NOVA — Industry & Products | AGORA Intelligence

Gemini 4 Carbon costs $2 per million input tokens and $10 per million output tokens until the end of November, then twice that. Google claims a cost per task "more than 2x" lower than the best rival model on three agent benchmarks and publishes a single price sheet, the introductory one; if Google calculated that advantage on introductory pricing, at the December price it falls close to parity. Carbon also uses more tokens than Argon, so the contract should be negotiated on post-promotion pricing and on cost per task.

NVIDIA releases GR00T 2 as an open model and makes public 100,000 of its 436,414 pretraining hours, about 23%. Adaptation takes half an hour on an RTX 5090, while training and industrial benchmarks run on Isaac Lab and Newton. For industrial buyers, the lock-in is likely to be NVIDIA's simulation pipeline and compute more than the model.

The Dynamo recipes NVIDIA published on November 8, 2026 for A.X K2, K-EXAONE 2.0, Solar Open 2 and Motif-3 all target the B200 GPU. The performance doubling, "up to 2x" according to NVIDIA, would at most mean serving the same traffic with half the GPUs on that single piece of hardware. The thesis is that sovereignty over the weights stays Korean while the exit cost shifts to compute.

Rebellions will connect its R200 NPUs to NVIDIA's NVLink Fusion for SK Telecom's data centers. In the rack described in the announcement, three of four layers belong to NVIDIA: the interconnect, the MGX architecture and the prefill GPUs. The Korean chip takes decode and is expected to be ready in 2028, 14 to 26 months after the announcement, while the Korean minister also seeks to secure GPU supply.

From December, with iOS 27.3, Siri, Writing Tools and Visual Intelligence in China will run on Alibaba's Qwen, because there jurisdiction picks the model. Trailing Huawei by 1.8 points in the third quarter, Apple is buying off-the-shelf compliance, and Alibaba gains the most durable access to Chinese users.

Google is giving K-12 schools separate controls for Gemini, Gemini notebook and Search, all off by default, while Search is still set to become an AI chatbot. New York's moratorium, which the city calls the broadest in the country, covers 0.4% of the students Google turned Gemini on for in August. That suggests the buyer's leverage lies in the clause written in at renewal.

The $15 billion in data center revenue Qualcomm is promising for 2029 is about 1.4 times the 2026 revenue reported by The Verge and almost six times the annualized pace of its Q1 2027 guidance. For boards, that figure is leverage in compute contract renewals today, before it is supply.

Pliant is opening B2B payments to Claude, ChatGPT and customers' own agents with an $80 million Series C: the most likely reading is that the agent becomes replaceable and value shifts to whoever holds the license and the liability. In the US, where part of the round is headed, that layer depends on regulated partners, so boards need to ask who signs in each country.

In NVIDIA's post on cities, Nemotron recurs in five of eight cases while prices and contract terms are missing from all of them. Service delivery stays with integrators such as ITMAX and 1MillionBot, so municipalities risk their strongest lock-in in the long-term service contract and shared infrastructure. All figures are claims by NVIDIA and its partners.

In Seoul, Siemens and Naver put payback on display in place of the pilot: 10,000 machines paid back in under three months according to Busch, and more than 200 robots in service for over four years according to Choi. Industrial AI lock-in thus shifts to the data connector, the barrier that 41% of respondents to the audience poll named as the first.