7213Ejaaz

2026 will be known as the Great American Manufacturing Revolution 2.0
DATA CENTERS
- colossus II (2GW) by elon
- Stargate clusters by sama
- fairwater by microsoft
- hyperion by meta
CHIP FABS:
- samsung (to build tesla ai5 chips, hbm and more)
- SK HYNIX (to build HBM and encroach on tsmc monopoly)
- tsmc ( their first site outside of taiwan)
- intel (for cpu and gpus)
HYPERSCALER CHIPS
- ai5 (tesla)
- openai custom asic
- google tpu
- amazon tranium
- nvidia blackwell, rubin, feynman
ENERGY
- nuclear sites foundations laid
- solar camps scale
- electric grid major expansion

happy sunday. sitting on my sofa hungover from too many post-christmas m&ms trying to digest everything going on in ai and tech
no matter how hard i try everything’s pointing to a crazy 2026.
models are now good enough to do a $250k/yr software engineers job (multiple if you run simultaneous agents)
spaceX is gonna ipo at $1 trillion+ and start training AI models in space
robots are finally good enough to go commercial starting with tesla / waymo autonomous robotaxis, tesla humanoid robots
turns out we’re in a ‘memory crisis’ in the wake of DRAM price 5X-ing, guess we were all wrong abt ai ca

openai hiring for a Head of Preparedness amidst expectation of self-improving ai
anthropic lead engineer tweeting about ai recursively improving itself (claude code)
robot company in china using its own robots to construct itself in the factory.
things about to get weird.
we’re accelerating towards a world where humans will have no idea how to do anything and become almost exclusively dependent on their ai.

Future of AI brain dump #1
just occurred to me that current chatbot llms are the equivalent of discovering pen and paper whereas the end game is ephemeral ai where an ai just generates the ideal user interface for you in realtime
but in order to get there i see a few intermediary trends playing out
1. llms start to naturally incorporate images, video and audio into their responses. google has already started to do this with 3 + nano banana.
2. llms start speaking to each of us in our preferred “style” eg if i’m trying to learn about something it might respond to me in the form of a Strateche

watching shrek whilst i try and understand context graphs for ai agents

if its not obvious yet, these ai models work very similarly to our brains, eyes and ears. llms process ingests words and spit out words. multi modal llms do the same with video, images and sound.
in the same way we ingest information to understand the world around us, progress up a career ladder, form relationships with humans around us - ai models are on a path to do the same and eventually a LOT better than us.
some look at this pessimistically, "we're killing our brains" but i choose to believe that if applied correctly, as a learning tool, ai is the ultimate symbiotic technology to ever pr

really feels like, despite all the chip talk, there is still insatiable demand for ai accelerators of all kinds. the complete opposite of everyone calling for a bubble.
nvidia gpus sold out into 2026
google tpus? pumping out 7M units next year and STILL looking for extra capacity
amazon tranium 3 same story.
neocloud providers? even their old gpus sold out.
as unsexy as it is the data center, semiconductor capex battle is very much one sided. 2026 will see much of the same.
as long as models require more data, more memory and more compute to grow intelligence… the capex bubble has larger ye

turns out the only thing that matters is how ‘buddy buddy’ you are with tsmc and the odds are very much in favour of nvidia
google tpus, amazons tranium chips may be better than nvidia BUT they can’t mass produce till 2027 because tsmc’s entire capacity is booked by jensen.
jensen knew this early on and has been playing the politics in taiwan to secure allo.
very clearly an opportunity to disrupt tsmc and eat their market share…
samsung’s probably our best hope.
happy sunday

biggest trend in ai next year will be world models.
LLMs are running out of data to train on, only way to solve this at scale is putting them in simulated worlds, running millions of real-world simulations and generating synthetic data.
google leads here (surprise surprise) with genie 3
tesla's a dark horse, most people don't realize full self driving runs on a world model simulating every possible car accident so that its avoids it.
world models will also be used to train agents and robots. one of the biggest drawbacks for robots today is theres no data for what a human sees and does.
real wo

saw the news about tesla’s driverless taxi going live as i sat in a $60 uber crawling 2 miles.
i cannot wait to see cost of getting from A-to-B trend to zero
the trial in austin tx already cut a 15-30 min drive down to $5-10, can you imagine this in nyc??
this is going to completely change the way people commute, own cars (i’ll just hop in a tesla) and work (cars become mobile office / entertainment room)

spaceX doesn't hit $1.5T on its own.
it relies on a cluster of companies: starlink, tesla and xai.
starlink provides the satellites to build data centers in space which will run tesla AI6 chips that train, inference xai's grok models.
elon's essentially built a protocol network where each layer of the stack is one of his companies.
insane vision

ai data centers in space
china buying H200s
tsmc capacity at breaking point
nvidia supply sold out
elon building his own fab
42MW supersonic gas turbine data centers
$52B tpus bought by anthropic
claude, , chatgpt users and usage record highs
HBM prices sky-rocketing
nvidia rubin vs. google tpu v8
we are very much in the industrialization era of AI.

check out the full episode @LimitlessFT here: