How Cheap AI Could Derail OpenAI And Anthropic's IPOs
The video examines whether OpenAI and Anthropic can justify their $800B+ IPO valuations as Chinese open-source models undercut them on price while matching frontier capabilities. Cohere CEO Aidan Gomez argues that regulated industries requiring secure, private deployments represent the durable premium market, but even that niche is contested by American competitors. The broader thesis is that AI pricing power is eroding faster than Wall Street's valuations reflect.
Summary
The video investigates whether OpenAI and Anthropic can sustain the trillion-dollar valuations they are seeking in upcoming IPOs, arguing that the market conditions underpinning their pitch are deteriorating rapidly. The core pitch — that these labs have durable pricing power akin to Microsoft or Google — is being undermined by two simultaneous pressures: cheap Chinese open-source models from labs like DeepSeek, Moonshot, and ZHIPU eating the low end of the market, and American competitors like Cohere, Nvidia, and Reflection AI targeting the high-trust enterprise segment.
On the cost side, the gap is stark. Benchmarking firm Artificial Analysis found that Anthropic's Claude Opus costs nine times more than the cheapest Chinese alternative for equivalent work. On OpenRouter, the largest AI traffic aggregator, Chinese models went from 1% of usage in 2024 to over 40% in 2025, with three of the top five models being Chinese. Chinese labs, constrained by U.S. export restrictions on Nvidia chips, were forced into algorithmic efficiency, producing smaller, cheaper, and increasingly competitive models.
The one remaining stronghold for American frontier labs is regulated, high-trust industries — banking, healthcare, defense, and grid operators — where Chinese models are simply not an option regardless of price or performance. Cohere CEO Aidan Gomez explains that his company exclusively targets these high-security enterprise settings, deploying models on-premises, air-gapped, even inside submarines with no internet connectivity, on as few as 2-4 GPUs. This compute-constrained environment structurally rules out massive frontier models like OpenAI's rumored 10-trillion-parameter system.
Gomez also highlights an emerging cyber security concern: as AI models increasingly write production software, the origin of those models becomes a national security issue. Using Chinese models to generate code for critical infrastructure risks the subtle introduction of vulnerabilities that could be exploited to shut down power grids or financial systems. This dynamic continuously raises the trust barrier for enterprise adoption and reinforces the case for private, democratic-aligned deployments.
However, even in this premium segment, OpenAI and Anthropic face American competition. Cohere, Nvidia's open-source Nemo Tron models, and startup Reflection AI are all building capable, efficient, trustworthy alternatives at a fraction of frontier prices. The video also notes that Elon Musk's decision to merge xAI into SpaceX — backstopping a $1B/month burn rate with SpaceX's revenues — signals that even the most aggressive AI investor felt a standalone AI lab needed diversification, a hedge OpenAI and Anthropic cannot make.
Gomez closes by expressing confidence in overall AI demand, calling it a rising tide, but acknowledges the market is shifting toward smaller, more efficient models as CFOs scrutinize AI spend at production scale. He also reflects on the public image challenges facing the industry, calling for more empathy and responsiveness to legitimate criticisms around energy consumption, artist compensation, and the pace of disruption.
Key Insights
- Artificial Analysis found that Anthropic's Claude Opus costs nine times more than the cheapest Chinese alternative to perform the same work, and Chinese models went from 1% to over 40% of usage on OpenRouter in a single year.
- Aidan Gomez argues that Chinese labs gained an efficiency advantage specifically because U.S. export restrictions on Nvidia chips forced them to innovate algorithmically, producing smaller and cheaper models that now rival American frontier labs.
- Gomez contends that as AI models increasingly write production code, the national origin of those models becomes a critical security issue — Chinese-origin models could subtly introduce exploitable vulnerabilities into critical infrastructure software.
- Cohere deploys models on as few as 2-4 GPUs in air-gapped, on-premises environments including submarines, which structurally excludes massive frontier models like OpenAI's rumored 10-trillion-parameter system from competing in that market.
- Gomez argues that even without model updates, the enterprise industry is still catching up to capabilities from 18 months ago, meaning demand for AI is not primarily constrained by frontier model performance but by enterprise adoption pace.
Topics
Full transcript available for MurmurCast members
Sign Up to Access