The Bit Wars: Why the Future of AI Might Be Measured in 1s and 0s

PrismML emerges from stealth with the world's first commercially viable 1-bit AI model, challenging the assumption that intelligence requires scale.

By Carry and Conquer Publications

On March 31, 2026, a Pasadena-based startup called PrismML stepped out of stealth with a claim that would have sounded absurd to most AI researchers just a few years ago: that you can run a powerful language model on a smartphone using less than a gigabyte of memory, with no cloud connection required. The company's flagship 1-bit Bonsai 8B model delivers reasoning performance comparable to full-precision 8-billion-parameter models while consuming 14 times less memory, running 8 times faster, and burning 4 to 5 times less energy. The founders did not achieve this by building a bigger model. They achieved it by asking a question the rest of the field had largely set aside - how much intelligence can you extract from a single bit?

The Man Behind the Math

Babak Hassibi has spent most of his career thinking about information at the theoretical level. Born in Tehran in 1967, he earned his PhD in electrical engineering from Stanford in 1996 under Thomas Kailath, then spent two years at Bell Labs before joining Caltech's faculty in 2001. He is now the Mose and Lillian S. Bohn Professor of Electrical Engineering and Computing and Mathematical Sciences, and his research spans information theory, signal processing, control theory, and machine learning. He was awarded the 2003 Presidential Early Career Award for Scientists and Engineers and a David and Lucille Packard Fellowship in the same year - recognitions that followed his work on the mathematical foundations of efficient communication.

PrismML is built directly on Caltech intellectual property developed by Hassibi and three co-founders: Sahin Lale, Omead Pooladzandi, and Reza Sadri, who serves as VP of Strategy. The company is backed by Khosla Ventures and Cerberus Ventures, with compute grants from Google and Caltech. Vinod Khosla, who made the investment, framed the thesis in stark terms: "AI's future will not be defined by who can build the largest datacenters. It will be defined by who can deliver the most intelligence per unit of energy and cost."

The years of mathematical groundwork show in what Hassibi describes as the core challenge: compressing a neural network without degrading its reasoning. Standard quantization - the practice of converting a high-precision model to fewer bits after training - typically involves meaningful performance tradeoffs. PrismML trained its models natively in 1-bit from the start, forcing the architecture to learn efficient representations rather than pruning an existing full-precision model after the fact.

A Physical Reckoning

The timing of PrismML's launch is not coincidental. The AI industry is facing a physical constraint it cannot simply spend its way past. Datacenters running today's frontier AI models require unprecedented amounts of power. RAND Corporation research estimates that AI datacenters could need 68 gigawatts of total power capacity by 2027 - close to the entire power capacity of California. The International Energy Agency projects that electricity consumption from datacenters, AI, and related compute could reach 945 terawatt-hours by 2030, roughly the total electricity consumption of Japan.

By early 2025, roughly $64 billion worth of US datacenter projects faced delays or opposition linked to concerns about energy use, water consumption, and local community impact. In Northern Virginia, Chicago, Atlanta, and Phoenix - the core US markets - vacancy rates fell to a record-low 1.6%, the tightest ever recorded, according to CBRE's North America Data Center Trends report for H1 2025. Asking prices for large-scale datacenter builds of 10 megawatts or more rose as much as 19% in that period. The bottleneck is no longer capital or technology. It is electrons.

Traditional AI scaling assumptions cannot hold against these physical limits. Each successive generation of frontier models demands more parameters, more training data, more compute, and more power - but the grid infrastructure to deliver that power is built on decades-old architecture and cannot expand quickly enough. Peter Perri III, Managing Partner at Jupiter Island Capital, captured the shift precisely: the primary scaling constraint has moved from chips to power generation. Hyperscalers are now signing deals to bring retired nuclear power plants back online and partnering with renewable developers and private equity to build multi-billion-dollar energy parks. These are not incremental infrastructure upgrades. They are structural responses to a model of AI development that is running into its physical ceiling.

What 1-Bit Actually Means

The conventional approach to neural networks stores each parameter - the learned values that encode a model's capabilities - as a 16-bit or 32-bit floating-point number. PrismML's Bonsai 8B stores every parameter as a single bit. In practice this means each weight can only take the value of 1 or 0, a binary representation of knowledge that would seem, on its face, to be too crude to preserve intelligence.

The fact that it works at all is a mathematical accomplishment. The key distinction from post-training quantization - where an existing full-precision model is compressed down to fewer bits - is that PrismML trains in 1-bit natively. The model learns, from the beginning, to represent the patterns in language and reasoning using binary weights. The architecture is not losing precision that was once there. It is finding a different mathematical path to equivalent capability.

Microsoft Research had explored related territory with its BitNet and BitNet b1.58 programs beginning in 2023 and 2024, using ternary weights with three possible values - negative one, zero, and one - rather than strict binary. Microsoft's open-source framework demonstrated that a 100-billion parameter ternary model could run on a single CPU at human reading speed. But Microsoft's largest commercially tested native model remained at 2 billion parameters, and the research team acknowledged that scaling to 100 billion while maintaining accuracy parity had not yet been demonstrated.

