Model routing is a fix for AI overspending. That's a problem for OpenAI and Anthropic

Corporate AI spending is coming under tighter budget discipline as companies shift from defaulting to frontier models toward model routing, which sends easy tasks to cheaper alternatives and hard tasks to premium models. Cognition says routine work can be 5x to 10x more cost efficient with adequate models, while Arvind Jain estimates about 95% of enterprise AI usage still runs on the most expensive frontier systems. The shift could pressure pricing power at OpenAI and Anthropic, even as frontier models retain value for the hardest jobs.

Analysis

This is less a demand slowdown than a margin-compression event for the AI supply chain. If enterprises move from “always-frontier” to workload-specific routing, the monetization stack bifurcates: premium model vendors retain the hard-edge workloads, but the high-volume, low-complexity traffic that subsidized their growth gets commoditized. That should pressure blended average selling prices over the next 2-4 quarters, especially for vendors whose usage growth depends on broad seat expansion rather than deeply embedded workflow value.

The second-order winner is infrastructure and orchestration, not raw model quality. Routing increases the importance of middleware, observability, evals, policy controls, and cost-management layers that help enterprises decide when to invoke which model. That is structurally supportive for Cisco’s AI-adjacent enterprise stack, but also for platform layers that sit between the enterprise and the frontier labs. The real bear case for the premium labs is that their take rate on enterprise tokens becomes more usage-efficient but less universal, which can force a re-rating from “all traffic grows at premium margin” to “only difficult traffic does.”

Timing matters: this should unfold gradually, not in a single-quarter air pocket. The near-term catalyst is board-level budget scrutiny and CFO KPI pressure, which can drive routing adoption faster than model quality improvements can offset it. The main reversal risk is if frontier models keep expanding their capability gap so aggressively that enterprises default back to simplicity over optimization; absent that, the secular path is toward more price discrimination and lower unit economics for premium inference.

The contrarian miss is that cheaper routing is not purely bearish for AI spend. If unit economics improve enough, total enterprise usage can expand materially because many tasks that were previously uneconomic become viable. That means the losers are likely the vendors with undifferentiated inference exposure, while the winners are the companies that can prove measurable output per dollar and own the workflow where routing decisions are made.

Key Decisions for Investors

• Short a basket of premium AI model beneficiaries on a 3-6 month horizon; prefer names most exposed to enterprise inference monetization and least differentiated via proprietary workflow lock-in. Risk/reward: asymmetric downside if routing adoption cuts blended pricing 15-25% before demand reaccelerates.
• Long CSCO or buy 3-6 month call spreads into the next enterprise spend guide revision; routing reinforces the value of control-plane, network, and governance infrastructure. Risk/reward: limited downside versus upside from AI budget reallocation into platform spend.
• Pair trade: long enterprise AI workflow/software enablers vs short frontier-model monetization proxies. Use a 6-9 month horizon and size for a gradual margin-compression thesis rather than an abrupt growth collapse.
• Buy medium-dated puts on high-multiple AI names that rely on broad token consumption rather than measurable productivity outcomes. Trigger: evidence of customer routing policies or usage-based budget caps in earnings commentary over the next 1-2 quarters.
• Avoid chasing any short in frontier labs until routing penetration is visible in actual billings; if unit economics improve faster than pricing decays, the entire AI spend pool can expand, making early shorts vulnerable to a 20-30% squeeze.