AI cost crunch as subsidized era ends
Key Questions
What financial losses has OpenAI reported?
OpenAI leaked a $38.5B loss amid the shift from subsidized models, with internal cost crunches also affecting Uber and Meta.
How are new models competing on cost?
Grok 4.5 undercuts rivals at $2/$6 per M tokens (90% cheaper per task), GPT-5.6 Terra runs at half prior cost, and Meta subsidizes Muse Spark 1.1 at a quarter of competitor pricing.
What infrastructure spending risks are highlighted?
A $5.5T CapEx supercycle carries $7T debt projection risk, with capital allocation conflicts potentially benefiting equity over bondholders.
How is the AI race shifting from size to efficiency?
Focus moves to cost-efficient routing, hybrid models, and cheaper systems like Perplexity's GLM integration and Ollama's Fortune 500 adoption.
What signals the end of the subsidized AI era?
Nvidia's revenue-sharing model, ZML's free inference server, and OpenAI paying $500K to bankers for training data indicate a pivot toward ROI and cost control.
OpenAI leaked $38.5B loss. Chinese models gaining US traction at lower cost, but China may ban exports. GPT-5.6 Terra tier at half GPT-5.5 cost. Grok 4.5 at $2/$6 per M tokens, undercutting Opus 4.8. Grok 4.5 #3 on CursorBench at 1/10th cost of Fable 5 Max. OpenAI GPT-5.6 Sol 54% token efficient on agentic coding. Meta Muse Spark 1.1 at $1.25/$4.25 per M tokens, beats Opus 4.8 and Grok 4.5 on OOD evals. Nvidia launches revenue-sharing model. ZML free multi-chip inference server. $5.5T CapEx supercycle risk. $7T debt projection. Hybrid model reduces costs. Cursor CFO council addresses ROI. Three giants shipped models in 24 hours, all leading with price — Grok 4.5 90% cheaper per task, GPT-5.6 54% token efficiency, Meta subsidizing Muse Spark 1.1 at quarter of rival cost. Meta AI cloud cost efficiency ($22B/GW vs $45B) challenges hyperscaler economics. GPT-5.6 Sol nearly matches Fable 5 at 1/3 cost, leads coding agent index. CVC investments show internal AI cost crunch at Uber and Meta. Shift from model size wars to cost-efficient routing: Perplexity GLM 5.2 integration, Benchmark 90% open-weight token prediction, Ollama Fortune 500 adoption. GPT-5.6 health intelligence push — Luna outperforms GPT-5.5 at 25x lower cost, targeting healthcare vertical. OpenAI paying Wall Street bankers $500K to train AI, not close deals — signals cost shift to training data. Capital allocation conflict in AI infrastructure spending: aggressive capex may benefit equity but risks bondholders.