2026-06-18 👁️ — views

Physical AI Energy — AV Fleet Charging, Dojo Power, and Humanoid Battery Life

AV fleet charging, Dojo training power, and humanoid battery life mapped as benchmark dimensions — energy cost is underweighted in Physical AI economics.

Article 114 in the Physical AI Benchmark Series — Physical AI Energy Infrastructure: AV Fleet Charging Demand, Dojo Training Power Consumption, Humanoid Robot Battery Life, and Why Energy Cost Is an Underweighted Variable in AV Economics

Physical AI is an energy-intensive industry. Unlike software AI, which runs on centralized data center hardware where energy costs are pooled across millions of inference calls, Physical AI systems operate in the real world — on public roads, in warehouses, and in factories — where energy consumption is per-unit, per-hour, and directly tied to the operational economics of each deployed machine. An AV fleet requires high-capacity depot charging infrastructure. Training compute clusters like Tesla Dojo and Google TPU farms consume significant electricity to produce the models that run on those vehicles. Humanoid robots carry onboard batteries that determine how many hours of useful work they can deliver before requiring recharge or battery swap. And in each of these cases, the cost of energy per unit of output is a variable that affects the economics of Physical AI deployment in ways that the industry has not yet fully internalized into its public benchmarking.

This article maps energy as a structured benchmark dimension for the Physical AI ramp. Section 1 quantifies AV fleet charging demand at scale. Section 2 analyzes depot charging infrastructure requirements and cost. Section 3 examines training compute energy for Dojo and comparable GPU clusters. Section 4 catalogs humanoid robot battery life constraints across current commercial and research platforms. Section 5 builds energy cost into the AV economics model across geographies with meaningfully different electricity price regimes. All figures are estimates from publicly available data or derived from disclosed specifications; figures labeled “(est.)” are directional calculations from first principles.

Section 1 — AV Fleet Charging: Energy Demand at Scale

The energy demand of an AV fleet scales linearly with vehicle count and driving pattern, but the absolute numbers are large enough to matter when planning depot infrastructure and assessing operational costs. The table below builds a fleet energy demand model from the disclosed vehicle counts and battery specifications of current and hypothetical AV deployments.

Several observations emerge from this demand model. First, Waymo’s current fleet at 1,100 vehicles in Phoenix consumes approximately 32 GWh per year — a non-trivial amount comparable to the annual energy consumption of roughly 3,000 average US households, but well within the capacity of a commercial utility to service at a dedicated depot. Second, Tesla’s hypothetical 10K-vehicle robotaxi fleet would require approximately 255 GWh per year — an order of magnitude larger. Third, the 100K-vehicle Cybercab scenario at 1.46 TWh per year is a material but manageable fraction of US electricity supply. What matters for fleet operators is not the national fraction, but the local grid capacity at specific depot locations — and whether the utility infrastructure at those sites can support the charging load.

The energy per vehicle per day figure is determined primarily by daily miles driven and battery capacity. A Waymo Jaguar I-PACE operating a full commercial day in Phoenix — where the warm climate enables extended hours — likely completes one near-full battery cycle per day. That 80 kWh figure is consistent with a vehicle driven 250-300 miles at approximately 0.27-0.32 kWh per mile (est.). Tesla’s Model Y Long Range at 70 kWh per day implies similar driving intensity. The Cybercab, targeting a smaller battery and lighter vehicle, might achieve lower energy per day at equivalent miles if vehicle efficiency improves as expected at lower weight.

Section 2 — Depot Charging Infrastructure Requirements

AV fleet operators must build or lease depot charging infrastructure scaled to their vehicle count and operational schedule. The choice between charging types determines both the capital cost of the depot and the operational flexibility of the fleet — specifically, how quickly vehicles can be returned to service after a charge cycle.

The optimal charging strategy for an AV fleet depends on the operational schedule. For fleets that can return vehicles to depot during a predictable overnight window — typically 2:00–6:00 AM when ride demand is minimal — Level 2 AC charging at 7-19 kW is the cost-effective base solution. A 90 kWh battery at 11 kW (L2) charges from near-empty to full in approximately 8 hours, fitting cleanly in an overnight depot window. Capital cost per stall of $3K-8K installed makes this the economically dominant choice for the base charging load.

DC fast charging at 50-150 kW becomes valuable for fleets operating near-continuously, where mid-shift top-up is necessary to maintain vehicle availability. A 30-60 minute charge window can add 50-80% battery from a 150 kW DCFC, sufficient to extend an afternoon shift. The capital cost of $15K-40K per stall is 5-10x higher than Level 2, so fleets should size DCFC capacity to the fraction of vehicles requiring mid-shift charging rather than the full fleet.

Ultra-fast 350 kW charging at $50K-100K per stall is primarily relevant for near-continuous operation scenarios — robotaxi fleets with very high demand density that cannot tolerate 30-60 minute charging windows. At current AV fleet scales, the operational need for sub-15-minute charging is limited; this infrastructure class will matter more as fleet sizes grow to tens of thousands of vehicles with 24-hour demand profiles.

