Relentless progress in solar

The entire community is currently experiencing a (justified) hype surrounding Claude Opus 4.5/Claude Code. However, if you listen to Dario Amodei and read his blog, you'll know that this is just the beginning and that significantly bigger leaps are to be expected very soon

Micron competitors Samsung and SK Hynix raising prices by 70% for customers like $MSFT $MU x.com/jukan05/status…

Opus 4.5 is incredibly impressive, but it's still trained and served in the previous compute paradigm. In 2026: - 1GW-class AI data centers start coming online from most frontier labs - Frontier models trained end-to-end on Blackwell/GB200 start landing On Blackwell, NVIDIA

Fascinating thought process, esp. when one factors in that US equities are ~65% of total equity market cap, & US big tech is ~40% of that number. A lot of wealthy people might find they are not as wealthy as they thought (& their lenders would need to find new collateral.) Via

after all, they have the largest cluster according to epochai

Memory Prices Surge ~50% in Q4 2025, Additional Increases Expected in H1 2026 - Memory Price Trend: The price of 64GB RDIMM, which was $255 in Q3 2025, surged to $450 in Q4 2025 and is projected to reach $700 by March 2026 - Changes in BoM Cost Structure: Rising memory prices

Thinking about the GDP impact of Innovation Platforms With disruptive technology, you get paid not once, but 4 times: 1 The technology motivates in new capital spending (as can be seen in this chart, and is we are experiencing now with datacenters). Financial capital that was

Macquarie’s calculation: - Over the next two years, the DRAM industry can only support 15GW of AI data centers. - A 1GW-scale AI data center configured with 400,000 GPUs (each with a TDP of 1,700W) requires more than 18,000 DRAM wafers per month, including both HBM and main

Very useful table of current leading open weight LLMs + closed frontier. Densely covers the ongoing Korean tournament arc. LG seems to be the clear favorite.

Chart of the Day: The dot-com boom ran on debt. The AI boom runs on balance sheets. That difference matters. Good take by Goldman Sachs.

out of sight, out of mind? there's no debt on balance sheets because they financially engineered it into SPVs. it's on pace to be the same amount of debt when .com crashed (~$1T) $112B openAI. $66B oracle. $40B microsoft. $30B meta. $20B xAI. $10B anthropic. etc

That 15M H100-equivalent number is already obsolete. Nvidia’s B300 delivers roughly 3.8x the performance of an H100. The B200 is about 2.5x. So every new Blackwell chip that ships is worth multiple H100s on this chart. 2026 US production alone is projected at 6.89 million

🚨Just In: Taiwan Semiconductor $TSM December revenue jumps over 20% Y/Y on strong demand for AI chips. TSMC's main customers include AMD $AMD, Nvidia $NVDA, Qualcomm $QCOM, Intel $INTC, Apple $AAPL, and Broadcom $AVGO.

Now this looks more like reality. Tesla, Figure and Agility together have delivered ≈3.4% of humanoid robots in 2024 and 2025. Charitably, if all «Others» are US-made*, then the US has a 13.5% market share, with China+Singapore having 86.5%. *they're not

Total compute is doubling every 7 months, and 60% are NVIDIA chips. Yes, even compute ist growing exponentially.

Given a trillion tokens of Opus 4.5 could we write a whole operating system and productivity suite from scratch. What about Opus 5? Opus 6?... Linux is about 30m lines of code/500m tokens to give you an idea..

Nvidia is so far ahead that even its 2 year old chips still outsell the rest of the industry combined.

1)18A Yield 65%-75% for laptop SOC chip 2)18A for low-end M-series processors will production in 2027 3)14A for low-end mobile A-series processors for iPhones in 2029 4)N2 Yield 70%-80% for ASIC mining chip 5)SF2 Yield less than 40% for mobile chip x.com/wallstengine/s…

thoughts after a month of daily claude code: - skill floor is low, ceiling is high - anything can be built in a day with the right prompts and .md setup. mostly everything is a skill issue now - startups in current format are cooked. mass extinction event. incentives aren't