Macro_Lin ｜ 市场观察员

写宏观，也写行业； 写定价，也写人和公司。 也写过去十五年在市场里， 我所经历过的繁荣、泡沫、恐慌与出清。 主要关注： 宏观如何传导到资产价格 A / 港 / 美股 / 行业周期 公司历史、资本运作与市场叙事 图表、长文，以及那些还没被充分定价的变化 希望把复杂问题讲清楚，把数据背后的逻辑讲明白。 关注宏观、周期、产业与资产配置的朋友，欢迎一起交流。

@curtuner 兆易的这个赛道本来就是beta生意，净利率有上限，最多20pct左右，给不了多高的估值；再加上目前fwd P/E早就透支了

@LinQingV 国内mcu龙头是兆易，产业趋势这么好但看k线好像也就那样，筹码结构真的挺重要的。

Bernstein on 华虹：AI and auto continue to drive structural growth, price hikes materializing. MCU was the fastest-growing segment in 1Q26. BCD (PMIC) demand remains one of the highest certainties, driven by AI servers, auto, and robotics, with notable order growth from overseas (US/EU) clients adopting a "China for China" strategy. Mgmt also confirmed the ongoing price hikes in NOR Flash, expecting a 10%-15% increase driven by the memory supercycle spillover. However, consumer electronics demand remains relatively weak due to the crowding-out effect from memory price surges.

@turingou 还有个方向 real-time human-in-the-loop

实话说我有点悲观，我不知道自己还能在 harness 上做些什么，当你想控制的马车几乎能自动驾驶去任何地方，控制它这件事本身就缺乏意义。做实时音频模型的创业公司也是一样，ChatGPT 只需要把自家的实时模型接入到 codex in ChatGPT 就分分钟能完成使用几乎双工的语音通道控制你家里 mac/windows/远程 SSH linux 主机上的 codex 从而控制任何计算资源。我觉得虽然今年还没走到一半，但对 harness startup 他们的故事已经结束了…

@Gxba4aUBU7d8TvO 1. 设备自主化的进程不会断；2. 先进设备基本不太可能放开。

@LinQingV lin，如果这次谈得好，能买到先进设备那不是完全利空这些公司吗？

中微公司，北方华创先拉爆了。

Really interested into their arch of LPU-like AI ASIC. My guess is this will be a clean-sheet design. Their current Hexagon NPU is a disaster for LLM workloads.

GS on 中芯国际和华虹： SMIC - 2Q26 Rev guidance beat; ASP improvement continued – Buy. 1Q26 Revenues ofUS$2.5bn (+11% YoY/ +1% QoQ) was in-line with GSe/ Bloomberg consensus, andin-line with management’s guidance of flat QoQ. 1Q26 GM of 20.1% was higher thanour/ street’s expectations of 19.0%/ 19.4%, and at the higher end of managementguidance range (18%-20%). The sequential revenue growth was supported by flat wafershipments and an ASP increase of +3% QoQ; management attributed the GMimprovement to product mix and higher ASP. 2Q26 guidance beat: 2Q26 revenuesguided to be +14%-16% QoQ, which is higher than GSe (+6% QoQ) and Bloombergconsensus (+7% QoQ); 2Q26 GM guidance was 20%-22%, in-line with GSe andconsensus (21.6%/ 20.5%). We maintain our Buy rating on SMIC, 12m TPs of HK$134 &Rmb241.60. Allen Chang Hua Hong - 2Q26 revenues +4%~6% with GM higher QoQ – Buy. Hua Hong expects2Q26 revenues to grow by 4%~6% QoQ with GM at 14%-16% (vs. 13% in 1Q26),showing an upward GM trend despite continuous capacity expansions. The 2Q26revenue guidance mid-point was 4%/3% below GSe/Bloomberg consensus, slightlylower than expected. However, the GM guidance beat at 14%~16%, comparing to GSe at13.0% and consensus at 14.6%. 1Q26 revenues (US$661mn, +22% YoY/ +0% QoQ) camein in-line with guidance; 1Q GM (13%, flat QoQ) was within the guidance range of13%~15%, while lower than our/Street estimates of 14.0%/ 13.5%. Operating lossnarrowed to US$20mn, vs. a US$45mn loss last quarter, which management attributedto a QoQ decline in labour costs. 12m TP of HK$152. Allen Chang

🚨 Japan’s power semiconductor integration still a long way to go? #Rohm reportedly says talks with #Toshiba and #MitsubishiElectric are taking longer, as negotiations over fab integration, R&D resources, and strategic control prove complex.💡More: pse.is/93nvge 🔗

Btw, the first qualified partner for Nvidia and most ODMs/OEMs for SSTs is Delta electronics Play wisely

Absolutely wild.

