Orchestration and Physical Infrastructure Define Future AI Value

The AI gold rush is moving from building models to orchestration. This shift suggests that the most valuable companies will not be those that build the smartest models, but those that own the interface where AI performs actual work. Aravind Srinivas argues that the era of the model as a product is ending. We are entering an agentic economy where the winner is determined by token value per watt. For founders and investors, competitive advantage now comes from building harnesses: orchestration systems that ground AI in real world tools and context. This reveals a non-obvious truth: infrastructure, specifically power and physical compute, is the true bottleneck of the next decade. Software-only strategies are becoming increasingly fragile.

The orchestration moat: why models are commodities

The conventional wisdom in AI is that the model is the product. Srinivas disagrees, arguing that models are rapidly becoming utilities. If you are just a reseller of model tokens, you do not have a sustainable business. The real value lies in the agent harness, which is the orchestration layer that pairs a model with tools, connectors, and rules to perform autonomous work.

"Without the harness you don't necessarily capture and convert the intrinsic intelligence in the model into valuable output tokens. The output tokens if you're literally just a reseller of model tokens you have no business because the model will get commoditized."

-- Aravind Srinivas

This creates a systems-level shift. The goal is no longer to reach a billion users with a chatbot, but to capture the token economy of power users who run autonomous agent loops. These loops, which perform root cause analysis or monitor systems 24/7, are where the revenue resides. Companies that orchestrate across multiple models, rather than relying on a single provider, gain an advantage by maximizing token value per watt.

The physical bottleneck: power as the new operating system

While software teams focus on fine-tuning, the real constraint on AI progress is physical. Srinivas points out that data center construction is currently hampered by power and permitting, not just chip availability. This creates a hidden consequence: companies that solve the physical infrastructure problem by securing power, land, and cooling will command higher value than software-only players.

The implication is that infrastructure is no longer just a cost center; it is a strategic asset. Srinivas suggests that Micron, as a supplier of High Bandwidth Memory, could become more valuable than Meta because they own the bottleneck. When the system is supply-constrained, power shifts from the service provider to the component manufacturer.

"Whatever is the bottleneck will command the price. AMD is doing really well because CPUs became a bottleneck again. Agent loops, agent harnesses are all running on CPUs... so suddenly there's a rise in enterprise CPUs. And the beneficiaries of these are like Intel and AMD."

-- Aravind Srinivas

The China paradox: export controls as a catalyst

Systems thinking reveals a counter-intuitive effect of U.S. export controls. By restricting access to top-tier NVIDIA GPUs, the U.S. may have forced China to become more efficient. Because they cannot rely on the brute force of the NVIDIA stack, Chinese labs are innovating at the memory-efficiency and architecture levels. They are building vertically integrated stacks that are less dependent on high-bandwidth memory.

This creates a long-term risk. The U.S. might be optimizing for a short-term lead while China develops more resilient, physically capable infrastructure. If the future of AI includes physical robotics and hardware-integrated models, China’s forced focus on physical-layer competence could become a lasting competitive advantage.

Key action items

• Shift from chat to agentic workflows: Stop measuring productivity by how many questions an AI answers. Start identifying repetitive, multi-step tasks like monitoring latency or automated reporting that can be delegated to an agent harness.
• Audit your dependency on frontier models: Evaluate where your application can use smaller, local, or fine-tuned open-source models. Relying solely on a single frontier model provider introduces significant API risk and cost volatility.
• Prioritize orchestration over model building: If you are building a product, focus on the harness, which includes the connectors, tools, and workflows that make the model useful, rather than trying to compete in training foundation models.
• Invest in physical literacy: For investors and large-scale operators, understand the total cost of operations for data centers. The winners will be those who secure power and infrastructure, not just those who buy GPU time.
• Adopt an attack mindset: Srinivas emphasizes that in this field, comfort is a death sentence. Use the 10-month timeframe mental model: if a project is scheduled for 10 years, ask what it would take to finish in 10 months to force radical efficiency.