Kilonode | AI Infrastructure @kilonode_ai

alberta holds over 90% of canada's planned ai data center capacity. the grid operator has handed out 1,200 megawatts against a queue of about 19,600, and the record peak is 12,800. one firm expects only about 28% of proposals to get built. does that queue look real?

nvidia says its pod-style data center unit can be running in 14 days. fair, but only on a site where power and water are already sorted. how long does it take to get a site ready? no vendor says. our rough estimate for power: kilonode.pages.dev/planner/?v=2

@shanaka86 fair read. 14 days only starts once the electricity is already there, and the grid gear to bring it takes years. big transformers arrive late 2028 to 2030, shortfall about 15 percent per Wood Mackenzie. how much announced capacity actually has a firm grid date vs just a permit?

you've probably seen 'brookfield and nextera to build a $100b ai hub on a cold war nuclear site'. its the paducah site in kentucky, picked by the energy department on jul 29. the power plan is gas (up to 2 gigawatts) plus batteries, not nuclear, and the target is 2032

amazon and google both just made it easier to run ai on chips that arent nvidia. here is the catch "easier switching does not create new chips. It only moves demand around the chips that already exist." full essay: kilonode.pages.dev/?v=2#journal/s…

small @dreamDEXSomnia testnet find: the bot kit order helper checks the tx hash then throws it away, so a bot cant link its own trades. i wrote my own order call to keep it, now every trade in my journal has an explorer link. reported it with 5 other issues

@nvidia unimicron's in your list; their chairman said last tuesday over 80% of chip-factory equipment lead times got longer, some tools up to 30 months. of this new capacity, how much is upgrades to lines you already run versus machines on a multi-year wait?

@nvidia neat. putting a whole rack together that fast is the easy half — the constraint is getting power and water locked in vs still in permitting, and that decides how many of these ship. how much of this pod's power and cooling is already firm vs still in planning?

read through the viral ai infrastructure posts of the last day and checked the big claims against filings and call transcripts. three needed a correction. we do our research with hark (@hark_labs). which claim fooled you this week?

@SemiAnalysis_ your crystal puller math inside a 20 GW solar plan is worth sitting with. bigger question: does that gear reach the US and get permitted in time? SpaceX AI compute was 1.4 GW, just the compute load, not the storage. how much of 20 GW survives once storage gets priced in?

anthropic's draft ipo filing reportedly says it has locked in over $600 billion of computing deals and could spend up to $750 billion if the deals still being negotiated close. when you read a big compute number, do you check how much of it is signed?

what i like most about my perps copilot is it can say no. it checks spread, price vs mark, funding and margin, then says GO, CAUTION or BLOCK. my first btc test on @dreamDEXSomnia got CAUTION bc longs were paying ~133% a year in funding. would you have taken it?

'china orders $35 billion of nvidia chips' skips what the chips are. per reuters it's over 3 million workstation graphics cards, about $11,400 each, chosen because export rules allow them. not data center chips. nvidia reportedly can fill only 5-10% of it

spacex's filing puts its ai computing at 1.7 gigawatts of power at the end of september, up from 0.6 a year earlier. thats filed. the extra 10 gigawatts in 2027 is musk saying on the call he's confident, not a promise. we track announced vs running: kilonode.pages.dev/tracker/?v=2

google's ceo didnt say cloud is sold out until 2028, whatever you saw online. asked about amazon's version of that line, sundar pichai called it reasonable and said a customer asking for 2027 compute now is 'incredibly difficult' to serve. still very tight, just not sold out

@jukan05 putting the cooling inside the chip stack helps but the heat still leaves the building. oracle's jupiter plant in new mexico uses closed-loop non-evaporative cooling, one-time fill about 2.5m gallons per building. permits and water decide what ships.

@SemiAnalysis_ this tracks with what meta confirmed. they run google's custom ai chips in production, and the open project they built together lets everyday ai code run on them. the vendor says cost per million tokens dropped about 44 percent in roughly a month on that shared software.

@SawyerMerritt the noncancelable contracts matter more than the headline number. broadcom's ceo recently said capital might be the biggest constraint on ai. so the question is whether the grid and construction can absorb it. will cloud providers have to turn away compute orders?

@elonmusk big headline, but a chip factory is only as fast as the rest of the chain. chip packaging is already so tight TSMC said it limits customers' growth, and the big three memory makers are mostly sold into 2027. the question is whether the supply chain keeps up.

built a thing for the @dreamDEXSomnia perps bounty: type a trade in plain english like "long btc 20 usdso 2x tp 2% sl 1%" and it checks the live market and tells you what could go wrong before placing it. first testnet round trip went through tonight github.com/Bonnie-debug/d…