▸ Siemens and Humanoid Inc. published the first production-grade humanoid metrics from a working factory floor — 60 tote moves per hour, 8+ hour uptime, >90% pick-and-place success at Siemens' Erlangen electronics plant — with the gap between 90% and the 99% manufacturing-grade threshold becoming the actual story this week.

▸ Bezos's Project Prometheus closed $10B at $38B and is raising a separate ~$100B fund to buy chipmakers, defence contractors, and aerospace manufacturers — bypassing chatbots and assembly-line robots to target upstream materials, prototyping, and engineering decisions where AI margins are largest and incumbent expertise is thinnest.

▸ SoftBank converted a Sharp LCD plant in Sakai into a battery factory for AI data centres, choosing energy over robotics manufacturing the same week it closes the $5.4B ABB Robotics acquisition — confirming that the power wall now constrains AI infrastructure as hard as the chip wall did three years ago.

A humanoid robot logged an eight-hour shift on a Siemens electronics factory floor in Erlangen, Germany, this month and Siemens published the metrics. It moved 60 totes per hour. It hit pick-and-place success above 90%. That is the news.

What happened. The HMND 01 Alpha — a wheeled humanoid built by Berlin-based Humanoid Inc. — ran autonomous tote-handling at Siemens' own electronics manufacturing site in Erlangen. The deployment is not a demo. It hit defined target metrics: 60 tote moves per hour, uptime exceeding eight hours, autonomous pick-and-place success above 90%. The robot runs on NVIDIA Jetson Thor at the edge, was trained in Isaac Sim and Isaac Lab, and integrates with Siemens' production systems through the Siemens Xcelerator portfolio — digital twin, PLC interfaces, fleet management, the industrial communication backbone that turns a robot from a feature into a coworker.

Why operators care. This is the first humanoid deployment from this stack with published factory-grade numbers. Until this month, "humanoid in a factory" almost always meant "humanoid in a video shot at a factory." Erlangen is different. It pairs a real-world throughput rate with a real-world failure rate, integrated into a real-world production system — a triangle most humanoid pilots have refused to publish in full.

Why investors care. The same triangle. Throughput, uptime, success rate, and integration depth are the four numbers humanoid valuations should be priced against. Erlangen offers all four. Apptronik's $5.5B, Figure AI's reported $39B, and Unitree's targeted $7B IPO valuation are now standing next to a published benchmark from a working competitor.

What could go wrong. Path Robotics' Andy Lonsberry told Manufacturing Dive this year that demos working 70% of the time will not survive in manufacturing — production-grade is 99-plus%. Erlangen sits at >90%. That is a real result. It is also nine percentage points below the bar Lonsberry just set. On a 24/7 line, the difference between 90% and 99% is the difference between a useful pair of hands and a system that breaks the throughput it is meant to add to. The Erlangen number is the floor of the next stage of work, not the ceiling of the last one.

The contrarian thread runs further. XPeng's head of AI and robotics, Mi Liangchuan, told a media workshop in Beijing this week that hardware reliability — specifically, dexterous hand reliability — is now the bottleneck holding IRON back from its 2026 mass-production target. The model layer is racing ahead of the actuator layer. Tokens decide. Torque is what is currently breaking.

Project Prometheus — $10B closed at $38B (23–24 April) + separate ~$100B "manufacturing transformation vehicle" in fundraise. Bezos, Vik Bajaj (ex-Google X), David Limp (CEO, Blue Origin) on the board. — Capital flowing from the lab to the factory floor.

Prometheus's existing $6.2B (raised at the November 2025 launch) funds a model layer purpose-built for manufacturing and engineering: pre-production, materials innovation, prototyping — the upstream decisions that determine whether a part costs $50 or $500 to make. JPMorgan, BlackRock, DST Global, and Arch Venture Partners participated in the $10B round. The separate $100B fund (WSJ broke the story in March; JPMorgan in preliminary talks via its Security and Resiliency Initiative; Bezos pitching Singapore and Middle Eastern sovereign wealth) is a different vehicle: an acquisition machine for chipmakers, defence contractors, and aerospace manufacturers. Not assembly-line robots. Upstream pre-production, materials, and engineering decisions inside fabs and factories where human expertise is thinnest. Context that frames the bet: hyperscalers will spend ~$700B on AI infrastructure in 2026; Amazon alone is committing $200B. Manufacturing contributes ~$16 trillion to the global economy, almost untouched by it. Bezos is filling the gap.

