2023至2025年，生成式數字AI掀起全民科技熱潮，AI聊天、圖文生成、智能辦公工具快速普及。但進入2026年上半年，生成式AI的應用普及進入平緩期。純軟體數字AI賽道增速持續放緩，價格戰在壓縮利潤，市場增長天花板逐步顯現。

而能夠感知現實環境、操控實體裝置的物理AI，正在迎來規模化商用拐點。

輝達CEO黃仁勳在2026財年股東大會提出，2026年是物理AI元年，數字AI只是產業上半場，物理AI才是數十年維度的核心增長主線。夏季達沃斯《2026十大新興技術》報告，也將能夠實現物理AI的世界模型列為未來五年重塑全球產業的頭號技術方向。

那麼，到底什麼是物理AI？為什麼它才能真正撐起AI產業的下半場呢？

From 2023 to 2025, generative digital AI sparked a nationwide tech boom, with AI chatbots, image-text generation tools and intelligent office software gaining rapid popularity. Yet in the first half of 2026, generative AI penetration has plateaued. Growth across purely software-driven digital AI tracks keeps slowing, squeezed profit margins from price wars are emerging, and the market is hitting visible growth ceilings. By contrast, physical AI — technology capable of sensing the physical world and controlling tangible hardware — has reached an inflection point for large-scale commercial adoption.

Jensen Huang, CEO of NVIDIA, stated at the 2026 fiscal AGM that 2026 marks the first year of physical AI. Digital AI merely represents the industry’s opening chapter, while physical AI will serve as the core growth engine for decades to come. The Top 10 Emerging Technologies of 2026 report released at the Summer Davos Forum also ranks world models powering physical AI as the top technology set to reshape global industries over the next five years.

So what exactly is physical AI? And why is it poised to anchor the second half of the AI industry’s growth story?

所謂物理AI，是能夠感知和理解現實世界物理規律、並操控實體裝置在真實環境裡自主執行任務的AI。簡單來說，傳統AI主要在電腦裡處理文字圖片，而物理AI能控制機器人、汽車在現實中行動，主要應用於具身智慧型手機器人、自動駕駛、工業製造等。

物理AI的核心價值，在於徹底打破了過去AI只服務於數字場景的邊界，第一次讓人工智慧的能力直接作用於工業生產、城市運行、生活服務等實體場景，打開了比虛擬AI大出數倍的市場空間。

不同於生成式AI主要聚焦內容創作、客服互動、演算法推薦等輕場景，物理AI的核心是“感知-決策-執行”的全鏈路閉環。通過視覺、觸覺、雷射雷達等多模態感測器採集物理世界的真實資料，再通過AI模型即時分析決策，最終直接驅動機械臂、機器人、自動駕駛車輛、工業裝置等實體終端完成操作。這套邏輯直接對準了實體產業的核心痛點——國內工業領域當前有超過8000萬個重複性高危崗位，相關場景的智能化替代具備廣闊的市場空間。

Physical AI refers to artificial intelligence that perceives and comprehends the physical laws of the real world, and controls physical hardware to autonomously complete tasks in real environments. Simply put, conventional AI mainly processes texts and images on computers, while physical AI directs robots and vehicles to act in reality, with major applications covering embodied intelligent robots, autonomous driving and industrial manufacturing.

Its core value lies in breaking the confinement of AI to digital scenarios. For the first time, AI capabilities are directly applied to physical sectors including industrial production, urban operation and daily services, unlocking a market several times larger than that of virtual AI.

Unlike generative AI, which focuses on lightweight scenarios such as content creation, customer service and algorithmic recommendation, physical AI features a full closed loop of "perception-decision-execution". It collects real-world data via multi-modal sensors including cameras, tactile sensors and LiDAR. AI models then conduct real-time analysis and decision-making to drive physical terminals like robotic arms, robots, autonomous vehicles and industrial equipment into action. This addresses a key pain point of tangible industries: over 80 million repetitive and high-risk jobs exist in China’s manufacturing sector, creating massive room for intelligent replacement.

從市場走勢來看，原來的數字AI僅能輸出方案、無法改造實體世界，是過去三年資本集中炒作的賽道，但紅利逐步見頂，賽道進入存量博弈階段。IDC最新資料顯示，2026年全球企業AI總支出約9400億美元，但其中純生成式軟件、MaaS模型服務增速明顯回落。國內大模型呼叫服務市場2026年營收預計186億元，行業玩家超6000家，同質化競爭下通用大模型API報價兩年降幅超70%，六成中小AI軟體企業無法實現穩定盈利。

相比之下，物理AI打通“感知—世界模型推演—實體執行—資料回流”閉環，機器人、自動駕駛、工業柔性產線、無人礦山都是其落地形態，剛性需求量大，營收將呈現長期增長趨勢。

In terms of market trends, digital AI, which only generates solutions instead of transforming the physical world, was heavily hyped by capital over the past three years. Yet its dividends have peaked, leading to stock-based competition. Latest IDC data puts global corporate AI spending at around $940 billion in 2026, with sharp slowdowns in pure generative software and MaaS growth. China’s large model inference service market is projected to hit 18.6 billion yuan in 2026 with over 6,000 industry players. Amid homogeneous competition, generic LLM API prices have plunged more than 70% in two years, and 60% of small and medium AI software firms fail to secure steady profits.

