News Bombshell: Intel-Nvidia partnership may kill Arc GPU

[Heartbreaker]

This looks like NVidia might be Intel iGPU going forward, and maybe even setting up for NVidia to take over Intel.

Intel and NVIDIA to Jointly Develop AI Infrastructure and Personal Computing Products

newsroom.intel.com

Intel to design and manufacture custom data center and client CPUs with NVIDIA NVLink; NVIDIA to invest $5 billion in Intel common stock
SANTA CLARA, Calif. – Sept. 18, 2025 -- NVIDIA (NASDAQ: NVDA) and Intel Corporation (NASDAQ: INTC) today announced a collaboration to jointly develop multiple generations of custom datacenter and PC products that accelerate applications and workloads across hyperscale, enterprise and consumer markets.

The companies will focus on seamlessly connecting NVIDIA and Intel architectures using NVIDIA NVLink – integrating the strengths of NVIDIA’s AI and accelerated computing with Intel’s leading CPU technologies and x86 ecosystem to deliver cutting-edge solutions for customers.

For data centers, Intel will build NVIDIA-custom x86 CPUs that NVIDIA will integrate into its AI infrastructure platforms and offer to the market.

For personal computing, Intel will build and offer to the market x86 system-on-chips (SOCs) that integrate NVIDIA RTX GPU chiplets. These new x86 RTX SOCs will power a wide range of PCs that demand integration of world-class CPUs and GPUs.
.

NVIDIA will invest $5 billion in Intel’s common stock at a purchase price of $23.28 per share. The investment is subject to customary closing conditions, including required regulatory approvals.

 

[ToTTenTranz]

IIRC people said the same thing about Kaby Lake G bringing the end of Intel graphics.

According to Ian Cutress, this is Nvidia making iGPU chiplets that connect to Intel IODs through NVLink.
It isn't Intel taking GPU IP from Nvidia to design a monolithic SoC, so tech is still pretty separate. It's like a Kaby Lake G.


The only thing I get from this deal is Nvidia got really scared of Medusa.

 

[Heartbreaker]

According to Ian Cutress, this is Nvidia making iGPU chiplets that connect to Intel IODs through NVLink.
It isn't Intel taking GPU IP from Nvidia to design a monolithic SoC, so tech is still pretty separate. It's like a Kaby Lake G.

This is MUCH bigger than Kaby-G, and technically nothing like Kaby-G. Kaby G was just a normal Radeon dGPU with it's own VRAM, connected via PCIe, basically all standard components.

This specifically says "chiplets", so it will be more like the GPU chiplet in new Sparc chip, that has a 600 GB/s C2C interface, so it can just share system RAM instead of requiring it's own VRAM. Basically it will let them create an Intel/NVidia version of Strix Halo.





Plus they are planning to "jointly develop multiple generations of custom datacenter and PC products".

Plus that $5Billion dollar investment.

This is NOTHING like Kaby-G.

 

[Saylick]

On the PC side, this will mostly be relegated to APU-style chips. This type of product isn't going to jive with the demographic that wants to be able to upgrade their CPU and GPU independently. The datacenter collaboration is more interesting with echos of Nvidia partnering with IBM in the past to integrate NVLink into the CPU, but IBM was largely not competitive since it wasn't x86-based. Since x86 is still a big portion of the datacenter market, I have to wonder if there's now a anti-competition case to be made since Nvidia GPUs are in such high demand that a forced bundling of an Intel CPU with it would crowd out the competition.

 

[adroc_thurston]

The datacenter collaboration is more interesting with echos of Nvidia partnering with IBM in the past to integrate NVLink into the CPU, but IBM was largely not competitive since it wasn't x86-based. Since x86 is still a big portion of the datacenter market, I have to wonder if there's now a anti-competition case to be made since Nvidia GPUs are in such high demand that a forced bundling of an Intel CPU with it would crowd out the competition.

it's not anti-competitive since AMD already beheaded Xeon and paraded its corpse around for years.

 

[adroc_thurston]

Also yeah, kids, Tegras SUCK.
They SUCK so MUCH.
Never Tegra.

 

[Saylick]

it's not anti-competitive since AMD already beheaded Xeon and paraded its corpse around for years.

