Edge computeEdge compute
Chips, SoCs, and firmware that let a robot run perception and control inference on-device, without a cloud round-trip.
4 companies 4 家公司 +3 named in conference materials, unverified +3 项来自会议材料，未经核实
system
Brain大脑
CE-NCE-N 52.3
Brain is the robot's thinking layer: the chips and software that decide what it does next. It's where investors are paying the richest premiums right now, even though our own scoring says it's the hardest layer on the board to actually solve with today's technology.
边缘算力、世界模型、VLA 策略与长程规划。共识给出最高倍数，本表却给出最低可解性。
Bottlenecks瓶颈 P04 Edge inference边缘推理 · P10 Predictive world models预测式世界模型 · P11 Generalisable VLA policies可泛化的 VLA 策略 · P12 Long-horizon planning & recovery长时程规划与恢复
What it isBrain is the cognition and control-software layer of the robot: everything that decides what the machine does next, from the chip at the edge to the model choosing the next action. Most industry frameworks (Morgan Stanley's own Brain/Body/Integrator split among them) treat “brain” as a single bucket, but this taxonomy splits it into four problems: P04, Edge inference, the on-robot chips and firmware that make safety-critical decisions without a cloud round-trip; P10, Predictive world models, the systems that forecast what happens next in the environment; P11, Generalisable VLA policies, the vision-language-action models that choose what the robot actually does; and P12, Long-horizon planning & recovery, the layer that sequences multi-step tasks and recovers from failure. Splitting it is the point: each has a genuinely different economic buyer and maturity profile, where edge silicon is sold to hardware integrators against a real bill of materials while the model layer is sold on a roadmap, so lumping them into one “brain” score would hide more than it reveals.
是什么大脑（Brain）是机器人的认知与控制软件层：决定机器人下一步做什么的一切，从边缘芯片到选择下一动作的模型。多数行业框架——包括摩根士丹利自己的“大脑/躯体/集成者”三分法——都把“大脑”当作单一板块处理，但本分类法将其拆分为四个子问题：P04，边缘推理，即让机器人无需云端往返即可做出安全关键决策的机载芯片与固件；P10，预测式世界模型，即预测环境下一步变化的系统；P11，可泛化的 VLA 策略，即决定机器人实际执行动作的视觉-语言-动作模型；以及P12，长时程规划与恢复，即串联多步任务并在失败后恢复的层级。拆分正是关键所在：这四者各自面对截然不同的经济买家与成熟度——边缘芯片卖给硬件集成商、有真实的物料成本可核算，而模型层卖的是路线图——把它们合并成单一的“大脑”评分只会掩盖差异，而非揭示它。
ThesisThe variant view: do not buy Brain as one position. On CE-N, P04, Edge inference ranks 2nd of 18 tracked problems site-wide (near the top of the board) while P11, Generalisable VLA policies, P10, Predictive world models, and P12, Long-horizon planning & recovery rank 6th, 11th, and 14th. The system's 52.3 average CE-N score masks that bimodal split rather than describing a uniformly middling layer. The re-rank on capital efficiency widens the gap further: P11, Generalisable VLA policies falls 9 places and P10, Predictive world models falls 8, and P12, Long-horizon planning & recovery scores 2.6 out of 5 on solvability (how tractable the problem is with today's methods), the lowest of any of the eighteen problems this index tracks. P04, Edge inference rises to 2nd of 18 on that same re-rank even as the model layer collapses. Score the chip and the model separately, because the evidence already does.
论点本表持有的是分歧观点：不要把“大脑”当作一个整体买入。在 CE-N 上，P04，边缘推理 在全站 18 个追踪问题中位列第 2，而 P11，可泛化的 VLA 策略、P10，预测式世界模型、P12，长时程规划与恢复 则分别位列第 6、第 11、第 14——该系统 52.3 的平均分掩盖的正是这种双峰分裂，而非描述一个整体中庸的板块。资本效率再排名进一步拉大了差距：P11，可泛化的 VLA 策略 下滑 9 位，P10，预测式世界模型 下滑 8 位，P12，长时程规划与恢复 的可解性得分仅 2.6/5，是本指数追踪的十八个问题中最低的。在同一次再排名中，P04，边缘推理 却上升至 18 个瓶颈中的第 2 位，而模型层却在坍塌。芯片与模型应分开打分，因为证据本身早已如此区分。
Products mapped to this system映射到本系统的产品
World models / Physical-AI foundation modelsWorld models / Physical-AI foundation models
Foundation models, VLA policies, and predictive world models that convert perception into action across embodiments, including a cluster of humanoid full-stack companies whose own product text names a VLA, world-action, or tactile-world-model component alongside their hardware.
