Intel Forecasts Foundry Break‑Even in 2027 as 14A Node Debuts

[AleksandarK]

Intel says its struggling Foundry division will finally break even in 2027, just as its advanced 14A process comes online. The announcement came during the J.P. Morgan Global Technology, Media & Communications Conference, where CFO David Zinsner outlined the plan to offset years of quarterly losses. Intel has been pouring billions of dollars into new fabs and equipment as it races to keep pace with TSMC and Samsung. Its first milestone is the 18A process, whose first product, codenamed Panther Lake, will arrive in client PCs late in 2025, with volume production following in 2026. Intel also plans to use 18A for its next Xeon "Clearwater Forest" server chips and offer it to a small set of external partners to prove the technology's readiness. "We have to start by using our own chips," Zinsner explained. "Once we show that Panther Lake and Clearwater Forest perform as expected, we'll see more interest from outside customers for 18A, 18A‑P, and then 14A."

He acknowledged that initial adoption is low, but expressed confidence that proven performance will drive committed volume. A key part of the strategy is the use of High‑NA EUV lithography at 14A. Zinsner admitted this will raise equipment costs at first, but he believes the improved transistor density and power efficiency will more than make up for the investment, boosting margins and factory utilization. Intel's path to break even also incorporates revenue from mature nodes such as Intel 16/12, advanced packaging services, and collaborations with UMC and Tower. Under its "smart capital" model, the company will balance internal wafer demand with third‑party work, maintaining flexibility and cost discipline. With a relatively modest external revenue target (for a cash-bleeding foundry), low to mid single‑digit billions per year, Intel Foundry aims to reach break‑even in 2027 and move into sustained profitability soon afterward. By 2027, Intel 14A node and ehnaced the 1.8 nm-class 18A-P(T) node will also complement these break-even efforts.



View at TechPowerUp Main Site | Source

 

[Vayra86]

What strikes me here is that Intel once again has so much promise ahead of us, and still delivered nothing but junk.

 

[Tek-Check]

They will barely be able to deliver one single die for Panther Lake by December 31st.

If they don't have enough volume for their own chips, how on earth they would produce more chips for other companies?

It's not the case that they have magically acquired dozens of High-NA EUV machines in recent months. ASML can deliver roughly 5-6 of those annually. Even if Intel hypothetically buys all, which is of course not the case as TSMC and Micron have them delivered already, how much time will pass until Intel have enough of those delivered and assembled for volume production? A reminder, they assembled only two by October 2024. Now, this might be three only.

Intel sets up second high-NA EUV machine | Electronics Weekly

Intel has completed the assembly of its second high-NA EUV machine for its 14A and subsequent processes, says ASML CEO Christophe Fouquet,and Intel exec

www.electronicsweekly.com

 

[usiname]

Didn't they said something in those lines an year or so ago?

 

[bgx]

lots of intel haters here...

Ever heard about ramping of production?

Let s wait and see. ArrowLake was not good.
Lunar lake was good for its intended target.

Let see how panther lake is ending up being, and then we can complain (or not)!
We dont have so long to wait anyway.

 

[ncrs]

lots of intel haters here...

Ever heard about ramping of production?

Let s wait and see. ArrowLake was not good.
Lunar lake was good for its intended target.

Let see how panther lake is ending up being, and then we can complain (or not)!
We dont have so long to wait anyway.

Neither Arrow nor Lunar Lake were actually made by Intel Foundry other than the 22nm base interposers. Compute/GPU/IO/SoC tiles were all TSMC.
It's hard not to be skeptical of Intel PR when it turned out to be hollow promises so many times before.

 

[Philaphlous]

I feel like Intel has a huge opportunity yet they seem to be unable to execute for years.....new CEO might be trying to change the internal dialogue but time will tell if it works or not. Investing in Intel right now seems like a huge gamble... I can foresee their stock going to $5/share but I can also see it going to $40/share depending on what they're able to do in the near future....

