20VC: OpenAI and Anthropic Will Build Their Own Chips | NVIDIA...

The countries that control compute will control AI, and you cannot have compute without energy, highlighting the foundational importance of these resources for AI dominance.

Jonathan Ross predicts NVIDIA could be worth $10 trillion in five years due to insatiable demand for compute, emphasizing the scale of the current AI infrastructure race.

The demand for compute is insatiable, with OpenAI and Anthropic stating they could double revenue with just double the inference compute, demonstrating the direct link between compute and AI service scalability.

Ross analyzes the current AI market not as a bubble, but as a high-investment phase akin to early oil drilling, where smart players are making significant returns despite the lumpiness of successes.

The AI market spend is characterized as "lumpy," with a few companies driving most of the revenue, mirroring the early days of oil exploration where instinct and a few big finds dictated success.

Ross questions investors about AI's future impact, finding no one certain it won't displace their jobs, which he uses to illustrate why hyperscalers are investing heavily to maintain leadership.

Ross shares an anecdote about using AI to deliver a custom feature in four hours with no human code writing, showcasing the value proposition of AI in rapid development and problem-solving.

The discussion explores whether major tech companies will move into chip manufacturing, with Ross emphasizing the immense difficulty and risk involved, citing failed efforts like Dojo.

The conversation touches on the concept of "Mag 7" companies needing to invest in chips to stay relevant, but Ross argues that successfully replicating NVIDIA's chip performance is a near-impossible task.

Ross clarifies that the investment in chips by AI companies is not an "infinite money loop" but a necessary spend on productive infrastructure that injects capital into the ecosystem.

The value increase in NVIDIA's stock is attributed to the belief that its AI compute will continue to drive revenue, fueled by the insatiable demand for more computational power.

The demand for compute is insatiable, with Ross suggesting that doubling compute for OpenAI or Anthropic could double their revenue within a month due to current rate limits and customer demand.

The importance of speed in AI applications is highlighted, drawing parallels to CPG margins and conversion rates on web pages, arguing that quick responses are crucial for user engagement and brand affinity.

Ross disputes the idea that AI latency is acceptable, stating that speed directly correlates to user engagement and competitive advantage, a lesson learned from building early internet companies.

The possibility of OpenAI or Anthropic verticalizing into chip manufacturing is discussed, with Ross acknowledging their ability but emphasizing the complexity of software and keeping pace with evolving technology.

NVIDIA's near-monopoly on HBM supply creates a bottleneck, incentivizing hyperscalers to explore building their own chips to secure necessary capacity and control their destiny.

The dominance of NVIDIA's GPUs is explained by how small performance edges translate into massive system value, making the chip cost negligible compared to the overall system deployment cost.

The HBM market presents a challenge due to NVIDIA's buying power, but a diversified supply chain for HBM manufacturers could create opportunities for competitors.

The capital cost and conservative nature of memory suppliers create a supply constraint for GPUs, even for a company with NVIDIA's financial power.

The cost of building data centers versus buying systems is debated, with data centers representing a larger, amortized cost over a longer period, but still a significant investment.

A faster payback period for chip investments is advocated for, suggesting that a three to five-year amortization cycle for chips might be too long given rapid technological advancements.

The value of a chip has two phases: initial deployment and ongoing operation, with the latter being more critical for profitability as long as operational costs are met.

The potential for excess compute supply due to rapid chip obsolescence is a concern, but long-term contracts and the cost of breaking them complicate this scenario.

The emphasis is on minimizing payback periods and operational costs to manage the lifecycle of chips effectively and avoid the risk of having underperforming, costly hardware.

Grok's value proposition is providing more compute capacity with a faster supply chain, a critical advantage when hyperscalers cannot meet demand.

The ability to provide more compute capacity is the primary value proposition, as it directly enables companies to serve more customers and therefore increase their revenue and lock-in.

Grok's supply chain advantage, with a six-month delivery timeframe for LPUs compared to NVIDIA's two-year GPU commitment, is a key differentiator that attracts hyperscalers.