20 companies 20 家公司 +3 named in conference materials, unverified +3 项来自会议材料，未经核实
Autonomy / planningAutonomy / planning
Autonomy software that plans and executes vehicle-level movement decisions rather than arm/body manipulation, spanning on-site autonomous aerial/aircraft operators plus additional autonomous-driving software named in the conference doc.
2 companies 2 家公司 +1 named in conference materials, unverified +1 项来自会议材料，未经核实
AI / robotics intelligenceAI / robotics intelligence
Lower-confidence, broader AI-for-hardware/robotics intelligence plays named in the conference doc that do not yet correspond to any linked on-site entity.
+3 named in conference materials, unverified +3 项来自会议材料，未经核实
Companies mapped to this system映射到本系统的公司
What the primary evidence shows
P04, Edge inference is the hardware and firmware that lets a robot make safety-critical, high-frequency decisions without waiting on a cloud round-trip: the chips, thermal design, and inference stack running on the robot itself. The named suppliers here have real, checkable financials, which is why this sub-problem carries the system’s highest confidence.
Ambarella’s edge-AI silicon: Q1 FY2027 revenue $100.4m, up 16.9% year over year, still GAAP loss-making ($18.1m) but non-GAAP profitable ($5.0m) — though GAAP gross margin fell 160 basis points year over year, the metric that actually tests pricing power.
NVIDIA’s Jetson AGX Thor (2,070 FP4 TFLOPS, 40-130W) is the reference platform design-win adopters (1X, Agility, Amazon Robotics, Boston Dynamics, Figure, Medtronic) are building against — but the lower-cost, higher-volume T3000/T2000 modules do not ship until Q1 2027, so any 2026 revenue thesis here is a thesis about developer kits and design wins, not shipments.
Horizon Robotics posted a large H1 2026 headline profit that its own filing attributes to a non-cash fair-value gain on a convertible loan, not operating performance — adjusted net loss actually widened to as much as RMB1.7bn over the same period.
P10, Predictive world models, P11, Generalisable VLA policies, and P12, Long-horizon planning & recovery are grouped as “the model layer” because the leading vendors are now architecturally merging them: NVIDIA’s next-generation GR00T N2 folds world-model prediction and action policy into one “World/Video Action Model” stack, a direction Unitree, 1X, and Ant Group’s Robbyant are all pursuing independently.
The capital chasing this layer is real and large — Physical Intelligence at $5.6bn (November 2025, with an $11bn round reported “in talks” for four-plus months since March 2026 without closing); Skild AI’s $1.4bn Series C at over $14bn (January 2026, led by SoftBank, backed by NVIDIA, Samsung, and Salesforce Ventures), roughly 3x its $4.5bn mark seven months earlier — but the evidence backing production readiness is thin. Skild’s headline deployment, on Foxconn’s Blackwell GPU-server line with NVIDIA, is described in every account as “early,” with no robot count, task list, or reliability metric disclosed.
The strongest technical counter-evidence — Google DeepMind’s Gemini Robotics 2 controlling a full humanoid across new embodiments in “a few hours with fewer than 200 examples,” with named hardware partners Apptronik, Franka, and Boston Dynamics — is real progress but still discloses no quantitative success rate.
Two independent academic benchmarks published in the last ten months, RoboDojo and RoboChallenge, put frontier generalist policies at roughly 12% and roughly 50% real-robot success respectively, against human baselines near 100%. That is the honest state of the art, not the 90%-plus figures found in unverified company claims.
What sell-side research adds
Morgan Stanley’s own internal TAM figures span $25tn (hardware only, by 2050), $5tn, $7.5tn, and “up to $50tn” for overlapping claims — an 8-10x spread from one house using the same top-down method on different denominators, which is itself the finding: treat the spread, not any single figure, as the evidence.
On P04, Edge inference, Barclays sharpens the framing directly, calling humanoids “as much a compute story as a robotics one” with large-scale economic deployment “closer to 2035 than 2030,” while Deutsche Bank corroborates independently that real industrial task success rates run “under 50% to over 99%” with payback “typically more than two years — not yet commercially viable industrially.”
On the model layer, the single most load-bearing new fact from this batch is architectural, not a benchmark: Morgan Stanley reports NVIDIA’s next-generation GR00T-N2 explicitly built on a “World/Video Action Model” architecture that folds P10, Predictive world models and P11, Generalisable VLA policies together, a direction Unitree’s own CEO frames the same way and that 1X, Ant Group’s Robbyant, and Rhoda AI are separately pursuing — a technical argument for scoring P10, Predictive world models and P11, Generalisable VLA policies jointly, though it says nothing about which direction a joint score should move, and no benchmark accompanies it.
