Jan 20, 2025 · 24-Hour AI Briefing: Gemini Explodes in Usage but Must Prove Retention, Zhipu Opens a Deployable “Hybrid Thinking” Model, and GTC 2026 Bets on Physical AI + Inference at Scale
Three updates from Google, Zhipu, and NVIDIA point to the same shift: AI is entering its second phase, where the hard problems are depth of use, reliability, scalable delivery, and cost curves—not just raw capability. Distribution got you “usage.” Now the market demands stickiness, enterprise-grade operations, and inference industrialization.
1. Gemini usage surges; Enterprise reaches 8M subscriptions across 1,500 companies, but depth and satisfaction remain open questions
Commentary:
Gemini isn’t just a standalone model—it “seeps” into daily workflows via Search, Gmail, Workspace, Chrome, and Samsung Galaxy devices. Google has proven reach and distribution. Phase two is whether usage evolves from “trial / shallow use” into “can’t live without it.”
8M enterprise seats and 1,500 companies look strong, but industry feedback highlights a classic tension: acquiring at scale can be easier than operating with precision. Enterprise buyers care not only about capability ceilings, but stability, cost control, and SLA-grade reliability.
So the critical metrics are renewal rate, active seat ratio, and how fast pilots expand from a department to company-wide deployment. By volume, Gemini’s numbers are impressive—now it has to win on depth and retention.
2. Zhipu releases and open-sources GLM-4.7-Flash: a “hybrid thinking” model with 30B total params and 3B active params
Commentary:
Zhipu’s signal is clear: turn “usable reasoning” into a lower-cost, deployable enterprise component instead of chasing only massive cloud-only models. With 30B total but 3B activated, it targets a practical balance between quality and efficiency.
This architecture can reduce inference compute and memory footprint while preserving expressiveness—well-suited for edge deployment, private cloud, and high-concurrency web services.
The risk is the stability of routing and “thinking” mechanisms: if complex tasks become inconsistent—too shallow when depth is needed—users will feel the boundary immediately.
Notably, Zhipu pairs free API access with an MIT license, lowering adoption friction for SMBs and independent developers—sharply contrasting closed, paid approaches.
3. NVIDIA confirms GTC 2026 (Mar 16–19, San Jose): Physical AI, AI Factories, and Inference are the three core themes
Commentary:
By centering Physical AI, AI Factories, and Inference, NVIDIA is signaling a pivot from “content generation” toward agent-driven interaction with the physical world—and toward industrializing inference delivery.
The attendee list spans global industry and academia, including major Chinese automakers/tech firms and US institutions and companies.
Even before product announcements, the theme itself is the message: NVIDIA believes the next growth wave comes from inference at scale, industrial-grade delivery, and real-world deployment. The market will care less about concepts and more about measurable improvements—throughput, latency, efficiency, TCO, and repeatable delivery. Are you excited for GTC this year?
Closing:
Gemini has already won on distribution-driven volume, but must prove retention and depth. Zhipu is betting on deployable, low-cost reasoning via open source. NVIDIA is pushing the narrative toward inference industrialization and Physical AI. In the next phase, the winners may be defined less by a single model’s peak and more by productization, reliability, and cost curves. Which path do you think compounds fastest—ecosystem distribution, deployable open-source reasoning, or industrial inference platforms?
Further reading (top AI events in the last 72 hours):