From "Digital AI" to "Physical AI" — Why the Industry Is Suddenly Paying Attention and the Scale of It

From late 2025 through the first half of 2026, conversations in Silicon Valley's investor community rapidly shifted from "what to bet on after LLMs" to "where in the physical AI stack to place your bets." According to multiple tech media outlets citing data from venture research firms PitchBook and Crunchbase, investment in the humanoid robotics sector surged from $239 million in 2022 to approximately $3.7 billion in 2025, with $2.37 billion deployed across 11 rounds in just the first few months of 2026. Crunchbase reported that of the $300 billion in total global venture investment in Q1 2026, AI absorbed 80%—and within that, AI operating in physical space, so-called Physical AI, was the primary driver of major new rounds.

Market size projections vary across research firms, but the estimates Goldman Sachs published at the end of 2024—subsequently updated—have become the industry benchmark. The firm revised its global humanoid robot TAM (Total Addressable Market) to $38 billion and 1.4 million units shipped by 2035, factoring in a 40% year-over-year decline in manufacturing costs. Adam Jonas and the tech research team at Morgan Stanley take an even more aggressive view, projecting annual revenues of $7.5 trillion by 2050 and a humanoid semiconductor TAM alone reaching $305 billion by 2045. Bank of America Global Research goes further still, suggesting that "by 2060, more people will own a humanoid than a car," sketching a scenario of 3 billion cumulative units deployed. The gap between the conservative estimate from MarketsandMarkets' 2026 report—a $15 billion Physical AI market growing at 47.2% annually through 2032—and the ultra-bullish scenario of $383 billion in 2026 presented by Future Markets Inc. using a broader scope is more than 250-fold. This largely reflects a definitional divide: whether "Physical AI" is understood as hardware-centric, or as "AI as environment" encompassing entire systems across agriculture, healthcare, defense, and logistics.

So why has so much capital flooded in all at once, and why now? The answer common to the leading GPs in Silicon Valley comes down to three points. First, general-purpose VLA (Vision-Language-Action) models are advancing in capability at the same pace as the evolution from GPT-3 to GPT-4, and over the single year spanning Physical Intelligence's π0 to π0.7, "generalization across different embodiments" became a concrete reality. Second, manufacturing costs have fallen dramatically—according to Goldman Sachs, per-unit production costs dropped 40% from $50,000–$250,000 the previous year to $30,000–$150,000, bringing the $20,000–$50,000 "break-even threshold" identified by McKinsey into range. Third, Boston Dynamics' Atlas (under Hyundai Motor), Tesla's Optimus, and Figure AI's Figure 03 all confirmed in 2026 that they are moving from demos to actual production lines, bringing the critical inflection point from "proof of concept to platform" suddenly much closer.

The Humanoid Stars: Figure AI, 1X, Apptronik, Boston Dynamics

As of 2026, the pure-play humanoid company that has raised the largest amount of capital is Figure AI. The company completed an additional raise of over $1 billion in its Series C in September 2025, reaching a post-money valuation of $39 billion. The round was led by Parkway Venture Capital, with participation from Brookfield Asset Management, NVIDIA, Macquarie Capital, Intel Capital, Align Ventures, Tamarack Global, LG Technology Ventures, Salesforce, T-Mobile Ventures, and Qualcomm Ventures. Total cumulative funding has reached $1.9 billion. Founder Brett Adcock (formerly of Archer Aviation and Vettery) ended his initial collaboration with OpenAI in February 2025 and fully shifted to his in-house VLA "Helix." At BMW's Spartanburg plant, Figure 02 loaded over 90,000 parts over 11 months, contributing to the production of more than 30,000 X3 vehicles, surpassing 1,250 hours of operation and achieving 99% accuracy. Its successor, Figure 03, began production in January 2026 at the company's own manufacturing facility "BotQ," and by May had reached a pace of one unit per hour, with The AI Insider reporting cumulative shipments of over 350 units.

