Herr Greenrush (e/acc)

Glenn is framing this Reuters article correctly. The most important part to take-away from this article is its px as part of Atlas-350 & that ByteDance is ordering it. Since Reuters report of production # keep going up every time, I would seriously just ignore any # they provide. Remember all the stories of DeepSeek V4 coming out in a week? Where is it? There are already 7 Atlas-350 vendors during the launch last week & Reuters report its targeting shipment in 2H? This is entirely ludicrous. Now, I wouldn't compare Ascend-950PR to H200, since the former is a lower cost option using cheaper/lower end HBM designed for inference. PR - (Prefill & Recommendation). Given the lower yield of producing HBM3 type of HBM (stack 8 die on top of each other, if yield of 1 die 90% -> 8 dies yield 43%. If you layer 4 die instead, yield 66%). Take a look at current inference speed of Minimax or GLM on Macbooks using VRAM. You don't need HBM3 to run inference on these models. Reuters article also fundamentally doesn't understand what "premium version" is. Ascend-950DT (Decode & training) - hence the support for 2 TB/s interconnect speed, 4 TB/s 144GB HBM & connected together in Atlas-950 thru OXC/OCS - fully optical SuperNode that can be connected w/ 63 other ones to form 1 ZFLOP SuperCluster, enough for multiple trillion parameter training runs. There are many reason why HW is doing things this way. 950 die is clearly smaller than 910C due to lower yield of SMIC N+2 vs TSMC N7. Training is harder than inference from networking + software pov. See where optical network tech is now wrt Hisilicon & such. 8-layer HBM is harder than 4-layer HBM, so you start of by producing 4-layer HBM, which is more than sufficient for inference. Inference demand is also much higher than training due to the recent OpenClaw phenomenon. As for production #. since China is expected to have close to 100k wpm of 5-7nm capacity by EOY. Let's say 25k is used for AI chips. And you get 80 good die per wafer (+ 2 die per chip), do your own calculation. HG Tech recently raised its projection of optical transceiver demand in China from 20 million in 2026 to 20 to 30 million. That's around 1/3 of global demand. Again, do your own projections. Things shouldn't be difficult to figure out if you put your mind into it. And no, EUV is not a must. It's a good to have.

Anthropic has been testing a new model called "Mythos" with certain customers: - a "step change" in AI capabilities, including "dramatically higher scores" in coding, academic reasoning and cybersecurity - "currently far ahead of any other AI model in cyber capabilities” - part of a new "Capybara" series of models, which are larger and more intelligent than Opus - more expensive to run than Opus; not yet ready for general release

CoWoP will not arrive as quickly as GF expected. CoPoS is the more pragmatic "near-term" solution.

Wait, what? According to Reuters, the 950PR is said to come in two versions: one using DDR and the other using HBM.