Set as preferred source
Insights from Physical AI and the Future of Robotics.
For most of the past two decades, the technology economy rewarded abstraction. Software separated growth from physical assets. Cloud infrastructure centralized compute. Digital products scaled without factories, warehouses or complex supply chains.
Physical AI reverses part of that logic.
The intelligence inside machines is improving quickly, but robots still have to operate in buildings, move physical objects, process information with very low latency and remain reliable around people. Unlike cloud AI, physical AI also depends on sensor quality, safety systems and training data that can represent a messy physical world. That pulls AI back into the constraints of the industrial economy: semiconductors, power, energy security, materials, integration and capital.
Silicon Valley Bank’s Physical AI and the Future of Robotics captures the scale of that shift. US venture investment in hardware is on pace to reach $120 billion in 2026. Excluding Anthropic and OpenAI, hardware is projected to account for nearly a third of US venture investment in 2026. The share of global venture investors putting at least 10% of their US deals into hardware has risen from 9% in 2022 to 22% in 2026.
That capital rotation matters. But the more important story is what is constraining it.
AI is leaving the cloud, and the economics change when it does
Cloud AI has benefited from a powerful model: concentrate expensive compute in data centers, spread utilization across many users and let falling inference costs improve economics over time.
A warehouse robot does not have the same flexibility.
A chatbot can tolerate a small delay. A machine moving near a worker or manipulating inventory cannot. As physical systems become more capable, more computation has to happen at or near the edge. That increases demand for onboard processing and high-bandwidth memory precisely when those inputs are already under pressure from data-center expansion.
The report estimates that Nvidia, Google, AMD and Amazon accounted for roughly 90% of global high-bandwidth memory consumption in 2025. US venture investment in edge-compute companies is on track to reach $3.4 billion across about 100 deals in 2026—both record levels—while investment in data-center companies is on pace to exceed $12 billion.
This creates a competition inside the AI economy itself. Cloud inference, agentic workflows and physical AI may serve different markets, but they draw from overlapping semiconductor capacity.
The constraint is no longer simply whether enough AI models exist. It is whether enough memory, packaging capacity, power and efficient edge hardware exist to deploy them economically.
Semiconductor resilience is also energy resilience
The geography becomes even more consequential once the semiconductor supply chain is examined beyond fabrication.
Taiwan attracts most of the attention because of its position in advanced chip manufacturing. But SVB mapped the primary manufacturing locations of more than 100 suppliers for some AI accelerator chips and found 57% in Japan. Those suppliers provide materials and components that sit well upstream from the finished processor.
The report then connects those semiconductor dependencies to energy.
Japan sources more than 90% of its crude oil through the Strait of Hormuz. During the 2026 disruption to the route, it drew on strategic reserves. Taiwan has a different vulnerability: LNG is critical to its power system, including the electricity supporting semiconductor fabrication. During the disruption, utilities bid aggressively for spot LNG, paying about twice previous rates per cargo to maintain supply.
That changes how supply-chain resilience should be interpreted.
Moving production away from one country does not necessarily remove concentration if the new production base depends on another chokepoint for energy. A semiconductor supply chain can look geographically diversified while remaining exposed through oil, gas, advanced materials or specialized packaging.
The broader shift away from China illustrates the same point. China’s share of US imports declined by nine percentage points between 2019 and 2025, while Taiwan, Vietnam, Ireland and Mexico gained share. US semiconductor manufacturing capacity is expanding as well, but the build-out takes years. TSMC increased its planned Arizona investment to $265 billion in July 2026 and announced its intent to build several additional logic fabs. Yet as of September 2026, only its first Arizona fab was in high-volume production. Fab 2 was targeting the second half of 2027, while Fab 3 and the initial construction of Fab 4 were underway.
The emerging model is less reshoring than redistribution of dependency.
The automation market is much less mature than the capital cycle suggests
The investment numbers can create the impression that physical automation is approaching broad adoption. The warehouse data suggests almost the opposite.
Across roughly 400 Prologis customers surveyed, 83% had not automated any of their operations. Even among larger companies, where automation penetration is materially higher, the median organization reported automation in only about 40% of its warehouse network. One-third had automated less than a quarter of their facilities.
AI is entering warehouses through a different route. Among respondents at facilities larger than 250,000 square feet, 46% reported using AI, compared with 28% at facilities below 100,000 square feet. Early use cases center on reporting and analytics, administrative workflow automation, safety monitoring, labor planning and exception management. The near-term opportunity may therefore be operational intelligence first and embodied automation later.
This is one of the report’s most important commercial signals.
The next warehouse-robotics market is therefore not primarily a replacement cycle in which new machines displace an installed base of older robots. Much of the addressable warehouse market remains untouched.
Historically, that has not been because operators failed to understand automation. Conventional systems work best in modern, high-throughput buildings with standardized workflows. Older facilities, variable SKUs, seasonal demand and frequent exceptions make fixed automation harder to justify.
Physical AI matters because it could change that equation.
A machine capable of handling variation rather than only repetition potentially expands automation into facilities that were previously uneconomic to automate. The value is therefore less about making an already automated warehouse marginally faster and more about increasing the share of the warehouse base that automation can reach.
Warehouses are exposing the gap between technical progress and commercial scale
The report’s warehouse data also puts some distance between the robotics narrative and operational reality.
Among the select large companies surveyed, established technologies are already reaching scale. Around nine in ten report fully scaled conveyor and sortation systems. Roughly three-quarters have scaled automated storage or goods-to-person systems, while about two-thirds have scaled autonomous mobile robots.
