Industry speculation regarding Apple acquiring PrismML has hit a major roadblock. Following reports from The Information that the tech giant was evaluating the startup's cutting-edge AI compression technology, Bloomberg's premier Apple analyst, Mark Gurman, has cast serious doubt on any deal actually happening.
According to Gurman, the chances of Apple buying PrismML or forming a deep partnership with them are incredibly low. This reality check comes right after the startup's bosses went on a big media tour, which likely ruined any actual talks with Apple.
The Fatal Mistake: Publicly Leaking Apple's Interest
The rumors peaked after PrismML boss Babak Hassibi went on CNBC and openly bragged that Apple was looking at his company's technology. He even said the early talks were going well.
But Bloomberg expert Mark Gurman quickly pointed out that talking to the media is a massive mistake. Apple is famous for keeping its secrets safe. Historically, if a small startup brags about talking to Apple, Apple immediately walks away from the deal. Gurman bluntly said that if Apple were actually serious about buying the company, the CEO would have kept his mouth shut.
Even without the PR slip-up, Apple has no real reason to buy PrismML. Apple already pours huge resources into its own AI training systems. In fact, the new Siri features coming in iOS 27 run on Apple’s own Foundation Models, which are already shrunk down and simplified from massive, high-end AI systems. Since Apple already has expert teams working to make models smaller and run efficiently on its own chips, buying an outside startup for the same technology is just a waste of money.
The Tech: Running 27-Billion Parameter Models on an iPhone
At the heart of the short-lived hype is PrismML's highly impressive quantization technology. The Caltech spinout specializes in extreme model compression, successfully shrinking massive neural networks so they can run directly on edge hardware. Just recently, PrismML demonstrated this prowess by compressing Alibaba's open-source Qwen 3.6 model from a massive 54GB down to a mere 3.9GB. This 1-bit, highly quantized build was able to run with all 27 billion of its parameters fully active on an iPhone 17 Pro.
The compression method simplifies how internal values are stored, squeezing 16-bit values into just one or three potential states. While this trade-off results in a minor single-digit hit to factual recall, the massive savings in memory and energy are incredibly appealing for on-device mobile AI. It explains why Apple engineers met with the firm initially, even if those talks are now dead in the water.
Read also
Key Points of the Rumor Breakdown
- Deal is Unlikely: Mark Gurman explicitly stated that Apple is highly unlikely to acquire or deeply partner with PrismML.
- NDA Backlash: Publicly speaking to CNBC about active evaluations essentially violated Apple’s unwritten rules of absolute secrecy.
- In-House Alternative: Apple is already successfully distilling its own models, such as the AFM 3 Core Advanced, reducing the need for external acquisitions.
- Proof of Concept Only: While PrismML's 27B model compression on the iPhone 17 Pro is technically impressive, Apple will likely develop similar techniques internally.
Ultimately, the brief flirtation between the two companies highlights the intense pressure Apple faces to run larger AI models locally to save on Private Cloud Compute server costs. However, those looking for a blockbuster acquisition to supercharge the iPhone's AI capabilities will have to look elsewhere. Apple is keeping its wallet closed on this one, choosing to rely on its own internal engineering to solve the on-device memory puzzle.
Loading
