Single long-form $ALAB (Astera Labs) conviction post sourced from his aibottlenecks.app dashboard. Frame: "the most boring chip in the rack has the highest gross margin in the rack." Retimers (signal-conditioning chips between GPU / CPU / PCIe fabric / memory-expansion path) are a ~3-supplier oligopoly at PCIe Gen 6 lane rates, and ALAB is sole-sourced into NVIDIA + hyperscaler platforms. Carries demian's signature conditional "trim trigger" in place of a fixed calendar date: *trim when Broadcom lands…
@demian_ai
@demian_ai
Captured source with a living profile. Source posts are leads until researched.
Profile
Why we follow
Coverage
Captured posts
Cost per token is a vanity metric You see input tokens and output tokens but rarely do you see the actual trajectory. A modern reasoning model does not generate the answer you read, it generates: - a private monologue - a draft - a critique - a verifier check - a tool call - a
30 likes · 5 retweets · 2 replies · 2026-06-02T01:01:50.000Z
Captured by scrape-follows from @demian_ai
Je suis Nebius
59 likes · 1 retweets · 0 replies · 2026-06-01T22:36:26.000Z
Captured by scrape-follows from @demian_ai
$FLNC leading the bottleneck book on the news today yesterday’s weekly recap called the rotation into deployment: power & grid cooled short term but the bottleneck thesis stayed live, with FLNC listed right in the stack to watch. today’s brief caught NVIDIA DSX Flex turning
64 likes · 7 retweets · 3 replies · 2026-06-01T15:27:26.000Z
Captured by scrape-follows from @demian_ai
The market is no longer buying AI exposure, it's buying scarcity by layer. The book I keep is up 111.83% over the last 6 months, while the S&P 500 is up 11.02% over the same window. That leaves the book with +100.81% alpha (only tracking 112 companies rn) But the headline
184 likes · 20 retweets · 8 replies · 2026-05-29T14:00:39.000Z
Captured by scrape-follows from @demian_ai
2026 is not the year AI got smarter, it's the year inference became the product. 3 things changed in twelve months and almost nobody framed them as a single shift. 1. The first change is that reasoning got billed by the second. OpenAI shipped o3 and o4-mini with explicit "think
110 likes · 15 retweets · 5 replies · 2026-05-29T01:56:11.000Z
Captured by scrape-follows from @demian_ai
Read-through of Dell's earnings: the beat matters less than the disclosure that Dell is now physically capped by components while demand keeps exploding — exec quotes blunt that H2 is supply-constrained, not demand-constrained. demian reads Dell's stated component hierarchy as "the map" of where margin pools migrate upstream, and slaps tickers on each layer. Same physics-down move as his dashboard. Tail: "rivals fumble, Dell feasts."
built a dashboard to track ai related bottlenecks:
1389 likes · 144 retweets · 50 replies · 2026-05-07T14:53:03.000Z
Captured by scrape-follows from @demian_ai
Receipts
External posts, videos, filings, and articles captured from this source.