They should get their Arc GPU's into the AI game.... that might be big for them?

 

[ncrs]

I feel like Intel has a huge opportunity yet they seem to be unable to execute for years.....new CEO might be trying to change the internal dialogue but time will tell if it works or not. Investing in Intel right now seems like a huge gamble... I can foresee their stock going to $5/share but I can also see it going to $40/share depending on what they're able to do in the near future....

They should get their Arc GPU's into the AI game.... that might be big for them?

Intel had AI products based on Xe - for big servers - Ponte Vecchio which was technically impressive with 63 tiles/chiplets but failed to perform adequately. Lackluster performance together with high complexity made Intel abandon that avenue and cancel its successor code-named Rialto Bridge. That is the design behind the Aurora supercomputer which was marred with problems due to it.
They also had smaller datacenter AI accelerators code-named Arctic Sound but that project didn't turn out well and the Intel Data Center GPU Flex series is based on desktop Arc dies instead. They also planned a successor called Lancaster Sound which got cancelled.
The last product is Gaudi AI chips from their Habana Labs acquisition which aren't popular either.
So you can't say that Intel didn't try, but in the end didn't score a big win with AI.

 

[Wirko]

Perhaps Intel shouldn't be buying chip grabbing equipment from that company.

 

[Squared]

Let s wait and see. ArrowLake was not good.
Lunar lake was good for its intended target.

Was Arrow Lake really not good? Arrow Lake on desktop is pretty disappointing for gaming but in laptops where it reaches clock speeds more similar to its predecessor, isn't Arrow Lake pretty good?

But Arrow Lake was originally promised for the Intel 20A node, which casts doubt over Panther Lake. Still, with Arrow Lake Intel sort of stopped talking about and being vague when mentioning it, whereas Intel is regularly insisting that Panther Lake or some client CPU is coming to 18A this year.

 

[redeye]

so go with a company that wishes for profit (intel) or go with a company that makes profit (TSMC) which one has a better chance not to raise prices?

 

[Tek-Check]

So you can't say that Intel didn't try, but in the end didn't score a big win with AI

It's not enough to try these days... execution is everything.

lots of intel haters here...
Ever heard about ramping of production?

There are no 'haters'. Nonsense.

We are analysing the situation and we need to be honest, including you.

Ramping up one thing, and ramping up on good, competitive and ready process nodes is another.

Years of delay on 10nm and then smaller nodes cost them dearly. That's why they were forced to use TSMC, to stay competitive.

They ramped up Intel 3 for entire data center. Still, it's not enough to ramp up. Zen5 EPYCs are up to 40% faster. Intel appears to be almost two generations behind with Xeons right now. How are you going to excuse that?

 

[Nhonho]

Either Intel rocks with a top lithography, or it rocks itself...

 

[Squared]

There are no 'haters'. Nonsense.

We are analysing the situation and we need to be honest, including you.

Ramping up one thing, and ramping up on good, competitive and ready process nodes is another.

Years of delay on 10nm and then smaller nodes cost them dearly. That's why they were forced to use TSMC, to stay competitive.

They ramped up Intel 3 for entire data center. Still, it's not enough to ramp up. Zen5 EPYCs are up to 40% faster. Intel appears to be almost two generations behind with Xeons right now. How are you going to excuse that?

Two generations behind? Does Intel Granite Rapids not exist? It's behind Zen 5 Epyc sure but not Zen 4 Epyc.

Welcome Back Intel Xeon 6900P Reasserts Intel Server Leadership

The Intel Xeon 6900P offers twice the cores, huge memory bandwidth and more connectivity. Welcome back Intel to the top end server CPU market

www.servethehome.com

 

[Tek-Check]

Two generations behind? Does Intel Granite Rapids not exist? It's behind Zen 5 Epyc sure but not Zen 4 Epyc.