In contrast, physical AI forms a complete loop of "perception-world model simulation-physical execution-data feedback". Its real-world deployments include robots, autonomous driving, flexible industrial production lines and unmanned mines. Driven by strong rigid demand, its revenue will maintain long-term growth.

過去五年，物理AI始終停留在試點階段，核心制約在於感測器成本過高、模型對複雜物理場景的適配能力不足。而在2026年的當下，這兩大瓶頸已經出現實質性突破，行業規模化普及的拐點正式到來。

一方面核心硬體成本在快速下探，國內供應鏈的技術突破給物理AI的大規模落地掃清了成本障礙。另一方面大模型技術成熟，讓物理AI的場景適配能力大幅提升。現在多模態大模型可以實現“通用感知”，部署到新的工業場景僅需一周的微調即可投入使用，適配效率較此前大幅提升。

從落地情況來看，物理AI在工業領域的應用正在加速滲透。根據產業生命周期的普遍規律，滲透率達到10%左右即標誌著行業從匯入期進入快速成長期。接下來物理AI將進入為期5-8年的高速普及期，市場規模的年複合增長率將保持在40%以上，遠高於當前生成式AI的15%左右增速。

Over the past five years, physical AI remained limited to pilot projects, held back by two core bottlenecks: exorbitant sensor costs and insufficient model adaptability to complex physical environments. In 2026, substantial breakthroughs have been made in both areas, marking the official inflection point for large-scale industry adoption.

For one thing, core hardware costs have dropped rapidly, and technological advances in domestic supply chains have removed cost barriers to mass deployment of physical AI. For another, mature large model technologies have greatly boosted its scenario adaptability. Modern multimodal large models deliver universal perception; they only require one week of fine-tuning to launch in new industrial scenarios, drastically improving adaptation efficiency compared with earlier solutions.

Deployment data shows accelerating penetration of physical AI across industries. Per standard industry lifecycle rules, a penetration rate of roughly 10% signals a shift from introduction to rapid growth stage. Physical AI is poised to enter a 5–8 year high-speed popularization phase with an annual compound growth rate (CAGR) above 40%, far exceeding generative AI’s current growth rate of around 15%.

全球權威機構Future Markets 2026年3月發佈測算報告，給出清晰的長期空間資料：2026年全球物理AI市場規模約3830億美元，到2040年將攀升至3.26兆美元，14年規模擴張超8倍，長期復合增速穩定維持40%以上。細分賽道增量同樣亮眼，摩根士丹利預測，全球機器人硬體銷售額將從2025年1000億美元增至2030年5000億美元，2040年有望觸及9兆美元。

而且，國內市場成長彈性更強。36氪研究院資料顯示，中國具身智能市場規模從2018年2133億元增長至2025年9150億元，2026年將正式突破兆元關口，年復合增速22.84%。當前國內物理AI整體產業滲透率不足5%，類比2018年新能源車起步階段，行業仍處在發展早期，增量空間充足。

In March 2026, leading global research firm Future Markets released projection data outlining its vast long-term market potential: the global physical AI market will reach approximately $383 billion in 2026 and surge to $3.26 trillion by 2040, representing over 8-fold expansion in 14 years with a sustained long-term CAGR above 40%. Segmented tracks also boast remarkable growth. Morgan Stanley forecasts global robot hardware sales will rise from $100 billion in 2025 to $500 billion by 2030, potentially hitting $9 trillion by 2040.

China’s domestic market exhibits even stronger growth momentum. Data from 36Kr Research Institute indicates China’s embodied intelligence market expanded from 213.3 billion yuan in 2018 to 915 billion yuan in 2025, and will top one trillion yuan in 2026 at a CAGR of 22.84%. China’s overall physical AI industry penetration stands below 5% at present, analogous to the new energy vehicle sector in its 2018 infancy, leaving massive untapped growth potential.

