The single most useful thing AI has done for this business was not writing
copy. It was finding people. I asked it to help me identify who actually sells
Bitcoin miners in volume, and it came back with direct contact information for
suppliers and brokers across the industry — the kind of list that normally takes
years of trade shows to assemble. Those introductions turned into real bulk
purchases, and bulk purchases are the reason there is used inventory on this
site at prices a one-machine buyer could not negotiate alone.
That is the honest version of the productivity story: it did not replace
anyone here, and it did not touch the part of the work that actually decides
whether a customer gets a good machine. It removed the delay between deciding to
do something and having it done.
What it actually changed, job by job
Concretely, in this operation:
Sourcing. Finding and qualifying volume sellers, which is
how lots like our 100-unit
S19 lots get bought in the first place.
Listing. Putting hardware on eBay used to be the bottleneck
— every machine needs a title, a spec block and honest condition wording. We
built a listing
tool around that problem and then gave it away free; it drafts the listing
from a photo or a product link, and the setup walkthrough is
here. Software that would once have justified a hire
is now something a small operation builds around its own workflow.
Documentation. Written procedures instead of improvisation.
Our six-step
bench test and our A/B/C grading
definitions exist as written standards because writing them down stopped
being a week-long job.
What it did not change, and this is the important half
Here is the same point in the specific. A used S19k
Pro arrives on the bench. What has to happen next is: power it under load,
watch all three hashboards report, read per-chip temperatures for long enough
that a marginal board has time to fail, check the PSU under sustained draw
rather than at idle, and decide whether the machine is an A, a B or a C. Every
one of those steps is a physical measurement with a real instrument, and the
answer determines the price and what we are willing to promise about it. The
procedure is written down in our
six-step bench test and the grades are defined in ASIC grading
explained.
No amount of language modelling contributes to that. What it does contribute
is that the procedure got written down at all — and a written standard is the
difference between grading consistently and grading by mood.
Nothing above touches the bench. No model can tell you whether a used
hashboard has three dead chips, whether a PSU is about to fail under load, or
whether a machine that reports 95 TH/s holds it for six hours in a warm room.
That is a person with a test fixture and a power meter, and it is the part
customers are actually paying for. It is also the reason we publish what is
wrong with a unit rather than rounding it up — see why mining hardware
fails for the failure modes we grade against.
AI is very good at the work that sits between decisions. It is no good at all
at the work that is the decision, and on a $3,000 machine the decision
is everything.
Execution speed is the variable, and there is evidence for it
The general form of this is measurable rather than anecdotal. A 2023 study
published by the National Bureau of Economic Research found workers using
AI assistance completed tasks up to 37% faster while maintaining or improving
quality — with the largest gains going to the least experienced workers, not the
most. That matches what happened here: the leverage showed up in the jobs nobody
was expert at, like writing a returns policy, not in the jobs we already did
well.
The same shift is visible in the aggregate. The U.S. Census Bureau’s Business Formation Statistics show new
business applications running sustained above pre-2020 levels, weighted toward
solo founders and micro-enterprises. Researchers at Stanford
HAI’s AI Index make the same point from the tooling side: capabilities that
were restricted to large firms are now available to individuals at marginal
cost.
The mechanism is boring and it is the whole story. Faster responses reduce
lead decay. Written procedures reduce errors. Same-day shipping decisions
prevent churn. Individually these are small; together they are the difference
between a margin and no margin.
Where it actively wastes time
An honest account has to include this part, because the enthusiastic version
of this article is everywhere and it is not useful.
Specifications. Ask a model for the power draw of a
specific miner and it will produce a confident number that is sometimes wrong by
several hundred watts. Every spec on this site is read off the unit, the rating
label or the manufacturer’s own published figure — never generated. A wrong
wattage on a $3,000 machine is a returned crate and an argument about who pays
freight.
Prices and market conditions. Anything that moves daily —
hashprice, difficulty, what a used S19 is worth this week — has to be measured
on the day. A model’s sense of “current” is whenever its training stopped.
Anything it cannot see. It cannot tell you whether the
hashboard in front of you has three dead chips. It cannot hear a failing fan
bearing. It has no opinion worth having about a machine it has not
measured.
The pattern in all three: it is good at the work that surrounds a decision
and bad at the decision itself. Used the first way it is genuinely a force
multiplier. Used the second way it is a confident liar, and in this business a
confident liar is exactly the failure mode that costs a customer money.
What a buyer should take from this
Not an opinion about AI. Two practical things about who you are buying
from:
Ask what was measured and what was assumed. Any dealer can
publish a spec table. The question that separates them is where each number came
from. Ours say — see any used listing,
where the efficiency figure is shown as arithmetic on the hashrate and wattage
rather than as a claim.
Be suspicious of round numbers and missing dates. A
profitability figure with no difficulty, no power price and no date attached is
decoration. That applies to what we publish as much as to anyone else, which is
why every page here that quotes hashprice states the day it was
measured.
Why a miner dealer is writing about this at all
Because it explains the pricing. A small operation that can source in volume,
test in-house and list without a marketing department can sell a tested,
firmware-tuned Antminer
S19k Pro for less than an operation carrying that overhead, and still stand
behind it. The efficiency is not a philosophy — it ends up in the price on the
listing.