Not sure if sarcastic, but if you're being serious, AMD's technical lead over Intel has historically never stopped their competition from using anti-competitive tactics to win out the market.

 

[adroc_thurston]

Not sure if sarcastic, but if you're being serious, AMD's technical lead over Intel has historically never stopped their competition from using anti-competitive tactics to win out the market.

Intel does not have the cash for anything anti-competitive.
And NV doesn't have the roadmap. It is funny.

 

[Saylick]

Intel does not have the cash for anything anti-competitive.
And NV doesn't have the roadmap. It is funny.

And if this Intel-Nvidia deal keeps up, Intel now has the cash and Nvidia has a CPU roadmap. It is funny how it works.

 

[branch_suggestion]

And if this Intel-Nvidia deal keeps up, Intel now has the cash and Nvidia has a CPU roadmap. It is funny how it works.

Yes, if.
If NV have a reputation for anything, it is being a terribad partner.

 

[adroc_thurston]

Intel now has the cash

They don't, $5B is pocket change given their capex.

and Nvidia has a CPU roadmap

Thing is, they already had one.
Spent years shrieking that aa64 is all you need. Now this? Seriously?

 

[Saylick]

Yes, if.
If NV have a reputation for anything, it is being a terribad partner.

Won't disagree with you here. Nvidia are known to be difficult to work with and will deflect blame and throw their partners under the bus when crap hits the fan.

They don't, $5B is pocket change given their capex.

Don't disagree.

Thing is, they already had one.
Spent years shrieking that aa64 is all you need. Now this? Seriously?

Again, they just investing everywhere because it's diversification. They literally have a mountain of money to spend. Allocating 5B here is peanuts compared to the 60B they allocated for buybacks last month.

 

[branch_suggestion]

Won't disagree with you here. Nvidia are known to be difficult to work with and will deflect blame and throw their partners under the bus when crap hits the fan.

Yeah they are currently doing it with MTK when everyone knows NV are to blame for N1X issues.
Tegra kids, not even once.

 

[adroc_thurston]

Again, they just investing everywhere because it's diversification. They literally have a mountain of money to spend. Allocating 5B here is peanuts compared to the 60B they allocated for buybacks last month.

Remember this isn't 5B, it's also a roadmap commitment.

 

[coercitiv]

The only thing I get from this deal is Nvidia got really scared of Medusa.

The way I read this deal is Nvidia got "asked" to contribute to Intel and they named their price: Nvidia GPU IP in Intel's mobile lineup. They essentially get a backdoor into the iGPU market without having to come up with a successful APU of their own. (which they'll continue to do anyway on the ARM side, betting on two horses is better than one)

 

[ToTTenTranz]

This is MUCH bigger than Kaby-G, and technically nothing like Kaby-G. Kaby G was just a normal Radeon dGPU with it's own VRAM, connected via PCIe, basically all standard components.

This specifically says "chiplets", so it will be more like the GPU chiplet in new Sparc chip, that has a 600 GB/s C2C interface, so it can just share system RAM instead of requiring it's own VRAM. Basically it will let them create an Intel/NVidia version of Strix Halo.

Given Intel's current architectures for consumer, I'd say it's more likely that the iGPU and CPU tiles become clients to the memory controller in the SoC tile.
Regardless, I was talking about the business / IP arrangement being closer to Kaby Lake G. Nvidia isn't sharing GPU IP with Intel, they're probably sending their GPU tiles for Intel to assemble and pack the full SoCs. Same thing happened with AMD sending them that Polaris/Vega hybrid chip.

Spent years shrieking that aa64 is all you need. Now this? Seriously?

Turns out running win32 apps compiled for x86 on ARM fast enough for a high-end windows experience wasn't going to happen anytime soon. Again.

The way I read this deal is Nvidia got "asked" to contribute to Intel

Erm... username checks out?

 

[coercitiv]

Erm... username checks out?

When it comes to Nvidia I always assume the most aggressive stance for a given move. My username came from a rather random event, one could say I was forced to use it I'll say this though: both my username and preference Philip K. Dick's novel (related to the avatar) are a scary fit for the times we live in.