Norway-based 1X Technologies, known as a portfolio company of the OpenAI Startup Fund, announced that pre-orders for its home humanoid "NEO" at $20,000 opened on October 28, 2025, and the first year's production sold out within five days. Although headquartered in Norway, the company has publicly announced plans to establish a factory in California capable of producing 100,000 units annually by 2027. According to Sifted, it has raised a cumulative total of approximately $130 million from EQT Ventures, Tiger Global, and the OpenAI Startup Fund. Additionally, as of September 2025, outlets including EqualOcean reported that the company was preparing a fundraising round targeting a valuation of over $10 billion at approximately $1 billion in size (no official close had been announced at the time of writing). NEO is designed to incorporate human teleoperation, and according to Sifted and The Robot Report, it employs a "Data Flywheel" model that accumulates training data by operating in buyers' living rooms.

Texas-based Apptronik, ahead of full-scale mass production of its Apollo bipedal robot, raised $520 million in an extended Series A round on February 11, 2026, with its valuation jumping to approximately $5 billion—roughly triple its previous level. Cumulative funding has reached approximately $935 million. According to CNBC and TechCrunch, Google and Mercedes-Benz continued as existing investors, while AT&T Ventures, John Deere, and the Qatar Investment Authority joined as new investors. Apollo is being piloted on Mercedes-Benz production lines and in GXO Logistics warehouses, and CEO Jeff Cardenas told automotive trade publication Automate that he expects "orders worth $1 billion" at approximately $80,000 per unit in 2027.

Veteran company Boston Dynamics unveiled the production-ready specifications of its all-electric new Atlas at CES 2026 on January 5, 2026. Parent company Hyundai Motor has announced a $26 billion investment in the United States and plans to build a factory capable of producing 30,000 Atlas units annually. On the same day, it was announced that Atlas would be equipped with the latest Gemini Robotics foundation model through a strategic partnership with Google DeepMind. All units shipping within 2026 have been allocated to Hyundai's Robotics Metaplant Application Center (RMAC) and Google DeepMind.

The rise of Chinese players cannot be ignored. According to data from research firm Omdia reported by Bloomberg in January 2026, of approximately 13,000 humanoids shipped worldwide in 2025, AgiBot Innovation (Shanghai) topped the list with 5,168 units, followed by Unitree Robotics and UBTech Robotics as Chinese companies. Unitree achieved 2025 revenues of 1.708 billion yuan (approximately ¥36.5 billion, up 335% year-on-year), driven by its G1 at a base price of $16,000, and in March 2026 filed for an IPO on the Shanghai Stock Exchange worth approximately $610 million, with a listing expected around mid-2026 as China's first publicly listed humanoid company. According to CGTN and Xinhua, UBTech plans to produce 5,000 units in 2026 and 10,000 units in 2027.

The Race for the "Robot Brain": Physical Intelligence, Skild AI, Gemini Robotics, and NVIDIA GR00T

While hardware companies are grabbing the spotlight, VCs see the real battleground in Silicon Valley as the race for dominance in "the robot brain = robotics foundation models." Physical Intelligence (known as π) is a San Francisco research company founded by former Google DeepMind researchers. In late 2024 it raised $400 million (approximately ¥62 billion) in a Sequoia-led round, followed in 2025 by a $600 million (approximately ¥93 billion) round led by CapitalG (Alphabet's growth fund) with participation from Lux, Bond, Redpoint, and Sequoia, reaching a valuation of $5.6 billion (approximately ¥868 billion). In March 2026, Bloomberg and TechCrunch simultaneously reported that "the company is in talks with Founders Fund and Lightspeed to raise an additional $1 billion (approximately ¥155 billion) at a valuation exceeding $11 billion (approximately ¥1.71 trillion)," though no official close had been announced at the time of writing. The company's flagship model π0, built on PaliGemma as a 3-billion-parameter Transformer, was announced in February 2025 and trained on over 10,000 hours of real-world data covering 7 robot body types and 68 tasks. Its open-source release of both code and weights made a significant impact on the industry.