@Munchi_xx 是的

@LinQingV 上一次的目标价是420，才没一个月又上修到436。

GS on 天孚通信：We are positive on TFC Optical and raise TP to Rmb436 to reflect ：(1)rising GS estimates on global optical modules TAM (Report link), and TFC as a key OE (Optical Engine) supplier, (2) incremental revenues from CPO scale-out and scale-up optical market (Report link), and(3) gradual expansion of optical module assembly business with higher content value. Following the 7% QoQ growth in 1Q26, we expect to see sequential QoQ growth ahead on improving optics chips supply and 1.6T products ramp up. We remain constructive on TFC Optical with its capabilities of offering a total solution (optical engine, FAU, ELS, optical module assembly etc.), benefiting from the growing optical module market and incremental CPO scale-out/scale-up opportunities.

@qcl92429686 是的，工程难度巨大，成本都排第二了。

@LinQingV 这玩意难度太大了，成本会巨高

$CBRS 假设 Cerebras 跟 CPO 结合起来，能否成为一台为极致推理而生的性能巨兽？ Cerebras WSE-3 的片上 SRAM 带宽是 21 PB/s，这个数字只对已经在片上的数据有效。一旦模型大到装不进单台 CS-3 的 44GB SRAM，就需要多台 CS-3 协同，activation 在机器之间流动。这段片间互联走的是 SwarmX 以太网 fabric，12 条 100GbE 链路，总带宽约 150 GB/s，跟片上 21 PB/s 差了超过十万倍。这是 Cerebras 部署 frontier model 时性能出现断崖的根本原因，也是 OpenAI 选择蒸馏小模型而不用 weight streaming 跑完整 GPT-5.3 的底层逻辑。 如果把 CPO引入 CS 系统，把光引擎直接封装到 WSE 的 package 上，片间互联带宽有望从现在的 150 GB/s 跳到几十 TB/s，提升两个数量级。电信号不用走长距离 PCB trace 再到外挂光模块，直接在芯片旁边完成电光转换，延迟更低，功耗更低，信号完整性更好。 跑一个万亿参数模型可能需要 20 到 30 台 CS 系统，权重全部常驻在各台机器的片上 SRAM 里不动，activation 通过 CPO 在机器之间高速流动。每台 CS 内部是 21 PB/s 的片上带宽处理几十层计算，跨机传一个几 MB 的 activation tensor 在几十 TB/s 的 CPO 下只需要亚微秒级延迟，基本可以被藏在计算延迟后面。系统的有效带宽会非常接近"全部在片上"的体验。 这种配置下 Cerebras 对 GPU 方案的带宽优势是碾压级的，NVIDIA 再怎么升级 HBM 也追不上 SRAM + CPO 的组合。对比 NVIDIA 刚收购的 Groq 多芯片方案也有数量级优势，Cerebras 每个节点是 44GB、21 PB/s 的整片晶圆，Groq 每个节点只有 500MB、150 TB/s 的标准芯片，跨节点通信频率差两个数量级。 工程难度非常大。在一整片 300mm 晶圆上集成 CPO 跟在常规芯片上做完全不同。光引擎的物理位置（晶圆没有传统意义上的 package 边缘）、WSE 本身 23kW 功耗旁边怎么保持激光器的温度稳定、CPO 光通道的良率怎么管理（WSE 的 compute core 可以靠冗余核补偿缺陷，光通道没有这个机制），每一个都是全新的封装工程问题。 这条路如果走通了，Cerebras 的 wafer-scale 架构就到了终极形态。片上 21 PB/s SRAM 带宽负责计算，CPO 负责多机扩展，权重常驻不动，activation 光速流转，一台专为推理而生的性能巨兽。这套系统在 decode 吞吐上可能没有理论对手。 推理是 AI 产业链里离收入最近的环节，谁的 token 更快更便宜，谁就吃到最大的商业化红利。尤其是高频交易、实时 Agentic 工作流、自动驾驶决策链这类对推理速度有确定性要求的场景，够用和极致之间的差距就是能做和不能做的区别。