How AI Actually Changed the Way I Run This Business
A first-person account of running a hardware business, not a
hardware guide. If you came for machines, start with the ASIC
miner buying guide or the used-miner buying
guide.
The single most useful thing AI has done for this business was not writing
copy. It was finding people. I asked it to help me identify who actually sells
Bitcoin miners in volume, and it came back with direct contact information for
suppliers and brokers across the industry — the kind of list that normally takes
years of trade shows to assemble. Those introductions turned into real bulk
purchases, and bulk purchases are the reason there is used inventory on this
site at prices a one-machine buyer could not negotiate alone.
That is the honest version of the productivity story: it did not replace
anyone here, and it did not touch the part of the work that actually decides
whether a customer gets a good machine. It removed the delay between deciding to
do something and having it done.
What it actually changed, job by job
Concretely, in this operation:
how lots like our 100-unit
S19 lots get bought in the first place.
— every machine needs a title, a spec block and honest condition wording. We
built a listing
tool around that problem and then gave it away free; it drafts the listing
from a photo or a product link, and the setup walkthrough is
here. Software that would once have justified a hire
is now something a small operation builds around its own workflow.
Our six-step
bench test and our A/B/C grading
definitions exist as written standards because writing them down stopped
being a week-long job.
What it did not change, and this is the important half
Here is the same point in the specific. A used S19k
Pro arrives on the bench. What has to happen next is: power it under load,
watch all three hashboards report, read per-chip temperatures for long enough
that a marginal board has time to fail, check the PSU under sustained draw
rather than at idle, and decide whether the machine is an A, a B or a C. Every
one of those steps is a physical measurement with a real instrument, and the
answer determines the price and what we are willing to promise about it. The
procedure is written down in our
six-step bench test and the grades are defined in ASIC grading
explained.
No amount of language modelling contributes to that. What it does contribute
is that the procedure got written down at all — and a written standard is the
difference between grading consistently and grading by mood.
Nothing above touches the bench. No model can tell you whether a used
hashboard has three dead chips, whether a PSU is about to fail under load, or
whether a machine that reports 95 TH/s holds it for six hours in a warm room.
That is a person with a test fixture and a power meter, and it is the part
customers are actually paying for. It is also the reason we publish what is
wrong with a unit rather than rounding it up — see why mining hardware
fails for the failure modes we grade against.
AI is very good at the work that sits between decisions. It is no good at all
at the work that is the decision, and on a $3,000 machine the decision
is everything.
Execution speed is the variable, and there is evidence for it
The general form of this is measurable rather than anecdotal. A 2023 study
published by the National Bureau of Economic Research found workers using
AI assistance completed tasks up to 37% faster while maintaining or improving
quality — with the largest gains going to the least experienced workers, not the
most. That matches what happened here: the leverage showed up in the jobs nobody
was expert at, like writing a returns policy, not in the jobs we already did
well.
The same shift is visible in the aggregate. The U.S. Census Bureau’s Business Formation Statistics show new
business applications running sustained above pre-2020 levels, weighted toward
solo founders and micro-enterprises. Researchers at Stanford
HAI’s AI Index make the same point from the tooling side: capabilities that
were restricted to large firms are now available to individuals at marginal
cost.
The mechanism is boring and it is the whole story. Faster responses reduce
lead decay. Written procedures reduce errors. Same-day shipping decisions
prevent churn. Individually these are small; together they are the difference
between a margin and no margin.
Where it actively wastes time
An honest account has to include this part, because the enthusiastic version
of this article is everywhere and it is not useful.
specific miner and it will produce a confident number that is sometimes wrong by
several hundred watts. Every spec on this site is read off the unit, the rating
label or the manufacturer’s own published figure — never generated. A wrong
wattage on a $3,000 machine is a returned crate and an argument about who pays
freight.
hashprice, difficulty, what a used S19 is worth this week — has to be measured
on the day. A model’s sense of “current” is whenever its training stopped.
hashboard in front of you has three dead chips. It cannot hear a failing fan
bearing. It has no opinion worth having about a machine it has not
measured.
The pattern in all three: it is good at the work that surrounds a decision
and bad at the decision itself. Used the first way it is genuinely a force
multiplier. Used the second way it is a confident liar, and in this business a
confident liar is exactly the failure mode that costs a customer money.
What a buyer should take from this
Not an opinion about AI. Two practical things about who you are buying
from:
publish a spec table. The question that separates them is where each number came
from. Ours say — see any used listing,
where the efficiency figure is shown as arithmetic on the hashrate and wattage
rather than as a claim.
profitability figure with no difficulty, no power price and no date attached is
decoration. That applies to what we publish as much as to anyone else, which is
why every page here that quotes hashprice states the day it was
measured.
Why a miner dealer is writing about this at all
Because it explains the pricing. A small operation that can source in volume,
test in-house and list without a marketing department can sell a tested,
firmware-tuned Antminer
S19k Pro for less than an operation carrying that overhead, and still stand
behind it. The efficiency is not a philosophy — it ends up in the price on the
listing.
If you came here for hardware rather than business operations, the
full catalogue is here, the
used
machines are here, and the honest version of what used stock does and does
not get you is in the
advantages of buying used.