2026年上半年，投融資資料明顯看出資本已經開始佈局。據統計，2026年上半年國內具身智能、物理AI全產業鏈共發生288起融資，披露總金額突破460億元；僅2至4月單月融資額分別達95.8億元、145.5億元、172.8億元，4月單月融資規模相當於2025年全年總額的38%。

而且，資金流向高度集中。前10家頭部企業拿走近242億元融資，千尋智能3個月完成4輪融資累計近45億元，投後估值突破200億元；銀河通用拿下25億元B+輪融資，創下國內人形機器人單筆融資紀錄。一級市場也出現明確趨勢，2026年上半年超50%融資流向世界模型、模擬等“機器人大腦”企業，硬體本體企業融資佔比僅12.8%。

Investment and financing data for H1 2026 reveals clear capital allocation toward physical AI. A total of 288 financing deals closed across China’s full embodied intelligence and physical AI industrial chain, with disclosed funding exceeding 46 billion yuan. Monthly financing from February to April hit 9.58 billion, 14.55 billion and 17.28 billion yuan respectively; April’s single-month fundraising equaled 38% of the full-year 2025 total.

Capital is highly concentrated among leading players. The top 10 enterprises secured nearly 24.2 billion yuan. QX Intelligent completed four fundraising rounds within three months raising almost 4.5 billion yuan, with post-money valuation topping 20 billion yuan. Galaxy Universal closed a 2.5 billion yuan Series B+ round, setting a record for the largest single financing of domestic humanoid robot firms. The primary market shows a distinct trend: over 50% of H1 2026 funds went to developers of robot "brains" such as world models and simulation platforms, while hardware manufacturers accounted for merely 12.8%.

從全球層面來看，2026年一季度物理AI初創企業融資總額超64億美元，歐美、中東產業資本持續加注。資本市場同步提速，年內已經有30余家機器人產業鏈企業啟動IPO，宇樹科技衝刺A股人形機器人第一股，擬募資42.02億元，頭部企業逐步進入業績兌現期。

從盈利確定性來看，物理AI技術壁壘極高，綜合覆蓋演算法、模擬、精密機械、感測器、電控多學科，單一軟體企業難以跨界入局，行業最終只會留存2至3家細分寡頭，龍頭企業能夠穿越經濟周期，具備十年以上長期持有價值。而且，商業模式上，物理AI採用“硬體整機一次性收入+模型訂閱+長期維運”組合，下遊客戶以工廠、車企、央國企為主，長單鎖定穩定現金流，盈利確定性遠高於依賴線上訂閱、客戶預算波動大的數字AI企業。

Globally, physical AI startups raised over $6.4 billion in Q1 2026 with continuous investment from industrial capital in Europe, the U.S. and the Middle East. Capital markets are accelerating as well; more than 30 robotics chain firms have launched IPO preparations this year. Unitree Robotics targets being the first A-share humanoid robot listed company, planning to raise 4.202 billion yuan. Top enterprises are entering a phase of solid revenue delivery.

Physical AI boasts extremely high technical moats spanning algorithms, simulation, precision machinery, sensors and electronic control, creating high cross-industry barriers for pure software players. Only two to three oligopolies will remain in each subsector, and leading firms can weather economic cycles with investment value spanning over a decade. Its hybrid revenue model consists of one-time hardware sales, model subscriptions and long-term maintenance services. End clients are mainly manufacturers, auto OEMs and central state-owned enterprises with long-term contracts securing steady cash flow, delivering far more predictable profitability than digital AI companies reliant on volatile online subscription budgets.

當前，業內將物理AI完整發展周期劃分為三階段。2026—2027年為佈局期，世界模型、模擬工具將率先兌現估值；2028—2032年高速增長期，工業機器人、高階自動駕駛批次落地，企業業績集中釋放；2033年後進入成熟期，人形機器人民用普及。預計完整產業浪潮將持續近20年。

從投資周期來看，短期1—3年內，數字AI垂直應用領域仍然存在結構性投資機會，適合波段佈局。但是放眼5—10年長期維度，物理AI全產業鏈是核心配置主線，承載長期資金，將孕育出兆市值科技龍頭。

可以說，AI的上半場，由數字AI完成資訊世界智能化改造，兌現短期流量紅利。而AI的下半場，將由物理AI打通數字智能與實體世界的壁壘，開啟實體經濟智能化革命。

The industry divides the full development cycle of physical AI into three phases. 2026–2027 marks the layout stage, where world models and simulation tools will first realize valuation growth. The high-growth phase runs from 2028 to 2032, featuring mass deployment of industrial robots and advanced autonomous driving alongside robust corporate earnings. Maturity arrives post-2033 with humanoid robots entering civilian mass use. The entire industrial boom is projected to last nearly two decades.

For investment horizons: Over the short term of 1–3 years, vertical digital AI applications still offer structural investment opportunities suitable for swing trading. From a 5–10 year long-term perspective, however, the full physical AI industrial chain stands as the core allocation track for long-term capital, poised to spawn tech giants with trillion-dollar market caps.

In short, the first half of AI’s evolution relies on digital AI to digitize the information world and capture short-term traffic gains. The second half will be driven by physical AI, bridging digital intelligence and the physical world to ignite an intelligent revolution across the real economy.    (無界社 Economic Views)

彙整鉅亨號中所有使用「圖文生成」關鍵字，所發布的內容 (依照時間排序，最新在前，舊的在後)`