How Cheap AI Could Derail OpenAI And Anthropic's IPOs

The IPO Ambitions and Market Position of OpenAI and Anthropic

• OpenAI and Anthropic are planning to go public with initial public offerings (IPOs) at valuations of over $800 billion, with the pitch that they will be the next major tech giants with lasting pricing power, but this pricing power is already being threatened by Chinese open-source models and American competitors 10s.
• The growth of these companies has been significant, with 80x growth per year if annualized, but the gap between Chinese and American models is closing rapidly, with Chinese models eating into the low-end market and American competitors targeting the high-end 1m30s.

Cost and Value Considerations in Enterprise AI Adoption

• As enterprises roll out AI projects across entire workforces, they are beginning to question the cost and value of AI models, and whether other models may be more performant and cost-effective, which could lead to a shift in spending 2m6s.
• Running a $10 million AI budget on Anthropic's top model, Claude, could burn through the budget in weeks, while running it on China's open-source model, Deepseek, could stretch the budget across most of a year, with cloud costs being nine times more than the cheapest Chinese alternative 3m30s.

The Rise of Chinese Open-Source AI Models

• Chinese labs, such as Moonshot, Xiaomi, Deepseek, and ZHIPU, have shipped open-source models that match or nearly match American frontier models on key benchmarks, and adoption is following the price, with three of the top five models on Open Router being Chinese 4m40s.
• The dominance of Chinese companies in open-source AI is due in part to their ability to develop smaller, more efficient models and algorithms, which has allowed them to close the gap with American companies despite being cut off from Nvidia's best chips 6m10s.

American Frontier Labs and Infrastructure Spending

• American Frontier Labs are spending hundreds of billions of dollars on AI infrastructure, but this is not necessarily a fundamental advantage, as most advances in AI come from algorithms and computer science, and Chinese companies are making significant strides in these areas 7m30s.
• The one area where American companies still have a stronghold is in trust-based industries, such as banking, healthcare, and defense, where customers are willing to pay a premium for Western, democratically-aligned technology, but even here, American competitors are building models that could squeeze OpenAI and Anthropic 9m40s.

Emerging American Alternatives to OpenAI and Anthropic

• Companies like Cohere, founded by Aidan Gomes, one of the authors of the paper that kicked off the modern AI era, are building smaller, more efficient models specifically for regulated industries, and have seen significant revenue growth, with Cohere's revenue increasing 6x last year 10s.
• Nvidia, a company that U.S. enterprise already trusts, is now shipping its own open source models called Nemo Tron, positioning them as an alternative to both Chinese options and the closed Frontier Labs, and has gained adoption from companies like Palantir, Salesforce, Servicenow, and CrowdStrike 1m2s.
• Reflection AI, a startup that just raised at a multibillion dollar valuation, is building open source frontier models as an American alternative to Deep Seek, and is going after the same gap as Cohere and Nvidia, which is to provide capable models at a fraction of frontier prices on infrastructure U.S. enterprises already trust 2m6s.

Challenges and Constraints for OpenAI and Anthropic

• OpenAI and Anthropic face not only a China problem but also an America problem, as they do not have the option to diversify their businesses like Elon Musk did by merging his AI company xAI into SpaceX, and will have to be judged by Wall Street on AI economics alone 3m45s.
• The market is moving in the direction of companies like Cohere, which builds models for regulated industries, and the demand for premium AI models is driven by the need for trust, security, and high-quality models, with companies willing to pay a premium to access Western, democratically aligned technology 6m15s.

Cohere's Focus on Secure and Enterprise-Grade AI Models

• Cohere's business model is focused on building models from scratch for the enterprise side, particularly in high-security settings such as grid operators, financial services, and government, where the data and systems that the models are accessing are national security concerns 9m20s.
• Cohere uniquely excels at secure deployments, whether on-premises or completely air-gapped, allowing for deployment inside a customer's data center or in high-security settings with no cyber risk, and this level of security provides a unique value proposition 10s.
• The high-security settings require deployment on extremely limited compute footprints, typically 2 to 4 GPUs, which presents a technological constraint that rules out massive models and necessitates a focus on right-sized models for the served market 42s.

Security and Deployment Challenges in AI

• OpenAI and Anthropic, two of the biggest frontier labs, are turning their attention to security with Mythos, but their massive models, such as those with 10 trillion parameters, are not suitable for deployment in compute-constrained environments 2m6s.
• The emergence of massive models capable of sophisticated cyber offense and defense use cases reinforces the significance of private deployments to prevent exposure of sensitive software and infrastructure to potential exploits 2m6s.
• The risk of exploiting vulnerabilities in software and infrastructure is higher than ever, making private deployment essential, especially in critical sectors like finance, healthcare, and energy 4m30s.

