I'm a bit confused. Aren't all these big tech companies worried about what's about to happen to their beloved AI bubble? Or I they just surrounded too much by hubris?

They do realise that one of the biggest funders of the bubble is the Gulf states and the military industrial complex.

But the Gulf states are currently on fire. They don't have very diversified economies besides oil and maybe tourism.

Due to the war, all airports and oil refineries are pretty much shut down. So no more petro dollar funding. Also, how are they, the US military going to even fuel their missiles and planes that are dropping bombs now?

Combined with the fact there is zero will for Americans to be in a war with Iran and repeat another Iraq. Nobody on the US is willing to be deployed and serve in a ground invasion force overseas. So they will have to resort to aerial bombardment alone but no war has been won that way.

Compared to Iran which are united as a people and see this as an existential threat to their homeland. They also have the geographical advantage of tunnels and have been preparing for this for decades.

I doubt Iran is going to even entertain any ceasefire and peace negotiations with Trump now after being bombed _twice_ now completely unprovoked in an effort to de-escalate their nuclear programme.

Seems like the very end of the military industrial complex, AI bubble and possibly the entire US economy as we know it. Which will likely have reverberating affects globally too.

We are with you Iran. Go forth and free us from our corrupt Epstein class regimes. :)

It's going to be really interesting seeing what Jensen Huang the pivot master does in the next 2 years.

An for a similar price, the SCHENKER KEY 18 Pro (E25) supports up to 256 GB RAM.

Quote from: AI / LLMs ? on December 18, 2025, 10:34:35This is good, because 96 GB would allow to fit and run AIs / LLM like

While it is nice and better than what the T-series allows (up to 64 GB RAM). I don't want to pay over 3000 (for the HX 370) (or 1700 for the lowest config) and be stuck with only up to 96 GB RAM (same for the P14s). If LENOVO didn't have the artificial limitation of only up to 96 GB RAM and increased it to 128 GB RAM, then I would have bought it.

But I'd still like to see up to 192 GB RAM support or the full 256 GB RAM support that this CPU/APU supports, tho:

Quote from: amd.com/en/products/processors/laptop/ryzen/ai-300-series/amd-ryzen-ai-9-hx-370.htmlMax. Memory    256 GB

(see under Connectivity)

For up to 192 GB or 256 GB RAM, right now this would mean 4 RAM stick slots (and probably slower than 5600 MT/s), as 64 GB per SODIMM slot is currently the highest available memory density.
SOCAMM2 / LPCAMM2 would also have been nice too, because under Connectivity, AMD states that LPDDR5x-8533 is supported by this CPU/APU.

I found that, while it doesn't look like it's build against MIL-STD 810G/H, the XMG EVO 15 (E25) with the same CPU supports up to 128 GB RAM and it's like half the price (and it has a 300 Hz 1440p display). Or 1/3 the price if choosing the cheapest Ryzen 7 H 255 CPU (still good, as it's a renamed 8745H), if one only cares about the 128 GB RAM (LLM inferencing speed will be the same as only 2-4 threads (2 threads is already quite close to the highest possible) are required to utilize the full memory bandwidth).
A Clevo laptop (first search result on a configuration page supports up to 128 GB RAM) could also be an option.

davidm, this is a non-AI specific business workstation laptop: If it had the Strix Halo chip (a 256-bit chip), the price would be (much?) higher, but I agree, this is maybe a missed opportunity, if the price wouldn't be much higher, but also this is not "ThinkPad P16s Gen 4 AI" edition, if it was and didn't have the Strix Halo APU chip, then I'd complain, too. (You are correct, Strix Halo would be over 2x the speed for LLMs: This is a classic 128-bit memory bus width, like over 90% out there, running at 5600 MT/s, if this was Strix Halo, it would be 2.85 times faster (2*8000/5600). But to reach 8000 MT/s, the RAM would be soldered, if no CAMM2 / LPCAMM2 would be used. CAMM2 / LPCAMM2 isn't exactly new so hopefully soon.
If people complain so much that the GPU is weak, then maybe Strix Halo is a missed opportunity, but again, it depends on the price, but for anything other than LLM inferencing, where an AMD GPU works, for finetuning, etc. a NVIDIA GPU would be needed/highly recommended (I think AMD has no real CUDA replacement so far).

