Microsoft’s CoPilot succeeded wildly for it’s intended purpose, but that purpose isn’t what you think it is. SemiAccurate knows it has failed for consumers, OEMs, and chipmakers so can we all finally admit it?
When the ‘AI PC’ craze was in full swing, SemiAccurate mocked the premise for several reasons. AI PCs were going to spark a refresh wave to make companies billions! Nope, didn’t happen. They are going to change the way we use PCs! Nope, didn’t happen. The wave of software coming really really soon is going to amaze you! Nope, didn’t happen. The power of the NPU will be used for other interesting things. Nope, didn’t happen. On all promised ‘benefits’, CoPilot has failed and failed miserably and cost all involved significant sums.
More on the cost in a bit, but what about the at wild success we mentioned in the first sentence? That part it easy, it is both an amazing success story and very profitable for Microsoft. Sure you will point to the numbers that show Microsoft’s CoPilot subscriptions are below miserable, no one likes it, and the forced slop is pissing off users all over the place. How is this successful? Easy, the point of CoPilot PCs has nothing to do with users or customers, it is all about doing one thing for Microsoft, lowering datacenter costs. On that front it has hit one out of the park, and we told you it would years ago.
Things did get quantifiably worse for users, that is the cost, but so what? Microsoft got what they wanted, the rest eat cake. Microsoft costs go down a lot and users get an even worse Windows 11 experience. We won’t beat the dead horse about, “Windows 11 got worse? How is that still possible?” then mix metaphors by saying the horse has left the barn barn before it died. Windows sucks and CoPilot is actively lowering the bar, but Microsoft saved a ton of money so why should they care? Do you have a practical choice when buying a new PC?
Bad as the user experience is, it does get worse. For chipmakers, they had to not only design the NPUs that no one uses, but put them in silicon. They aren’t small, eyeball math on questionable die plots says in range of about 5-10% of the total die. Tossing around hypotheticals, you could get four more cores, maybe 1/3 more GPU, or save 10%+ of the chip cost. Remember, yield is not linear with die size so savings could be significant.
In short all of these other things to do with that area are actually useful for the user. The ‘AI’ features are used rarely if at all, and are usually faster on a GPU albeit using a negligible bit more power. So you trade off less cores, a weaker GPU, or higher cost for a few minutes of battery life a week, what a win for users! As we pointed out earlier, Microsoft doesn’t give a sh*t about users, you included.
“But what about always on stuff like Recall?” I hear you bleat. Technically you are correct, if you used Recall without an NPU it would take that minutes of battery life a week and turn it into minutes of battery life lost a day. When asked about running Recall on a GPU, it works, people have hacked it to run and there is no performance penalty but chipmakers say it will drain the battery a lot. When asked to quantify that statement in percentages or time, they all come back with a likely lie about them not having measured that, and yes we did ask them directly. Quite telling.
Then there is Recall itself. It is SemiAccurate’s position that anyone who uses it is insane. Sure it CAN have benefits on occasion but the downsides are not worth the risk. First of all, the security sieve that is Microsoft software means that there is a high likelihood that malware will get onto your machine. Think about what happens if they steal the Recall database. Corporate espionage? Your golden tool awaits. And then there is the legal system, when divorce lawyers figure this one out, game over. Then again you can turn it off, right? Yup. And when the opposing counsel ask why you destroyed evidence… That is when you say you turned it off long ago, and they counter with, “prove it”. Good luck people.
So in the end, AI PCs are either a screaming success or an abject failure, which one depends on whether or not you are Microsoft. Their goal was achieved, they no longer pay for CoPilot searches, those datacenter cycles are offloaded to the users. For the users, you get quantifiably worse results, higher costs, and borderline useless hardware whose area could be used for things that actually benefit users. The software itself is mostly still absent and the slop that has come is mostly useless or worse, but could be done just as well on non-NPU blocks. Microsoft doesn’t care though, how do we know? They enforced minimum requirements for CoPilot badging in 2024, then dropped any requirement increases. Why? Their goals were met, we lost, they won. Can we just end the hype and get on with real work now?S|A
Charlie Demerjian
Latest posts by Charlie Demerjian (see all)
- AI PCs Have Failed Miserably, Can We Just Admit It? - Mar 11, 2026
- Adata Shows Off Quad-Ranked DDR5 CUDIMMs At High Speed - Feb 13, 2026
- Qualcomm Desperately Tried To Hide Snapdragon X2 Elite Power Use - Feb 11, 2026
- Cybersight Does Biking Glasses Right With Zenith - Feb 9, 2026
- Qualcomm Lost Two Key People But What Is The Impact? - Feb 3, 2026
