Why We Were Wrong About Apple.
We bet that Apple's research would set the pace for personal, on-device AI. The trend arrived. Apple's research did not drive it, and Apple ended up renting the frontier.
In brief
We were wrong about the protagonist, not the trend. Personal AI momentum built through 2025, but it was driven by small open-weight models and rival assistants rather than Apple's research. Apple Intelligence underwhelmed, the Siri upgrade slipped to 2026, and in January 2026 Apple agreed to pay Google roughly $1 billion a year for a custom Gemini model to power Siri.
The Prediction
Strong results from Apple's on-device research accelerate momentum around personal on-device AI.
What was the original call?
The 2024 State of AI Report noted that Apple was publishing credible research on efficient on-device inference and small models, and predicted that strong results would accelerate momentum around personal on-device AI. The prior looked reasonable: two billion devices, custom silicon, a privacy posture built for local processing.
The trend half of the call was right. Personal AI and capable on-device models did gain momentum through 2025. The protagonist half failed: Apple's research was not what accelerated it.
What actually happened?
Apple Intelligence shipped in stages and landed softly. The rebuilt Siri, the product that would have proven the thesis, slipped repeatedly into 2026. Meanwhile the actual accelerants of on-device AI came from elsewhere: open-weight small models that run on consumer hardware and an NPU-equipped device cycle across the industry.
Then came the concession. In January 2026, Apple confirmed it would pay Google roughly $1 billion a year for a custom 1.2 trillion parameter Gemini model to power the new Siri, running under Apple's Private Cloud Compute so the data stays in Apple's trust boundary. The company with the strongest on-device distribution in the world chose to rent the intelligence.
The scorecard grades the prediction a miss on exactly that ground: on-device AI momentum arrived, but Apple's own research did not meaningfully drive the trend.
Why did we get it wrong?
We inferred product velocity from research output. Apple's papers were real, but publishing efficient-inference research and shipping a frontier-quality assistant are different organizational capabilities, and we scored the first as evidence of the second.
We also underweighted how much frontier scale would matter for assistant quality. The personal AI race was won by whoever had the best large model to distill from or rent, not the best on-device optimizations. Small models became genuinely useful, but the intelligence people wanted in an assistant kept living at a scale Apple had not built.
Why does the miss matter?
It is the cleanest example of a failure mode that shows up elsewhere in our track record: right trend, wrong protagonist. Like the humanoids miss, it collapses two forecasts into one, what happens and who makes it happen, and the second is much harder. The pattern is documented in the accuracy analysis.
There is also a specific lesson about Apple. Its historical strength is adopting late and integrating well, not leading research categories. Predicting Apple as the research protagonist ignored Apple's own operating history. Buying Gemini may prove an excellent product decision; our forecast about how the momentum would arrive was still wrong.
What we are watching now
Whether the Gemini-powered Siri actually ships well and moves personal AI usage on iPhone, and whether Apple's in-house models close the gap enough to bring the workload back inside.
On the trend itself, the interesting frontier is how capable small open-weight models get: releases like gpt-oss-20b already run on 16 GB consumer hardware, which keeps the original thesis alive even though the actor changed. We were wrong about Apple; on-device intelligence is doing fine.
Frequently Asked Questions
Why did Apple pay Google for Gemini?
Apple's in-house models were not ready to power the rebuilt Siri on the timeline customers expected, so Apple agreed to pay Google roughly $1 billion a year for a custom 1.2 trillion parameter Gemini model, served inside Apple's Private Cloud Compute.
Was Apple Intelligence a failure?
It shipped, but the assistant that mattered slipped to 2026 and the flagship intelligence is now rented from Google. Our miss is narrower than a verdict on Apple: the prediction claimed Apple's research would drive industry momentum, and it did not.
What actually drove on-device AI momentum?
Small open-weight models and the NPU hardware cycle. Releases like gpt-oss-20b brought near-frontier reasoning to 16 GB consumer machines, and every major chipmaker shipped dedicated AI silicon. The trend our prediction pointed at was real; the credit went elsewhere.