Sub-Creation: Gaming Built the Infrastructure. Robotics Just Moved In.

What Tolkien, Game Engines, and a $700M Startup Reveal About Where Physical AI Moats Form

Bob McGrew – former Chief Research Officer at OpenAI, the person who led the team that taught a robot hand to solve a Rubik's Cube entirely inside a simulation – just named his new company Arda. In Tolkien's legendarium, Arda doesn't mean Middle-earth; it means the world as a whole.

McGrew was one of the earliest engineers at Palantir. That company is named after Tolkien's seeing stones. His co-founders Jakob Frick and Alex Mark are also Palantir veterans. The company is reportedly raising $70 million at a $700 million valuation from Founders Fund and Accel, with Khosla and XYZ participating.

That would be trivia on its own. But it's part of a pattern.

Palantir. Anduril (Aragorn's reforged sword). Mithril Capital. Valar Ventures. Narya Capital. Erebor Bank. Varda Space Industries. Samuel Arbesman's running list now tracks over 25 Tolkien-named tech companies, clustered heavily around the Thiel network. Lux Capital alone has five Tolkien-named portfolio companies. The question that matters isn't why Tolkien. It's why this specific mythology for these specific companies.

(If none of these references mean anything to you, bear with me; the pattern matters more than the lore)

What made Tolkien singular was the seriousness of his world-building as much as the stories themselves. He spent nearly 60 years constructing Arda – inventing languages, geologies, histories, and internally consistent rules. His 1947 essay On Fairy-Stories articulated the concept of "sub-creation": the idea that a creator builds a Secondary World with its own coherent logic, a complete reality you can step inside.

Whether they'd frame it this way or not, these founders are reaching for the same idea Tolkien spent his life on: constructing a coherent secondary world, complete enough to act inside. McGrew's Arda watches factory floor video, builds computational models of physical processes, and uses that understanding to train robots. He is constructing a digital Secondary World from observations of the primary one, then using it to reshape reality. The resemblance to sub-creation is hard to miss.

But Tolkien would also have recognised the danger. His most powerful artefacts come with a warning attached: the palantíri reveal truth, but never the full picture; the Ring amplifies power at a cost no wielder can anticipate. The gap between the world you build and the world you inhabit is the oldest story in mythology. And it maps onto the central technical problem in robotics: sim-to-real transfer. The moment your beautifully constructed simulation meets the mess of physical reality.

The Thesis

Some of the most defensible companies in physical AI are likely to be the ones building the simulated worlds that robots learn in, rather than the robots themselves. The gaming industry spent three decades building critical pieces of the infrastructure stack that physical AI now depends on. Hardware is commoditising – Barclays estimates humanoid production costs have compressed from roughly $3 million to under $150,000 in the past decade, and Chinese volume manufacturing is pushing that further. But simulation environments calibrated for specific industrial domains, video-based world models trained on proprietary data, and the data flywheels they create compound in ways that hardware cost curves don't. They do not get cheaper for your competitors just because they got cheaper for you.

For an early-stage investor, the question is not which robot to back. It's which layer of the simulation stack to enter.

The Pipeline Nobody Talks About

Everyone knows NVIDIA made GPUs for gamers before AI researchers repurposed them; that narrative is exhausted. What's far less appreciated is how deep the gaming-to-physical-AI pipeline actually runs, and how many of its foundational technologies were built by people who started in game engines, demoscenes, and MMO server rooms.

The clearest example comes from Finland. The demoscene there seeded an outsized share of the talent that later shaped NVIDIA's graphics and generative AI research.

Future Crew, founded in 1986, was the most celebrated PC demoscene group in the world. Their members went on to found Remedy Entertainment and Futuremark. From the broader Helsinki ecosystem, Hybrid Graphics emerged – and when NVIDIA acquired it in 2006, they gained Timo Aila, a researcher who had built the Umbra occlusion culling library that ships in most modern AAA games. Aila went on to play a central role in NVIDIA's ray-tracing research, including the design of RTX hardware units. Then he pivoted to machine learning. His colleagues Tero Karras and Samuli Laine, from the same Helsinki University of Technology ecosystem, authored StyleGAN – the architecture that proved AI could generate photorealistic human faces. NVIDIA established an entire research lab in Helsinki to capture this talent pipeline. The path was demoscene optimisation → game rendering middleware → GPU hardware design → generative AI.

