The AI industry's opulence and leverage mirror conditions before the 1998 collapse of the hedge fund LTCM. A peripheral market shock could cause a domino effect, leading to a sudden, dramatic failure of a major AI player that currently seems invincible.
Enterprises shouldn't lock into a single AI lab like OpenAI or Anthropic. Instead, they need a multi-model strategy using evals and routing to leverage different models for different tasks, ensuring they aren't trapped when the 'seasons' inevitably change.
Companies are designing expensive, multi-year infrastructure plans based on today's AI models. Because model architectures evolve so rapidly, these highly optimized but inflexible systems will quickly become suboptimal, leading to a huge waste of capital on underperforming hardware.
The biggest opportunity in applied enterprise AI is not industry-specific solutions, but horizontal platforms deployed with a localized go-to-market strategy. This 'Uber playbook' focuses on winning geographies rather than verticals, a non-consensus approach.
Aggressive token budgeting prevents employees from experimenting and discovering new AI-native workflows. Companies that overly restrict usage will not see the productivity gains needed to reshape their business and will ultimately be outcompeted by those that encourage more liberal use.
An isolated LLM is relatively harmless. The real danger for enterprises is when an LLM is connected to other systems via APIs, creating a vector for sensitive data to be exfiltrated. This makes governance of integrations, not just the model, the critical security focus.
According to SemiAnalysis, multiple major Chinese AI labs are signaling to inference providers that their next frontier models will not be open source. Instead, they will be available only through licensing, suggesting a rapid decline in the open-source movement for top-tier models.
The initial 'give me everything' hype cycle for enterprise AI is over. Buyers now demand clear ROI and cost justification. This shift from broad experimentation to budget reconciliation will have significant downstream impacts on the entire AI vendor ecosystem.
Many viable AI product ideas are currently impossible because the cost of search makes their unit economics unworkable. The next wave of innovation will be unlocked not just by better models, but by bringing search costs down another order of magnitude.
Unlike the volatile LLM space, the world of physical AI—robotics and autonomous systems—diffuses more slowly and linearly. The inherent safety risks of operating in the real world prevent the explosive growth and subsequent crashes seen in software, creating a more stable development trajectory.