Have you ever tried to do some math with benchmarks? On the face value, Granite Rapids is often on par with or faster than some Zen4 EPYCs systems.
So far, so good. Only two years behind, aka one generation. Now, devil is always in detail. What does it take for Granite Rapids to compete with Zen4 Genoa? Let's find out.

1. Single 9680P is 37% faster, on average, than single EPYC 9654, but... per core Xeon scores 6.50 and EPYC scores 6.34 points. Xeon is also almost three times more expensive (!), uses 25% more power and faster RAM. How does it look like now to you?

Xeon 9680P	EPYC 9654
128 cores	96 cores
500W	360-400W
RAM 6400	RAM 4800
$12,460 (Intel); was ~$17,000	$4620 (Newegg)

2. How many cores does Xeon 6980P 2P system need to achieve the parity with Zen4 EPYC 9684X 2P system? It takes 256 Xeon cores and 192 EPYC cores. The same performance, on average, with 33% more cores, 20% more power and 20% more expensive.

With CPUs, especially server chios, you need to take into account more factors, such as performance per core, per Watt, cost of rack ownership, etc.
That's why I said that current Xeons are almost two generations behind. Would you agree?

 

[Squared]

Have you ever tried to do some math with benchmarks? On the face value, Granite Rapids is often on par with or faster than some Zen4 EPYCs systems.
So far, so good. Only two years behind, aka one generation. Now, devil is always in detail. What does it take for Granite Rapids to compete with Zen4 Genoa? Let's find out.

1. Single 9680P is 37% faster, on average, than single EPYC 9654, but... per core Xeon scores 6.50 and EPYC scores 6.34 points. Xeon is also almost three times more expensive (!), uses 25% more power and faster RAM. How does it look like now to you?

Xeon 9680P	EPYC 9654
128 cores	96 cores
500W	360-400W
RAM 6400	RAM 4800
$12,460 (Intel); was ~$17,000	$4620 (Newegg)

2. How many cores does Xeon 6980P 2P system need to achieve the parity with Zen4 EPYC 9684X 2P system? It takes 256 Xeon cores and 192 EPYC cores. The same performance, on average, with 33% more cores, 20% more power and 20% more expensive.

View attachment 399820

With CPUs, especially server chios, you need to take into account more factors, such as performance per core, per Watt, cost of rack ownership, etc.
That's why I said that current Xeons are almost two generations behind. Would you agree?

1. You say the Xeon 9680P is 37% faster, has 33% more cores, and uses 25% more power. That suggests more performance per core and less power per core, suggesting Granite Rapids is actually a bit ahead of Zen 4 Epyc. (The TDP isn't entirely representative though, I think in testing Xeon's power consumption is a little worse than the TDP difference.) Prices aren't super relevant; theoretically price is set by supply and demand so it's difficult to infer where the prices could go under different market pressure. And those prices seem cherry-picked. That price for the Epyc 9654 is from a third-party seller on Newegg, "ADP Tech". Newegg charges $8,199.99. CDW charges $7,941.99 for the OEM version and $30,371.99 for the regular version. High-end server CPUs seem to be barely available at retail these days and instead nearly always being sold in OEM systems. So I think it's impossible to infer anything from these store prices. Also regarding the faster RAM, that's a generational advantage. Xeon brought support for this faster RAM to market before Epyc.

2. Regarding the comparison with the Epyc 9684X, Phoronix did benchmark both in a 2P system. In the "Geometric Mean" result the Xeon was 11.8% faster. Now 33% more cores ought to be more than 11.8% faster in multithreaded tests, but the Phoronix tests included lightly threaded tests and some outliers that suggest an optimization problem (testing a 1-year old series against a new release at the time). Moreover as a task scales up to more threads, the data synchronization overhead increases so it's reasonable to expect performance not to scale perfectly for a single task. Lastly, the Xeon did win in the less threaded workloads as well. https://www.phoronix.com/review/intel-xeon-6980p-performance/11

So no. I don't think Xeon is 2 generations behind. It was released after Zen 4 and just before Zen 5 Epyc and it performs between Zen 4 and Zen 5. Since the performance leans closer to Zen 4 and the timing closer to the of release closer to Zen 5, I'd put Xeon around 0.7 generations behind.