PrismML's claim is to have crossed a different threshold: a commercially viable, natively 1-bit model at 8 billion parameters that matches the benchmark performance of leading full-precision 8-billion parameter models, including Meta's Llama3 8B. At 1.15 gigabytes of memory, the Bonsai 8B requires one-sixteenth the memory of its full-precision equivalent. Ion Stoica, co-founder of Databricks and a professor at UC Berkeley, offered a systems-level read on what this unlocks: reducing models to 1-bit representations changes the optimization equation, enabling AI systems that operate efficiently at the edge and scale economically in the cloud simultaneously.

Amir Salek, founder of Google's TPU program and a Cerberus Ventures investor in the company, extended that argument to hardware architecture itself. Memory footprint and bandwidth demand have been two of the most persistent bottlenecks in AI inference design. A model that requires a sixteenth of the memory is not simply running on different hardware - it is opening the design space for a new generation of inference chips built around entirely different assumptions about what AI workloads look like.

The Edge Unlocked

For the past decade, serious AI capability has required a cloud connection. Inference - the act of running a model to generate a response - has demanded hardware that consumer devices cannot carry. That constraint has shaped product design across an entire industry: AI assistants call home to servers, AI features disappear when connectivity drops, and privacy-sensitive use cases carry inherent risks because data must leave the device.

PrismML's smallest model, the 1-bit Bonsai 1.7B, runs at 130 tokens per second on an iPhone 17 Pro Max with a memory footprint of just 0.24 gigabytes. The 4-billion parameter model hits 132 tokens per second on an M4 Pro laptop chip. These are not demonstration benchmarks at reduced quality. According to PrismML's published performance data, the models maintain accuracy competitive with full-precision equivalents across standard intelligence benchmarks. Bill Jia, VP of Engineering at Google for Core ML/AI, observed that when advanced models can run on constrained devices, it reshapes system design end to end.

The applications that become viable with persistent on-device intelligence span a significant range. Robotics and autonomous systems require real-time inference without the latency penalty of cloud roundtrips. Wearables need intelligence that can operate continuously without draining a battery through radio transmission. Medical devices need AI that works in environments where connectivity cannot be guaranteed and patient data cannot leave the premises. Industrial embedded systems need reasoning capability in locations where datacenter connectivity is impractical or impossible.

The Bonsai models are released under the Apache 2.0 license, free to download and deploy. The open-source release signals PrismML's confidence in the commercial position: the value is not in keeping the weights proprietary but in building the ecosystem around a new architecture.

Intelligence Density as Competitive Moat

The framing PrismML has chosen is deliberate and strategic. The company's website describes its mission as building "ultra dense intelligence" and frames its competitive metric not as intelligence per model but intelligence per bit. The Bonsai 8B delivers more than ten times the intelligence density of a full-precision 8B model by that metric. This is not a claim about raw performance. It is a claim about a new unit of competition.

The implications for the structure of the AI industry are significant. The current model concentrates power at the frontier: the labs that can afford the largest training runs, the most GPUs, and the most expensive researchers control the most capable systems. Deployment is a secondary concern, solved by infrastructure at scale. What PrismML's approach suggests is a different layer of value - one that sits between the frontier labs that produce raw intelligence and the devices and systems where that intelligence ultimately needs to run.

If 1-bit compression can be made to work reliably at scale, then the control point in AI deployment shifts. Frontier labs remain relevant for training capability into models. But the companies that can compress and redeploy that capability across constrained hardware - efficiently, without performance loss - occupy a strategic position that is difficult to replicate through raw compute. Industry metrics are already reflecting this shift. The leading emerging measure in datacenter circles is tokens per watt per dollar, a ratio that rewards efficiency at least as much as capability.

Vinod Khosla's bet, and the thesis that PrismML has built its launch around, is that this layer matters enough to anchor a company. The winners in the coming era of AI deployment may not be determined by who trains the largest models. They may be determined by who can fit the most intelligence into the smallest space.

What Comes Next

PrismML is clear that 1-bit is not an endpoint. Hassibi describes it as a starting point - a new computational paradigm from which further research can build. The mathematical theory his team developed at Caltech enables compression without reasoning loss at the 8-billion parameter level. The next question the company is implicitly asking is whether that holds as scale increases, and whether the architecture advantages compound at larger model sizes.

The broader ecosystem is watching. Existing quantization research has generally treated efficiency as a postprocessing problem - something you apply to a model after it has been trained in full precision. PrismML's architecture-first approach inverts that assumption. If the approach proves durable at larger scales, it changes the calculus for how frontier labs think about training. A model designed to be 1-bit from its inception may require a fundamentally different training process, different hardware optimization, and different benchmarking standards than the 16-bit paradigm that has defined AI progress for the last decade.

The compute grants from Google and the validation from Bill Jia suggest that at least some of the hyperscalers are watching PrismML's progress with more than casual interest. Datacenter operators who can improve hardware utilization and reduce energy consumption at the model level have a direct advantage in an era where the primary scaling constraint is power. The bit wars have started, and the question of how much intelligence fits in the smallest representation may turn out to be among the most consequential in the industry.