---
title: "What an interesting week ..."
url: https://stacklist.com/card/8ef9548d-f75c-44cc-b90b-397541904f94
source_url: "https://www.linkedin.com/posts/markruddock_what-an-interesting-week-what-genuinely-share-7484045671281389569-ElNJ/?utm_source=share&utm_medium=member_ios&rcm=ACoAAAI21ZsBNnZPaKuTab7nquKLCveUW7o-1DE"
stack: https://stacklist.com/stack/614ecd6d-12cd-4195-bf9c-3def755c89b2
summary: "Mark Ruddock analyzes three structural shifts in AI: Chinese open-weight models (Kimi K3) reaching frontier performance at lower costs, TSMC identifying advanced packaging and power as the new bottleneck, and enterprise value moving from model quality to deployment and proprietary learning loops. The frontier advantage is collapsing from years to months, making workflow ownership and data control the new competitive moat."
tags: "ai-models, open-source, enterprise-strategy, compute-infrastructure, market-analysis, proprietary-data, deployment-economics"
key_entities: "Mark Ruddock (person), Satya Nadella (person), Moonshot (organization), TSMC (organization), OpenAI (organization), Anthropic (organization), Blackstone (organization), Kimi K3 (technology), CoWoS packaging (technology), open-weight models (concept), frontier models (concept), proprietary learning loop (concept), China (location)"
classification: "analysis"
content_hash: "sha256:89dc8f4c07981127e7d1c24ab6627085aad1c3749034c6a23233acbd59698799"
acp_version: "0.2"
token_counts_approximate: 977
visibility: public
agent_accessible: true
status: "final"
---

# What an interesting week ...

Mark Ruddock 1d Report this post What an interesting week ... What genuinely mattered: Three things happened this week that are structurally more important than the noise around them. First, a Chinese lab shipped what is now the largest open-weight model ever built (Moonshot's Kimi K3), and it landed proximate to the top of the world on independent leaderboards at about half the price of the closest US frontier model. Second, TSMC reported a record quarter and told us, in the plainest language it has ever used, that the binding constraint on AI is no longer GPUs — it is advanced packaging and power, with lead times now stretching into 2027. Third, Satya Nadella published an essay arguing that enterprises using frontier APIs are "paying for intelligence twice" — once in tokens, and again by handing the labs the proprietary know-how that makes those enterprises valuable. Each of these is a different face of the same shift. What surprised me: The speed of open-weight convergence. We have been saying "the gap is closing" for a year, but K3, permissively licensed and undercutting closed models by 2–3x on price, is a different statement from "closing." The frontier is no longer a moat measured in years. It is measured in months, and for many workloads, in weeks. What you should stop believing: Stop believing that model quality is where durable advantage lives. For the median enterprise use case, model quality is becoming table stakes, and it is deflating in price faster than almost any input in the history of software. The scarce resources moved down the stack to compute and up the stack to distribution, proprietary data, and the workflow you own end-to-end. What you should start paying attention to: The economics of deployment and the ownership of the learning loop. OpenAI is buying forward-deployed engineering firms. Anthropic partnered with Blackstone on implementation. Nadella is telling CIOs to guard their "exhaust" — the corrections and tool-use traces that make a model better at your job. The next trillion dollars of value is being fought over in the layer between a raw model and a business outcome, not in the model itself. If you remember only three things six months from now: - Open weights went from "credible alternative" to "co-frontier." Kimi K3 is the marker. The price per unit of intelligence is collapsing, and China is driving the deflation with permissive licensing. - The bottleneck is physical, not algorithmic. TSMC's CoWoS packaging and grid power — not model architecture — now gate how fast the industry can grow. Whoever controls compute controls the tempo. - The moat moved to workflow, data, and distribution. The value-capture battle is now about deployment and owning the proprietary learning loop, not about who has the best base model. Carpe Agentem View C2PA information 25 2 Comments Like Comment Share Copy LinkedIn Facebook X Rodrigo Madriz 1d Report this comment Running my little data normalization project across homogeneous signals I've noticed a parallel issue that seems to handicap the training of all models'. No exception. They all rely on SEO ranked results. That can quickly blur authoritative "truth" based on how many pages down they search. And based on all my testing they seldom go "beyond page 1" of results and right there you can imagine all the landmines that can tamper with social constructs. It's both fascinating since building your own proprietary thorough index of the sub-segments web you care about could create large opportunities, but on the not so good side you can already smell is how "truth" is structurally a matter of how deep your intent to find it is. Like Reply 1 Reaction 2 Reactions Val Bercovici 55m Report this comment Brilliant extraction of key signals from this week’s usual AI noise Like Reply 2 Reactions 3 Reactions See more comments To view or add a comment, sign in