Neura Robotics × AWS — strategic agreement, 21 April. Cognitive robotics deployment in Amazon fulfilment centres + Amazon SageMaker training integration. — The hyperscaler infra layer for robotics is now a three-way contest.

Amazon will deploy Neura's cognitive robots in select fulfilment centres. Neura Gym training environments integrate with SageMaker. AWS positions directly against Microsoft's Azure Physical AI Toolchain (also live this week from Hannover) and NVIDIA's Cosmos-anchored stack — three different bets on where the platform layer for embodied AI gets won.

Schaeffler × VinDynamics — strategic MOU, 21–22 April, Hanoi. Tier-1 German motion-tech supplier (110,000 employees, 250 locations, 55 countries) supplying high-precision actuators and planetary gearboxes to Vingroup's humanoid arm VinDynamics. — Industrial Tier-1s entering the humanoid component stack.

Schaeffler's first humanoid partnership in APAC, alongside existing Europe, China, and US relationships. The actuators and planetary gearboxes are the muscles and joints — they convert motor torque into the precise, repeatable movement a humanoid needs to grip a tote without crushing it or drop a screw without missing the slot. That is the exact layer XPeng publicly admitted, three days later, is its 2026 bottleneck. Joint data collection on actuator performance feeds Schaeffler's predictive maintenance services. The pattern matters: the same Tier-1 industrial suppliers that built much of the EV powertrain stack (Bosch, Schaeffler, Continental) are now positioning to do the same for humanoid actuators. Reliability and volume are not where startups win.

Hannover Messe goes sovereign. Deutsche Telekom's Industrial AI Cloud — built in Germany on NVIDIA infrastructure — launched as one of Europe's largest dedicated AI factories. Agile Robots, SAP, Siemens, PhysicsX, and Wandelbots are running workloads on it. EDAG is running its industrial metaverse platform on it. Dell, IBM, Lenovo, and PNY are showcasing NVIDIA-accelerated systems for edge-to-data-centre deployment. The framing is unambiguous: a third doctrine — alongside US private capital and Chinese state capital — is now operational in the physical AI infrastructure race.

KION, Accenture, and Siemens are building digital-twin warehouses for GXO. GXO is the world's largest pure-play contract logistics operator. The work uses NVIDIA's Mega Omniverse Blueprint to train and validate fleets of Jetson-based autonomous forklifts before they touch the warehouse floor. The under-discussed implication: forklift autonomy is the near-term humanoid alternative for warehouse operators — same labour problem, lower cost per unit, no bipedal complexity, working today.

NVIDIA's stack is now the platform layer behind every other story this week. Cosmos 3 generates synthetic training data and cross-environment generalisation. Isaac GR00T N1.7 powers vision-language-action reasoning for multi-step tasks. Newton 1.0 — the open-source physics engine — went GA, with rigid and flexible body simulation for dexterous manipulation. Toyota Research is customising Cosmos for its own world model. FANUC and Fauna Robotics are pushing CAD-to-OpenUSD as a robot-design pipeline. FieldAI, Hexagon, Linker Vision, and Skild AI are running on the Physical AI Data Factory Blueprint.

Policy & Regulatory Radar — EU AI Act Digital Omnibus trilogue, 28 April. Brussels meets to close the Digital Omnibus package before the 2 August 2026 General-Purpose AI obligations bite. The Commission's December 2025 proposal slips Annex III high-risk compliance to 2 December 2027 and Annex I to 2 August 2028. Parliament's IMCO committee votes its position 13 May. For physical AI vendors: industrial robotics under the Machinery Regulation timing remains unsettled; AV and drone autonomous-decision systems are likely Annex III high-risk regardless of the slip. Compliance documentation by Q3 2027 is the planning target, not the 2028 backstop. (Single-source reporting; verify Commission text before relying on specific dates.)

XPeng publicly admits hardware is the bottleneck. Speaking at a 24 April technical workshop in Beijing, Mi Liangchuan — XPeng's head of AI and robotics — acknowledged that AI progress on the IRON humanoid is now ahead of hardware reliability. The first-generation dexterous hand has reliability issues; XPeng is running multiple R&D tracks rather than locking a single design. Production-grade IRON will use a triple-Turing chip architecture running VLA 2.0, the same stack as XPeng's autonomous vehicles. A dedicated data factory is going up in Guangzhou. The 84-joint architecture and five-degree-of-freedom waist are locked. The actuators are not.