"the device does get rather loud, and it runs pretty hot. Additionally, the GPU performance is weak for a workstation"

Thank you for a great summary and saving my time. Stopped reading here.

Skimmed the article for "halo," "395," nothing. It's a bit of a joke. It can take up to 96GB RAM, yet the 395+ has 128GB of much faster RAM that generally makes it up to twice as fast for many operations and effectively gives it a mid range GPU with massive fast RAM. What is Lenovo thinking, and why is nbc not pointing it out? Disappointing.

QuoteDifferent from the older P16s Gen 2 AMD, the RAM is not soldered. Instead, the Lenovo ThinkPad P16s G6 has two SO-DIMMs for up to 96 GB DDR5-5600 memory.

This is good, because 96 GB would allow to fit and run AIs / LLM like Gpt-Oss-120B quite easily (natively ~4-5-bit per weight aka Q4 quant), actually even quite larger ones quantized down to 4-bit to 5-bit per weight (below 4-bit per weight and the performance may degrade exponentially) (e.g.: if a 180B parameters LLM at full FP16 = 360 GB, then Q8 = 180 GB and Q4 = 90 GB, this may just fit) (or use a smaller LLM, but use a higher bit-per-weight quant, like Q8).
With no dGPU I'd like this laptop to stand out a bit more tho: Like supporting up to 128 GB RAM and having the CAMM2 / LPCAMM2 RAM standard, which would allow the RAM to run at a faster MT/s, still being upgradable, and give it a higher memory bandwidth: LLMs' token generation (tg) speed linearly depends on the MT/s value (5600 MT/s to 8500 MT/s = 50% faster token generation). RAM can also go bad, and if it was soldered, one couldn't replace and repair it.
But: Having no dGPU means slower prompt processing (pp) (the larger the input, the slower) (by the same ratio that a dGPU have a higher FPS value than this iGPU, so would be the prompt processing). With dGPU, partically offloacing (or if you have a small LLM, then fully) a LLM to the GPU's VRAM would also speed up the token generation (tg).

QuoteWithout a dedicated graphics chip, it simply can not compete.

Quotebad GPU performance (for a workstation)

Despite what I said above, it depends on the user' use-case. Not every workstation has to have all features (100% AdobeRGB coverage, ECC RAM, dedicated workstation GPU with ECC VRAM, ..).

Since it's a workstation, maybe the question if ECC RAM is supported could be answered. (Technically adding ECC functionality would be cheap, so it's an artificial marketing segmentation at best.)

QuoteOLED not available with Ryzen 9

This is weird. If this laptop' IPS had display P3 coverage, then I'd say OLED would only be required for gaming (due to its fast pixel response times), but this has no dGPU, so no OLED is fine and also OLED consumes more power? and would need to be 120 Hz to make really sense for gaming, but then it would also require an even faster dGPU to more likely reach the 120 Hz.
Static content, which may happen more often on a workstation, on OLED may still be an issue, idk tho, because OLED, as you say, is available on the Ryzen 5 and Ryzen 7 configurations.

I think the author meant the most powerful P16s models.

I would have bought this laptop if it had a better screen than 1200p; for a 16-inch display, that resolution is quite low.
We'll have to wait for Gen 5 to see what Lenovo does with thr powerful AMD version of this model.

Maybe in a few years, we'll see people trying to replace the panel of the full-screen assembly. (If it is possible)

QuoteThe Lenovo ThinkPad P16s Gen 4 AMD is the most powerful Lenovo laptop with a 16-inch screen

...after P1 G8 and P16 Gen 3.