Qualcomm chief executive Cristiano Amon says some leading AI companies want phones able to run models with at least 100 billion parameters continuously by 2028, Stocktwits reported. Running a model of that size continuously would demand substantial memory and create heat that could limit performance.
Key points
- Amon says AI companies want a 100-billion-parameter model to operate continuously on a phone by 2028.
- Wccftech reported his account that AI companies are building smartphones for such models.
- A 4-bit version would require at least 50GB of RAM, Wccftech estimated, before allowing memory for the operating system and other uses.
- Amon expects agents to draw on information held on phones while working alongside cloud computing.
Amon’s 2028 target for phone models
Amon described the 100-billion-parameter requirement as a request from companies at the forefront of AI, rather than a specification for a Qualcomm product. Those companies want the model available on the device continuously, he said. Wccftech reported his account that AI companies are building smartphones intended to run models of that size.
The target is larger than the on-device model support described for Qualcomm’s current premium chip. AI Affairs reported Qualcomm’s launch of the Snapdragon 8 Elite Gen 6 Extreme with support for 30-billion-parameter models. Amon’s account concerns what prospective phone makers are seeking by 2028, rather than a capability of that chip.
Amon expects software agents to become more closely connected to a phone’s operating system and applications than the standalone AI tools available on many handsets. He said an agent could use apps and services on a device to act for its owner. “There’s going to be a battle of the agents,” he said.
His argument for processing more AI work on a handset rests on the information it holds. Amon pointed to personal context, app data and sensors as inputs that could make an agent more useful. Local processing could also make agents more private, according to Stocktwits. He expects phones and cloud systems to work together on AI tasks, rather than one replacing the other.
A 100-billion-parameter model’s memory requirement
Memory is one obstacle to the proposed handset capability. Wccftech estimated that a 100-billion-parameter model compressed to 4 bits would need at least 50GB of RAM on a phone. The handset would also need memory for its operating system, applications and the model’s working context.
Wccftech described current DRAM supply and prices as constraints on adding that much memory to smartphones. It said some manufacturers had retreated from higher-RAM versions of their phones, after Android devices with 24GB of LPDDR5X memory had been available a year earlier. The publication judged a move beyond 16GB unlikely soon.
Reducing the precision of the model could lower the memory requirement. Wccftech calculated that a 2-bit version would still need 25GB solely for the model, and said compression beyond 4 bits can substantially reduce quality, particularly on reasoning tasks. It also described storing some model components in flash memory as an alternative, while warning that slower storage could reduce inference speed.
Continuous operation brings a separate heat constraint. Wccftech said sustained use of a dense model could cause a phone to limit performance as it warms. Its discussion of memory and heat places conditions on the 2028 capability Amon described: the model must fit alongside the rest of the phone’s software and keep running at a usable speed.
Amon’s case for premium smartphones
Amon linked more capable on-device AI to a possible reason for consumers to replace their phones. He characterised the handset market as driven by replacements. Stocktwits reported that memory costs were adding pressure to device prices. Those costs matter to a proposal that calls for substantially more RAM in a handset.
He said the premium and high-tier segments continue to grow as a share of the overall smartphone market. Amon expects AI capabilities to strengthen the case for an upgrade, with agents taking on tasks across the apps and data already on a phone.