 

[Heartbreaker]

Regardless, I was talking about the business / IP arrangement being closer to Kaby Lake G. Nvidia isn't sharing GPU IP with Intel, they're probably sending their GPU tiles for Intel to assemble and pack the full SoCs. Same thing happened with AMD sending them that Polaris/Vega hybrid chip.

The only commonality is that they aren't transferring GPU IP. Why would they?

But this is a broad multifaceted partnership, rather than just buying one chip, and again Kaby-G was no more integrated, than any third party GPU chips are integrated with the CPU in any laptop. Kaby-G was a sad joke.

NVidia is supposed to get Custom CPU flowing in their direction as well, plus the collaboration is multigeneration.

With chips flowing both ways, nearly any imaginable configuration of CPU-GPU/AI integration is possible.

 

[ToTTenTranz]

and again Kaby-G was no more integrated, than any third party GPU chips are integrated with the CPU in any laptop.

This is just not true. Kaby Lake G had the Radeon GPU, HBM2 and Intel CPU all on the same MCM. The Vega M was exclusive to Intel, and Intel were the ones supplying the driver updates.

NVidia is supposed to get Custom CPU flowing in their direction as well, plus the collaboration is multigeneration.

With chips flowing both ways, nearly any imaginable configuration of CPU-GPU/AI integration is possible.

You're assuming a whole lot out of Nvidia purchasing <5% of Intel stock and a pinky promise of future collaboration that is at least 2 years away (though more probably 3 or 4).

 

[Heartbreaker]

This is just not true. Kaby Lake G had the Radeon GPU, HBM2 and Intel CPU all on the same MCM. The Vega M was exclusive to Intel, and Intel were the ones supplying the driver updates.

It was a standard circuit board using PCIe 3.0 X8 bus, connected over copper traces, between CPU and GPU. They could have place them separately on the laptop MB, or AMD could have used the GPU without one of their own CPU as it was just a standard copper PCIe bus connecting them. Yes it was made exclusively for Intel but it was folly and recognized as such pretty much from the start.

You're assuming a whole lot out of Nvidia purchasing <5% of Intel stock and a pinky promise of future collaboration that is at least 2 years away (though more probably 3 or 4).

You are free to just disbelieve everything they said... It looks rather big to me.

 

[marees]

I believe this partnership effectively kills of all low end discrete laptop GPUs

Desktop GPUs can reuse tiles (like AT3 & AT4)

 

[coercitiv]

Ian Cutress weighs in on the subject:

Kaby-G wasn't a partnership. It was Intel commissioning a chip from AMD, buying it, then packaging it as its own product. It was an Intel product - Intel messaged, Intel sold.

We can't be sure of the final scope of this partnership, but the original framing/setup is clearly different.

 

[igor_kavinski]

Now all Intel needs is a partnership with ARM to license a tiny chiplet of x86 cores so ARM designs become perfectly x86 compatible.

 

[Keller_TT]

There are some contradictory statements here and those trying to downplay it.
I see it as a rather low hanging fruit that benefits both Intel and NV. WoA, even if it has matured, hasn't shifted customer confidence from leaving Wintel. NV can just invest pocket cash knowing that AMD hasn't really dented Intel's PC share even with Intel at its weakest (as I had predicted). AMD laptop share has actually shrunk QoQ in 2025. It would also satisfy NV GPGPU clientele for DCs wanting their trusted x86 platform.

The latest figures for Windows PC share has Intel owning 76%, with Strix barely making a splash. A price competitive Intel + NV will trounce AMD's greed for mainstream "Copilot+ AI PC" paper tiger parts on top of rather poor availability. I think Intel with NV chiplets will flood the market than wondering where the heck is Strix Halo.

Intel is not as hopeless as AMD was with Bulldozer before Zen, and really only Zen 3 when they edged ahead. Intel has pretty good cores with more to come. The ?? has been about Intel's grandeur fab ambitions not delivering. If 18A can stay in touch with TSMC, this will put it back on track.

I don't care a hoot for the products though to be honest. I'm fully committed into Tenstorrent and RISC-V for my own needs. Will shortly take delivery of a QuietBox with Blackhole chips. I am available to join the porting work necessary for Mojo to work with TT's MLIR. But I did invest on 2xRadeon Pro W7900 48GB cards for my GenAI 3D passion project. Came in at under $8000 even after tax, and definitely worth it.