Lightnews — Scholar-powered news

Ruo Shui

@ruoshuiresearch.bsky.social

1.9K followers 200 following 6K posts

stay hydrated https://ruoshui.substack.com/

ruoshui.substack.com

Posts Media Videos Starter Packs

Pinned

Ruo Shui @ruoshuiresearch.bsky.social · Jul 1

The Loss of an Objective Reality

trend: chaos dominance and wealth gap + AI/GLP-1

ruoshui.substack.com

1 11

Ruo Shui @ruoshuiresearch.bsky.social · 8h

millennial Stacy Rasgons at this point

Ruo Shui @ruoshuiresearch.bsky.social · 8h

A stacked bar chart titled “Pure Foundry Market Share (Q1 2024 – Q2 2025)” shows the quarterly revenue share percentages of major semiconductor foundries. The data illustrates a clear trend of TSMC (Taiwan Semiconductor Manufacturing Company) expanding its dominance in the global foundry market through mid-2025.

Axes and Structure:
X-axis (horizontal): Quarters from Q1 2024 to Q2 2025.
Y-axis (vertical): Revenue Share (%).

Each bar represents 100% of total foundry revenue, divided among companies by color:

🔴 TSMC (red)
🔵 Samsung Foundry* (blue)
🟦 SMIC (dark blue)
🟦 UMC (light blue)
🟧 GlobalFoundries (orange)
⚫ Others (gray)

Market Share by Quarter:
Quarter TSMC Samsung Foundry SMIC UMC GlobalFoundries Others
Q1 2024 63% 11% 6% 6% 5% 9%
Q2 2024 65% 10% 6% 6% 5% 9%
Q3 2024 66% 9% 6% 5% 5% 9%
Q4 2024 69% 8% 6% 5% 5% 8%
Q1 2025 68% 9% 6% 5% 4% 8%
Q2 2025 71% 8% 5% 5% 4% 8%

TSMC’s market share rose steadily from 63% to 71%, capturing more than two-thirds of global foundry revenue by mid-2025.

Samsung Foundry declined from 11% to 8%, reflecting competitive losses in advanced-node production.

SMIC, UMC, and GlobalFoundries maintained relatively stable but small shares (4–6%).

The “Others” category remained consistent around 8–9%, representing minor regional and specialty foundries.

Note: Samsung’s figures include foundry services for its internal logic IC business.

Source: Counterpoint Research, Q2 2025.

Ruo Shui @ruoshuiresearch.bsky.social · 8h

still blame the blunt tbh

Ruo Shui @ruoshuiresearch.bsky.social · 8h

NVIDIA Blackwell Raises Bar in New InferenceMAX Benchmarks, Delivering Unmatched Performance and Efficiency

NVIDIA Blackwell swept the new SemiAnalysis InferenceMAX v1 benchmarks, delivering the highest performance and best overall efficiency. InferenceMax v1 is the first independent benchmark to measure to...

blogs.nvidia.com

Ruo Shui @ruoshuiresearch.bsky.social · 8h

InferenceMAX™: Open Source Inference Benchmarking

NVIDIA GB200 NVL72, AMD MI355X, Throughput Token per GPU, Latency Tok/s/user, Perf per Dollar, Tokens per Provisioned Megawatt, DeepSeek R1 670B, GPTOSS 120B, Llama3 70B

newsletter.semianalysis.com

Ruo Shui @ruoshuiresearch.bsky.social · 8h

GitHub - InferenceMAX/InferenceMAX

Contribute to InferenceMAX/InferenceMAX development by creating an account on GitHub.

github.com

Ruo Shui @ruoshuiresearch.bsky.social · 8h

it's open source

InferenceMAX by SemiAnalysis

A platform for monitoring and analyzing inference performance.

inferencemax.semianalysis.com

Ruo Shui @ruoshuiresearch.bsky.social · 8h

gonna bookmark this

A dark-themed scatterplot titled “Token Throughput per All-in Utility MW vs. Interactivity” compares GPU architectures based on compute efficiency and responsiveness when running Llama 3.3 70B Instruct (FP8, 1K/1K). The visualization originates from SemiAnalysis Datacenter Industry Model and highlights different GPU families’ trade-offs between throughput efficiency and latency.

Axes
Horizontal axis (x-axis): Interactivity (tokens per second per user, tok/s/user) — from 0 to 225.

Vertical axis (y-axis): Token throughput per all-in utility megawatt (tok/s/MW) — from 0 to 3.8 million.

Color Legend (GPU Architectures)
🟩 H100 — green
🟦 H200 (TRT) — lighter blue
⚪ B200 (TRT) — white
⚫ B200 — gray
🔴 MI300X — bright red
🔴 MI325X — red
🔴 MI355X — dark red

B200 (TRT) (white line) achieves the highest overall throughput, peaking near 3.8 M tok/s per MW at low interactivity levels.

H200 (TRT) (light blue) follows next, roughly 2.8 M tok/s per MW.

Standard H200 and H100 lines (dark blue and green) trail progressively lower, around 1 M tok/s per MW mid-curve.

AMD’s MI300X family (reds) shows reduced efficiency, typically under 1 M tok/s per MW, with steeper decline as interactivity rises.

At lower-latency workloads, Nvidia’s optimized TensorRT-LLM (TRT) configurations for B200 and H200 yield the best energy efficiency (more tokens per MW) and sustain higher throughput before tapering off with user concurrency. AMD’s MI series performs less efficiently, reflecting architectural and software-stack maturity differences.

Source: SemiAnalysis Datacenter Industry Model; power ratings noted per GPU (e.g., B200 2.17 kW, H200 1.73 kW, MI355X 2.65 kW).

