AI PCs will dominate shipments by 2026, but not because of demand AI PCs are forecast to account for 43 percent of PCs by 2025, says tech market sales number cruncher and consultancy Gartner, and by 2026, an AI device may be the virtually the only laptop a big business can choose. Shipments of AI PCs are estimated to be 43 million units in 2024, an increase of 99.8 percent from 2023 and 114 …
Just as Microsoft mandated that in order to qualify for the Vista Ready sticker PCs had to have a GPU, only to then abandon the 3D effects that required it in the first place. Now on my company laptop the task manager is reporting that same GPU has been sitting at 0% utilization all morning...
I'm pretty sure every XP-capable device out there had a GPU, otherwise you wouldn't have been able to see the pretty colours and know when you were clicking on things. Heck, even servers have a GPU to be able to drive their (usually) VGA ports. Yes, it's normally a GPU built into the BMC/IPMI controller, but it's still a GPU.
I think what you're trying to get at is having a GPU with a minimum specification capable of running the OS desktop effects, which was also not new for XP - XP required an SVGA environment with a minimum monitor size of 800x600, although they recommended XGA with a minimum of 1024x768.
In fact, Vista had the same basic requirements as XP - an SVGA environment with 800x600. For Aero effects you also needed your GPU to be WDDM 1.0-compliant, have 32 bits per pixel, and have DirectX 9.0 support and Pixel Shader 2.0 support.
The difference is that the term "GPU" means that it has modern 3D processing capabilities. Prior to Vista, there were many systems that only had 2D graphics capabilities, or extremely limited 3D acceleration (1990s-class). Most 3D graphics had to be handled in software. Aero effects required hardware 3D processing at the DX9 level. XP's prettiness wasn't hardware-accelerated. Servers maintained their use of 2D graphics controllers for a very long time, and even now the Matrox G200 controller they use is just barely technically 3D-capable, designed in 1998, and wouldn't even support Aero.
That's what I have (according to lspci) in my Xeon based server, I was surprised to see the onboard graphics as Matrox G200. I used to buy those for Windows 98 and 2000 back in the day, they were great desktop cards. Nowadays, at least at the datacenter I use, those servers don't ever get a display hooked up. They are provisioned remotely through a web interface (choose from serveral linux images) and if you lose control of it, remote KVM console (through IPMI interface).
It was funny that until pretty recently, you could still see ATI graphics chips in Device Manager on some new servers. And what really sucked with those older chips like the G200e is that Windows Server 2016 and up (maybe even 2012 R2, I forget) just used the generic SVGA driver for them, so they ran at 800x600 and you couldn't increase it even for a 1280x1024 attached monitor until you installed the driver from the server vendor. They're so old Windows stopped including even a half-decent driver but they're still the most common display chip. Can't even get something designed this millennium/century/decade. There are doctors with full degrees younger than these chips.
Wonder if HP / Dell etc would get the message if their NON AI PC's keep selling well vs the more expensive AI ones with features that no-one wants / is using
Or if large customers start asking for more non AI kit as they do not want a 5/10 pct price increase for no benefit of the customer
We recently picked up some discounted Lenovo touchscreen devices that the market had obviously declined.
For anyone doing serious AI work, the processor probably the smallest part of the story: you need lots of memory and fast access (zero memcopy is nice) to it. This why MacBooks with 64GB are popular for it. But, given the price of such kit, I don't see it selling in volumes for a while.
Looks like there is a market opportunity for Raspberry Pi style desktops and laptops. I suspect with much working being done “in the cloud”, the majority of corporate desktops don’t actually need to run much more than an office suite and Team//Zoom video streams.
Okay these won’t be suitable for AutoCAD etc., but…
Stock up on non AI PCs/laptops and warehouse them, to use or sell. They will work better (speed/battery per $) than the next decades worth of AI PCs, hobbled with NPUs and hiked by inflation/Trump's tariffs. Lawsuits for AI snafus are going to be lucrative and you don't want to be on the wrong end of one.
Which is the heart of the problem with the idea that everybody will want one. It's like selling cars with retractable wings because somebody might work out a way to make them fly in the future. All the companies who've invested fortunes in AI are desperate to get that money back, even though it has no use for the average person on a local PC, and the fact that the general reaction is "I don't want it" should be a warning to those planning to make us all pay for it anyway that some companies will undercut them on price and outcompete them by not offering it.
> That killer app may not be here now, but could be around the corner
Oh, but it is here!