The "hardware lottery" concept suggests that models are designed for existing hardware, creating a loop that benefits incumbents and makes it hard for new entrants to compete without a faster iteration cycle.

For new chip market entrants, a faster development and deployment cycle is essential because customers will not design their models for chips that won't be available for two years.

The current market prediction is that the US has a two-to-three-year advantage in the "away game" of AI, emphasizing the importance of speed and allied collaboration.

Despite past underestimations of compute needs, companies are now trying to overbuild capacity, but still often fall short due to the rapid growth of AI demand.

AI's impact on product quality differs from SaaS; AI allows for continuous improvement on each query by spending more compute, meaning more compute directly leads to better products.

The assumption that efficiency is the sole focus for future AI models like GPT-5 is questioned, as cost-sensitive markets require different optimization strategies than performance-driven ones.

The current focus on efficiency for models like GPT-5 is driven by the need to penetrate cost-sensitive markets like India, where low-cost AI access is paramount.

The misconception that Chinese AI models are cheaper to run is debunked; US-based OSS models are significantly more cost-effective for inference despite potentially higher training costs.

The cost to train an AI model is amortized over its inference usage, making cheaper inference a key competitive advantage, especially when scaling to a global audience.

The US holds a substantial compute advantage, enabling more intensive model training that lowers inference costs, giving it a strategic edge in the AI race.

The subsidy potential of the Chinese government for AI compute is acknowledged, but the US and its allies can compete by having more energy-efficient chips, crucial for countries with limited power capacity.

The "away game" for AI leadership is won by having better chips, which provide an advantage in regions where building massive energy infrastructure is not feasible, solidifying the US's position for the next few years.

The idea of open-sourcing models is discussed, with the conclusion that while brand strength can initially drive adoption, prompt compatibility and cost efficiency are key for wider uptake.

Prompt reusability is a significant factor in adopting open-source models, allowing users to transition to premium models without losing their existing work.

The intense energy requirements for AI compute are addressed, with nuclear power being efficient but renewables also playing a crucial role, and the key strategy being to locate compute where energy is cheap.

The US's fear of "mistakes of omission" drives its aggressive AI investment, contrasting with Europe's caution and legislative focus, which could lead to being left behind.

Europe could unlock significant energy potential through renewables like wind in Norway and nuclear power, but fear and a lack of decisive action are hindering progress.

Japan's rapid progress in building a two-nanometer fab and allocating substantial funds to AI shows a determination to compete, serving as a model for Europe.

The speed of building renewable energy infrastructure like wind turbines is a challenge, suggesting that governments or hyperscalers should fund these projects to meet AI's energy demands.

Collaboration with regions like Saudi Arabia, which are investing heavily in data centers and energy, could offer a solution for Europe's energy needs for AI.

The high cost of permitting in the US and Europe for infrastructure like nuclear power plants is a significant hurdle, potentially slowing down AI energy deployment.

The core takeaway is that controlling compute means controlling AI, and compute requires energy, making energy infrastructure a critical factor in global AI leadership.

Europe's current lag in the AI race is significant, but acting now could allow it to catch up, especially by leveraging allied strengths and focusing on energy solutions.

Model sovereignty alone is insufficient to win in AI; abundant compute is the ultimate determinant of AI performance, as even a superior model will underperform if it lacks sufficient computational power.

The question of whether European models like Mistral can compete hinges on the availability of sufficient compute, not just on their origin or ownership.

Grok, despite being a compute provider, aims to avoid competing with its customers by not developing its own AI models, creating a safer ecosystem for developers.

While companies like CoreWeave offer on-demand GPUs, their finite allocation limits their ability to fully address the massive demand for compute.

As AI models mature, inference becomes more critical, and specialized hardware like Grok's LPUs may offer advantages over GPUs for this specific task.

NVIDIA is expected to continue selling GPUs due to their established brand and the virtuous cycle between training and inference, where more inference requires more training.

The inference market is evolving rapidly, with the unexpected rise of language-based AI making it accessible to a wider audience and driving faster adoption than initially predicted.

