal teams can innovate, and acquiring when founders are building on ClickHouse in strategic future categories.
ClickHouse has acquired Langfuse, a Berlin-based company with three founders, to bolster its capabilities in agent observability. Aaron Katz believes this technology will be essential for every enterprise.
Aaron Katz prioritizes revenue growth for his company, stating he worries less about expenses like coding agents if sustained revenue growth is observed. He believes that covering a broad service area and accelerating the roadmap justifies the spending.
Aaron Katz draws parallels between ClickHouse's sales approach and those of DataDog and Snowflake. He initially favored DataDog's product-led growth (PLG) but now recognizes the need to layer in an enterprise sales motion, learning from his Salesforce experience.
NVIDIA is making significant moves in the AI space, acquiring Poolside for $12 billion and investing $1 billion in Mercor and Perplexity's latest funding rounds. The Poolside acquisition includes licensing their model factory for $6 billion and moving 109 engineers to NVIDIA's Nemo Tron team. This strategic expansion aims to bolster NVIDIA's AI ecosystem and full-stack capabilities.
OpenAI's CFO, Sarah Fry, has confirmed that the company is definitely going public this year. This announcement comes as Anthropic is also pursuing an IPO, indicating a significant shift for major AI players into the public markets.
An investor letter revealed that a company was unable to raise $2 billion for GPU acquisition, forcing them to accept a $6 billion acquisition by NVIDIA. The letter detailed the immense capital intensity required for AI model development, suggesting that competing at the frontier is becoming increasingly difficult for many.
The conversation touches on the implications of NVIDIA's acquisitions and investments, suggesting a trend towards consolidation in the AI infrastructure market. While NVIDIA's strategy benefits existing investors and founders, questions remain about the survival prospects for smaller AI startups.
The discussion highlights that despite challenges in raising capital, companies can still achieve significant value through acquisitions in a rapidly growing market. The focus shifts from standalone profitability to the cash extractable from acquirers who possess capital, GPU access, and market position.
Showing the latest 15 of 29 covered episodes.