3 2

Ruo Shui @ruoshuiresearch.bsky.social · 14h

that's consensus basically. Jensen > everything

at least he seemed visibly irritated when talking about H-1B

Ruo Shui @ruoshuiresearch.bsky.social · 14h

still needs a push though. an angry tweet. another NVDA headwind

1 1

Ruo Shui @ruoshuiresearch.bsky.social · 14h

just the usual opex expiration stuff

1 1

Ruo Shui @ruoshuiresearch.bsky.social · 14h

window next week 🙏

Ruo Shui @ruoshuiresearch.bsky.social · 15h

describes the "all them deals and market is flat on the week" sentiment

Ruo Shui @ruoshuiresearch.bsky.social · 15h

big boy MOC

Ruo Shui @ruoshuiresearch.bsky.social · 15h

don't see that...

Ruo Shui @ruoshuiresearch.bsky.social · 15h

A 3D surface chart visualizes the U.S. Treasury yield curve over time from 1990 through 2025 (Q3). The chart shows how interest rates at various maturities (from short-term to long-term Treasury securities) have evolved across economic cycles, including recessions, yield curve inversions, and monetary tightening phases.

Axes and Structure:
Horizontal axis (front to back): Time, spanning from 1990 on the left to 2025 on the right.

Horizontal axis (side to side): Treasury maturities — from 3-month bills on the left side to 30-year bonds on the far right.

Vertical axis: Yield percentage, ranging from 0% to 9%.

The chart surface is a semi-transparent mesh in gradient green tones, representing the shape of the yield curve at each point in time.

The surface peaks correspond to high-interest-rate periods (e.g., early 1990s, mid-2000s, and 2022–2023).

Valleys mark low-rate environments, notably after the 2008 financial crisis and during the COVID-19 pandemic (2020–2021).

The inversion of the yield curve—where short-term rates exceed long-term rates—is visible as dips or folds in the surface, particularly around 2000, 2007, and 2023, all historically associated with pre-recession signals.

The visualization highlights the cyclical nature of U.S. yield curves, showing alternating periods of steepness and inversion over 35 years. As of 2025, yields remain elevated and relatively flat, suggesting a persistently tight monetary policy environment following the post-pandemic inflationary surge.

Source: U.S. Treasury, FRED, NYTimes, PolicyViz; First American Calculations (Q3 2025).

Ruo Shui @ruoshuiresearch.bsky.social · 15h

SlopTok doing SlopTok things

Ruo Shui @ruoshuiresearch.bsky.social · 7d

"Slop is distinguished from spam by the fact that slop is brought into being by demand rather than supply."

John David Pressman @jdp.extropian.net · 7d

I've written an essay reflecting on how I use the word slop and some theory of what distinguishes slop from spam and kitsch.

minihf.com/posts/2025-1...

Ruo Shui @ruoshuiresearch.bsky.social · 15h

"too much" is gonna be relative

George Pearkes @peark.es · 15h

Whenever we end up building too much data center capacity…it hasn’t happened yet.

(Bloomberg) -- Microsoft Corp.'s data-center crunch will continue for longer than the company has previously outlined, underscoring the software giant's struggles to keep up with cloud demand.
Many of Microsoft's US data center regions are experiencing shortages of physical space or servers, according to people familiar with the company's internal forecasts. New subscriptions for Azure cloud services are restricted in some crucial server-farm hubs, including Northern Virginia and Texas, through the first half of next year, said the people who requested anonymity to discuss internal forecasts.

A lack of servers to rent out to customers has been a recurring concern for cloud providers in recent years. During the last six quarterly earnings calls, Microsoft has said it was unable to meet all of its customers' cloud demand. Amazon and Google have described similar constraints.
A Microsoft spokesperson said that a majority of Azure services and regions in the US "have available capacity so that existing customers with deployed workloads can continue to grow." In some cases of unplanned demand spikes, the company will introduce "capacity preservation methods" to balance customer demand across its data center fleet, the spokesperson said.

1 1

Ruo Shui @ruoshuiresearch.bsky.social · 15h

getting crowded out by OpenAI

Carl Quintanilla @carlquintanilla.bsky.social · 15h

* MICROSOFT STRUGGLING WITH LACK OF SUPPLY FOR AZURE CLOUD

$MSFT

Reposted by Ruo Shui

Christian Hubicki @chubicki.bsky.social · 20h

New humanoid dropped this morning. Figure 03.

Pretty video. Pretty robot. Very pretty demo.

How reliable? How reliable? How reliable? As far as I'm concerned that's the whole ballgame.
youtu.be/Eu5mYMavctM?...

Introducing Figure 03

YouTube video by Figure

youtu.be

3 1 5

Ruo Shui @ruoshuiresearch.bsky.social · 18h

next year isn't going to be fun

Wendy Edelberg @wendyedelberg.bsky.social · 21d

Since immigration’s effect on job growth is in the news:
@taraelizwatson.bsky.social & @stanveuger.bsky.social and I explain how breakevens are affected by current immigration policy. There’s lots of uncertainty but my preferred estimates from the paper are in this table.
www.aei.org/wp-content/u...

1 1 4

Ruo Shui @ruoshuiresearch.bsky.social · 19h

the fact that alt text makes images searchable is underappreciated

1 3

Ruo Shui @ruoshuiresearch.bsky.social · 20h

still crack up when COST helps hold up QQQ

Ruo Shui @ruoshuiresearch.bsky.social · 21h

a young boy is standing in front of a basketball hoop .

ALT: a young boy is standing in front of a basketball hoop .

media.tenor.com

Ruo Shui @ruoshuiresearch.bsky.social · 21h

Why China’s AI breakthroughs should come as no surprise

From DeepSeek to MiniMax, China’s generative AI surge wasn’t sudden - it was designed. Years of coordinated policy, tech diffusion, chip innovation and cultural optimism have built an AI ecosystem des...