Zephry

I want to talk about yesterday, particularly DeepSeek designing their own inference chip. Roughly a year into development, still pre-tapeout, in talks w/ design houses, foundries, and memory suppliers. Semis sold off hard. Nvidia's China revenue is still zero under export controls, and those same controls block DeepSeek from leading-edge foundries, which is a binding constraint for any inference ASIC since serving throughput is memory bandwidth bound, not FLOPS bound. China does have a domestic HBM path in CXMT, but it is rather degraded. CXMT is targeting HBM3 mass production by the end of 2026, reportedly still in qual with yield concerns, while SK, Samsung and Micron are already shipping HBM4 for Rubin. That is a 1-2 generation gap on the exact component that determines TPS/$. So the realistic outcome is a domestic inference chip on a trailing node with trailing memory, competing against Huawei's Ascend for China's $50B internal market. Not against Blackwell or Rubin. But I am starting to see a pattern, and I think we all saw this coming at some point. OpenAI unveiled their ASIC, Jalapeño with $AVGO. Anthropic is evaluating its own silicon, and now DeepSeek. Every frontier lab is vertically integrating because inference is the fastest growing workload and cost/tok is decided at the hardware layer! The HBM stacks, the advanced packaging, the switch silicon, and the MW all still get bought, by more buyers than before. I think the one name that needs to read this carefully is indeed Nvidia. DeepSeek itself is not the threat, but that pattern is. If every major buyer of accelerators ends up designing its own inference silicon, $NVDA keeps the training moat (breachable) but watches its customers carve up the inference TAM, and inference is where the volume growth is. Margin compression on a pie that is still expanding. Still irrelevant for the buildout.

Humanity has not prevailed ( • ᴖ • ｡) AWTF Heuristic is now over and OpenAI has completely demolished human competitors. In the end, humans performed quite well, but the performance of OpenAI was outstanding. I'll post my detailed thoughts about this in a few days, since this deserves a longer commentary.

CHINA’S GOVERNMENT HAS APPROVED PURCHASES OF 200,000 NVIDIA H200 CHIPS BY DEEPSEEK, BYTEDANCE, ALIBABA AND OTHERS FOR TRAINING PURPOSES.

Does this market have enough money to do an SK Hynix deal at reasonable prices? We couldn't even do Rivian with a couple of points of where it was... We have to be careful

Goldman Sachs: June 2026 Update Supply/Demand Compared to their May report, they have updated: > NAND and DRAM from "tight" to "very tight" supply in Q1-Q2 of 2027. > InP from "very tight" to "tight" in CY27 > T-Glass from "balanced" to "tight" in CY27. Highlights: > DRAM Pricing Surge: Like-for-like prices for DRAM are projected to skyrocket by 250–300% in 2026, with Average Selling Price (ASP) including product mix jumping 300–350%. > Memory & T-Glass Supply Relief Delayed: The supply/demand balance for DRAM/NAND and T-Glass is expected to remain "Very Tight" longer than previously expected, extending all the way through Q2 2027 before potentially easing slightly to "Tight" in Q3 2027. > InP Substrate: Like-for-like pricing for 2026 has been updated to 15–20%. > Foundry (TSMC): Stays severely constrained through 2027. > Advanced Substrates & Materials: ABF Substrate, PCB, and CCL face persistent, severe shortages. This is reflected in heavy price hikes, with CCL and ABF Substrate seeing price increases of 30–50% in 2026 and climbing even higher (45–60%) in 2027. > Optical Components: Optical Cables and Optical Devices remain severely supply-constrained through the end of 2027, maintaining steady double-digit price increases. > Global Power & Analog Semis: Shifting from "Balanced" or "Easing" in early 2026 to "Tight"by late 2026/early 2027. Note: Local China supply for these sectors is expected to remain "Balanced". > Silicon Wafers (300mm): Starting out "Easing" in early 2026, but projected to transition to "Tight" by 2027. > Local China Foundry & Equipment (SPE): Supply conditions for local Chinese foundries and Semiconductor Manufacturing Equipment (SPE) remain mostly "Balanced" with minimal price fluctuations (0–5%). > MLCC & Tantalum Powder: Enjoying an "Easing" or "Balanced" state in early 2026 before settling into a stable "Tight" market for the remainder of the forecast period.

IT'S OVER

OK this is a serious flex the image doesn't make a whole lot of sense (where is she scanning from? she doesn't point at the poster and this isn't a real camera interface) but QR replication indicates a powerful underlying model

Crazy forecast

TRUMP: IRAN CEASEFIRE AND MOU ARE OVER I THINK

Filing: Chinese AI model maker Z․ai is seeking to raise ~$4B from the sale of 19.8M shares at ~$202 to ~$216 each, after its stock jumped 1,500% since January (Bloomberg) (Visit Techmeme dot com for the link and full context!)