Pittsburgh-based Skild AI was founded in 2023 as a spinout by Carnegie Mellon University professors Deepak Pathak and Abhinav Gupta. In January 2026, it completed a $1.4 billion (approximately ¥217 billion) Series C led by SoftBank, with a valuation of $14 billion (approximately ¥2.17 trillion) — nearly triple the approximately $4.5 billion (approximately ¥697.5 billion) valuation from seven months prior. Investors include Nvidia NVentures, Jeff Bezos's Bezos Expeditions, Samsung, LG, Schneider Electric, Salesforce Ventures, Lightspeed, Felicis, Coatue, and Sequoia. According to Crunchbase News, cumulative funding surpassed $2 billion (approximately ¥310 billion) within 18 months. The company's "Skild Brain" is an omni-bodied foundation model capable of operating diverse robot forms — quadrupeds, humanoids, tabletop arms, mobile manipulators — without retraining, and can adapt to limb loss, wheel lock, and sudden load changes. In March 2026, through collaboration with Foxconn and NVIDIA, the company publicly announced — via outlets including Hoodline — the deployment of its model on assembly lines producing NVIDIA Blackwell GPU servers in Houston, Texas. TechCrunch has reported that the company's 2025 revenue reached $30 million (approximately ¥4.65 billion).

Google DeepMind announced "Gemini Robotics" in spring 2025 — a Vision-Language-Action model built on Gemini 2.0 with "action" added as an output modality — and on April 14, 2026 released "Gemini Robotics-ER 1.6" with enhanced spatial reasoning. At CES 2026, the company announced a strategic partnership with Boston Dynamics, with plans to integrate the model into the new Atlas. Deployments to Apptronik's Apollo and Agile Robots' humanoids have also been confirmed, and Google is effectively positioning itself to compete at the layer that provides "a common brain across multiple robot bodies."

NVIDIA aims to "span both hardware and the brain as a platform." CEO Jensen Huang was set to consecutively unveil a Physical AI strategy framework he calls the "Five-Layer Cake" at GTC (San Jose, USA) in March 2026 and GTC Taipei at COMPUTEX on June 1, 2026, having already integrated three layers: Cosmos (world foundation model), Isaac Sim/Lab (simulation), and GR00T (foundation model for humanoids). As of March 2026, GR00T N1.7 is commercially available as early access, with N2 slated to ship by end of 2026; the company claims it more than doubles the success rate on novel tasks in unknown environments compared to conventional VLA models. Cosmos was trained on over 200 million curated video clips and has surpassed 2 million cumulative downloads. In January 2026, TechCrunch went so far as to describe NVIDIA as "trying to become the Android of generalist robotics."

Tectonic Shifts in Adjacent Industries: Simulation, Semiconductors, Actuators, and Teleoperation

For Physical AI to descend into the real world, not only the brain and hardware but the surrounding technology stack must be in place. Most critical is the simulation/world model layer that compensates for the scarcity of real-world data. Cosmos and Isaac Lab 3.0 on NVIDIA Omniverse (equipped with the new physics engine Newton 1.0) provide the infrastructure for training complex dexterous movements through reinforcement learning. In its latest Big Ideas 2026, a16z defines "learned representations of physical dynamics, architectures of embodiment, simulation and synthetic data foundations, expansion of sensory manifolds, and closed-loop agent coordination" as the common primitives of Physical AI, arguing that players such as Generalist AI (GEN-1), World Labs (the world model initiative by Fei-Fei Li et al.), Wayve (for autonomous driving), and Cosmos (NVIDIA) serve as cross-cutting layers.

The semiconductor layer is also dramatically booming. As noted above, Morgan Stanley estimates the humanoid chip TAM at $305 billion (approximately ¥47.3 trillion) by 2045, naming NVIDIA, Qualcomm (an investor in Wayve, focused on automotive/edge AI inference), AMD, and Arm as the primary beneficiaries. Indeed, in April 2026 Wayve raised an additional $60 million (approximately ¥9.3 billion) from AMD, Arm, and Qualcomm, expanding its total Series D to $1.2 billion (approximately ¥186 billion) and bringing total committed capital to $1.5 billion (approximately ¥232.5 billion), accelerating commercial robotaxi deployment in the UK, US, and Japan.

