Harish (@HarishMuk) profile picture

AI infrastructure. NYC. Prev: AI/ML@Apple

New York
Joined August 2021
Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)Photo by Harish (@HarishMuk)
AMD is the listed end user for a planned Pennsylvania data center campus. first building alone is 160MW, with fuel cells as primary power. chipmakers are locking up capacity for themselves now datacenterdynamics.com/en/ne…
9
A flash model is cheap to serve, so the same GPUs can push far more tokens. Question is whether Tencent is serving it on its own hardware or through US providers.
OpenRouter model usage is slated to go up 60% this week. It is almost all driven by Tencent's Hy Flash model. This is unprecedented. First, in a world that is compute constrained, how does supply increase so quickly? Second, why the sudden increase in Tencent models?
25
the 40% margin is the real variable here. route simple requests to smaller models, and price by the value of the work instead of per token. a coding agent that does days of engineering work can carry a far higher price
FT: Anthropic would need about $602B in yearly revenue to earn a 20% return on its roughly $1T in compute commitments, according to an estimate in the FT’s Unhedged newsletter. A 20% return on $1T of capital, at a 40% operating margin and 17% tax rate, requires about $602B in annual revenue, more than Alphabet’s $471B or Apple’s $478B and two-thirds of the $908B earned by all 76 software and IT services companies in the S&P 1500.
19
Most big US grid operators just filed data center hookup rules at FERC. PJM wants one 1-year review for load and generation. CAISO and MISO add new service for loads that can cut use on demand (Utility Dive). flexible data centers get the fast lane utilitydive.com/news/pjm-rto…
3
2
76
DOE just put up a $2.8B conditional loan for a 106-mile Wyoming line that doubles power flow between the eastern and western grids (WyoFile). 1.8 GW of bridge between two grids that barely talk. good start, but construction doesnt begin until 2030 wyofile.com/feds-back-4b-pro…
20
SEIA/Benchmark: US energy storage just passed 200 GWh installed. 14.3 GWh went in during Q3 alone, and Texas led the year with 12.3 GWh of utility-scale. texas isnt doing it for the climate. batteries just pencil there powermag.com/u-s-exceeds-200…
19
Texas data center load actually running went from ~4,000 MW to 5,115 MW in one month. approvals frozen since August while ERCOT restudies ~200 GW of requests. Dec 17 is the date to watch for AI power in Texas nitter.gq/VantixResearch/status/…
ERCOT just released its October report, and the most important slide is about data centers. Texas has approved 9,456MW of large loads, mostly data centers, to connect to its grid. For nine straight months, about 4,000MW of that was actually running. In October it jumped to 5,115MW, the biggest monthly increase on the chart. The other half of the chart is what matters. Total approvals barely moved all year, from 8,786MW last December to 9,456MW today. And there have been zero new approvals since August. Texas froze new large-load hookups while it re-studies a queue of roughly 200GW of requests, and the results don’t go to regulators until December 17. So the data centers already approved are ramping fast, but only about 4,300MW of approved room is left before the cap hits. Meanwhile, supply is flooding in from the other side. Batteries are now the biggest category in ERCOT’s queue at 160GW, with about 87GW scheduled for 2027 and 2028 alone. Gas projects mostly don’t show up until 2029 to 2032, which is how long new turbines take. The forward market is pricing exactly that. 2027 Texas power is trading near $41/MWh, down from about $65 a year ago. 2031 is near $58, barely changed. Cheap now, tight later. Record demand, frozen approvals, and a wave of batteries add up to one thing: the near-term Texas power trade isn’t about generators getting paid more. It’s about who gets through the approval gate on December 17, and who doesn’t need the grid’s permission at all.
1
53
Texas paused data center permits in September. first big one out the door since: 35 gas turbines, ~750 MW, for SpaceX's $24B Starbase campus. building your own power plant is turning into the default plan for AI campuses nitter.gq/EvaFox/status/21228309…
Texas has approved the air permit for SpaceX's $24 billion supercomputing data center at Starbase 🇺🇸 • The Texas Commission on Environmental Quality issued the permit on Monday, November 16. 
• It covers 35 natural gas turbines producing 750 megawatts for an off-grid power plant on the 444-acre site. 
• The data center is part of SpaceX's plan to build 10 to 20 gigawatts of AI compute capacity next year.
1
47
Frenetic pace of news cycle in the past 2 weeks > Learjet 60 > Midterms > Jensen > FAA ground stop > Starship Catch > Coming up this week: Anthropic IPO
1
107
more than 90% of china's big solar manufacturers are losing money, according to the IEA. module prices are already up 10-30% from 2025 and its still not enough. cumulative losses hit almost $5.5 billion in 2025, and 70+ companies have gone under since 2024. consolidation is coming. the risk the IEA flags: those same factories start dumping into batteries next pv-magazine-usa.com/2026/11/…
41
Everything launched this week. and its still tuesday. >Amazon: Ring Robots >Anthropic: Claude Fable 5.2 & Mythos 5.2; Claude Whiteboards, Sheets & Data + interactive connectors > Apple: $49.99 Intelligence Usage Packs >ByteDance: Seedream 5.5 >Google: Gemini 4 Carbon; TPU 8tx & 8i Max; Gemini Robotics 2.5; Gemini for healthcare >Meituan: LongCat-2.5 (open-sourced) >Meta: Muse Ember >n8n: Nod (open-source personal agent) >NVIDIA: GR00T 2 >OpenAI: Businesses in ChatGPT (Business Hub + Business Agents)
Now in beta: Businesses in ChatGPT. Claim your business, manage your information, and get insights on how people engage with you. Eligible businesses can also build a Business Agent to answer customers’ questions and help them explore what you offer. openai.com/index/introducing…
108
If you want any of these, DM me.
And electrical equipment, cooling, networking and batteries
61
1/8 I spent years building the software businesses run on (Microsoft Dynamics), then years on AI (Apple, then founding Memori Labs). Now I bring these together to put AI into real businesses. Turns out the hard part of business AI isn't the AI. It's the business. 🧵
7
1
104
8/8 AI adoption is a behavior change problem. Very few open a new tab for AI. Put it inside the habits people already have, or it dies quietly.
8
7/8 Every correction is learning opportunity for AI. Feed human edits and feedback back into prompts, rules and examples. An agent without learning is a confident new hire on day one, every day.
10
6/8 If you're not measuring it, you're guessing. Build evals from your real cases before you ship. Track accuracy, edits and escalations every week. Evals, short for evaluations, are tests for an AI system.
17
5/8 AI drafts, humans approve. Keep a person on anything touching customers or money until the error rate is zero.
7
4/8 Your data is the bottleneck. Messy records and tribal knowledge will sink any AI investment. Make one person responsible for pulling together the data, relationships and knowledge. It doesn't need to be fancy. Just legible.
7
3/8 Most AI fails at the system of record. The real business lives in the CRM and ERP. If your AI can't read from and write back to them, it's a demo, not a tool.
8