Understand the world. Leave the datacenter behind.
We're raising $500,000 to build AI that predicts how a world responds to action. It's grounded in real simulation and engineered to run on hardware a small team can own, starting with autonomous testing for games.
Raising
$500KPre-seed
Runway
18 moTo seed-ready evidence
First market
GodotAutonomous game testing
First hardware
16 GBOne local GPU
boulderresearch.aiOctober 2026 · Draft for discussion
Boulder Research01 · Summary
01 / The short version
AI can describe the world. It can't yet predict it.
Ask a language model what happens when a knife is pushed off a counter and it will say the knife falls. Ask a world model the same thing and it has to produce the fall: the slide, the tip, the rotation, the bounce. One states the outcome. The other has to produce the process. That second ability is what robots, simulators, and autonomous testing need, and it doesn't come from scaling the first.
The problem. Predicting observations costs orders of magnitude more than predicting text, errors compound instead of staying local, and the standard metrics reward plausible-looking frames over correct physics. Today's answer is the datacenter.
Our thesis. Scale isn't the bottleneck. Representation and systems are. A hybrid model that imagines with a learned network, checks against a real engine, and manages memory like an operating system can do useful work on one GPU.
Why games first. Games are the one domain where you can reload a state and try a different action. That gives counterfactual data for free, a decision-level metric (did it find the bug?), and paying customers.
The ask.$500K over 18 months to ship a Godot testing pilot, publish a decision-level benchmark, demonstrate interactive rollouts on a 16 GB GPU, and reach seed-ready customer evidence.
“The language model states the outcome. The world model has to produce the process.”
Robots, simulators, digital twins, and automated testing all depend on the second kind of model. Today it's built by the few organizations that can afford to brute-force it in a datacenter. We think the bottleneck is what the model holds while it predicts, not how big it is, and that a small team with the right systems skills can attack it directly.
Boulder Research · Confidential draft2
Boulder Research02 · The problem
02 / Why the datacenter is the default
A frame is not a token.
The language-model playbook (collect everything, predict the next unit, scale) worked because text is already compressed by the humans who wrote it. Video arrives raw. The arithmetic from our working paper:
1.5M
Tokens' worth of bytes in one raw 1080p frame. A single frame outweighs a million-token context.
0.2sec
Of video held by a 131,072-token context. The same window holds a novel for a language model.
~1,250×
Attention comparisons for one 20-step diffusion frame vs. one token at a 4K context.
200GB
KV cache for one minute of rollout (reported by WorldKV), more than a single B200 holds.
33GB
Cache for a ten-second rollout. One session fills a 32 GB device, so batching collapses.
3.6×
Slower than real time: a naive robot-control path takes 5.7 s for each 1.6 s action chunk (DreamZero).
Industry fixes exist (distillation, step reuse, split guidance, quantization) and stack to roughly 38× on a GB200. They shorten or hide the same loop. None changes what the model holds while it predicts, which is where the cost comes from.
03 / Why more compute doesn't fix it
Compute buys fidelity. Not correctness.
The loss
Pixel and diffusion losses measure distance to one recording. A cup landing a centimeter to the left is physically correct and still penalized, so models learn to blur between futures.
The units
Video encoders discard detail inside each 8-pixel cell and each 4-frame stride. Contact and velocity live exactly there, so no larger model downstream can recover them.
The errors
A wrong token is local and recoverable. A drifted pixel feeds the next frame. A ten-second rollout is 240 predictions, each built on the last one's mistakes.
“A blurry model that gets the physics right beats a sharp model that gets it wrong. The planner consumes the physics.”
Boulder Research · Confidential draft3
Boulder Research03 · Thesis
04 / Our thesis
Four problems. One company built around them.
Our paper closes by naming four open problems that would change what world models can do. None of them is a matter of more compute. Boulder Research is built to work on all four at once:
01
The paperA state that can be both read and drawn.
Hybrid state. The game engine holds explicit, readable state (positions, rules, inventory). The learned model handles appearance and fast imagination, and is periodically re-anchored to the engine so drift can't compound.
02
The paperCounterfactual data with a measured gap.
Games replay. The real world can't be reset to try a different action, but a game can. Save states give unlimited "what if I'd done X" trajectories, which is the data planning needs and recording can't provide.
03
The paperA metric that tracks decisions, not frames.
Bugs found. Our product metric is the decision metric: did the agent complete the level, find the crash, reproduce the stuck state? It's scored by the engine, not by how the frame looks.
04
The paperA cache that keeps evidence, not frames.
Systems runtime. Memory that grows with facts instead of rollout length, tiered across GPU, RAM, and NVMe. The paper calls this "the one item a reader with a systems background is best placed to solve." That's our team.
