China’s recent technology narrative turned sharply inward this week as a high‑profile consumer‑rights broadcast and a string of industry announcements exposed both the fragility and accelerating ambitions of the AI and robotics ecosystem.
State broadcaster CCTV’s annual 3·15 consumer rights programme aired an investigation that illustrated how bad actors can deliberately “poison” the inputs feeding large AI models. Investigators bought a software package marketed as a “GEO optimization system” that fabricated a product called the Apollo‑9 smart band, automatically generated multiple promotional posts and seeded them across e‑commerce and other internet channels. When journalists queried several mainstream Chinese large models using the product name, responses ranged from correct rejection to confidently wrong endorsements — a reminder that models absorb and amplify low‑quality or deliberately misleading internet content.
The episode is important because it demonstrates a practical attack vector against models that rely on web scraped or retrieval‑augmented knowledge: adversaries can manufacture believable artifacts on the open web, engineer their propagation, and thereby distort what models regard as credible. Different model providers reacted unevenly — some flagged the product as dubious, others refused to answer, and at least one treated the fabricated brand as a plausible niche offering. The result was a live test of model provenance and the limits of current content‑quality safeguards.
At the same time, Elon Musk announced that Tesla will start construction within days on a wafer fab, Terafab, to produce its next‑generation AI chips. Tesla’s motive is straightforward: the company is designing a fifth‑generation autonomous‑driving chip (AI5) and says third‑party suppliers cannot meet its future volume needs. The planned plant is described as larger than Tesla’s car “gigafactories” and reflects a broader industry shift toward vertically integrated compute stacks when demand and geopolitical pressure collide.
These two stories — about poisoned data and sovereign chip capacity — expose the two technical constraints now shaping AI deployment. On the front end, models are only as reliable as the information they ingest and the provenance systems that trace it. On the back end, companies are racing to secure chip capacity and reduce exposure to supply bottlenecks. Both developments incentivize firms that can control more of the stack: data, algorithms, compute and hardware.
The robotics sector supplied a third data point. At the China Appliance and Consumer Electronics Expo (AWE 2026), Shenzhen‑based Songyan Power (松延动力) showcased humanoid models dubbed “Little Bumi” and “Little Naughty N2,” announced a consumer delivery on the show floor and unveiled a partnership with JD.com for retail scenarios. Songyan says it scaled from order backlog to mass production in months by tightening its supply chain, building a local production base and self‑developing key components — an encouraging sign for firms that can combine systems engineering with manufacturing agility.
Market responses were measured. China’s AI and robotics ETFs slipped amid broader A‑share volatility, but trading patterns suggested rotation rather than wholesale investor exit: smaller AI plays saw turnover and accumulation, while the large robot portfolio experienced lighter flows. Investment houses at AWE urged a “barbell” strategy that balances domestic demand capture and export‑ready product lines, hinting that investors still see a long horizon for hardware‑enabled AI growth despite short‑term noise.
Taken together, these developments illustrate an industry transitioning from proof‑of‑concept novelty to operational reality — and encountering real‑world trust, supply and scaling problems. The immediate policy and commercial responses will matter: improved data provenance, provenance‑aware retrieval, provenance labelling and tougher platform moderation can blunt poisoning attacks, while more aggressive on‑ and near‑shoring of semiconductor capacity will reconfigure capital flows and supplier relationships. For investors and policymakers, the choice is becoming clearer: back vertically integrated players who can guarantee both the data and the compute, or accept higher risks in a fragmented stack.