Thinking Machines debuts Inkling, a giant open model it admits is not the best

Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first AI model, Inkling. The model is open-weight, allowing developers and companies to download and use it. The article frames the launch as part of a broader manifesto emphasizing experimentation (“keeping the weirdness alive”).

Analysis

The market is likely to misread this as a pure open-source headline, but the real mechanism is margin compression at the model layer and revenue expansion at the infrastructure layer. If a credible team outside the incumbent frontier labs can ship usable open-weight models, pricing power migrates away from API-only vendors and toward whoever owns distribution, compute, and enterprise workflow integration.

Near term, the biggest loser is not necessarily the best-known model company; it is any mid-tier AI vendor whose product is mostly a wrapper around third-party models. Those names face a 1-3 month risk of multiple compression if investors conclude model access is becoming interchangeable. By contrast, cloud and GPU providers should see a longer-duration benefit as self-hosting and fine-tuning increase token consumption even when per-token pricing falls.

The contrarian point is that open-weight can be net bullish for total compute demand. Enterprises often choose local control for security and cost predictability, which raises inference volume and shifts spend from software licenses to infrastructure. The thesis breaks if the model underwhelms technically; if it benchmarks near frontier, the pricing pressure on closed model APIs becomes a 6-18 month structural issue rather than a one-day sentiment event.

Key Decisions for Investors

• Long NVDA vs. short C3.ai (AI), 1x beta-adjusted, for a 1-3 month relative-value trade: open-weight adoption should lift GPU utilization while commoditizing the valuation premium in pure AI narrative names; cover if AI shows accelerating bookings or better gross margin discipline.
• Overweight SMH on weakness over the next 1-3 months: the best second-order beneficiary is not the model lab but the compute stack, especially if enterprise self-hosting drives inference demand; risk/reward improves if the sector sells off on 'open-source fear.'
• Prefer AMZN or MSFT over pure-play model vendors for a 6-18 month horizon: both monetize higher AI workload intensity through cloud consumption, while model-layer pricing gets competed down; falsify this if cloud AI revenue growth decelerates despite rising public model usage.
• No immediate directional trade in private frontier-model names; treat this as a watch item and wait for evidence of API price cuts, benchmark parity, or enterprise deployment wins before shorting the model layer.