On the mechanical components side, Hyundai Mobis (a subsidiary of Hyundai Motor Group) supplies actuators for Atlas, while ABB (industrial robotics division) agreed in October 2025 to be acquired by SoftBank Group for $5.375 billion (approximately ¥833.1 billion). SoftBank CEO Masayoshi Son positions the deal as the core of the next phase realizing "Physical AI = the fusion of robotics and ASI (Artificial Super Intelligence)," with completion expected in mid-to-late 2026 pending regulatory approval. On the battery side, sources such as Battery Tech Online analyze that "locomotion accounts for 70% of power consumption, while computational load has expanded to 20–25%." Even Figure 03's 2.3 kWh pack yields only 3–4 hours of actual operation, making peripheral equipment premised on fast charging or battery swapping indispensable to seamlessly cover an 8-hour shift.

In the data collection layer, teleoperation (remote control) has been positioned at the core of the Data Flywheel cycle of "humans operate → data becomes training material." According to industry reports from sources such as Labellerr, the total cost of teleoperation data collection as of 2026 stands at $118–$200 per hour (approximately ¥18,000–¥31,000), requiring 1–10 minutes of skilled operator time per trajectory. 1X is scaling in-home teleoperation with NEO at scale, while Mentee Robotics, Skild's Foxconn factory, Apptronik's Mercedes-Benz factory, and Figure's BMW factory all follow the same "deployment = data acquisition" pipeline — making the number of deployment sites synonymous with competitive advantage for each company.

Silicon Valley VCs' True Feelings: a16z's "Rewriting the Industrial Stack" Thesis

The narrative that Silicon Valley's major GPs are articulating publicly is converging beyond surface-level bullishness into a structural thesis. Andreessen Horowitz (a16z) released two flagship investment frameworks in succession from late 2025 into 2026—"Frontier Systems for the Physical World" and "Big Ideas 2026: Physical AI and the Industrial Stack"—defining robotics as "the most literal embodiment of the logic by which AI systems perceive, reason, and act physically." Three focus domains were proposed: (1) robot learning, (2) autonomous science centered on materials science and life sciences, and (3) new human-machine interfaces such as BCIs, silent speech, and olfactory digitization. Marc Andreessen himself, in a May 2026 Joe Rogan podcast appearance and a Latent Space conversation, stated that "no one fully understands the scaling laws for world models and robotics," emphasizing the degree of uncertainty that justifies concentrated deployment from mega-funds. a16z is directing capital toward implementation-heavy deals including Mind Robotics, Anduril, and Physical Intelligence (reportedly under consideration for participation).

Sequoia Capital led Physical Intelligence's Series B and added to its position in Skild AI's Series C, while continuing to invest in Agility Robotics—covering the full "foundation model × embodiment × deployment destination" stack. On the firm's podcast Training Data, NVIDIA's Jim Fan discussed "fast and slow thinking in robotics," sketching a scenario in which Yann LeCun-style world model thinking and Sutton-style reinforcement learning thinking are blended together. Founders Fund launched its largest-ever $6 billion Growth IV in March 2026, and has been reported to be participating in Anduril's $5 billion Series H (valued at $61 billion, co-led with a16z) and a round under discussion with Physical Intelligence—reinforcing the Peter Thiel-aligned "hard tech × defense" orientation.

Khosla Ventures assembled its $3.5 billion Fund XIII in 2024, listing humanoid robotics, nuclear fusion, and AI infrastructure as its official mandate. Founder Vinod Khosla, in April 2026 interviews with Bloomberg and Fortune, went so far as to say "the humanoid robot business will surpass the automotive industry within 20 years" and "by 2030, AI will be capable of performing 80% of all jobs," articulating a scenario in which "physical AI rewrites the economy in a deflationary way."

What these discussions share is a framing of physical AI not as "merely an extension of smart chat" but as something built on "new operating models, industrial infrastructure, and capturable data moats"—a thesis about rewriting the industrial stack across sectors like automotive, construction, logistics, and energy that were historically hard for VC money to reach. When Coatue and Sapphire Ventures stated in their 2026 outlooks that "AI is expanding across every layer of the stack," they were expressing the same view in different words.

