A startup co-founded by Caltech professor Anima Anandkumar has built a model that processes 5 trillion data points in a single prompt, a scale that the company says far exceeds conventional large language models. Accelerated Understanding Inc, founded by Anandkumar and Benedikt Jenik, is pitching an Accelerated Understanding neural operator physics AI architecture that replaces the transformer blueprint that has dominated the field since 2017. The company says its system was trained on 1 trillion tokens and handles more than 5 trillion tokens at inference, with 1 trillion parameters in pre-training.
Accelerated Understanding neural operator physics AI: Neural Operators Replace the Transformer Stack
The model is built on neural operator architecture rather than the transformer framework, a technical choice that the founders argue better suits physics-heavy workloads. Neural operators learn mappings between entire functions rather than discrete tokens, allowing the system to operate in four dimensions across 3D space plus time. Anandkumar, who previously served as director of machine learning research at NVIDIA, said the approach lets the model ingest continuous physical fields directly. “The transformer was designed for language,” Anandkumar said in a statement. “Physical phenomena are not language. We needed a different primitive.”
Scale Rivals the Largest Models Ever Built
Accelerated Understanding said its model was scaled to 1 trillion parameters in pre-training, placing it in the same weight class as the largest foundation models ever built. During inference, the system can process 5 trillion data points in a single prompt, while training data reached 1 trillion tokens and inference-time data exceeded 5 trillion tokens. The company has not disclosed the compute footprint or training cost. The numerical footprint compares with publicly disclosed parameter counts from OpenAI, Google DeepMind and Anthropic, whose flagship models are also reported in the hundreds-of-billions range.
Founders Turned Down Bezos-Backed Project Prometheus
Anandkumar and Jenik co-founded Accelerated Understanding Inc after declining an early role at Project Prometheus, the Jeff Bezos-backed AI venture that launched in late 2025 with a reported $6.2 billion in commitments. The founders declined to disclose the size of their own funding round, but said their decision to build independently was driven by a desire to focus the technology on scientific and industrial use cases. “We were offered a seat at the table,” Jenik said. “We decided to build a different table.”
Target Applications Span Energy, Chips, Robotics and Weather
The startup is targeting five vertical markets: energy optimization, chip design, robotics, weather prediction and medical innovation. In energy, the company is working with utilities on grid load forecasting and turbine placement. In chip design, the model is being used to simulate thermal behavior across multi-die packages. Robotics partners are using the system to predict contact dynamics in soft-material manipulation. Weather prediction and drug discovery clients are also in active pilots, though the company declined to name them. The four-dimensional nature of the outputs, spanning 3D space plus time, is positioned as the differentiator for problems where classical numerical solvers remain too slow for real-time use.
Physics-First Approach Could Reshape Industrial AI
Analysts tracking the foundation-model sector say the launch marks one of the first credible attempts to scale a non-transformer architecture to trillion-parameter territory. Most industrial AI today relies on fine-tuned transformers or domain-specific solvers that do not learn from data at scale. A general-purpose physics engine trained at this scale would compete with both, and could displace simulation software that has changed little in two decades. Investors will watch whether Accelerated Understanding can convert the technical benchmarks into paying enterprise contracts before larger labs field competing systems. The bet behind the Accelerated Understanding neural operator physics AI is that the next decade of foundation models will be defined less by language and more by the laws of physics, a wager the founders are now asking customers to test with their own production workloads, and the broader market will judge the future of the Accelerated Understanding neural operator physics AI on whether that bet pays off.
The timing of the launch lands at a moment when enterprise buyers are questioning whether transformer-based systems can deliver reliable performance on structured scientific problems, where hallucinated outputs carry regulatory and safety consequences that text-generation errors rarely do. Major cloud providers have begun marketing “physics-aware” AI services, but most are wrappers around numerical solvers rather than natively trained architectures, leaving room for a specialized entrant to differentiate on accuracy and latency. Accelerated Understanding has signaled that enterprise pilots will be evaluated on benchmarks for partial differential equations, turbulence modeling and electromagnetic propagation, areas where small percentage gains in fidelity translate into measurable cost savings for industrial customers. The company’s pricing model has not been disclosed, though founders have suggested a consumption-based structure tied to simulated physical units rather than token counts, an unusual framing that could complicate procurement cycles but may appeal to engineers accustomed to licensing traditional simulation software. Whether that commercial approach resonates with chief information officers, who have spent two years standardizing on token-priced APIs from frontier labs, will likely shape the first wave of adoption contracts and determine how quickly the broader foundation-model sector takes physics-native architectures seriously. This is what makes Accelerated Understanding neural operator physics AI a meaningful shift in how the space evolves.
Source: CryptoBriefing — Accelerated Understanding Inc launches new AI model that ditches transformers for neural o