UBTECH ships 1,079 humanoids and books RMB 1.3B in 2026 orders. UBTECH was the global leader in full-size humanoid shipments in 2025, deploying Walker S2 units on factory lines at BYD and FAW Group at an average price of RMB 760,000. The Walker S2's three-minute autonomous battery swap is the operationally significant detail — operational continuity inside an industrial shift. UBTECH's 2026 plan: 5,000 humanoid deliveries, 10,000 by 2027.

AgiBot crosses 10,000 cumulative humanoid shipments — half in the last quarter. AgiBot rolled out its 10,000th cumulative humanoid in late March, with roughly half delivered in the prior three months. TrendForce projects Chinese humanoid production grows 94% in 2026, with Unitree and AgiBot together capturing close to 80% of domestic shipments. The talent layer reflects the capital layer: UBTECH is offering RMB 18M (~$2.5M) for a chief scientist. Beijing's Big Fund III made its first embodied intelligence investment last year. State capital is the floor under the Chinese sector valuation.

Japan — SoftBank chooses batteries over robotics for the Sakai site. SoftBank is converting part of the former Sharp LCD plant in Sakai, Osaka — acquired in 2025 for ¥100B (~$676M) and currently being built out as an AI data centre — into one of Japan's largest dedicated battery factories. Production within five years; gigawatt-hour annual output projected. Per Bloomberg, executives weighed using the site for robotics manufacturing and chose energy. The same week SoftBank closes the $5.4B ABB Robotics acquisition, it is signalling that the next constraint on AI infrastructure is power, not robots.

Singapore — MPA commits SGD$100M (~$78.7M) to autonomous maritime infrastructure. The Maritime and Port Authority and the Singapore Maritime Institute launched the 2026 Maritime Technology and Research Roadmap during Singapore Maritime Week. Four priority areas: autonomous port operations, smart ships, intelligent port services, alternative energy. MPA and PSA Singapore issued an Expression of Interest for autonomous container feeder vessels between Tuas and Pasir Panjang. Tuas Port is targeted as the world's largest fully automated container terminal. Around 80% of global trade moves by sea — Singapore's autonomy gains ripple far beyond the strait.

South Korea — MSIT acknowledges the physical AI gap publicly. At the World IT Show 2026 (22–24 April, COEX, 460 companies, 17 countries), Second Vice Minister Ryu Je-myung convened a roundtable in which Korean industry leaders — Maum AI, Movensys, MakinaRocks, Pebblous, plus academic representatives — explicitly named gaps in technology, data, and industrial deployment versus global leaders. Korea has the highest robot density on earth at 1,012 robots per 10,000 workers. The hardware density does not translate automatically into Physical AI leadership, and the Ministry just said so on the record.

NVIDIA's National Robotics Week briefing surfaced Mimic Robotics' video-action model, which pairs a pretrained internet-scale video model with a flow-matching action decoder — replacing the static image-language backbone of conventional vision-language-action systems with video-learned physical dynamics. The reported result: 10x sample efficiency and 2x faster convergence on real-world manipulation tasks. The relevance is direct — data scarcity is the most-cited bottleneck in scaling humanoid deployments, and XPeng's data factory in Guangzhou points at the same problem from the opposite end. (NVIDIA-attributed; underlying paper not independently verified at time of publication.)

The week's signal arrives in three layers, and they fit together cleanly.

The Brains layer just got pointed at factories. Prometheus's $10B + $100B is explicitly aimed at chipmakers, defence contractors, and aerospace manufacturers — not chatbots, not assembly-line robots, not the agentic platforms Silicon Valley is currently obsessed with. Schaeffler's first APAC humanoid deal sits inside the same thesis from the supplier side. The hyperscaler contest above (AWS, Azure, NVIDIA) decides which platform owns the inference layer; Bezos is already two layers down, buying the manufacturing base.

The Brawn layer just got its first published factory-grade benchmark: Erlangen's 60 totes per hour at >90% success — and immediately got told by XPeng that hardware reliability is the actual constraint on the next leap. Schaeffler signing in Hanoi the day before XPeng's admission is not coincidence. It is the supply chain repositioning around a problem the AI press still mostly ignores.

The Batteries layer just stopped being a metaphor. SoftBank turning a Sharp LCD fab into a battery plant — and explicitly choosing energy over more robotics capacity in the same building — confirms that AI infrastructure now competes for grid capacity the way data centres compete for chips.

Physical AI Weekly is a free weekly brief for VCs, founders, and hardware decision-makers tracking what is actually deploying in the Physical AI ecosystem.

physicalaiweekly.com • X / @PhyAIweekly47 • LinkedIn