It just isn't an app, but a "virtual app". One that doesn't actually exist (hence virtual) but still drives the Marketing Departments to promote it.
And just as lies that the news media continually recirculates become accepted "truth", sooner or later these phantom, non-existent, apps will take on a life of their own.
The great thing about software that doesn't exist, is that it is bug-free. Though with some smart advertising and promotion, you will still need to refresh your old PCs every three to five years, just to keep up with the minimum spec. for these mythical applications.Just in case one day, somebody does actually create one.
Do remember that there is a big push to having software run in someone's cloud (to generate perpetual income). So what software running locally would need AI hardware ?
The marketing people need to work that one out.
Also: what is the extra energy requirement of this extra hardware -- even if it is not used ?
will be asking their supplier,
"How can I disable all this AI crap? We need real data not made up stuff."
"Sorry sir, Microsoft mandated that it can't be disabled."
"See that door? Go through it and don't come back."
AFAIK, most businesses don't want this shit. Imagine working on the End of Year accounts and the AI decides that your £1M profit is wrong and should be a £3M loss?
Not only will you be looking for another job but the company will be sued by investors for reporting false numbers.
IMHO, MS will make it mandatory for all PC's from the end of 2026.
But this is nonsense. Cold Reality Check follows.
Today, businesses upload tens or hundreds of thousands of expense receipts, annually. These receipts are categorised, by some combination of OCR and current GOFAI, with an appallingly high error rate. 10% or more. Then, it’s all passed back to either (non-financial) employees to check and fix, who will categorise it any damn way they get reimbursed for their own out-of-pocket expenses and DGAF. And/or, the task is passed to Mechanical Turk in the Philippines, who also have a horrendously high error rate. I’m not saying most large companies accounts are *entirely* fictitious. Ok, I am actually saying that.
If any of these AI could actually process expense claims at even a 2% error rate, it would wipe the floor with the accuracy companies achieve now.
What a strange outlook. All these businesses have to have their accounts audited, and sometimes again by the tax man. The board's collective arse is well and truly on the line if the tax man finds any irregularities. Don't know where you've worked, but everywhere I've worked has been anally strict about properly filing of expenses. I can easily see that being the norm. Can't say the same for your perspective.
Well, clearly the boards arse is *not* on the line, if they fail to notice double-recorded expense claims. Nor do auditors catch it, even when it when it happens routinely, and at scale over decades. The words “Post Office” spring to mind.
What was unusual at the Post Office, was recovering their accounting errors from the affected employees (yes, I know, not employees), using a private police force. Not the fact of the errors.
I suspect they are asking the wrong question, ie “what high-performance hardware do we need to add, for the user client to be able to do thing X”. Whereas, we should be re-visiting the age-old ping-pong tradeoff thin vs thick client, which depends on “Is CPU or connectivity the expensive bit”.
Right now, 10-100 Mbps over WiFi, real world, is almost free. These models are executing at *kilobits per second*. Even image-based AI is executing at less than one frame per second; and even if you want to hook up a real-time webcam, you’d *still* be looking at 2Mbps streams.It seems barn-door obvious that the horsepower should be server-based, whether the server is in the office or not. Particularly since these models really, really like being in streaming-mode parallelised with about 16-32 users per hardware unit.
I say “wrong question”, and yeah obviously Intel are trying to sell PC units, as being the right thing for their balance sheet. But ultimately it will be the wrong thing for their balance sheet, because other companies will offer the better technical tradeoff, which will be much better value.
"The long term trend undoubtedly points towards an onslaught of AI PCs as the inclusion of an NPU propagates down to lower-tier PCs and supply eventually finds itself in a position where producing processors without an NPU becomes cost prohibitive."
Many years ago I asked my boss why our new desktops had no CD-ROM drives, given that it was pretty much impossible to buy a PC without one at the time. He investigated. It turned out that the IT bods were paying the supplier extra to remove the drives before shipment! Can't have employees playing CDs while working.
-A.
If you buy an early AI PC, you won't have anything particularly useful to do with it, and you'll get hardware that just barely meets Microsoft's performance requirements to be called an AI PC. Then if AI stuff really does manage to find a use for common users in several years, that hardware will no longer be high-performance enough to run the tasks well since AI will have evolved significantly by then, and you'll need to replace it anyway. Or we'll go back to being able to just run it on general CPU cores. But more likely the hardware will just never get used at all and the money spent will have been wasted.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