China's National Strategy and Geopolitics: The Future of Hegemony

What is directly shaking Silicon Valley's optimism is China's national strategy. In its 15th Five-Year Plan announced in March 2026, China elevated "Embodied Intelligence" to one of the top 10 emerging industrial tracks, positioning it in the highest strategic category alongside nuclear fusion. According to analysis by MERICS, The Diplomat, and the IFR (International Federation of Robotics), this enables mobilization of the National AI Industry Investment Fund (totaling 60 billion yuan, approximately 1.32 trillion yen), along with local matching funds and state-owned venture capital. Reuters reported that from late 2024 through early 2025, the Chinese government injected over $20 billion (approximately 3.1 trillion yen) into the robotics industry through subsidies, loans, tax credits, and state funds. China's industrial robot installations in 2024 reached 295,000 units, accounting for 54% of global installations, with cumulative operational units surpassing 2 million.

The Ministry of Industry and Information Technology (MIIT) launched the Humanoid Robot and Embodied Intelligence Standardization Technical Committee in December 2025 and announced a national standards framework covering the entire industry lifecycle by March 2026. China is also aiming for a leading role in IEC international standards for elderly care robots. CNBC, TechCrunch, and Bloomberg all assessed that "China is far outpacing U.S. players in manufacturing speed and price, leveraging both hard and soft advantages through its EV supply chain," citing estimates that Unitree alone ships approximately 36 times more units annually than Tesla and Figure combined.

At the same time, geopolitical tensions have cooled cross-border investment. According to Reuters, major U.S. pension funds are reducing their exposure to Chinese AI robotics stocks, while regulations on inbound and outbound investment are tightening on both sides. SoftBank's acquisition of ABB's robotics business for $5.375 billion (approximately 831.1 billion yen) can also be read as a strategic repositioning by Japanese capital seeking to balance between China and the United States via European assets. In Japan, the quiet adoption of humanoids in the service industry has begun, with JAL (Japan Airlines) commencing actual humanoid operations at Haneda Airport from May 2026.

Tone of reporting across newspapers and forecasts from specialist analysts

The tone of media coverage has shifted from "demo videos and flashy procurement news" dominant through 2025 to a phase of "soberly examining the transition from PoC to platform" entering 2026. Bloomberg's January 8, 2026 article "Chinese Firms Dominated Global Humanoid Robot Shipments in 2025" centered on Chinese shipment volumes surpassing Tesla and Figure, while Wall Street Journal-affiliated media prompted a shift in perspective toward "value over volume." TechCrunch, through its coverage of Skild AI's funding round in January 2026, Apptronik's round in February, and Physical Intelligence's fundraising observations in March, has consistently noted that "consolidation is advancing at the foundation model layer." Reuters and CNBC frequently frame coverage around the binary of "China's national strategy vs. the free competition of U.S. tech giants," and ETF providers such as KraneShares have popularized the phrase "the race from Pilot to Platform."

Analyst forecasts also vary widely. Goldman Sachs Research notes "significant demand exists in structured environments with current technology (EV assembly, parts sorting, etc.)," projecting shipments of 50,000–100,000 units in 2026 and over 250,000 by 2030. Morgan Stanley's Adam Jonas and colleagues, in their January 2026 "The Humanoid 100" report, consolidated the entire humanoid value chain (foundation models, semiconductors, actuators, reducers, vision sensors, energy) into 100 stocks. Bain & Company takes a cautious stance, arguing that "in the near term, the majority of deployments should be assumed to require teleoperation." McKinsey's *Crossing the Chasm* report identified current costs ($150,000–$500,000) as still too high for mass-market adoption, concluding that reaching $20,000–$50,000 is a prerequisite for widespread uptake.

In a January 21, 2026 press release, Gartner predicted that "fewer than 20 companies will reach production-stage deployment for manufacturing and supply chain by 2028," anticipating a realistic shake-out in the industry. Conversely, Bank of America's global technology team published a report in March 2026 presenting an extreme long-term scenario in which "more people will own humanoid robots than cars by 2060." ABI Research characterizes 2026–2027 as an inflection point at which the major barriers of regulation, safety, and ROI will be largely resolved.