Edit: regarding the chart you showed, it comes from this Phoronix article, and it's inconsistent with Phoronix's test result of the same CPUs in the article I shared.

https://www.phoronix.com/review/amd-epyc-9965-9755-benchmarks/14

 

[Pizderko]

From intel 2015:

Since the early 1980s, most electronics have relied on the use of complementary metal–oxide–semiconductor (CMOS) transistors.
However, the principles of CMOS operation, involving a switchable semiconductor conductance controlled by an insulating gate, have remained largely unchanged, even as transistors are miniaturized to sizes of 10 nanometres.
We investigated what dimensionally scalable logic technology beyond CMOS could provide improvements in efficiency and performance for von Neumann architectures and enable growth in emerging computing such as artifical intelligence. Such a computing technology needs to allow progressive miniaturization, reduce switching energy, improve device interconnection and provide a complete logic and memory family.
Here we propose a scalable spintronic logic device that operates via spin–orbit transduction (the coupling of an electron’s angular momentum with its linear momentum) combined with magnetoelectric switching.
The device uses advanced quantum materials, especially correlated oxides and topological states of matter, for collective switching and detection.
We describe progress in magnetoelectric switching and spin–orbit detection of state, and show that in comparison with CMOS technology our device has superior switching energy (by a factor of 10 to 30), lower switching voltage (by a factor of 5) and enhanced logic density (by a factor of 5). In addition, its non-volatility enables ultralow standby power, which is critical to modern computing.
The properties of our device indicate that the proposed technology could enable the development of multi-generational computing.

So today is the year 2025 - Where is MESO (magneto electric spin orbit) and everything they promised?????

 

[kondamin]

1. You say the Xeon 9680P is 37% faster, has 33% more cores, and uses 25% more power. That suggests more performance per core and less power per core, suggesting Granite Rapids is actually a bit ahead of Zen 4 Epyc. (The TDP isn't entirely representative though, I think in testing Xeon's power consumption is a little worse than the TDP difference.) Prices aren't super relevant; theoretically price is set by supply and demand so it's difficult to infer where the prices could go under different market pressure. And those prices seem c

wasn't the point of granite rapids that it had a load of accelerators that drastically increased the performance of certain tasks
Something that isn't reflected in general benchmarks?

 

[Squared]

wasn't the point of granite rapids that it had a load of accelerators that drastically increased the performance of certain tasks
Something that isn't reflected in general benchmarks?

I tend to think that the point of a CPU is general purpose compute and that more focused instructions are just meant to enhance the overall general-purpose performance.

 

[kondamin]

I tend to think that the point of a CPU is general purpose compute and that more focused instructions are just meant to enhance the overall general-purpose performance.

Sure for the end user that holds true, but if you need a mail server and you can pick one that's awesome at doing that vs one that's does a decent job at everything.
I know what i would be putting my money towards for that server.

 

[Squared]

Sure for the end user that holds true, but if you need a mail server and you can pick one that's awesome at doing that vs one that's does a decent job at everything.
I know what i would be putting my money towards for that server.

Sure, but Intel wants a large share of the server CPU market. You don't get that by making a CPU that is really great at being mail server and merely alright at everything else. And Granite Rapids outperforms Zen 4 Epyc in many benchmarks, not just the mail server ones. And I haven't checked but I imagine it still loses to Zen 5 Epyc in mail server* workloads.

Edit: Certainly great mail server performance is part of Intel's plan to win sales. But I think it's a part of the plan, and not the central point of Granite Rapids.

*Mail server here being a metaphor for the tasks Granite Rapids has dedicated instructions for.