The massive adoption of AI, with 10% of the world's population being weekly active users of services like ChatGPT, is astounding, but compute remains a bottleneck for quality and broader language support.

The AI ecosystem relies on data, algorithms, and compute; improving any one of these can enhance AI, but compute is the most accessible and consistently improving factor.

The demand for compute is underestimated because adding more compute directly improves AI quality and user experience, leading to increased demand and economic growth without traditional bottlenecks.

The economy's strength is predicated on a shift from labor spend to AI, with AI expected to create labor shortages rather than widespread unemployment due to increased productivity and new industries.

AI is predicted to cause massive deflationary pressure, leading to reduced costs for goods and services, and consequently, people working less and retiring earlier.

AI will create new jobs and industries that are currently unimaginable, similar to how agriculture's automation freed up labor for new economic pursuits.

The US can benefit from AI advancement due to policies that facilitate energy production and infrastructure development, which is crucial for scaling AI compute.

The concept of "vibe coding" is examined as an enduring skill, analogous to reading and writing becoming universal, enabling non-programmers to create functional tools.

Margins in the compute business are important for stability and weathering market volatility, but the primary focus should be on customer value and solving unmet needs.

The strategy for margins is to keep them as low as possible for business stability while retaining the ability to increase prices if needed, leveraging volume to drive profit.

Jevon's paradox applies to compute: increased availability leads to increased sales as costs decrease, creating a virtuous cycle of innovation and adoption.

Businesses should focus on customer needs and differentiation rather than solely on the bottom line, as solving unique problems leads to customer loyalty and revenue growth.

AI's impact is expanding the Total Addressable Market (TAM) by making products easier to use and enabling new functionalities, leading to overall revenue increases even if per-unit costs decrease.

The value of AI lies in its ability to deliver tangible benefits, as evidenced by private equity firms seeking cheap AI compute to improve their businesses' bottom lines.

The potential for AI to add more labor to the economy by enhancing compute and AI capabilities is unprecedented, suggesting a significant shift in economic paradigms.

A short-term speed bump in AI development could have a significant multiplier effect on the economy due to the concentration of value in major AI companies.

Economic downturns can be challenging but also create opportunities for robust businesses to emerge, highlighting the cyclical nature of markets.

The economy's unpredictability stems from human behavior and the impact of predictions on those predictions, making forecasting difficult.

A major problem in AI is the dispersion of talent, where engineers can raise vast sums to start new companies, hindering the growth of existing startups by fragmenting expertise.

The war for tech talent is more aggressive than ever, but unlike sports, there are no salary caps or limitations on the number of startups, allowing for unlimited team creation.

Google is praised for its turnaround and AI DNA, with Gemini showing success in adoption, though consumer product integration needs further refinement.

OpenAI has established a significant distribution advantage, making it challenging for incumbents like Google to catch up, despite their efforts to iterate on AI products.

The enduring value of OpenAI and Google in the AI landscape is likely, with Anthropic potentially carving out a niche in areas like coding.

Engineers are shifting between AI tools like Sourcegraph, Anthropic, and Codex based on what is currently the "best tool," indicating low switching costs and a dynamic competitive environment.

The rapid switching of AI tools by cutting-edge engineers raises questions about the enduring value of specific platforms if switching costs are low.

Both OpenAI and Anthropic are considered undervalued, as their R&D efforts are expanding the market itself, not just competing within a finite space.

The bull case for AI labs is that they will achieve similar market value to current tech leaders and potentially expand the "Mag 7" into "Mag 9" or more.

Successful tech companies naturally tend to move up the value stack and subsume parts of their customers' businesses, a tendency that AI labs may also follow.

Grok's decision not to build its own AI models is a strategic choice to avoid competing with its customers, ensuring a safe infrastructure for developers, though this could be a potential risk.

Grok's recent $750 million raise at a nearly $7 billion valuation was driven by investor confidence in their ability to execute and the positive margins on their hardware sales.

Hardware companies like Grok have an advantage with positive margins on their products, unlike software companies whose margins can vary depending on the model and operational costs.