Full-scale practical implementation cases: deployment at BMW, Mercedes, Hyundai, Amazon, and airports

The Figure 02 deployment at BMW Spartanburg symbolizes the turning point from PoC to platform. BMW Group has positioned this pilot as a "showcase for the European expansion of physical AI," beginning test deployment of Hexagon's wheeled humanoid AEON at its Leipzig, Germany plant in April 2026, with plans to expand to a pilot in high-voltage battery assembly and parts manufacturing processes by summer. BMW has also established a "Center of Competence for Physical AI in Production" in Leipzig, serving as the command center for rolling out AI and robotics across factories worldwide.

Mercedes-Benz has deployed Apptronik's Apollo for "intensive logistics (parts transport and initial quality inspection)" and is gradually expanding toward fully autonomous operation. Hyundai Motor Group plans to deploy Boston Dynamics' new Atlas at its own factories (including Hyundai Robotics Metaplant America) in 2026, with full-scale operation planned at its Georgia plant from 2028. On the Chinese side, XPeng, Xiaomi, and UBTech are advancing deployment on EV lines, with Automotive Manufacturing Solutions reporting that more than 20 mainstream automotive OEMs have moved to invest in or adopt humanoid robots.

In logistics and warehousing, Agility Robotics' Digit has demonstrated the transport of over 100,000 totes at GXO Logistics, Spanx, and Mercado Libre centers. In its Series C, led by WP Global Partners with participation from SoftBank, Amazon, DCVC, and Playground Global, the company raised $400 million (approximately ¥62 billion), bringing total funding to $641 million (approximately ¥99.4 billion) and reaching a valuation of $1.75 billion (approximately ¥271.2 billion). Amazon reached a scale of operating 1 million robots globally by mid-2025, combining its in-house models—Sequoia (integrated automation platform), Sparrow (articulated arm, suction-type capable of handling approximately 65% of SKUs), Proteus (autonomous mobile), Robin, and Cardinal—and has launched full-flow automation at its next-generation center in Shreveport.

In the services sector, physical AI applications are expanding across Anduril's autonomous drone Roadrunner and the aviation and defense fields. In May 2026, the company raised $5 billion (approximately ¥775 billion) in a Series H co-led by Thrive Capital and a16z, reaching a valuation of $61 billion (approximately ¥9.46 trillion). Anduril's Arsenal-1 factory (in Ohio, 5 million square feet, with a $1 billion investment of approximately ¥155 billion) is becoming an iconic hub for physical AI manufacturing. At airports, JAL implemented a humanoid at Haneda Airport in May 2026, entering an operational verification phase.

Autonomous driving remains one of the "largest commercial deployments" of physical AI. Waymo completed a fundraise of approximately $16 billion (approximately ¥2.48 trillion) in 2025 at a valuation of $126 billion (approximately ¥19.53 trillion), covering 11 U.S. cities and over 1,400 square miles commercially, and aims to expand to 1 million weekly trips and 20 new cities (including London and Tokyo) by the end of 2026. As previously mentioned, Wayve signed an MoU with the UK government in May 2026 and has partnered with Uber to begin L4 commercial trials in London within 2026.

Remaining challenges: autonomy, safety, battery, data, cost

The challenges behind the enthusiasm remain significant. The first barrier is the "autonomy gap." As both Bain & Company and Morgan Stanley point out, most current humanoid demos should be assumed to involve teleoperation, and cannot be considered fully autonomous unless explicitly stated otherwise. According to industry analyses such as Robozaps, even benchmarks trained on 1 million trajectories and 217 tasks achieve only a 78% success rate, far short of the 95% threshold required for unattended operation. Out-of-distribution object drop rates due to grasping failures also occur at a rate of 5–15%.

The second barrier is battery and power. The 2.3 kWh pack in the Figure 03 limits practical operation to 3–4 hours, with locomotion consuming over 70% of power, while the computational load share is expanding to 20–25% as systems become more agentic. Handling an 8-hour shift requires a trinity of automated battery swapping, fast-charging infrastructure, and lightweight material design.