Company-level sell-side data cuts both ways on Unitree: its IPO financials (5,500 humanoids shipped in 2025, RMB1.7bn revenue, roughly 60% gross margin, roughly 37% adjusted net margin) are real, but a separate Morgan Stanley note puts only 4% of China’s 12,000-plus 2025 humanoid shipments in industrial or logistics use against 42% in R&D and education — the same “shipments are not deployment” caution this system’s other primary evidence already carries.
And the largest new entity in the entire batch is structural rather than technical: SoftBank Group is consolidating twenty robotics bets, including its Skild AI stake, into “Robo Holdings” (over $13bn disclosed equity value) alongside a pending acquisition of ABB Robotics — the same related-party pattern already flagged elsewhere on this site (Schaeffler/Agility, Foxconn/Agility) now showing up at the model-layer capital level too, since the same balance sheet increasingly sits on multiple sides of brain-layer deals.
The investor read
Do not buy Brain as one position. P04, Edge inference behaves like a picks-and-shovels hardware layer — second-highest capital efficiency of eighteen tracked bottlenecks, checkable financials, real design wins independently corroborated by the sell-side batch — even though the volume ramp is a 2027 event, not a 2026 one.
The model layer is the opposite profile: it commands the industry’s richest and most internally inconsistent narrative (Morgan Stanley’s own TAM figures span an 8-10x range), yet P11, Generalisable VLA policies, P10, Predictive world models, and P12, Long-horizon planning & recovery score the lowest solvability of any of the eighteen bottlenecks this index tracks (2.7, 2.8, and 2.6 out of 5).
Value capture is the second risk, independent of whether the technology works: within 24 hours of Spirit AI taking the RoboChallenge leaderboard, it open-sourced the winning model, and AgiBot has separately released a free million-trajectory manipulation dataset. If the frontier policy and the frontier training data are both given away, the model layer may not be where the value settles even once the reliability numbers improve — and SoftBank’s roll-up of twenty robotics bets into one holding company is a reminder that a meaningful share of the capital chasing this layer may be circular rather than a market test.
Evidence quality tracks this split directly: P04, Edge inference and P11, Generalisable VLA policies carry this system’s strongest sourcing; P10, Predictive world models is moderate; P12, Long-horizon planning & recovery — arguably the least glamorous and least funded of the four — has the system’s thinnest evidential record and its lowest solvability score. Size conviction to the evidence tier, not to the headline valuation.
一手证据显示什么
P04，边缘推理 是让机器人无需等待云端往返、即可做出安全关键、高频决策的硬件与固件——运行在机器人本体上的芯片、散热设计与推理栈。这里的供应商有可核查的真实财务数据，因此该子问题在系统内置信度最高。
