And there are other FPGA chips with even bigger price gaps — some are literally 10x more expensive in the US and Europe compared to China.
There's a bunch of reasons for this. Besides what I already mentioned in Dylan's comment, the biggest thing is, it's basically just how the market works — AMD has different sales strategies across different regions.
@dylan522p first off, this 47dr1156 chip isn't even $1,000 in China right now — it's more like $800. Yeah, it did hit $1,000 at one point last year, but that was the peak. Second, this thing isn't mainly for military use — it's mostly civilian/commercial stuff. And honestly, you can't
first off, this 47dr1156 chip isn't even $1,000 in China right now — it's more like $800. Yeah, it did hit $1,000 at one point last year, but that was the peak. Second, this thing isn't mainly for military use — it's mostly civilian/commercial stuff. And honestly, you can't really blame AMD for the price difference. Unlike AI chips where the manufacturer sells straight to customers, FPGAs go through a ton of distributors and middlemen, so it's not all on AMD.
Kimi K3 is now officially available on Bedrock. Bedrock’s self-hosted model adopts a tiered commission structure — the higher the tuning level, the larger the share.
China’s open-source models are gaining growing global recognition, and Kimi’s ARR will keep rising.
docs.aws.amazon.com/bedrock/latest…
Citi and Morgan Stanley's Analysis of CPO's Impact on CCL
Citi: Concerns over CPO-driven material downgrades are overstated: current sample boards still use M8-class CCL with approximately 40–50 layers. Citi expects gradual, selective downgrades rather than a wholesale shift to M4. Lower material grades do not necessarily reduce PCB value, as more complex power delivery, routing and manufacturing requirements continue to support value-added content.
Morgan Stanley: CPO presents a credible long-term risk to high-end CCL demand, but its impact on the 2026–2028 demand outlook should remain limited. Cost, yield, reliability, serviceability and thermal challenges constrain adoption, particularly in scale-up architectures. Morgan Stanley therefore maintains its high-end CCL supply-tightness thesis and views CPO-driven pullbacks as potential accumulation opportunities.
Two weeks.
That's how long it took to go from GLM-5.3-Flash's first run on domestic accelerators to serving all of its production traffic, with 3.2× end-to-end throughput along the way.
What I keep thinking about is who did much of the work: an Infra Agent powered by GLM-5.3.
Computer Use could become a major driver of demand for multimodal AI. It requires a model to continuously observe the screen, take action, and interpret the updated interface to decide what to do next—effectively sampling frames throughout the process. A long-running task can therefore involve processing a large volume of high-resolution screenshots.
There is a practical reason for this approach: many applications still lack comprehensive command-line interfaces or APIs, and plenty of operations remain difficult to perform through those interfaces. Using a keyboard and mouse to interact with a graphical interface remains the most universal fallback in the digital world.
In that sense, keyboard-and-mouse interaction is the digital equivalent of a humanoid robot. Just as physical tools and environments were designed around the human body, software has long been built around human clicks, typing, and visual feedback. By adopting the same interaction patterns, AI can use existing software without waiting for every product to be redesigned for it.
The hardest part, however, is more than recognizing an interface or locating a button. It is maintaining a consistent goal across a long sequence of actions, understanding the outcome of each step, and recovering when something goes wrong. If every step has a 90% success rate and the steps are independent, a task requiring 40 consecutive successful steps has an overall success rate of only about 1.5%.
The core bottleneck is therefore the capability of the underlying foundation model, particularly its ability to reason and execute across long sequences. Multimodal understanding allows the model to interpret the current state; strong reasoning helps it determine what to do next. Continuous feedback and error correction determine whether it can actually finish the task.
I had my “touch of AGI” moment yesterday.
I recorded a video of myself talking through a problem: how to improve maintenance of tangled hair in the shower drain.
It was a ~4 min video where I explained the problem and also used a digital caliper to take measurements.
I asked
185 Followers 7K FollowingFounder of Ganzo: Multi-Agents with local, remote, and distributed inference on a secure, anonymous, safe, private, uncensored Agentic Blockchain @ganzoai
161K Followers 268 FollowingThe AI Lab behind GLM models, dedicated to inspiring the development of AGI to benefit humanity.
https://t.co/gOw7WpwkJt
https://t.co/ot9sGVXU8x
1.8M Followers 7K FollowingHost: @twistartups @theallinpod @thisweeknai; I invest in 100 startups a year @launch & @founderuni [email protected] for life
1.1M Followers 63 FollowingIt's time to build.
https://t.co/A9eTFq6Xbx
Posts are not investment advice or an advertisement for investment services. See https://t.co/nX2FtaLE06.
1.0M Followers 215 FollowingOnly on X, don’t trust fake accs
AI/Semi Supply Chains Research
Nothing is investment advice. No paid promos; may trade/hold names disc, views my own.
5.4M Followers 4 FollowingOpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6LgzPA
7.1M Followers 1K FollowingBig stories and breaking news as they are published on https://t.co/EYmAcRLBHv. Register here to access free articles: https://t.co/NRg2hritkA
163K Followers 1K FollowingSemiAnalysis
AI Infrastructure Research and Consulting
DMs are open for consulting, quotes, or to talk shop,
Opinions my own
55K Followers 832 FollowingResearch 📃 & News 🗞️ | Semiconductors & Tech | Sharing my knowledge & insights with the world 🌍 | Deep dives, research, discord in my substack!📍| PMs Open 👇