The third barrier is safety and regulation. ISO 10218:2025 and ANSI/A3 R15.06-2025 form the foundation, with ISO 25785-1 for dynamically stable robots currently being drafted. The EU is moving toward mandating cybersecurity requirements and third-party assessments under the AI Act (2025) and the Machinery Regulation (applicable from 2027). Japan has implemented JIS B 8433-1/2, while the United States is expected to see a "patchwork" of voluntary standards such as UL 3300 and state-by-state ordinances in the absence of federal legislation. McKinsey identifies "safety systems for fenceless operation" and "shift-equivalent operational continuity" as the two key requirements for establishing ROI.

The fourth barrier is cost structure. The current range of $150,000–$500,000 cited by McKinsey is still 3–25 times higher than the $20,000–$50,000 threshold needed for mass-market adoption. Whether Tesla's Optimus ultimately reaches below $20,000, 1X's NEO reaches $20,000, and Unitree G1 reaches $16,000 will determine the "tipping point for the home/consumer market."

The fifth barrier is geopolitics and data fragmentation. Driven by Chinese state subsidies and U.S. export controls and inbound investment screening, the physical AI supply chain—covering specialty magnets, reducers, semiconductors, and high-end sensors—is increasingly bifurcating into two poles. The number of cases subject to prolonged review by both CFIUS (the Committee on Foreign Investment in the United States) and China's SAMR (State Administration for Market Regulation) is growing, and the SoftBank/ABB deal is also said to be contingent on regulatory approval from the EU, China, and the United States.

From the Second Half of 2026 to 2028──What Happens Next

The most closely watched near-term milestone is Jensen Huang's GTC Taipei at COMPUTEX keynote on June 1, 2026. NVIDIA is expected to unveil its "Five-Layer Cake" strategic framework for physical AI, along with the official release of GR00T N2, Cosmos Reason 2, and Isaac Lab 3.0, while further expanding partnerships with more than 110 robot-brain developers, industrial automation firms, and humanoid companies. Tesla is set to begin mass production of Optimus Gen 3 at Fremont starting in the summer of 2026, targeting hundreds to thousands of units within the year; Musk's stated "annual capacity of one million units" remains a nominal goal for end-2026, with real demand expected to ramp up in 2027–28. Tesla is also reported to be planning general consumer sales of Optimus by end of 2027.

On the hardware side, 1X plans to operate a mass-production factory in California capable of 100,000 units per year by 2027; Apptronik anticipates commercial orders for Apollo worth $1 billion (approximately ¥155 billion) in 2027; and Hyundai plans humanoid production at a scale of 30,000 units at its Georgia plant in 2028. On the foundation-model side, the closing of Physical Intelligence's additional funding round, the full activation of Skild Brain on Foxconn lines, and the integration of Google Gemini Robotics-ER into Atlas are key events for the second half of 2026. Gartner projects that fewer than 20 companies will advance to mass-production stage in manufacturing and supply chain by 2028, and the field of surviving players is already narrowing.

Also drawing attention is a scenario Morgan Stanley outlined in its December 2025 outlook: that by sometime in 2026, one of the major tech giants — Meta, Google, Apple, Amazon, or OpenAI — will formally announce a robotics initiative. Amazon absorbed key talent from Covariant in 2024, then faced turbulence when an antitrust filing was submitted to the FTC in 2025, yet continues to invest seriously in robotics foundation models as an extension of its fulfillment automation efforts. Google's current central strategy is a layer approach — supplying models to hardware companies via DeepMind — while Microsoft is positioned mainly through indirect exposure via NVIDIA, Wayve, and Anthropic. Multiple outlets have reported observations that OpenAI is shifting from its external dependencies through 1X, Physical Intelligence, and (formerly) Figure back toward an in-house robotics initiative, though as of this writing there has been no official announcement.

Synthesizing from a Silicon Valley VC perspective, the period from the second half of 2026 through 2028 will see three simultaneous structural shifts: the dominant players in the foundation-model layer will be determined, the hardware layer will transition to mass production, and regulatory and safety standards will be internationally harmonized. As ETF provider KraneShares puts it, physical AI has moved to "a race from Pilot to Platform." In the near term, an unstable phase will persist in which hype and the shakeout of certain companies coexist — but the very fact that the "body, brain, nerves, and bloodstream" of robotics are converging simultaneously marks the industrial inflection point from digital AI to physical AI.