Different game. Tenstorrent is going for open-source, low-cost, scale-out — RISC-V based, fully open stack, affordable hardware. Not competing on raw inference speed with Cerebras at all. Solid value proposition on TCO and accessibility, though the AI accelerator itself feels a bit middle-of-the-road — no single architectural standout. Where Jim Keller really shines is the high-performance RISC-V CPU side. That might end up being the more consequential part of Tenstorrent’s story.

@LinQingV How does it compare to Tenstorrent's offering?

China’s Yangtze Memory will submit its IPO application in China next month now that its expansion into DRAM has been greenlit by Beijing amid the Memory Supercycle, media report, adding its new Wuhan Fab (Phase 3) dedicated to LPDDR DRAM will be ready for production by end-2026. The listing is expected either on the Shanghai Star Market or in Hong Kong, valuation RMB 160 billion to 300 billion (US$23.6B to $44.2B). YMTC currently competes in the NAND Flash memory market. $SSNLF $HXSCL $MU $SNDK #Kioxia digitimes.com.tw/tech/dt/n/shwn…

你说得对，CPO 跟现有 SwarmX fabric 比，延迟改善不了一个数量级。多台 CS-3 部署 frontier model，activation 带宽需求本来就低，150 GB/s 传几 MB 的 activation 确实够用。单用户 throughput 相比单台部署会下滑不少，但 pipeline parallelism 填满之后系统总 throughput 会起来。CPO 更大的价值可能不在 activation 这条路径上，而是在 MemoryX 到 WSE 的数据通路，让权重不用全部常驻片上也能跑出合理的 decode 速度，降低多机部署的成本门槛。这条路如果走通，Cerebras 的 TAM 就不再被 44GB SRAM 框死在延迟敏感的 niche market 里，而是能进入通用 frontier model 推理市场，跟 GPU 集群正面竞争大客户的推理工作负载。

"片间互联带宽有望从现在的 150 GB/s 跳到几十 TB/s" 没用的，片间互联传播的是activation tensor，BW需求极低，即便是150GB/s也是overkill，带宽利用率就是1% 量级，batch很大的情况下好一些。所以提高到几十TB/s也没有啥用 片间互联重要的是latency，这一点方面CPO没有太多改善 cerebras standalone做推理，成本上就是个灾难，只能做niche market，高成本换高速度 等Rubin LPU起来了，这块市场会被进一步蚕食，因为能做到400 token/s而且throughput仍然还可以的程度了

Good point. MemoryX/SwarmX is real and the weight streaming I mentioned is exactly that. The issue is training vs inference. Training has large batches, high arithmetic intensity, so MemoryX streaming latency gets hidden behind compute. Inference decode is the opposite: one token at a time, reading the full weight set each time, ~2 FLOPs/byte. SwarmX physical layer is 12 × 100GbE, ~150 GB/s total. That's the hard bottleneck, and it's why OpenAI distilled Codex-Spark to fit in 44GB on-die SRAM rather than weight streaming full GPT-5.3. You're right that the endgame is likely MemoryX + CPO hybrid, not pure CPO. CPO could accelerate both activation flow between CS-3s and the MemoryX-to-WSE data path. That combination is what actually unlocks frontier model decode at speed.

Solid analysis on the CPO path, but worth flagging your missing the MemoryX/SwarmX disaggregation play Cerebras already publicly demoed last year, weight streaming from external MemoryX storage with SwarmX broadcasting to a cluster of CS-3s, weights never resident on-chip during training, activations stay local. That's the existing answer to the "model too big for 44GB SRAM" problem you're framing as unsolved. CPO on top of that would be additive (faster activation movement during inference), not the only architectural escape hatch. The trillion-param scenario described is closer to a MemoryX + CPO hybrid than a pure CPO rearchitecture.