Tesla’s Supercharger network is a structural advantage for the robotaxi scenario: 250 kW V3 Superchargers are already deployed at thousands of locations, providing a distributed mid-shift charging option that a commercial fleet operator would have to build from scratch. This embedded charging infrastructure represents meaningful value for the Tesla robotaxi model that does not appear on any current balance sheet as an AV-specific capital investment.

Section 3 — Training Compute Energy: Dojo and TPU Clusters

The energy consumed in training the AI models that run Physical AI systems is a distinct cost category from operational fleet charging. Training energy is a capital expenditure analog — it is spent upfront to produce a model capability that then deploys across the entire fleet. Understanding training energy cost contextualizes the investment required to build better autonomous driving models and identifies energy efficiency as a competitive dimension in training compute.

The most important observation from the training energy table is that even at the high end of estimates ($4.4M/year for a single ExaPOD), training energy cost is a relatively modest fraction of the total investment required to develop a competitive autonomous driving system. Tesla’s Dojo investment is measured in billions of dollars of hardware capital; the annual electricity bill for running that hardware is an order of magnitude smaller than the hardware depreciation cost. Training energy cost matters at scale — when multiple ExaPODs run continuously for years — but it is not the binding constraint in training compute economics.

What does matter is FLOP/Watt efficiency — how many useful training floating-point operations a system delivers per watt of electricity consumed. Tesla designed the Dojo D1 chip with this metric in mind, aiming to outperform commodity GPU clusters on the specific computational patterns required for video-based autonomous driving training. If Dojo achieves meaningfully better FLOP/Watt than H100 clusters on these workloads, the advantage compounds with scale: at 10 ExaPODs running for 3 years, a 2x FLOP/Watt advantage translates to either 2x more training at the same cost, or the same training at half the electricity and cooling cost.

Google’s use of TPU v4 pods for Waymo training benefits from Google’s renewable energy commitments, which reduce the effective carbon cost of training even if the electricity rate is similar. Carbon cost is not yet a significant financial variable in AV training economics, but it may become relevant as corporate sustainability commitments translate into actual energy procurement costs.

The practical implication for the Physical AI benchmark is that training compute energy is worth tracking as a dimension of competitive efficiency, but it is not — at current scales and price levels — a primary economic variable in the physical AI race. The returns to better training (safer, more capable autonomous systems) dwarf the cost of the training energy consumed.

Section 4 — Humanoid Robot Battery Life: The Operational Duration Constraint

For humanoid robots, onboard battery life is a direct operational constraint that determines the commercial viability of any deployment scenario. A robot that requires recharging every 90 minutes cannot work a continuous factory shift without either multiple battery swaps or extended downtime. The current state of humanoid robot battery life represents one of the most significant gaps between research-grade capabilities and commercial deployment requirements.

The battery life constraint is more severe than the table suggests, because the runtime figures typically represent laboratory or lightly-loaded conditions. A humanoid robot performing high-torque manipulation tasks — lifting boxes, operating tools, climbing stairs — draws significantly more current than one performing light inspection work or standing relatively still. Real-world operational runtime under commercial workloads may be 30-50% lower than the headline battery life figures (est.).

The gap between current humanoid battery life (1-4 hours for research platforms; potentially 4-8 hours for commercial-target platforms) and the 8+ hour shift-length operational requirement has two potential solutions: better battery density, and hot-swap battery systems. Better battery density — more kWh per kilogram of battery weight — is a technology development challenge that shares characteristics with EV battery progress. Hot-swap battery systems require mechanical design that allows a depleted battery pack to be quickly replaced with a charged one without disassembling the robot — a solvable engineering problem, but one that adds mechanical complexity and cost to the robot design.

The energy density of the battery pack is also constrained by the robot’s weight budget. A humanoid robot that adds 5 kg of battery to achieve 8-hour runtime changes its weight distribution and affects the joint torque requirements for locomotion — potentially requiring actuator upgrades that partially offset the benefit. Battery life improvement in humanoid robots is therefore a system-level optimization problem, not simply a battery chemistry problem.

Tesla’s Optimus targets factory environments where power outlets are available, potentially enabling tethered operation for stationary tasks. This sidesteps the battery constraint for the specific use case Tesla is initially targeting (repetitive assembly tasks with limited mobility) but does not solve the mobility case where untethered operation is required.

Section 5 — Energy Cost as an AV Economics Variable

The geography of electricity pricing creates meaningful variation in AV fleet operating costs that has not been systematically incorporated into public AV economics models. Commercial electricity rates in the United States vary by roughly a factor of four between the lowest-cost markets (parts of Texas and Arizona) and the highest-cost markets (California). For an AV fleet operating at scale, this geographic variation creates real cost differences that affect where AV deployment is economically attractive.

