The initial wave of AI investing was defined by the economics of pure abundance - funding foundational models that could instantly scrape, summarize, and generate creative content at a marginal cost approaching zero. But as the industry hits the wall of online data exhaustion and attempts to transition from the Reading Age to the autonomous Doing Age, a structural bottleneck emerges.
The scraped, extractive data that fueled the last cycle is finite. What replaces it is not another corpus to mine, but data generated from first-principles physical constraints, proprietary calibration against real-world outcomes, and the self-generated record of every simulated attempt, failure, and retry - a supply that compounds with use rather than depleting. In enterprise software, deep tech, and physical automation, the ultimate constraint is no longer a model's capacity to generate a plausible output, but an organization's ability to recreate reality faithfully enough to trust what happens inside it. Intelligence alone does not eliminate uncertainty - the real world is full of edge cases and failure modes that make perfect prediction impossible, and the closer AI moves toward real-world action, the more expensive a wrong answer becomes. When generative data loops and code become functionally free, the premium on value and risk flips entirely: from prediction toward verification. We are entering the age of the Simulation Premium: an era where market dominance can be captured not by those who generate the most output, but by those who can build the most accurate synthetic version of the world to test it in - before it ever touches the real one.
Physics
An LLM predicts the next most likely token, but tokens alone cannot compute thermal warpage or model structural fatigue across ten thousand load cycles. When the cost of a hallucination shifts from a broken line of code to a melted multi-million-dollar silicon wafer, the market is attracted towards something far stricter: mathematical certainty.
It is a categorical shift in what "correct" means: instead of relying on merely plausible outputs, emerging companies like Axiomatic AI are building systems that combine AI-generated engineering reasoning with formal proof frameworks such as Lean to verify physical constraints and design logic before simulation or manufacturing, reducing costly iteration cycles and engineering time. Through tools like Lemma and Axiomatic Measurement, agents are being developed that can read scientific papers, construct physics-consistent models, and actively operate physical laboratory equipment.
Vinci approaches the problem from a different angle, treating geometry and physics as a unified data layer so chip designers can quickly model heat and mechanical stress in 3D packages without building complex simulation grids. Its physics-driven AI platform runs thousands of verified semiconductor simulations in hours rather than weeks, backed by $46M from Khosla Ventures and Eclipse and already validated by more than half of the world's top 20 semiconductor companies. The approaches diverge, but the underlying logic is identical: the value is not in the output itself, but in the simulated environment rigorous enough to guarantee it.
Agent Training
Crossing from the Reading Age to the Doing Age requires more than better models. It requires environments where models can prove themselves before they are trusted with anything real. The industry calls these environments rollouts - simulated end-to-end task executions where an agent navigates enterprise software, fails, retries, and receives a reward only upon verifiable completion. The simulation is not the product. It is the precondition for any product worth trusting.
Deeptune’s growth reflects the market's growing recognition of this shift. The critical infrastructure layer of the agentic stack is not simply the model itself, but the synthetic arena where models demonstrate they are ready for deployment - stress-tested against realistic failure conditions before they ever touch a live system. The fidelity of that arena matters as much as the model inside it: an agent trained in a low-resolution simulation of enterprise software will fail the moment it encounters the complexity of the real thing. The gym is not the athlete - but without it, there is no athlete worth trusting. [EDIT: Mercor to acquire Deeptune]
Adversarial Simulation
Applied to security, the same logic becomes even more urgent. As agents gain the ability to reason across complex systems and chain together minor weaknesses into full attack paths, legacy scanning tools and periodic penetration tests cannot keep pace. Finding vulnerabilities is not the same as preventing breaches - the missing layer is continuous adversarial simulation.
Emerging companies like RunSybil are automating the role of the human red team - reasoning through enterprise infrastructure the way an adversary would, autonomously and continuously, before anyone else can. The defensive posture shifts from reactive reporting to active simulation of what an attacker would do next. What makes this valuable is not the automation alone, but the fidelity of the synthetic threat environment it operates within - the more accurately it replicates real enterprise infrastructure, the more reliably it surfaces what an actual adversary would find. Like physics engines and agent training arenas, the value is in the fidelity of the synthetic threat environment, not the report it generates.
Why This Isn't Just a Feature
The natural objection is that simulation will simply become another feature inside existing AI platforms. If foundation models can reason about anything, why wouldn't they eventually generate the environments themselves?
The answer is that simulation is not defined by the ability to create an artificial world. It is defined by the ability to create one accurate enough that decisions made inside it transfer reliably to reality. The difficult work is not generating a synthetic environment; it is continuously encoding the edge cases, constraints, failure modes, and domain-specific knowledge that separate a convincing simulation from a useful one.
A semiconductor simulation company does not win because it can visualize a chip. It wins because its models understand thermal behavior, material constraints, manufacturing limitations, and decades of engineering knowledge. An agent training platform does not win because it can recreate a software interface. It wins because it captures thousands of subtle workflows, exceptions, and failure states that determine whether an agent succeeds in production.
This creates a fundamentally different type of moat. The value compounds through proprietary environments, proprietary failure data, and increasingly accurate models of reality. Every simulation run improves the system's understanding of what can go wrong - a feedback loop generic models cannot easily replicate, and one that turns the companies who own it into gatekeepers between possibility and deployment.
Who Actually Wins?
This also explains why simulation will likely be dominated by vertical platforms rather than one universal environment. The challenge is not creating a synthetic world; it is defining the rules, constraints, and failure modes that make that world useful. A semiconductor simulator, autonomous vehicle environment, and enterprise software training ground require fundamentally different representations of reality. As systems become more general, verification becomes exponentially harder. The winners will be the companies that choose a domain narrow enough to model deeply, but valuable enough that owning that environment becomes indispensable.
Across physics, training, and security, the same structural dynamic is emerging. The companies best positioned to capture durable value are not generic tooling layers, but domain-specific simulation platforms that embed proprietary synthetic environments directly into existing workflows. Simulation quality compounds - closing the gap between synthetic environment and physical reality is itself the hard problem, and the accumulated domain knowledge required to do it continuously is precisely what makes these platforms difficult to replicate, and generic tooling commoditized as models improve.
The enduring winners will be the companies that become the arena itself - the environment standing between AI-generated output and real-world consequence, and one that competitors cannot simply prompt into existence. The Simulation Premium is not a feature. It is the new infrastructure layer of the Doing Age - and it will be built by those who understand that the most valuable thing in an AI-powered world is not the output, but the environment that proves it.