The newer layers remain much earlier. Around three-quarters are still evaluating or piloting computer vision and scanning technologies, while close to two-thirds are piloting robotic picking and palletizing.
The distinction is not simply technical maturity. It is economic reliability.
Once installed, automation often exceeds expectations on reliability and uptime, while labor savings broadly meet expectations. Throughput is slightly weaker than expected. The largest disappointments are ease of implementation and flexibility during peak periods, while cost and payback remain the biggest stated barrier to scaling.
Humanoids make that tension especially visible. Venture investment has accelerated, but three in four surveyed technology heads identified humanoids as the most overhyped technology relative to current operational value.
That does not invalidate flexible robotics. It suggests the market has greater conviction in the need for flexibility than in the specific form factor that will deliver it.
The winning machine may walk on two legs, roll on wheels or look nothing like a person. What matters economically is whether it can work inside existing human-designed environments without requiring expensive rebuilding around the robot.
Capital is distinguishing strategic hardware from ordinary hardware
The funding market is already making its own distinction.
US venture investment in supply-chain AI and hardware reached roughly $25 billion on a trailing-12-month basis, while deals above $500 million accounted for 84% of 2026 capital. Much of that funding is flowing into autonomy and physical-AI platforms such as Waymo, Figure, Skild AI, Physical Intelligence, Zipline and Applied Intuition.
The distribution matters because physical technology needs time. Hardware development, certification, manufacturing and deployment create longer capital cycles than software alone.
At the same time, geopolitical relevance is beginning to affect valuation. Defense-oriented US hardware startups are commanding a 31% pre-money valuation premium over the broader VC-backed hardware market in 2026.
That premium suggests that investors expect government demand and security priorities to matter. It does not prove that these companies are insulated from normal commercial cycles. It does show that strategic relevance is entering the valuation environment.
Yet liquidity remains uneven. In the report’s exit data, nearly half of supply-chain companies exited by Series A, and 70% of exits were M&A transactions. Among transactions with disclosed terms, the median valuation was about $56 million. The market is separating companies that solve narrow operational problems from platforms that can generalize across environments.
Generalization is becoming both a technical capability and a capital-market distinction.
The infrastructure problem goes well beyond chips
Physical AI is also developing inside a broader industrial system already under stress.
The dependencies extend well beyond semiconductors. Aluminum is used across vehicles, batteries and industrial equipment. Resins and plastics depend on petrochemical feedstocks derived from natural gas processing and crude oil refining. Nitrogen fertilizer faces a separate but related exposure: ammonia is the starting point for mineral nitrogen fertilizers, and just over 70% of global ammonia production is based on natural gas.
Shipping connects these markets. Disruption through the Strait of Hormuz pushed up tanker freight rates, marine fuel prices and war-risk insurance costs while threatening a route that carries roughly one-third of global seaborne fertilizer trade.
The cost pressure has spread across transportation. In August, US producer prices were 77.8% higher than a year earlier for diesel fuel and 66.8% higher for jet fuel. Prices for truck, water and air freight transportation were also up 14.3%, 16.7% and 7.6%, respectively.
US consumer inflation, which reached 4.2% in May 2026, eased to 3.4% in August. But the US CPI energy index remained 16.3% higher than a year earlier.
These are not separate stories.
Semiconductors, EVs, logistics, agriculture and automation share exposure to many of the same energy, transportation, petrochemical and industrial-material supply chains. An energy or shipping shock can raise fuel and freight costs, constrain petrochemical feedstocks, disrupt fertilizer production and trade, and change manufacturing economics across otherwise unrelated sectors.
That interconnectedness makes physical AI more exposed to physical bottlenecks than the software cycle that preceded it. Software could often scale around constraints in the physical economy. Physical AI has to scale through them.
What changes strategically
The emerging robotics cycle is likely to reward a different combination of capabilities than the previous AI cycle.
Model intelligence remains important, but compute efficiency, access to memory, energy resilience and deployment economics increasingly determine where that intelligence can create value.
The 83% non-automation rate in the Prologis survey also reframes the warehouse opportunity. The largest market may not sit inside the most advanced facilities, but across the much larger base of warehouses where traditional automation has never delivered adequate flexibility or payback.
Semiconductor resilience is becoming inseparable from energy resilience. Japan’s oil exposure and Taiwan’s LNG costs show why chip diversification cannot be measured only by the number of fabrication locations.
Capital markets are drawing another boundary. The 31% defense-hardware premium and concentration of supply-chain AI and hardware funding in very large rounds suggest that geopolitical relevance, generalizability and the ability to finance long development and deployment cycles are increasingly shaping which companies can scale.
Physical AI therefore represents more than another application layer for artificial intelligence. It is bringing technology back into the industrial economy.
Software determines what machines can understand. But chips, power, materials, infrastructure and integration will determine where those machines can scale — and where the economics ultimately hold.
Key risk signals
- Semiconductor energy exposure: Taiwan’s LNG dependence and Japan’s reliance on Hormuz-linked oil create vulnerabilities outside the semiconductor supply chain itself.
- Compute competition: Robotics, data centers and agentic AI increasingly compete for HBM, advanced packaging and power.
- Deployment economics: Much of the warehouse market remains unautomated because implementation complexity, facility constraints and exception handling still weaken payback.
- Capital concentration: With 84% of supply-chain AI and hardware venture funding flowing into deals above $500 million, scale is becoming increasingly dependent on access to large pools of capital.
- Cross-sector input shocks: Energy and shipping disruptions can propagate through petrochemicals, fertilizer, freight, EV and technology manufacturing simultaneously.
