The Founders Who Walked Away From Bezos: Accelerated Understanding Launches Physics AI That Skips Transformers

Over dinner at an upscale restaurant in greater Los Angeles in late 2024, Vik Bajaj — who would go on to co-found Project Prometheus with Jeff Bezos — slid a formal offer letter across the table. It proposed that Caltech professor Anima Anandkumar and her collaborator Benedikt Jenik come lead Prometheus’ AI effort: a combined 35% equity stake, a combined annual salary of $1 million rising to $2 million after three months, and more than $2 billion in committed Series A and B financing from investors including Bezos himself. Reuters reviewed the letter. The pair declined.

On August 25, 2026, they unveiled what they built instead: Accelerated Understanding Inc. (branded “Au”), an enterprise-focused physics AI company whose models are built on neural operators rather than Transformers — and which says its system ingested a physical field of more than 5 trillion continuous values in a single inference pass during testing. That is roughly 5 million times the number of tokens flagship models from Anthropic and Google typically process at once, though the two figures measure fundamentally different things.

Prometheus, for its part, closed a $12 billion Series B in June 2026, targeting AI systems for automating the manufacture of complex physical systems. Anandkumar and Jenik kept building on their own terms.

What they actually built

The headline number deserves unpacking, because it is not directly comparable to a language model’s context window — and understanding why gets to the heart of what Accelerated Understanding is attempting.

A language model’s context counts discrete tokens: words, subwords, or characters mapped to a fixed vocabulary. What Au’s system processes is a spatially and temporally distributed physical field — the temperature at every point in a 3D volume at every time step, the pressure gradient across a chip’s thermal surface, the geological stress through a subsurface formation. Each value is a continuous measurement, not a token. The company’s own materials describe three distinct figures that coverage has tended to collapse into one: 1 trillion parameters in pre-training, 5+ trillion context elements at inference, and a 35-trillion-parameter scaling experiment.

Four claims define the architecture:

• Direct 4D prediction. The models are 3D in space and predict the full time rollout in one shot — 4 dimensions at once. Flattening space would lose detail; autoregressive step-by-step prediction would let errors compound.
• Native super-resolution. The models are resolution-invariant — any level of detail can be swapped in during training or inference, tested up to 5 trillion context without patching or sub-sampling tricks.
• Multiple physics domains in one model. Fluid dynamics, heat transfer, electromagnetic propagation, and structural mechanics are trained into the same network. The company reports observing “cross-physics uplift” — a universal model outperforming models trained on individual domains, mirroring what scale did for language.
• Directional feedback. An experiment tells you what happened, not why or how to improve. The models provide a direction of improvement, which powers a simulate → understand → improve loop for invention and discovery.

The company has been training frontier physical AI models for over a year — hundreds of pre-training runs, model sizes up to 1 trillion parameters, scaling experiments up to 35T, and 2–6 petabytes of data in an average training run.

Why Transformers can’t speak physics

The Transformer architecture, introduced by Google researchers in 2017, processes sequences of discrete tokens. It treats the world as sequence prediction, which works extraordinarily well for language. It works less well for physics. Temperature, pressure, velocity, and material stress are continuous variables defined at every point in space and every moment in time. No tokenization scheme preserves the continuous spatial relationships that govern how they interact — a Transformer must first discretize those fields onto a fixed grid, losing resolution independence and requiring retraining when resolution changes.

Neural operators take a different path. First introduced by Anandkumar’s research group in 2020 and formalized in a Journal of Machine Learning Research paper, they learn mappings between function spaces rather than between finite-dimensional vectors. Feed in an initial condition and the operator returns the output function at any resolution without retraining — a property called discretization invariance that classical numerical solvers share trivially (use a finer mesh) but standard neural networks do not.

The most widely deployed variant before Au was the Fourier Neural Operator (FNO), which approximates kernel-integral computation with a Fast Fourier Transform, capturing global physical structure at O(N log N) cost. Anandkumar’s group applied it to build FourCastNet in 2022 — an AI global weather model that ran tens of thousands of times faster than conventional numerical weather prediction while matching or exceeding its accuracy, and which now runs at premier weather agencies. Anandkumar spent five years at NVIDIA as Senior Director of AI Research convincing Jensen Huang that physics-trained AI could replace numerical solvers; his response, as she recalled to Reuters, was characteristically unambiguous: “I want it to eat all their lunches.”

Au extends the FNO lineage in two directions: native 4D operation and multi-domain training. If the cross-domain claim is validated, it would be a qualitative step beyond FourCastNet (atmospheric physics only) and beyond PhysicsX (strong results in specific engineering domains, no claim of general cross-domain generalization).

Four markets, one model

Accelerated Understanding has identified chip design optimization, robotics, weather prediction, and geological/energy analysis as its primary commercial domains. The pitch to each is the same: R&D is bottlenecked by lab work and slow numerical simulation, and a real-world experiment can tell you what happened but not how to improve it. A model that simulates physics in seconds, at arbitrary resolution, with directional feedback, compresses design cycles that currently take days or weeks per iteration.

There is competition. Startups linked to Yann LeCun and Fei-Fei Li are pursuing “world models” that understand physical space better than text-based AI — though most of those efforts are video-prediction architectures, which flatten 3D space into 2D frames and predict autoregressively. Anandkumar’s framing of the philosophical divide: “The language-centric view of intelligence is humans at the center. Putting physics at the center is a nature-centric view.”

Reason for caution

The launch comes with real caveats. The company has not published a technical paper, released reproducible evaluation code, disclosed benchmark comparisons against established simulation software, or announced a named production customer. A third-party technical analysis published at launch noted that a single inference sample at the claimed largest context produces roughly 22 terabytes of output data, implying distributed storage and GPU infrastructure beyond any single server. Anandkumar confirmed hardware partners have supplied compute clusters but declined to name them; NVIDIA did not respond when Reuters asked whether it was backing the venture.

None of this makes the launch empty — Anandkumar’s research lineage (FourCastNet, the neural-operator formalization itself) is documented and peer-reviewed. But enterprise buyers should treat the 5-trillion figure as a stated scale claim from a newly public company, not an independently validated result.

The wager, though, is clear. If intelligence is getting cheaper and more abundant, the bottleneck shifts from generating ideas to executing them — and execution in the physical world runs through simulation. The founders who walked away from a $2 billion safety net are betting that the next foundation model isn’t a chatbot. It’s the universe, modeled directly.