Or consider Erwin Coumans.

Coumans built the Bullet physics engine at Sony Computer Entertainment R&D around 2003. Bullet went into AAA games, then Hollywood visual effects – earning him a Scientific and Technical Academy Award. He moved through AMD, then to Google Brain, where he created PyBullet – the Python bindings that became the default physics engine for reinforcement learning research. OpenAI built Roboschool entirely on Bullet, democratising RL by removing the MuJoCo licence barrier. Google's landmark quadruped locomotion paper – the one where a robot learned to walk by imitating animals – was built on Coumans' work. He now works at NVIDIA on Omniverse. Game physics → movie VFX → RL research → real robot deployment → NVIDIA's simulation backbone. One person, one codebase, twenty years – from PlayStation to NVIDIA's simulation backbone.

The borrowings go deeper still. Behaviour trees – the decision architecture most modern game NPCs use – emerged in games in the early 2000s, with Halo 2 providing one of the earliest well-known implementations. Researchers at KTH Royal Institute of Technology in Stockholm later formalised them for robotics, and they are now standard in UAV control and multi-robot coordination. A* pathfinding, refined obsessively by the game industry, underpins autonomous navigation. John Carmack – who invented BSP trees for Doom and popularised the fast inverse square root for Quake – now runs an AGI startup. He has argued repeatedly that without early 3D games creating demand for powerful GPUs, the modern AI boom might never have happened.

Three Layers, One Investable

The simulation pipeline breaks into three layers, and only one of them is investable at seed.

The bottom layer – physics engines – is commoditised. MuJoCo is open-source under Google DeepMind. PyBullet is open-source. PhysX is open-source under NVIDIA. Drake is open-source under MIT and Toyota. There is no venture-scale business in building another physics engine. Value accrues not to the engines themselves,but to domain-specific frameworks built on top of them.

The middle layer – digital twins, rendering, synthetic data platforms – is being captured by the NVIDIA-Siemens axis. Siemens unveiled its Digital Twin Composer at CES 2026, and early adopters like PepsiCo reported 20% throughput increases and 10–15% capex reductions. A new horizontal digital-twin platform now runs into incumbents with far deeper distribution and ecosystem control.

The top layer is where the seed opportunities live. Video-to-sim. Domain-specific world models. The core idea it that rather than painstakingly hand-building simulation environments, learn a model of the physical world directly from video. NVIDIA, Google DeepMind, Meta, Runway, and World Labs (Fei-Fei Li, $1 billion raised in February 2026) are all building variants. The compounding advantage comes from the combination of domain-specific data, task-specific models, validation pipelines, and the feedback loop from real deployments – not any single variable in isolation.

The numbers make the leverage concrete. NVIDIA's GR00T Blueprint generated 780,000 synthetic robot training trajectories – equivalent to nine continuous months of human demonstrations – in just 11 hours, improving model performance by 40%. That is the power of simulation. But that power is only as good as the domain specificity of what you are simulating.

A simulation environment tuned for automotive welding lines does not transfer to pharmaceutical packaging. The edge cases, failure modes, and physics parameters are domain-specific and hard-won. Once a startup has calibrated its simulation for a specific industrial vertical, the switching costs are real – replicating months of calibration and validation with a competitor is expensive and carries genuine operational risk. The pattern is familiar from vertical software: once the workflow becomes domain-specific, the moat compounds faster than outsiders expect. Domain-specific simulation creates the same kind of defensibility that vertical SaaS does in healthcare or legal.

