Did DeepSeek or Microsoft burst the AI bubble last week? It is actually both and SemiAccurate told you this would happen months ago.
Lets start with a bit of background. Last year Microsoft announced the AI/CoPilot PC scheme, a terrible idea that only looked good when compared to it’s horribly botched implementation. The first implementation on Qualcomm silicon flat out didn’t work, (Authors Note: I have had one since days after launch and it has never gotten past the boot loop errors, and not for a lack of trying) and didn’t sell even with a free large screen TV to bribe customers. Most of the promised software didn’t work or was delayed, users are still looking for something that doesn’t scare them, and progress is not confidence inspiring. Then Microsoft intentionally screwed over Intel and AMD, but to be fair they also screwed over Qualcomm. One big happy family there, eh?
You might recall in our first piece on AI PCs, SemiAccurate said in no uncertain terms that Microsoft doesn’t give a shit about AI, they are just trying to offload their data center costs. The plan is simple, take the rather awful ‘AI’ ‘features’, note the quotes, it is neither one much less both, and push them to a platform with far less performance. Sure the results will be quantitatively worse but, well, the user pays for them rather than Microsoft. Win/Lose but Microsoft wins and you lose so smile and take it. Then Microsoft can sell you an ‘upgrade’ to the marginally less crappy version for a fee. Win/Lose once again.
Here is where we say SemiAccurate was wrong earlier, in the piece above we said they would sell it back to you implying an _OPTIONAL_ upcharge. Silly us, Microsoft not abusing their monopoly power? We should have known better. Instead they just jacked up the price of Office 365 by *30%*, no opt out, and directly said it was to pay for AI. No one wants this garbage, even usually reliable Microsoft oriented sites are not happy. Read the comments.
So Microsoft has come clean, AI is about a 30% cost adder to their products. Those tens of billions they promise to invest have to come from somewhere, and in this case that somewhere is you. And you can’t say no but you can turn it off. For now. With quite a bit of effort. AI is, as ZDNet put it, a total disaster. Luckily there is a silver lining, instead of being called Office 365, it is now Microsoft 365 Copilot, soon to be renamed Copilot Bing Copilot 365 Copilot with AI.
Users seeing the true cost of ‘AI’ features is starting a rebellion, that is the first pin prick in the bubble. The second is what DeepSeek did to Nvidia or at least to their stock. You might recall that DeepSeek is claiming to be on par with the best that OpenAI et al can do, plus or minus a bit. Fair enough but the big bang is that they claim to have done it at a cost several digits less than what the big boys did it for. On older hardware. In less time.
Sure they are directly saying they need more and current hardware to take things to the next level, and SemiAccurate believes this. Whether they get the hardware or not is irrelevant, what they did on a purported shoestring is the point. We are aware of the caveats from being a flat out psi-op to building on what others did, but again that isn’t the point. Whatever numbers you ascribe to their cost and time, it is clear that DeepSeek did what the big boys did with a fraction of their resources.
Why does this burst the bubble? Nvidia is invincible when it comes to AI. AMD has proved to be a credible upcoming player after a few decades of trying, Intel is floundering and can’t figure out why they are lost, and several startups show promise. That said none are in the same league as Nvidia, nor does anyone, inside, outside, or on Wall Street think Nvidia has anything to worry about.
Invincibility has an upside, that is their stock. To say it is inflated is, well, do I really need to go into it? Nvidia has monopoly pricing power, abuses it to crush potential competition, and is 100% supply bound. What used to be a $500 graphics card is now a $40,000 AI card, but if you actually want one you probably have to buy a server with several GPUs for a couple of digits more. Just kidding, wait in line for a year or so if you aren’t buying but the rack. Monopolies backed by invincibility has its pricing privileges.
So the Nvidia stock keeps rising because there is no credible challenge to the dominance out there. Until DeepSeek. If they can do it, someone else can. DeepSeek may flame out or take the crown, it doesn’t matter, the fact that they DID what they did matters. Until last week, it was impossible, you needed racks and racks of multi-million dollar servers to train the next model. Now you don’t. So what else is lurking?
If DeepSeek did what they did with ‘only’ an order of magnitude less resources rather than two or three orders, does it matter? Will the next upstart do better? We are now in a scenario where there are dozens of startups, some known and some unknown, nipping at the ankles of the giant. Two weeks ago they were brushed off, not credible, now they are real and a threat. Invincibility bubble burst.
So where does that leave us. On the consumer side the costs of adding ‘AI’ to devices are pretty clear, 30% if you use Office, no opting out. Consumers don’t want it and don’t want to pay for it. Microsoft has huge investments to cover so readers like you are the goose and they want pate. Demand for the back end isn’t strong if it exists outside of a few big players either, but you have to do it on Nvidia hardware for enormous costs, right? Not any more.
It will take a while for things to trickle down to the pricing level, right now we are at the point where alternatives to Nvidia are credible, beating them can be done. This means companies pitching solutions not involving Nvidia GPUs suddenly get their phone calls returned. That is where it all starts. It will take a while to be visible to the public and 99 out of 100 players will probably fail, but those three or four that survive will profoundly change things. Two weeks ago this was impossible, today it is real. AI will continue to move on even if users don’t want it, but the hype bubble is burst and the costs are hitting home.S|A
Charlie Demerjian
Latest posts by Charlie Demerjian (see all)
- This is the NEW SemiAccurate site - Feb 16, 2025
- Why Are Intel’s Production Costs So High? - Feb 11, 2025
- Did DeepSeek Or Microsoft Pop The AI Bubble? - Jan 28, 2025
- Another Suitor Talking To Intel - Jan 21, 2025
- Sources Say Intel Is An Acquisition Target - Jan 17, 2025