The new paradigm, Reinforcement Learning, is an inference heavy game.

And China wants more chips to play.

This has included export control violations, HBM smuggling, national investment on a scale larger than the Manhattan project, and activating the full might of Huawei: 🧵

1/7 Export controls have been effective at limiting the Chinese ecosystem's access to compute. China was expecting more than a million H20 chips, which has better inference performance than an H100. This would have enabled larger-scale model serving and faster RL progress.

2/7

Three years after the CHIPS Act pledged $52B to reboot U.S. chips, Clay, NY still looks like this—just a “Coming Soon” sign in an empty field.

Micron, Amkor, and SK Hynix are mired in NIMBY fights and two-year permits while Asia pours concrete in weeks.

1/7 Time is money, delays cost ≈ $5M per day on a $20B fab. Red tape, not lithography, is the new yield limiter. Meanwhile politicians pat themselves on the back for cardboard models.

2/7

What’s the profitable component in the Nvidia AI server supply chain?

If your answer is the “GPU”, then you are dead wrong. In fact, the most profitable component within Nvidia’s server is the rail kit made by “Kingslide”!

This is how hyperscaler procurement people describe it: “Kingslide is King”

1/4 So, what are rail kits and why are they so important to AI servers.

Server rail kits are not like the rail kit you find in your IKEA furniture. These were introduced to servers to enable much an easier maintenance procedure. Each hyperscaler has their own dimensions and requirements for the rail kit as well as a long qualification cycle for new vendors.

For each AI server, the rail kit becomes even more important as each H100 node is north of 80kg (200 lbs), significantly heavier than a CPU server.

Hence, the ASP of the AI rail kit is 10 times higher than that of a CPU server, which brings us to our next point.

2/4

TSMC has committed to invest $165b in the US. While this is margin dilutive, we see a path for TSMC to achieve 65% Gross Margins by 2030. This would continue a trend of higher margins over the past two decades.

There are three core drivers of future gross margin expansion:

- Leading-edge pricing power and higher utilization
- Advanced Packaging margin improvement
- Removal of temporary gross margin headwinds

1/5 We estimate TSMC is currently running its Taiwan wafer fabs at +60% Gross Margin. TSMC should be able to leverage its quasi-monopoly status at the leading-edge to drive ASP increases 3% above cost increases.

Pricing power and higher utilization should drive Taiwan wafer fabs to the high-60% range.

2/5

BREAKING NEWS: The upcoming MI450X IF128 could potentially break the CUDA MOAT and might even be better than NVIDIA's VR200 NVL144!

This will be AMD's first-ever rack scale architecture!

We explain the architecture details below🧵👇

1/7 In H2 2026, @AMD will release the MI450X IF128 which will be their first ever rack scale architecture.

It uses Infinity fabric protocol over ethernet for the scaleup domain to connect 128 GPU packages with over 1.8TByte/s unidi bandwidth per GPU.

2/7

DeepSeek released another model: this time, it is able to do what most reasoning models can't.

Below we explain how they did this 🧵👇
1/6 Most reasoning models can not do mathematical proofs due to their reliance on heuristics and how they are trained.

DeepSeek released a model specifically designed to do mathematical proofs, and it does that exceedingly well.
2/6

The newly launched Huawei ClusterMatrix 384 is China’s competitor to @nvidia ‘s GB200 NVL72

It networks together 5x more AI chips together than Nvidia’s flagship system

Below, we will explain the networking architecture & aspects it is competitive with Nvidia on 🧵👇
1/7 Each CloudMatrix 384 System is 16 racks

- 4 scale up networking racks
- 12 compute racks where each rack has 32 AI chips

All 384 chips are connected together with 350GByte/s high speed networking to allow for AI chips to work together on the same AI model
2/7

Despite limited success, mobile stands out as Intel's most expensive venture outside its core business.

Intel threw $25B+ on mobile-related R&D and acquisitions, cumulatively, according to SemiAnalysis’ estimates. Intel's ongoing foundry investments dwarf in front of mobile investments by a a big margin.

While Intel pulled the plug on mobile in 2019, we would like to highlight its mobile story briefly in this thread. Luckily for Intel, mobile investments happened in its heyday, having a limited impact on its core business financials. Intel's ongoing foundry investments dwarf in front of its mobile investments by a considerable margin.

Intel is not a stranger to big bets outside its core CPU business. Some include mobile processors (baseband and apps processor), Optane memory, 3D NAND, FPGA, software, AI accelerator, discrete GPU, WiMax, ADAS, wearables, drones etc.

There’s a lot of fear mongering surrounding high purity quartz (“HPQ”) and Spruce Pine, NC following the devastating flooding from Hurricane Helene
The area contains the purest form of natural quartz, but the significance of supply disruptions from the mines is exaggerated
1/8

https://twitter.com/FossilLocator/status/1840218595541778791

First - inventory levels for raw wafer companies are currently low, but even then, there’s ~3 months of DIO at Globalwafers and Siltronic and 8 months of DIO at SUMCO. Existing inventory is a buffer.
Can mining activities restart within 3 months?
Probably.
2/8

1/4 - Chevron deference has just been struck down by the US Supreme Court. What is it and how does it affect semiconductor companies? ⬇️
Under that 40-yr-old legal doctrine, US federal agencies had the power to create their own rules & regulations when a law is ambiguous. In our industry, this is particularly relevant for technology export controls – agencies were in the driver’s seat and didn’t have to worry about being challenged in Court. This is now over, and the power has been handed back to the Court system after the Supreme Court’s ruling in Loper Bridge Enterprises v Raimondo. 2/4 - The US Commerce Department’s Commerce Control List is a set of items subject to export controls and includes a plethora of Semiconductor products and manufacturing equipment. Currently over 500 pages long, items on the list are identified in exacting detail by Commerce Department officials with subject matter expertise without direct input from Congress, and free from any serious worries of court challenges. Updating the list to close loopholes and adapt to rapidly evolving technology is a perpetual game of whack-a-mole.