Trust and Origin Concerns in AI Model Adoption

• The trust barrier to enterprise adoption of AI models is increasing, and the origin of the models is becoming a concern, with potential risks associated with using models from untrusted sources, such as those from China 6m15s.
• While Chinese models can be useful in low-sensitivity settings or for startups looking for low-cost options, they may pose risks in high-sensitivity environments due to potential backdoors, and hyperscalers like AWS may promote them for certain applications 8m40s.
• The preference is to use democratically supported and aligned technology, and companies like Cohere are contributing to a more democratic open-source approach, with the goal of continuing to make this technology available, especially for less secure settings, where cheaper options may be sufficient 10s.

Market Shift Toward Efficiency and Cost Optimization

• The calculation for enterprises is changing, particularly for those not in industries that require high security, as open-source models are getting closer to the frontier and the cost of using large models is becoming a constraint, leading to a shift in the market towards more efficient and smaller models 1m42s.
• The need for efficient models is driven by the compute bottleneck, as there is not enough compute to support the use of massive models, and companies like Cohere are building smaller, more efficient models that are good enough for specific use cases 2m6s.
• The market is entering a phase where CFOs are looking to optimize expenses spent on AI models, and companies will need to find ways to use smaller, more efficient models to reduce costs, without shifting to a Chinese tech stack 3m15s.

Global Contributions to Open-Source AI and Model Development

• The gap between American and Chinese AI models is closing rapidly, with countries like Canada, France, and Germany also contributing to the development of open-source models, and companies like Cohere, Nvidia, and Reflection are working on open-source options 4m30s.
• China is setting itself ahead by distilling large models, which is against the terms of service of many models, giving them a shortcut to catch up, while other companies are building independent stacks from scratch, requiring more effort and work to build competitive models 6m0s.

Efficiency and Infrastructure Trends in AI

• The trend in the market is moving towards more efficient models, with a big wave of companies seeking to reduce costs and make things more efficient, and the huge infrastructure spend by hyperscalers may not be justified, as companies like Cohere are building specialized models on just a few GPUs 8m40s.
• The revenue of AI companies has been growing rapidly, with a six-fold increase in the past year, and this growth is expected to continue, driven by virtually insatiable demand, 10s.
• As the infrastructure for AI expands, there will be a push towards efficiency, resulting in more AI being used with fewer GPUs, and the hyperscalers, such as OpenAI and Anthropic, are building the necessary infrastructure to serve this demand, 42s.

Market Fragmentation and Deployment Preferences

• However, the hyperscalers will not capture the entirety of the market, as on-prem deployments and private data centers are becoming increasingly popular due to security concerns, and are expected to account for half of the market, 1m30s.
• The demand for frontier models, such as those developed by OpenAI and Anthropic, is expected to be extraordinary, and the industry is still in the early stages of exploring the applications of AI in the enterprise sector, 2m6s.

Pricing Models and Business Strategies in AI

• The pricing of AI models is not expected to increase due to compute constraints, as companies do not charge customers for the compute used in private deployments, and instead provide the software and let customers purchase their own compute, 4m10s.
• The company's business model is focused on providing the software and platform for AI models, and does not involve building infrastructure, such as data centers, but rather integrating with existing tools and data to automate workflows, 5m20s.

Reflections on AI's Public Image and Industry Challenges

• The coauthor of a famous paper that started the modern AI age reflects on the current state of the industry, including infighting among top figures and growing public backlash, and believes that the industry has mishandled the public image of AI, 7m30s.
• The development of AI technology is expected to have significant impacts on people's lives, both professionally and as consumers, and it is crucial that this technology is developed and communicated in an accessible way, allowing the public to try it, test it, and provide feedback 10s.
• There are concerns and skepticism about AI technology, including resistance from the creative industry, with some people feeling that it may not be beneficial, and to address this, it is essential to better compensate artists and ensure they can contribute to the development of the next generation of models 1m20s.

Societal and Economic Impacts of AI

• The adoption of AI technology also raises concerns about the energy crisis, as it requires a significant amount of energy to power, which could lead to increased costs for consumers, making it a priority to invest in energy infrastructure to prevent price increases 2m6s.
• Criticism of AI technology is seen as a positive thing, as it allows for awareness of its weaknesses and shortcomings, and being aware of the potential consequences is essential for the long-term success of the technology 2m40s.
• The narrative around AI technology is also important, with some people feeling that there is an image problem, and it is necessary to have empathetic, kind, and thoughtful spokespeople who can respond to criticism with action to mitigate the potential downsides 4m10s.

Responsibility and Leadership in AI Development

• Companies like OpenAI, led by Sam Altman, and Anthropic, are at the forefront of AI development, and it is essential for them to be responsive to criticism and to work towards addressing the concerns and skepticism surrounding AI technology 5m20s.