There is a bear case worth taking seriously. The sim-to-real gap remains genuinely difficult for deformable objects, contact-rich manipulation, and tactile sensing – areas where simulation fidelity still falls short. NVIDIA's stack is expanding relentlessly, and any startup building on Omniverse needs a credible answer for what happens when NVIDIA moves into their layer. And the Covariant story is instructive: a company that raised $222 million, built a genuinely impressive robotics foundation model, and then saw Amazon hire its founders, take about a quarter of the team, and license the core technology in a reverse acqui-hire that looked like a disappointing outcome relative to Covariant's last private round. Big Tech has deployment data advantages – Amazon runs over 750,000 robots – that standalone startups struggle to match. The bull case for simulation moats requires acknowledging these risks, not ignoring them.

Arda: The Case Study

At Palantir, McGrew helped shape the Forward Deployed Engineer model – embedding engineers at customer sites to bridge the gap between what the product does and what the customer needs. At OpenAI, he led Dactyl, which trained a robot hand entirely in simulation using domain randomisation – randomising mass, friction, gravity, dimensions to make the learned behaviour robust enough to transfer to physical hardware without fine-tuning. On Sequoia's podcast, he noted that what took OpenAI years on a single Rubik's Cube task now takes companies like Physical Intelligence months across a wide range of problems, because foundation models and simulation infrastructure have matured. His conclusion: robotics is at the end stages of being a research challenge, and months or small-digit years away from commercialisation.

Arda's approach inverts traditional sim-to-real. Instead of building simulations and hoping they transfer to reality, it watches reality – factory floor video – and builds understanding from observation. The video model learns manufacturing processes, then uses those learned representations to train and coordinate robots across full production cycles. Co-founder Augustus Odena brings deep model-building credentials from Google Brain and Adept AI, including work he describes as foundational to modern reasoning techniques. Palantir veterans Frick and Mark round out a team that combines frontier AI research with enterprise deployment experience.

The reshoring macro tailwind is real. According to the Reshoring Initiative, 88% of reshored US jobs in 2024 were in high or medium-high tech sectors.

But the hard question – the one a good seed investor should ask – is about GTM, not technology. Early manufacturing customers will not buy the grand "lights-out factory" vision. They will buy scrap reduction, OEE improvement, and changeover speed. Factories are heterogeneous, under-instrumented, risk-averse, and politically complex. At a reported $700 million valuation, investors are pricing McGrew's track record and the belief that video-to-sim creates a fundamentally different competitive position. That might be right. But in manufacturing, model quality alone is rarely enough; distribution, workflow integration, and trust decide who actually gets adopted. This is a tension I would not resolve. I would name it, watch it, and evaluate the team's ability to navigate it.

The European Edge

Europe has a better claim on physical AI than it is usually given credit for. The advantage is not just talent, but it's proximity to real industrial systems.

Wayve in London – fresh from a $1.5 billion financing at an $8.6 billion valuation, backed by Eclipse, Balderton, and SoftBank – shows Europe can produce physical AI companies at frontier scale. 1X Technologies, founded in Norway and backed by OpenAI and EQT Ventures, has announced a shared intent to facilitate rollout of up to 10,000 humanoid robots across EQT's portfolio companies by 2030. And General Intuition in Geneva raised a $134 million seed round to bridge gaming data – roughly 2 billion annual gameplay clips from its Medal platform – directly into embodied AI training. That last one might be the purest expression of the gaming-to-simulation-to-robotics thesis on the European continent. The real edge here is not romance about old industry. It is access: factories, suppliers, automation stacks, and the messy operational data that world models need in order to matter.

The Oldest Story in Mythology

Whether intentionally or not, the founders naming companies after Tolkien's world are pointing at the same idea: in physical AI, the moat forms less in hardware than in domain-specific world-building – in simulation, in sub-creation.

But Tolkien's deepest lesson – one that even founders who've spent careers on the problem can underestimate – is that the gap between the world you construct and the world you inhabit is where things break. The palantíri showed truth, but never the full picture. The Ring granted power, but at a cost the wielder could never anticipate. The sim-to-real transfer problem is, it turns out, as old as mythology itself.

The founders worth backing aren't just building worlds – they're obsessing over what happens when those worlds collide with reality, armed with domain-specific data no one else has. McGrew named his company "the world". The question is whether his Secondary World will be faithful enough to the primary one.

That is the bet.