Capability Moves Forward, Costs Become Visible: Two Curves of the AI Industry

Capability Moves Forward, Costs Become Visible: Two Curves of the AI Industry

A handful of signals from the field point to the same structural shift: model capability keeps advancing, while the bill, the boundaries of responsibility and the way AI gets used are becoming visible. What follows is not a list of headlines, but six signals assembled into one picture.

1. Frontier labs are already betting on the next generation

Boris Power, OpenAI's Head of Applied Research, says 80 to 90 percent of the company's research already goes toward GPT-7, GPT-8 and beyond; improvements within a single generation are deliberately framed as short-term bets. He also says model performance is not the biggest problem. That is worth noting for teams in China: once capability stops being the main bottleneck, the remaining bottlenecks tend to sit in process, trust and usage patterns.

2. Agents do the work in development, but humans still decide

A research team analyzed 769 task logs from building its own AI model: AI agents supplied up to 55 percent of method proposals, but humans made more than 85 percent of final decisions, and a third of the tasks would not have been attempted without AI. The authors warn that more agent activity does not mean responsibility has shifted — it does not automatically mean better outcomes either.

3. Embodied AI skips the dedicated control layer

Researchers from Stanford and Caltech had a humanoid robot powered by GPT-6 Astra independently tidy up an unfamiliar kitchen. Their HomeBody system skips a specially trained control layer, letting the language model call directly into modular skills such as grasping and navigating. The value of this signal is not that a robot can clean a kitchen, but the architectural choice: treat the language model as a dispatcher instead of training yet another controller.

4. Small models move down into narrow scenarios

Nvidia released Nemotron 3 Diarization, a free 100M-parameter model that identifies in real time who is speaking at any given moment in a conversation, distinguishing up to eight speakers. Compared with general models in the hundreds of billions of parameters, this class of specialized small model sits closer to deployable scenarios: meeting notes, customer-service QA, multi-person collaboration.

5. Commercial loops and cost bills arrive together

Google has started a limited test in India that lets users buy from Walmart-owned Flipkart through Gemini and AI Mode. It covers select products and users, with a broader rollout planned for later in October.

The bill is arriving from another direction too: Blue Cross Blue Shield says hospital use of AI tools led to an additional $942 million in healthcare spending over a two-year period.

6. Reading the two curves together

Placed side by side, these signals show an asymmetry: iteration on the capability side keeps accelerating, while measurement, accountability and scenario validation on the cost side are still early. Agentic commerce is a "limited test"; frontier labs are extending their research horizon two generations out; and in healthcare, the cost of usage has already been quantified by an insurer at $942 million. The real divergence is not about whose model is stronger, but about who can first make the process, the accounting and the responsibility legible.

Conclusion: four action items for teams in China

1. Separate "proposal rights" from "decision rights" in R&D workflows. Using the 55 percent and 85 percent figures as a reference, count agent contributions in process metrics while keeping final decisions and quality accountability explicitly human.
2. Favor specialized small models when choosing scenarios. Real-time speaker diarization at the 100M-parameter scale is easier to get working in narrow settings such as meetings and customer service than a general large model, and easier to keep per-call costs under control.
3. Build AI cost attribution early. The $942 million case shows that without independent accounting, AI-driven spending shows up as one large bill that is hard to break down after the fact.
4. Agentic commerce is still in limited validation. The Gemini and Flipkart test covers select products and users, which suggests payment, inventory, after-sales and compliance flows are not fully worked out; do not commit resources on the assumption that it is already fully rolled out.

Sources

• The Decoder — OpenAI says 80 to 90 percent of its research already targets GPT 7 and beyond: https://the-decoder.com/openai-says-80-to-90-percent-of-its-research-already-targets-gpt-7-and-beyond/
• The Decoder — AI agents do more of the work in model development, but humans still make the decisions: https://the-decoder.com/ai-agents-do-more-of-the-work-in-model-development-but-humans-still-make-the-decisions/
• The Decoder — Researchers plug GPT-6 Astra directly into a robot and let it clean up an unfamiliar kitchen: https://the-decoder.com/researchers-plug-gpt-6-astra-directly-into-a-robot-and-let-it-clean-up-an-unfamiliar-kitchen/
• The Decoder — Nvidia drops a free 100M-parameter model that identifies up to eight speakers in real time: https://the-decoder.com/nvidia-drops-a-free-100m-parameter-model-that-identifies-up-to-eight-speakers-in-real-time/
• TechCrunch — Insurers claim AI is already increasing healthcare costs: https://techcrunch.com/2026/09/26/insurers-claim-ai-is-already-increasing-healthcare-costs/
• TechCrunch — Google tests buying from Walmart-owned Flipkart through Gemini and AI Mode in India: https://techcrunch.com/2026/09/26/google-tests-buying-from-walmart-owned-flipkart-through-gemini-and-ai-mode-in-india/

Related tools

• 模型选型顾问