Without a doubt, the Big Tech is still fumbling with the AI rollout. Alphabet’s Gemini AI was so embarrassing that Elon Musk described its inner workings as “‘insane racist, anti-civilizational programming clear to all’. Microsoft’s Copilot struggles with integration, as October’s Gartner survey showed that employees overwhelmingly struggle to integrate it into their daily routines.
Most recently in January, Amazon’s AI chief Rohit Prasad told FT that AI confabulation is “still an open problem in the industry”. Apple (NASDAQ:AAPL) is no stranger to AI hallucinations as the company halted its AI-generated news alert feature mid-January.
This caused some concern, as Apple Intelligence is supposed to be one of the main premium features to draw in more customers. This was evident by the AI focus on “Pro”, starting from iPhone 15 Pro and Pro Max onward. In December’s SellCell survey, users were largely dissatisfied, with 73% saying the AI feature doesn’t add much value to the iPhone experience.
Nonetheless, AI development is rapid and likely to improve. With the released iOS 18.3 update and macOS Sequoia 15.3 on January 27th, Apple enabled AI features by default, minus the aborted news summary.
Now that DeepSeek optimizations are on the open-source table, more space is open for Apple to offer on-device AI. Although Apple is perceived to be late to the AI game, the fact is that everyone is still on the same AI confabulation boat. But it seems there was wisdom in Apple waiting for the AI dust to settle.
Apple’s Cautious AI Approach May Pay Off
While Microsoft (NASDAQ:MSFT) relies on fine-tuned OpenAI’s large language model (LLM), and Meta (NASDAQ:META) leverages open-source Llama, Apple has been working on Multimodal Large Language Models (MLLMs), starting with MM1 as pre-training research phase.
MM1 variants can reach up to 64B parameters, emphasizing image interpretation and the mix of text and image data. This makes sense, as Apple Intelligence would primarily utilize the phone’s hardware. It is also notable that Apple bought DarwinAI startup, which focuses on smaller, more cost-efficient AI models.
So far within Big Tech, Apple has spent the least on AI annually, only ~$10 billion in 2023, or about $20 billion from 2019 to 2024.
However, Apple’s spending seems to follow a more holistic approach, starting from the new M5 chip that utilizes TSMC’s cutting-edge 3nm node (N3P) process. Alongside M5, these more efficient chips will process AI tasks complementary to cloud processing, with Apple’s first server AI chip created in partnership with Broadcom (NASDAQ:AVGO). Code-named Baltra, these server chips for AI workloads should come online by 2026.
This is the hardware baseline for on-device generative AI. DeepSeek’s optimizations represent software baseline, which Apple can use free of charge. After all, DeepSeek-R1 has an MIT licence permissive for commercial use, modification, distribution and private use.
In other words, it turned out that Apple was correct to adopt a wait-and-see approach instead of going all-in like Meta, Alphabet (NASDAQ:GOOGL) and Microsoft. And while Apple can integrate DeepSeek and DeepSeek-like models on its own terms, Microsoft pays for OpenAI’s services.
Why Is On-Device AI Superior?
Even before the ChatGPT-instigated AI hype, Apple invested in the long game. This started with 2017’s A11 Bionic chip on iPhone X, specifically designed for co-processing machine learning (ML) algorithms. The aforementioned M5 is a massive upgrade from that Neural Processing Unit (NPU), aimed to enhance image recognition, on-device ML, voice recognition, natural language processing, translation and other AI-tasks.
By their very nature, such tasks are inherently best-conducted on-device. Not only is the processing faster (without server lag and congestion), but Apple can provide privacy. Presumably, as users utilize AI tasks over time, the on-device AI would improve even more, all with the privacy in mind that Apple users expect.
Moreover, if AI tasks and features have to be done real-time, it is ideal to have low-latency work independent of cloud infrastructure. Suffice to say, this on-device AI approach should give Apple a significant competitive edge as both Alphabet and Microsoft rely on their expensive cloud infrastructure.
“If anything, DeepSeek’s smaller model is more of a validation of Apple Intelligence since it will rely more heavily on a local on-device AI approach increasingly based on edge technology,”
Thomas Husson, Forrester VP principal analyst to Fortune magazine
But until Apple’s AI server chip goes into production, and until AI confabulation is near-zero, users will have to be satisfied with Apple’s partnership with OpenAI.
The Bottom Line
Unsolved technical problems with AI’s reasoning have given Apple pause, and rightfully so. But in the meantime, Apple has been paving the road for more cost-effective AI models to complement the company’s upcoming hardware.
In the meantime, although Alphabet’s smartphone share is miniscule, the company unrolled Pixel 9 phones with Google’s Tensor G4 chip, which appears to be “falling behind competitors” and Gemini being as unreliable as ever.
It is also uncertain if Alphabet can overcome its ideological capture problem, which may cripple its AI-training efforts. On the other hand, Apple’s serious on-device AI competition is likely to come from Samsung (KS:005930) and Chinese manufacturers. Huawei aims for Q1 2025 for its AI-optimized 910C chip.
Given that DeepSeek is China-born, it may be the case that Chinese phone makers end up winning the AI race long-term. This is one of the main reasons why USG ramped up export controls so drastically just before President Trump’s inauguration.
***
Neither the author, Tim Fries, nor this website, The Tokenist, provide financial advice. Please consult our website policy prior to making financial decisions.