Can tech companies learn to love cheaper AI models?

AI buyers are shifting toward cheaper models, with Brian Armstrong projecting 80% of workloads could move to models that are 99% cheaper within 12-18 months. A Harvey/Fireworks AI test reportedly cut inference costs by 3x without reducing quality, suggesting enterprise users can reduce spend materially by routing tasks to smaller models. That could pressure big labs such as OpenAI and Anthropic by reducing inference demand and raising questions about frontier-model ROI.

Analysis

The market is likely underestimating how quickly AI spend can reprice from a frontier-model monoculture to a multi-tier procurement stack. The second-order effect is not just lower token bills; it is a change in budget allocation from model quality to orchestration, routing, and evaluation layers that decide when to spend on premium inference. That shifts value away from the “model as product” framing and toward infrastructure that can dynamically route tasks, compress context, and enforce quality gates.

The biggest near-term losers are the labs whose revenue mix is most exposed to default premium usage, because a small reduction in per-task model size can overwhelm underlying workload growth. That matters most over the next 3-12 months, before IPO narratives can fully re-anchor on training moat or consumer scale. If enterprise buyers learn that 70-80% of use cases are “good enough” on cheaper systems, the market will start discounting frontier training ROI, which could also tighten VC funding for model labs and adjacent application startups that were built around expensive inference assumptions.

The more interesting winners are not necessarily the cheapest model providers, but the distributors of demand: inference platforms, orchestration software, and vertical apps with strong workflow control. They benefit from model arbitrage and can monetize routing efficiency, usage analytics, and governance. A hidden risk is that lower inference costs may not expand demand proportionally if customers simply do fewer calls or shorter prompts; in that case, the whole AI capex chain, from GPUs to networking, sees a slower utilization ramp rather than a clean substitution boom.

Contrarianly, the move may be less bearish for frontier labs than headline takes imply if the premium tier becomes more concentrated and defensible. If only the top 10-20% of tasks require frontier capability, those calls could remain highly priced and strategically important, while cheaper models commoditize everything below them. That would compress breadth of usage but preserve pricing power at the top end, making the real trade a barbell: short undifferentiated model exposure, long picks-and-shovels routing and workflow layers.

Key Decisions for Investors

• Short the most inference-revenue-sensitive AI bellwethers on rallies over the next 1-3 months via put spreads in the next 2 earnings cycles; favor names where valuation already implies sustained premium-token growth. Risk/reward: limited carry cost, asymmetric downside if enterprise mix shifts faster than expected.
• Long a basket of AI infrastructure enablers with routing/orchestration exposure over 6-12 months (e.g., SNOW, DDOG, MDB, NET) versus short frontier-model proxies; the thesis is that value migrates from model intelligence to control plane. Use equal-dollar pair sizing to reduce broad AI beta.
• Buy call spreads on vertical AI application names with clear workflow ownership and pricing power over the next 6 months; they can capture lower COGS without needing frontier-model spend to remain elevated. Prefer businesses where inference is a pass-through cost rather than a core margin driver.
• Reduce exposure to pure GPU/AI-capex beneficiaries on any strength if utilization data starts to soften over the next quarter; the risk is not lower adoption, but lower tokens per workflow. Watch for commentary on context length, call frequency, and routing to cheaper models as leading indicators.
• For higher-conviction expression, pair long model-agnostic inference/routing software against short one or two premium-model names for a 6-12 month horizon; target a 1:2 risk/reward with tight discipline around earnings and product-launch catalysts.