7 retweeted
Revenue per GPU can increase as token prices decline if per GPU token throughput increases enough.
So both Anthropic revenue *and* margins can go up as token prices go down *if* GPU prices go up less than the product of (token price decline * per GPU throughput increase).
But I have super enjoyed all the commentary to the contrary!
7 retweeted
$ANT Revenue per compute dollar rose 5.2×
The bigger story is what happens when AI moves from answering questions to completing workflows.
Latest LRMI report is up!
lrmi.ai/p/anthropics-s-1-and…
7 retweeted
Samsung has accepted Nvidia’s request to have TSMC manufacture the base die for NVHBM (Nvidia’s custom HBM based on next-gen HBM4E), a concession to the world’s biggest HBM buyer, media report. Samsung joins SK Hynix and Micron in using TSMC’s 3nm base die to ensure consistent performance and reduce validation work across suppliers.
Among the three big memory chip makers, only Samsung has historically supplied its own base die for HBM and will continue to do so for standard products and other clients. It is also validating whether it can supply its own version of the NVHBM base die in case demand exceeds TSMC’s capacity. $NVDA $TSM $SKHY $MU $SSNLF #Samsung #semiconductors en.sedaily.com/finance/2026/…
7 retweeted
Up to 8x faster, and now you know why ⚡️@openai GPT-6 Astra Ultrafast runs on NVIDIA
Read the story: nvda.ws/3W2bZ5H
7 retweeted
An NVIDIA GPU is a high-yield asset.
An H100, bought on the secondary market in Sep 2025 for ~$19k:
Resale today: ~$20.5k
Net rent earned: ~$13.5k
Total: ~$34,000, +76% in 1 year.
A B300, bought in Jul 2026 for ~$54k:
Resale today: ~$51k
Net rent earned: ~$14k
Total: ~$65,500, +21% in under 3 months
What allows these assets continue to earn? Resale value has held as rental rates have increased and demand has grown across all hardware generations.
Resale rates have held within −5% to +7%, while Ornn's H100 index is $2.91/hr, +49% from a year ago, and our new B300 index tracks at $11.32/hr, up 66% from June. H100 marketplace utilization is 87% today. B300 is at 90%.
Ornn's forward curves predicts that in 2 years, by Sep 2028:
H100 ~$63,500 (+227%)
B300 ~$186,000 (+244%)
NVIDIA Hopper and Blackwell hardware both hold incredible residual value.
7 retweeted
Taiwan’s August export orders soared 71.4% year-on-year to a fresh record high US$102.96 billion, media report, beating the 63% growth expected. AI-related demand led, but a surge in custom ASIC chip design orders caused August’s upside surprise, said W.C. Huang, director-general of Taiwan’s statistics bureau.
Orders for ICT (information and communications technology) rose 100.5% year-on-year to $34.21B, while electronics +83.9% to $45.75 billion.
Orders from the US led all others, +88.9% to $42.13B, while ASEAN +70.5% to $18.35B.
September export orders are expected to top August’s, with $106B to $109.1B the official range. Third quarter orders are seen as high as $310B and full-year 2026 orders could reach $1.1 trillion. $GOOGL $AMZN $META $MSFT $AVGO $MRVL $TSM #mediatek #GUC #Alchip #semiconductors
7 retweeted
TSMC could create a new “Moore’s Law for CPO,” doubling bandwidth every two years as it accelerates the rollout of co-packaged optics (CPO) to meet the soaring bandwidth requirements of AI data centers, according to a media report.
TSMC will focus on three areas to scale CPO:
• Per-lane speed: Increase single-channel speeds from 200G today to 400G and beyond
• Optical channel density: Expand channel counts from 16 today to 32, then 64 and beyond
• WDM (wavelength-division multiplexing): Enable each optical fiber to carry more wavelengths simultaneously
The report forecasts that doubling bandwidth every two years could take CPO bandwidth from a 3.2T baseline today to around 410T by 2040 – 128x higher. $TSM ctee.com.tw/news/20260914700…
7 retweeted
As I wrote a few weeks back on my substack, the compute paradox is occurring in real time right now.
Long live the Neoclouds.
$CRWV $NBIS $IREN $CORZ
lrmi.ai/p/the-compute-parado…
7 retweeted
The price of intelligence is collapsing. The price of compute is not.
That divergence may be one of the most important signals in AI infrastructure right now.
I wrote about what it means for neoclouds, token economics, and CoreWeave’s hidden financial convexity.
$CRWV $IREN $NBIS $CORZ $NVDA $META $MSFT
lrmi.ai/p/the-compute-parado…
7 retweeted
TSMC is raising prices 5% - 10% on all leading-edge 7nm and below chip manufacturing processes (N7/5/4/3/2 etc.), which account for ~75% of TSMC’s total revenue, reports @tculpan on Culpium, adding: “Since early this year, business development and sales teams at TSMC have been told by senior leadership that they need to find ways to charge more, according to multiple sources…” $TSM $INTC $SSNLF #Samsung #semiconductors culpium.com/p/tsmc-clients-h…
7 retweeted
75TB of likely SK Hynix $DRAM
7 retweeted
Fun fact for the day!
GE Vernova $GEV "We had more orders in Electrification in the first quarter of 2026 than we had in all of 2025." The Electrification backlog went from $9B at end of 2022 to $42B by Q1 2026.
Sure hope you have Electrification exposure to your portfolios!
7 retweeted
$NVDA's Networking Business appears to be doing quite well.
The networking business alone generated $14.9 billion in revenue at +199% year-over-year.
I believe the perception gap is not about whether NVIDIA is "winning AI." It's about whether the market has fully re-rated NVIDIA from chip company to AI infrastructure operating system, with networking, storage, and the AI factory as the new unit of computing analysis.
Networking alone could be a re-rating driver for $NVDA
7 retweeted
Samsung Electronics has started shipping next-gen HBM4e memory chip samples to clients, media report, 12-layer high-bandwidth memory 4e. The new chip offers more than a 20% speed increase over HBM4, was made using Samsung’s 1c DRAM manufacturing technology, and holds a logic base-die made with Samsung Foundry 4nm. $HXSCL $SSNLF $MU #Samsung #SKhynix #semiconductors #HBM reuters.com/world/asia-pacif…
