Twelve months ago, US frontier models controlled roughly 70 percent of AI traffic. Today, Chinese open-weight providers, led by DeepSeek, Z.ai, Moonshot, and Minimax, account for 45 to 61 percent of top-tier model traffic on OpenRouter, with DeepSeek alone processing more tokens than Google, Anthropic, or OpenAI individually.
In this Big Story edition, Ray Rike and Peter Buchanan unpack how this shift happened, how US labs and regulators are responding, and three scenarios for how the closed versus open weight competition plays out for enterprise AI buyers.
Key topics discussed:
The pricing collapse driving enterprise migration. DeepSeek made a 75 percent price cut permanent in May, bringing its V4 Pro model to a fraction of a cent per million tokens versus $2.50 per million for GPT 5. Minimax delivers GPT 5.5 class coding performance at 5 to 10 percent of the cost. This is why Uber, Microsoft, and Walmart are now implementing formal usage governance on frontier models rather than treating cost control as temporary.
The Mythos and Fable shutdown as a trust event. The 18-day suspension of Anthropic's top models over export control concerns spooked global enterprise buyers who realized mission-critical workloads could be cut off without warning. This single event accelerated the adoption of open-weight alternatives and pushed allied governments to invest in sovereign AI capacity.
Distillation attacks and the IP leakage problem. Anthropic accused Alibaba's Qwen lab of running a large-scale adversarial distillation campaign, using tens of thousands of accounts and tens of millions of exchanges to extract agentic reasoning capability from Claude. This reframes the security conversation from model safety to unauthorized technology transfer, which is a distinct and arguably bigger risk for any enterprise relying on proprietary model capability as a moat.
Cybersecurity parity is closing faster than expected. Multiple Asian labs, including Z.ai's GLM 5.2, Beijing based 360 Security, and Japan's Sakana AI, now claim benchmark performance approaching Anthropic's Mythos model on vulnerability detection and both offensive and defensive cyber tasks, often at significantly lower compute cost. This weakens the safety and capability gap argument that has justified restricting access to frontier models.
Real deployments have moved from theory to production. Coinbase cut AI spend in half after migrating to Z.ai and Moonshot's Kimi models, even as token usage grew. Cursor shipped a coding tool built on Kimi, with a Grok-based version reportedly imminent. Andreessen Horowitz estimates 80 percent of its portfolio companies already use open-weight models in production AI products.
Three scenarios for how this settles, and why the decision belongs at the board level. The hosts outline bifurcation (premium closed models for regulated use cases, open weight for commodity workloads), export control entrenchment (Washington treats the Fable ban as a template rather than a one-off), and capability convergence (the rationale for unilateral bans erodes as the performance gap closes). All three are already visible simultaneously, which means enterprise AI architecture decisions, including primary and backup model orchestration, are becoming strategic decisions that belong with the CEO and board, not just the technical team.
Why does this podcast episode matter for enterprise executives selecting models, especially for agentic AI deployments?
The vendor you choose today may not be the vendor you can use tomorrow, for reasons that have nothing to do with model quality. Regulatory risk, geopolitical exposure, and pricing volatility are now first-order variables in model selection, alongside capability and cost. Any agentic AI architecture built on a single model provider carries concentration risk that didn't exist a year ago. Building orchestration flexibility with primary and backup models, and understanding the true economics behind token pricing, is quickly becoming a board-level governance question rather than a procurement detail.
For the metrics and framework to instrument and report ROI on your AI investment, get the Big Book of AI Metrics at benchmarkit.ai under Media.
Podden och tillhörande omslagsbild på den här sidan tillhör
Ray Rike. Innehållet i podden är skapat av Ray Rike och inte av,
eller tillsammans med, Poddtoppen.