Ambarella 的边缘 AI 芯片：2027 财年第一季度营收 1.004 亿美元，同比增长 16.9%，GAAP 口径仍亏损（1810 万美元），非 GAAP 口径已盈利（500 万美元）——但 GAAP 毛利率同比下滑 160 个基点，这才是真正检验定价权的指标。
英伟达的 Jetson AGX Thor（2070 FP4 TFLOPS，功耗 40-130W）是设计中标客户（1X、Agility、Amazon Robotics、Boston Dynamics、Figure、Medtronic）正在围绕其开发的参考平台——但成本更低、出货量更大的 T3000/T2000 模块要到 2027 年第一季度才交付，因此任何针对 2026 年的营收论点，谈的都是开发套件与设计中标，而非实际出货。
地平线机器人（Horizon Robotics）2026 年上半年的高额账面利润，其自身财报归因于一笔可转换贷款的非现金公允价值收益，而非经营表现——同期经调整净亏损实际上扩大至最高 17 亿元人民币。
P10，预测式世界模型、P11，可泛化的 VLA 策略、P12，长时程规划与恢复——被归为“模型层”，是因为主要厂商正在架构层面将它们合并：英伟达下一代 GR00T N2 将世界模型预测与动作策略折叠进同一个“世界/视频动作模型”技术栈，宇树科技、1X 与蚂蚁集团旗下 Robbyant 都在各自独立追求同一方向。
追逐这一层的资本规模真实且巨大——Physical Intelligence 估值 56 亿美元（2025 年 11 月），一笔 110 亿美元的融资自 2026 年 3 月以来被报道“洽谈中”逾四个月仍未落定；Skild AI 14 亿美元的 C 轮融资使估值超过 140 亿美元（2026 年 1 月，软银领投，英伟达、三星与 Salesforce Ventures 参与），较七个月前的 45 亿美元上涨约三倍——但支撑量产就绪的证据薄弱。Skild 的标志性部署——在富士康与英伟达合作的 Blackwell GPU 服务器产线上——各方报道均称其为“早期”，未披露机器人数量、任务清单或可靠性指标。
技术层面最有力的反证——谷歌 DeepMind 的 Gemini Robotics 2 控制一个完整人形机器人，“数小时内、不到 200 个示例”即可适配新本体，并有 Apptronik、Franka、Boston Dynamics 等硬件合作伙伴具名——确属真实进展，但同样未披露量化成功率。
过去十个月内发布的两项独立学术基准测试 RoboDojo 与 RoboChallenge，分别将前沿通用策略的真实机器人成功率定为约 12% 与约 50%，而人类基线接近 100%。这才是行业的真实现状，而非未经验证的公司声明中动辄 90% 以上的数字。
卖方研究补充了什么
摩根士丹利自身内部给出的总市场规模数字，从 25 万亿美元（仅硬件，到 2050 年）到 5 万亿美元、7.5 万亿美元，再到“高达 50 万亿美元”不等，针对相互重叠的论点——同一家机构用同样的自上而下方法、套用不同的分母，得出跨度达八到十倍的数字，这一分歧本身就是发现所在：应把这种跨度、而非任何单一数字，当作证据来看待。
在 P04，边缘推理 方面，巴克莱把这一论点表述得更尖锐，称人形机器人“既是一个算力故事，也是一个机器人故事”，大规模经济性部署“更接近 2035 年，而非 2030 年”；德意志银行则独立印证，真实工业任务成功率“从不到 50% 到超过 99%”不等，回收期“通常超过两年——在工业领域尚不具备商业可行性”。
在模型层方面，这批研究中最具分量的新事实是架构层面的，而非某个基准分数：摩根士丹利报告称，英伟达下一代 GR00T-N2 明确建立在把 P10，预测式世界模型 与 P11，可泛化的 VLA 策略 折叠在一起的“世界/视频动作模型”架构之上，宇树科技 CEO 本人也采用同样的框架，1X、蚂蚁集团旗下 Robbyant 与 Rhoda AI 也在各自独立追求同一方向——这是把 P10，预测式世界模型 与 P11，可泛化的 VLA 策略 联合评分的一个技术论据，但它并未说明联合评分应朝哪个方向调整，也没有配套的基准测试。
公司层面的卖方数据在宇树科技身上呈现两面性：其 IPO 财务数据（2025 年出货 5,500 台人形机器人，营收 17 亿元人民币，毛利率约 60%，经调整净利率约 37%）是真实的，但摩根士丹利另一份报告指出，中国 2025 年逾 1.2 万台人形机器人出货量中，仅有 4% 流向工业或物流用途，42% 流向研发与教育用途——这与本系统其他一手证据已经提出的“出货不等于部署”告诫如出一辙。
而这批研究中最大的新增实体，是结构性而非技术性的：软银集团正把二十项机器人相关投资（包括其持有的 Skild AI 股权）整合进“Robo Holdings”（已披露股权价值超过 130 亿美元），同时还有一项对 ABB 机器人业务的待完成收购——这与本站在别处已经标记过的同一种关联方模式（Schaeffler／Agility，富士康／Agility）如出一辙，如今也出现在了大脑层的资本层面：同一张资产负债表越来越多地同时出现在大脑层交易的多个方向上。
给投资者的启示
不要把“大脑”当作单一头寸买入。P04，边缘推理 表现得像一条卖铲人式的硬件层——在全站 18 个瓶颈中资本效率排名第二，财务可核查，设计中标真实存在，且已被卖方研究批次独立印证——尽管其放量出货是 2027 年的事件，而非 2026 年。
模型层则呈现相反的画像：它拥有行业里最丰富、内部也最不一致的叙事（仅摩根士丹利一家给出的总市场规模数字跨度就达八到十倍），但 P11，可泛化的 VLA 策略、P10，预测式世界模型、P12，长时程规划与恢复 在本指数追踪的十八个瓶颈中可解性得分最低（5 分制下分别为 2.7、2.8、2.6）。
价值捕获是第二重风险，与技术是否成熟无关：Spirit AI 夺得 RoboChallenge 排行榜榜首后 24 小时内即开源了夺冠模型，AgiBot 也另行发布了一个免费的百万轨迹操作数据集。如果前沿策略与前沿训练数据都是免费的，即便可靠性数字持续改善，模型层也未必是价值最终沉淀的地方——而软银把二十项机器人投资整合进一家控股公司，也提醒我们：追逐这一层的资本中，相当一部分可能是循环资本，而非市场检验的结果。
证据质量与这一分化直接对应：P04，边缘推理 与 P11，可泛化的 VLA 策略 是本系统内证据最扎实的两项；P10，预测式世界模型 为中等；P12，长时程规划与恢复（四者中或许最不起眼、投入最少）证据记录最薄弱，可解性得分也最低。信念的强弱应锚定于证据层级，而非头条估值。
Contents目录
brain/