9/
here's another angle most of the posts on my timeline have missed.
not all inference workloads look like what etched is optimizing for. regulated buyers, sovereign workloads, anything with an audit trail have materially different requirements from centralized latency-sensitive inference.
that's a separate problem the industry has barely started building for.
1
23
7/
cluster-scale memory.
a whole cluster of chips behaves like one shared memory pool. chips stop wasting cycles waiting to talk to each other (less idle time).
this attempts to solve the bandwidth issues during the decode phase.
both are legitimate hardware wins.
1
21
6/
low-voltage inference.
run the chip at lower voltage so it produces less heat per calculation. that lets them pack way more math onto the hardware without cooking it.
this is solving for the heat issue in the prefill phase.
1
21
5/
now that we understand what models do during inference and the limitations of gpus, here are the two technical innovations they're claiming that are worth sitting forward for:
1
17
4/
instead, etched went specialist. they hard-wired the transformer architecture directly into silicon.
much faster in theory, but only if transformers stay dominant.
their founder acknowledged this in 2024: "if transformers go away, we'll die. but if they stick around, we're the biggest company of all time."
1
21
like many of you, the @Etched announcement has dominated my feed today. with $800M raised, and $1B in contracts for their custom silicon for AI inference - they've come out of the gates with a bang.
for those of you who aren't hardware engineers or ML infra specialists, here's the simplified version of what they're claiming and why it matters:
2/
when a model answers you, it does two things (prefill & decode):
1. it reads and processes your prompt, which is a raw math problem. (prefill)
2. then it writes the response, which is a memory bandwidth problem. (decode)
1
21
when do we stop slapping "ai powered" on everything? at some point it's just a given that everything is using ai... nobody says 'digital camera' anymore, it's just a camera. what's the next buzzword that's gonna get slapped on everything? agentic? quantum?
14
if LeCun is right.. the funniest outcome of the world models thing is we kill the token as an idea and keep it as a pricing line item. rip. metering latent compute and still charging you per "token" like nothing happened
43
The first waitlist spots are filling up quickly. Visit the OpenPC website and see how much your space could earn you.
This black box pays its host up to $2,000 a month. You give it a corner of a garage or basement, an outlet, and an internet connection. We handle everything else, and AI companies pay you for what it produces.
40
a box in a garage pays $2,000/month. wall-mounted unit. runs 24/7. we cover the hardware, install, and electricity. you do nothing. just give it a spot on the wall.
3/
the numbers:
– $0 upfront, ever
– $24,000/year ($2,000/month)
– electricity reimbursed
– no lock-in, we pick the box back up if you cancel
mortgage covered. AC bill drops. real money, not tokens. qualifying the first wave of hosts now.
1
22
4/
see your real monthly number in 30 seconds.
plug in your location, power, and connection.
sign up if the math works.
→ openpc.io
16
We're hiring!
Apply now at openpc.io/
We're hiring 10,000 landlords for the AI economy.
No tenants. No leases. No phone calls at 2am.
The job: rent us a corner of space you're not using.
The pay: up to $2,000/month, every month, paid directly.
↓ Apply + how it works below
16