Grok aims for low margins for business stability while retaining the ability to increase prices if market demand necessitates it, leveraging volume for profit.

In five years, NVIDIA is predicted to retain significant revenue share but a smaller portion of chip sales as custom chip development increases and customers gain more power to choose alternatives.

NVIDIA's brand value allows them to charge premium prices, but this can make them less competitive as customers prioritize business success over brand loyalty when making purchasing decisions.

Customers with significant spending power, like those in the top 35-36 clients in the AI market, will increasingly make decisions based on business success rather than brand name.

The prediction is that NVIDIA will be worth $10 trillion in five years, driven by its established position and the ongoing demand for AI compute.

Grok's potential to reach $10 trillion valuation is linked to its ability to produce near-unlimited compute without the supply chain constraints faced by others.

Grok's competitive advantage lies in its ability to produce compute at scale and cost-effectively, addressing the market's primary need beyond just brand reputation.

Grok's innovation in chip design, particularly with SRAM integration, allows for more efficient multi-user capabilities on a single hardware unit, optimizing system-level performance.

The cost of SRAM per bit is inherently higher than DRAM due to its architecture, but Grok's system-level approach and massive chip deployment reduce overall cost per token.

Grok optimizes compute at a world-scale distribution level, load balancing across multiple data centers globally to provide tailored solutions based on geographical needs and model performance.

Grok has the capacity to significantly increase its supply chain build-out rate, having raised substantial funds and being oversubscribed, indicating a readiness to meet extreme demand.

Grok's cost per token, especially with its speed advantage, is highly advantageous, allowing them to charge less than the market while still building a stable business through volume.

The focus for Grok is purely on execution and satisfying compute demand, not on immediate profitability or public offerings, reflecting a commitment to core business growth.

The biggest misconception about NVIDIA is that its CUDA software ecosystem is an insurmountable moat for inference, as specialized hardware and developer adoption of alternative platforms are challenging this.

If founding Grok today, Ross would not pursue chip development due to the long lead times, advocating instead for faster-moving areas of the AI ecosystem.

The temporal moat of chip development, requiring years from design to production, is a significant barrier to entry, making it difficult for new chip providers to compete with established players.

Larry Ellison's meteoric rise with Oracle is attributed to brilliant business decisions and a willingness to move aggressively and embrace risk, a strategy Ross advocates for in the AI space.

When everyone else is fearful about AI, it's an opportunity for smart investors to be greedy, and vice versa, emphasizing contrarian investment strategies.

Investors should be greedy where moats exist, and in the early stages of AI investment, focus on identifying companies that have the potential to build strong, defensible moats.

Ross's approach has shifted from preserving optionality to intense focus, believing that saying "yes" to fewer, more strategic opportunities leads to better business outcomes.

Elon Musk's Grok and XAI may succeed by differentiating their offerings, as AI markets are expected to diverge, with different companies excelling in specific niches.

AI markets will eventually diverge, with companies like Anthropic focusing on coding and others potentially specializing in different AI applications.

Companies that do not differentiate themselves risk dying, as seen in the tech landscape where clear business focuses are essential for long-term survival.

Microsoft's AI strategy, leveraging its OpenAI partnership and ability to deploy compute, positions it well, while Amazon's lack of AI DNA is a concern despite its compute capacity.

Amazon's lack of AI DNA is identified as a weakness, contrasting with Google and Meta, which have historically strong AI foundations, and Microsoft's strategic acquisition of AI capabilities through OpenAI.

Ross is excited about the future of AI, viewing LLMs as a "telescope of the mind" that will expand human intelligence and reveal new truths about the universe.

The biggest problem in AI is the difficulty in attracting and retaining top engineering talent due to excessive funding for new startups, which fragments expertise.

The rapid adoption of AI by 10% of the world's population is astonishing, but compute limitations and language support are holding back even wider use.

AngelList is recommended as a modern platform for venture funds, offering an all-in-one software solution and a dedicated service team to manage fund operations efficiently.

.tech domains are suggested for tech founders to secure the ideal startup name without compromise, signaling innovation and a focus on technology.