Cloud-led Future · Intelligence-driven Hotline +86 21 6228 0217 中文 | EN

Embodied AI Enters an Order-and-Cost Verification Phase as Capital Rotates Back to Chips

Overview

Three signals this week point in the same direction while landing in different places. First, an organisational shift in the general-model approach. According to foreign media reports confirmed by the company, OpenAI has restructured its stand-alone robotics team: robotics researchers moved into the core research organisation led by the chief research officer, while applied robotics and hardware staff moved into the applied organisation led by the president. The company says robotics remains a long-term core direction and the change is intended to accelerate progress. Reports note that early in the year the team expected advances to come from video models, but that general models’ reasoning and multimodal capabilities have since allowed some robotic tasks to draw directly on general-purpose capability. The company has yet to show its own humanoid robot and aims for a first public demonstration in 2027.

Second, commercial commitments are becoming tied to quantified conditions. Logistics operator GXO and humanoid robot company Apptronik signed a letter of intent to expand trials around goods-to-person and person-to-person picking, tote and container handling, packing and sorting. Field trials are planned to expand from 2027, and if agreed productivity, safety and operational metrics are met, the parties intend to drive deployment of up to 1,000 Apollo robots globally over four years, with initial North American deployment expected in the second half of 2027, using a robotics-as-a-service model. Importantly, the 1,000 units are a conditional future deployment intention, not a signed purchase order or delivered volume.

Third, capital flows are rotating. A financial media report on 15 November noted diverging investor views on embodied-AI hardware and models, with funds withdrawing from whole machines and models and rotating back into semiconductors, concentrating on embodied-AI chips. Disclosed transactions in the same period include a 500 million yuan angel round for a company building a unified big-brain/small-brain system-on-chip for robot-side inference, with a development and validation platform planned for 2027 and a first chip in 2028; roughly 1.7 billion yuan raised in October by a high-compute chip maker in two rounds; a 500 million US dollar round valuing a German humanoid and robotic-arm maker at 8.5 billion US dollars, with its plant expected to produce around 2,000 units in 2026; and an over-100 million yuan round for a domestic construction-robot firm whose plastering robots cover roughly 300-350 square metres per day, with 2026 product orders expected to grow more than 100 percent.

Background & Interpretation

1. Technology Evolution and Context

Embodied-AI architecture is evolving from a single dedicated model toward layered collaboration. Robots traditionally required separately trained vision, motion and control models. As general models gain reasoning and multimodal capability, high-level task planning, semantic understanding and exception handling can potentially reuse general-purpose capacity, while real-time response, motion safety, force-control precision and environmental adaptation still depend on dedicated models and low-level control. The restructuring reflects exactly this reassessment: folding robotics research into the core model organisation does not eliminate the dedicated control layer, but shifts the centre of competition upward to general planning and downward to reliable execution. Whether a stand-alone team remains necessary depends on whether the interface between the two can be standardised.

2. Core Drivers and Underlying Mechanisms

The first mechanism is a shift in validation from demonstration to unit economics. Tying deployment scale to productivity, safety and utilisation, and using RaaS to lower customers’ upfront capital outlay, pushes suppliers to carry longer-horizon operating and maintenance risk; competition moves from whether an action can be performed to cost per task, continuous operating hours and failure rates. The second is compute migrating from cloud training to on-device deployment. Latency, offline availability and power constraints make edge inference chips a distinct category, and unified big-brain/small-brain SoCs attempt to carry visual perception, model inference and motion control on one die - the capital rotation into chips is effectively pricing this bottleneck. The third is differences in supply-chain maturity. Technology routes for whole machines and models have not converged and valuations lack comparable anchors, whereas chips, sensors and reducers have clearer mass-production timelines and customer validation paths, so capital concentrates upstream where claims are verifiable when uncertainty rises.

3. Market Structure and Value-Chain Impact

Division of labour is likely to deepen. General-model vendors supply high-level planning, control systems and real-time execution remain with specialists, edge chips and sensors set physical limits, and operators define product specifications backward from scenario data and efficiency metrics. For whole-machine makers, valuation logic shifts from technical narrative to order quality - conditional intentions, pilot contracts and volume purchases need separate disclosure, and conflating them amplifies expectation volatility. For chip makers, validation and launch timelines in 2027-2028 imply limited near-term revenue, so financing amounts should not be treated as industrialisation progress. For logistics, construction and manufacturing users, RaaS and narrow-scenario robots provide low-barrier experimentation; the construction-robot path of stabilising a single process through digitalised workflow before adding environmental understanding shows that commercialisation need not start from a general model.

Implications & Outlook

Near term, the industry enters a period of dense disclosure of order and cost data, with attention moving from launch demonstrations to verifiable operating metrics. Over the medium term, if reuse of general models for high-level planning is validated, value in dedicated models and whole machines may concentrate in data, control and system integration; if edge compute fails to meet power and cost targets, mass-production timelines will be constrained. Variables to watch include whether field-trial productivity and safety metrics are published, per-task pricing under RaaS contracts, tape-out and measured performance of edge SoCs, and the ratio of delivered units to intentions in disclosed orders. On the capital side, whether the rotation from machines to chips proves durable or merely a defensive phase also matters.

Takeaways for Industry Participants

Investors should verify intended units, pilot units and delivered units separately, and use cost per task, utilisation and long-run stability as valuation inputs rather than parameter counts or demonstration videos. Machine and solution vendors should disclose conditional clauses and acceptance criteria proactively to reduce expectation mismatch, and include maintenance, spare parts and remote diagnostics in product cost models. Chip and component firms should define compute, latency and power targets around real operating conditions and build validation platforms with machine makers early. Application users should specify efficiency baselines, safety responsibility and data ownership in pilot contracts, retaining the right to scale deployment in stages against metrics rather than locking in unvalidated volume at once.

This column compiles industry information and shares technical perspectives; it does not constitute investment advice.