Why the rise of open source AI isn’t hurting Anthropic … yet

The article argues “open source won’t hurt frontier,” claiming enterprises increasingly shift production to lighter/open-source models while frontier labs retain discovery—supported by recent token-share data. On Vercel, DeepSeek leads token volumes at ~33% over the past week, but Anthropic still represents over half of overall token spend (despite slight share decline as prices rise). On OpenRouter, DeepSeek V4Flash processes ~5.3T tokens weekly versus Opus handling ~2T, yet Opus’s average token cost (~$1.37 per million vs ~6 cents) implies frontier models likely still capture most spend, suggesting frontier economics remain resilient as usage broadens.

Analysis

The key market mechanism is not displacement but expansion: cheaper production models lower the cost of deployment, which tends to widen the addressable workload base rather than shrink total AI compute demand. That is constructive for NVDA because the highest-value part of the stack still requires frontier-grade training and tuning, while the broader rollout of lighter models increases inference cycles and keeps accelerator utilization elevated. The bigger loser is likely not a chip vendor but AI software names whose value proposition depends on persistent API premiumization; once routing becomes price-sensitive, moat shifts from model access to workflow control.

Near term, the catalyst is less about token-share headlines and more about capex commentary from hyperscalers and frontier labs over the next 1-3 months. If spend keeps rising while mix shifts toward cheaper models, NVDA’s demand base remains intact; if pricing pressure starts outrunning token growth, the market will begin to question how long AI infra margins can stay above cycle averages. Over 6-18 months, the real risk is commoditization forcing ROI discipline across the AI stack, which could compress multiples for application-layer beneficiaries before it meaningfully dents compute demand.

The consensus seems to be over-reading open-source as a demand destruction event. A better framing is that open-source is a margin recycler: it pushes routine workloads down the cost curve, but frontier models retain the premium tokens tied to discovery, difficult reasoning, and new use cases. SBUX and LRRIF have no direct trade implication here.

Key Decisions for Investors

• Add NVDA on 3-5% pullbacks over the next 2-6 weeks; use a 1-3 month call spread to define risk. The bet is that mix shift to cheaper models expands total inference volume faster than it compresses AI capex demand.
• Relative-value: long NVDA / short QQQ or XLK for the next 1-2 months if AI remains the only durable earnings-growth pocket. This isolates AI infra outperformance from broader mega-cap multiple noise.
• Do not short frontier AI exposure purely on open-source token gains. Set a falsifier instead: if frontier spend share drops materially below ~40% and hyperscaler capex guides flatten, then reduce NVDA exposure.
• No action on SBUX or LRRIF from this tape; the article does not create a credible second-order revenue or margin read-through for either.