Boulder Research · Confidential draft4
Boulder Research04 · Product & platform
05 / First product
Break games before players do.
The customer. Independent and mid-sized studios on Godot, the fastest-growing open-source engine. Capital is arriving (W4 Games raised an $18M Series B and signed with Tencent), but testing tools haven't kept up.
The job. Hand us a build. Autonomous agents explore it and return reproducible failures (crashes, stuck states, regressions), each with the action trace, state snapshot, and video.
Why it compounds. Every run generates action-labelled, failure-labelled trajectories, the scarcest data in the field. The largest public robot corpora are around 500 hours. We generate ours by running games.
Path
When
Revenue
Managed testing We run it, you get reports
Now
Paid pilots, retainers
Godot plugin + CI Self-serve testing
Next
Subscriptions, compute
World-model API Rollout & state prediction
Later
Usage, adapters
Custom systems Robotics, IoT, digital twins
Explore
Engagements, licensing
06 / Leaving the datacenter behind
If it runs on a desk, it runs anywhere.
Rather than fund a pretraining run (published benchmarks cite more than 4,160 H200 GPU-days), we adapt open checkpoints and spend our effort where the paper's arithmetic says the cost lives. Its sensitivity analysis ranks the knobs: patch size (quadratic), then resolution, then frame rate, then rollout length. We turn them in that order and add the engine as the source of truth.
Tier 0 is the design target and first milestone. Tiers 1–2 are scale-out paths, not dependencies.
Adapters
An 84 MB task adapter on a 2.6 GB base: one model, many environments, no duplicated weights.
Cache tiers
KV state streamed across GPU, host RAM, and NVMe, then recalled before it's needed, with periodic re-anchoring.
Kernels
Custom kernels and engine integration. Our team has shipped models onto hardware vendors said couldn't run them.
Boulder Research · Confidential draft5
Boulder Research05 · The round
07 / Use of funds · proposed allocation
$500K. Mostly people and GPUs we own.
Category
Amount
Share
What it buys
Engineering team
$300,000
60%
Two founders full-time for 18 months, plus a part-time systems contractor
Compute
$90,000
18%
Owned Tier 0/1 GPUs, plus cloud bursts for adaptation and evaluation
Godot integration with save-state replay: counterfactual rollouts on demand
First Tier 0 rollout benchmark: memory, latency, and drift on one 16 GB GPU
Months 3–6
Prove it finds things
Evidence-cache prototype: memory grows with facts, not frames
Decision-level benchmark v0: levels completed and bugs found per GPU-hour
Three Godot design partners signed
Months 6–12
Prove customer value
Paid pilots on external builds with reproducible failure reports
Measured comparison against scripted and manual testing on cost and coverage
Interactive Tier 0 target (480p / 20 FPS) measured and published
Months 12–18
Prove it's a company
CI integration and self-serve testing workflow
Publish the paper and open the decision benchmark
Seed-ready: paid pilots, unit economics, and a second environment beyond games
Boulder Research · Confidential draft6
Boulder Research06 · Team & evidence
09 / Team
Built by people who make models fit.
Adhitya Mohan
ML systems & robotics
Reverse-engineered the Rockchip RK3588 NPU to deploy a vision-language model the vendor toolchain couldn't compile (SHARD, EuroMLSys '26), and built an open compiler and runtime for it.
Ported all 130 LIBERO manipulation tasks to MuJoCo Warp: thousands of parallel robot environments on one $700 GPU.
Sim-to-real RL at AMD Research & Advanced Development, with upstream ROCm and PyTorch fixes. MS CS, CU Boulder (BASIL lab).
Chris Kellogg
AI product & infrastructure
AI-native full-stack architect who takes concepts to running, cloud-hosted systems without handoffs.
Built and runs Miniloader, a local-AI platform for running agents and models on your own hardware.
Track record across media, sports tech, law, and real estate, including large-scale data and AI analysis systems for entertainment clients.
10 / Evidence and honesty
What exists, and what we'll prove.
Exists today
Agents completing a full Super Mario World level (video available)
Godot integration work and GPU kernel experiments
A working paper on world-model foundations, costs, and open problems
Team experience shipping models onto constrained and heterogeneous hardware
This round proves
That the system finds commercially meaningful bugs
That studios pay for the workflow
That Godot-specific integration materially improves cost or capability
That interactive rollouts meet the Tier 0 target on a 16 GB GPU
Hardware and performance figures in this brief are targets or published third-party numbers unless stated otherwise. We'll report measured results as we produce them: bugs found, coverage, runtime, memory, hardware cost, and reproducibility.
11 / The ask
$500K to build AI that understands our world, without the datacenter.
We're looking for investors who want the full paper and the demo videos, and for Godot studios who want to be design partners. info@boulderresearch.ai