The energy cost differential of $300-450 per vehicle per year between Phoenix/Texas and San Francisco is meaningful but not dominant in the AV cost structure. From Article 113 of this series, the total cost per ride is dominated by vehicle amortization, depot operations, and remote assistance staffing — energy represents approximately $0.04-0.06 per mile, or $0.20-0.30 per 5-mile ride. At 50,000 miles per vehicle per year, energy cost totals $2,000-3,000 in Phoenix versus $5,000-6,500 in San Francisco (est.) — a real difference but not the primary economic driver.

What makes energy cost strategically significant is its relationship to other geographic variables. Phoenix has low electricity costs, favorable weather for continuous AV operation, lighter urban driving complexity, and permissive regulatory environment. San Francisco has high electricity costs, fog and rain that challenge sensor systems, complex urban driving conditions, and historically stringent AV regulation. The energy cost differential is one component of a broader geographic economic advantage that compounds across multiple cost categories simultaneously.

For depot infrastructure, the electricity rate at the depot location also affects the economics of DCFC versus Level 2 charging. At $0.06/kWh (Phoenix), a 90 kWh battery costs $5.40 to charge from empty — a trivial variable cost. At $0.23/kWh (PG&E commercial), the same charge costs $20.70. For a fleet of 1,000 vehicles completing one full charge cycle daily, this is a difference of approximately $5.5M versus $20.7M in annual charging electricity cost (est.) — a $15M annual difference that is not trivial at the fleet scale Waymo is targeting.

The implication for Physical AI benchmarking is that energy cost is a relevant geographic variable in deployment economics, most significant when evaluating whether AV economics work at a specific location and scale. Phoenix is not just an operationally favorable market because of weather and regulatory environment — it is also one of the cheapest places in the US to charge an electric vehicle fleet at commercial rates.

Section 6 — Why Energy Is an Underweighted Benchmark Dimension

The Physical AI benchmark framework tracks what matters for commercial viability. Energy cost has historically been underweighted in AV economics analysis for a coherent reason: at current AV fleet scales (hundreds to low thousands of vehicles), energy cost is genuinely small relative to vehicle amortization, depot operations, and remote assistance staffing. A Waymo vehicle cost of $100K-200K amortized over 300K-500K lifetime miles produces a per-mile amortization cost of $0.20-0.67 — far larger than the $0.003-0.013/mile energy cost in any geography. It is rational to focus on the larger cost categories first.

What changes the relative importance of energy as scale increases is not the per-vehicle energy cost — that remains roughly constant — but the absolute fleet-level energy expenditure, the depot infrastructure capital requirements, and the interaction between electricity pricing and geographic deployment strategy. A 100,000-vehicle AV fleet consuming 1.46 TWh per year at $0.20/kWh costs $292M annually in electricity. At $0.07/kWh, the same fleet costs $102M. The $190M difference is material to the economics of AV at national scale.

Training compute energy follows a different trajectory. As AI models grow larger and training runs become longer, the electricity cost of training increasingly factors into the comparative cost of deploying more compute versus using more data or more efficient architectures. The FLOP/Watt efficiency of different training hardware becomes a meaningful competitive differentiator as training energy costs approach tens of millions of dollars per year per organization at frontier-model scale.

For humanoid robots, battery life is not an underweighted variable — it is widely recognized as a critical constraint. What is underweighted is the system-level energy economics: the cost of the energy per productive task-hour, and how that cost compares to the human labor cost being displaced. A humanoid robot that consumes 2 kWh over a 4-hour shift at $0.10/kWh costs $0.20 in energy per shift — a negligible variable cost. The binding constraints are battery life (operational duration), battery replacement cost over the robot’s lifetime, and the capital cost of the robot itself — not the electricity cost per kWh.

The benchmark conclusion is that energy is a second-order variable in current Physical AI economics — important to track, meaningful at scale, but not the primary differentiator between success and failure at current fleet sizes. The primary variables remain vehicle cost, remote assistance ratio (VPO), and fleet utilization for AVs; battery life, actuator reliability, and task performance for humanoid robots; and FLOP/Watt efficiency for training compute. Energy becomes a first-order variable as Physical AI scales from thousands to hundreds of thousands of deployed units.

Section 7 — Energy Benchmark Scorecard

The Physical AI energy benchmark will be revisited as fleet scales grow, as Dojo FLOP/Watt data becomes more transparent through Tesla’s disclosures, and as humanoid robot battery life data from commercial deployments becomes available. Energy is not the lever that determines the outcome of the Physical AI race in the near term — but it is the variable that determines long-run operational economics at scale, and it deserves a dedicated position in the benchmark framework.

Note: All figures labeled “(est.)” are directional estimates based on publicly available specifications, EIA electricity rate data, and first-principles modeling as of mid-2026. AV fleet sizes, Dojo power consumption, and humanoid robot battery specifications are not uniformly disclosed by operators; estimates are derived from available data and should be treated as order-of-magnitude benchmarks. This article does not constitute investment advice.

Sources

• Tesla energy consumption data — Tesla ↗
• Waymo fleet operations — Waymo ↗
• NVIDIA H100 power specifications — NVIDIA ↗
• Unitree G1 specifications — Unitree ↗
• US commercial electricity rates — EIA ↗

Tags