Where the Wheels Are Coming Off: Physical Limits That Will Shape AI for Business

TL;DR: AI’s bottlenecks are physical — not just software. Limited chip supply, rising energy and cooling costs, and the need for real-world data are real constraints that will determine who wins with AI agents and AI automation. Leaders should treat AI as an industrial system: secure compute capacity, tighten agent governance, pilot alternative model architectures for physical tasks, and plan workforce transitions where automation complements scarce labor.

Why this matters for leaders

Executives often frame AI as a software line item: hire the models, plug in the data, and watch productivity rise. That mindset misses a hard truth — AI runs on hardware, power, and messy real-world inputs. When chips, electricity, cooling and geopolitics become limiting factors, strategy shifts from “more models” to “how to get and use scarce physical resources.”

For sales, operations and product teams, that means an AI program can stall not because an algorithm failed but because there isn’t enough silicon, enough affordable power, or enough validated field data to make autonomous systems safe and compliant. Those are procurement and risk problems as much as technical ones.

Chip supply and the EUV bottleneck

Advanced chips require a specialized manufacturing step called extreme ultraviolet (EUV) lithography — the machines that etch the smallest, fastest, most power-efficient processors. ASML is the only company that makes these EUV systems, creating a strategic chokepoint for modern semiconductor capacity.

“The industry is in a phase of rapid chip production growth, but physical production capacity will limit supply for years.” — Christophe Fouquet, ASML

Translation: think of compute like highway lanes. Software wants more lanes instantly, but building the bridges and tunnels (chip fabs and EUV machines) takes years and massive capital. Expect meaningful chip supply constraints for roughly two to five years as capacity ramps — a window that matters for procurement and competitive positioning.

Energy, cooling and the thermodynamic limit

Compute is cheap only if power and cooling are cheap. “Flops-per-watt” — compute throughput per unit of power — is now often the real metric CEOs should watch. When efficiency hits thermodynamic limits, the conversation moves from chips to energy contracts, data center location, and even speculative ideas like orbital data centers that trade terrestrial cooling for radiative heat rejection in space.

“Demand is massive and growing fast; co-designing chips, software, and models yields big energy-efficiency advantages.” — Francis deSouza, Google Cloud

Co-design means building chips, software stacks and models together so every layer is optimized for energy efficiency. In plain terms: you get more useful work from each watt consumed. Hyperscalers are already committing capital and backlog to these designs (Google Cloud reported over $20 billion in quarterly revenue and a rapidly growing backlog), which gives them a practical edge on flops-per-watt.

The data bottleneck for physical autonomy

For robotics, drones and autonomous vehicles, compute is necessary but insufficient. Certain training signals — rare corner cases, sensor failure modes, regulatory interactions — can only be captured in the real world. High-fidelity simulation helps, but it doesn’t fully replace messy live data.

“For physical autonomy, certain training data can only be collected in the real world—simulation cannot fully replace it.” — Qasar Younis, Applied Intuition

That reality reshapes strategy in two ways. First, firms building physical AI must plan long-term data capture pipelines and field validation budgets. Second, nations and regulators care about who controls those systems: sovereignty concerns mean some countries will resist foreign-controlled robot fleets or autonomous logistics operating within their borders.

Architecture choices: scale-oriented LLMs vs. energy-based models

There is no single winning model architecture for every problem. The industry is bifurcating:

• One track doubles down on scale: ever-larger LLMs powering broad-purpose AI agents and automation across knowledge workflows.
• The other favors alternate inductive biases — built-in assumptions that make a model better at certain tasks — like energy-based models (EBMs) that learn physical rules rather than predicting the next word.

“Reasoning isn’t inherently linguistic; models that learn physical rules rather than next-token prediction may better suit robotics and chip design.” — Eve Bodnia, Logical Intelligence

EBMs can be much smaller (Logical Intelligence has models on the order of ~200 million parameters) and run thousands of times faster than top LLMs. They’re also easier to update incrementally instead of retraining huge models — a useful property when you need rapid patches for safety-critical automation.

Practical implication: pick the right model for the job. Use scaled LLMs and agent stacks for language-heavy workflows (customer support, sales enablement, knowledge work). Use EBMs or other physics-aware architectures for robotics, control systems, chip layout and domains where rule-governed behavior matters.

Agentization and enterprise governance

AI agents — digital workers that act on behalf of users — are moving from novelty to core enterprise capability. Platforms like Perplexity Computer position agents as programmable workers with connectors into corporate systems.

“Agents should be controllable at fine granularity—enterprise admins must set precise permissions and require explicit approvals when agents act.” — Dimitry Shevelenko, Perplexity

Control mechanisms that matter:

• Connector permissions: enable read-only access by default; restrict write access to approved flows.
• Plan-and-approve: require agents to submit a proposed action plan for human sign-off before execution.
• Audit trails and explainability: log agent decisions, data sources, and approvals for compliance and troubleshooting.

These controls let enterprises adopt AI automation and AI for sales without creating runaway data or regulatory exposure.

Quick wins this quarter

• Negotiate multi-quarter capacity commitments with cloud vendors and ask for flops-per-watt guarantees where possible.
• Audit all planned agent connectors and set conservative defaults (read-only; require human approval for writes).
• Start field-data capture pilots for any autonomy use-case; prioritize logging, edge-case capture and redundancy.
• Run a 90-day experiment comparing an LLM-driven agent vs. a targeted, smaller physics-aware model for one production task.

Five-point checklist for C-suite

• Secure compute capacity
  Lock in cloud/colocation commitments and evaluate hybrid options to hedge chip supply constraints.
• Measure energy efficiency
  Track flops-per-watt and factor energy costs into ROI for AI projects.
• Govern agents
  Implement connector permissions, approval flows, and audit logs before scaling agent-based automation.
• Choose architectures by use-case
  Map problems to model families: LLMs for language; EBMs/physics-aware models for control and safety-critical tasks.
• Plan workforce transitions
  Identify roles where automation complements scarce labor (mining, trucking, farming) and where it may replace routine office tasks; invest in reskilling accordingly.

Scenarios and strategic posture

• Optimistic: Chip fabs and EUV capacity ramp faster than expected; energy innovations and co-designed stacks reduce marginal costs. Outcome: broad enterprise adoption accelerates.
• Baseline (most likely): 2–5 years of constrained advanced-chip supply. Energy costs and cooling remain important drivers. Outcome: winners are those who secure capacity and optimize for energy efficiency.
• Pessimistic: Geopolitical export controls or supply chain disruptions extend constraints; countries restrict foreign-controlled physical AI. Outcome: fragmented AI ecosystems and higher costs for cross-border deployments.

Questions for your leadership team

• Do we have multi-quarter compute and energy contracts for our top AI initiatives?
• Which AI use-cases need real-world data collection, and what’s our budget and timeline for that capture?
• Have we defined connector-level permissions and approval gating for any agent we intend to deploy into production?
• Which problems should we pilot with smaller, physics-aware models instead of scaling an LLM?
• How will automation affect our entry-level hiring pipeline and what reskilling programs are we funding?

Final verdict

AI agents and AI automation can deliver big gains across sales, operations and product teams. But the race for scale is bumping into real-world physics and geopolitics. Treat AI like an industrial system: secure scarce hardware, govern digital workers tightly, choose architectures based on the problem, and plan workforce transitions where automation fills labor gaps rather than just cuts costs.

Short-term focus — capacity commitments, energy-aware architecture choices, and rigorous agent governance — separates costly experiments from durable advantage. Leaders who recalibrate their procurement, security and R&D practices now will convert the promise of AI for business into sustained value.

Notes: Remarks summarized here reflect public comments and a panel discussion featuring leaders from ASML, Google Cloud, Applied Intuition, Perplexity and Logical Intelligence at a recent industry conference. Data points referenced include Google Cloud’s reported quarterly revenue and backlog figures as discussed publicly by company leadership.