NoDiligence
7 Posts
NoDiligence
rates. chips. compute.
2 Following
3 Followers
12 Liked
Posts
Portfolio
All
Quotes
Verified
AMD has promised to pay $30.28bn for wafers and packaging it has not received. it made $11.54bn of revenue in the quarter it filed that number.
AMD does not own a factory. everything it sells is manufactured by somebody else, mostly tsmc, and the way you get a production slot two or three years out is to promise you will take the output and pay for it whether you end up needing it or not. that promise is not borrowing and it does not appear on the balance sheet. it is one line in a footnote of the 10Q.
which was $3.99bn in march 2024 and it is $30.28bn now, 7.6 times bigger. revenue roughly doubled over the same eight filings. so the promise went from about two thirds of one quarter's sales to two and a half quarters of them. AMD spent $1.2bn on its own PP&E in the half year and committed $30bn to everyone else's.
that “unconditional” in the note is doing a lot of work. it basically locks in one side. AMD owes the supply chain whether or not the mi355x racks that went to saudi arabia get repeated, and nothing in there puts a customer on the hook on matching terms.
at two and a half quarters of sales, one weak quarter turns reserved capacity into parts AMD has to accept and cannot move on. intel and samsung run that same inventory risk, but they own the plant and write it down over years. AMD is carrying it with no fab and nothing to spread it across.
a purchase obligation is a promise to buy. so nothing in the note says who the eventual customer is, or whether there is one. there should be an order book visible.
AMD does not own a factory. everything it sells is manufactured by somebody else, mostly tsmc, and the way you get a production slot two or three years out is to promise you will take the output and pay for it whether you end up needing it or not. that promise is not borrowing and it does not appear on the balance sheet. it is one line in a footnote of the 10Q.
which was $3.99bn in march 2024 and it is $30.28bn now, 7.6 times bigger. revenue roughly doubled over the same eight filings. so the promise went from about two thirds of one quarter's sales to two and a half quarters of them. AMD spent $1.2bn on its own PP&E in the half year and committed $30bn to everyone else's.
that “unconditional” in the note is doing a lot of work. it basically locks in one side. AMD owes the supply chain whether or not the mi355x racks that went to saudi arabia get repeated, and nothing in there puts a customer on the hook on matching terms.
at two and a half quarters of sales, one weak quarter turns reserved capacity into parts AMD has to accept and cannot move on. intel and samsung run that same inventory risk, but they own the plant and write it down over years. AMD is carrying it with no fab and nothing to spread it across.
a purchase obligation is a promise to buy. so nothing in the note says who the eventual customer is, or whether there is one. there should be an order book visible.
I'm really bullish on NVIDIA if this acquisition goes through. even if it doesnt, we now know that NVIDIA is going aggressive with their operations. all in all, they are in the right direction.
$NVDA
$NVDA
NoDiligence
·
--
hugging face told nvidia no at $7bn last year. nvidia now wants the whole thing at $12.9bn.
NVIDIA went into the 2023 round with google, amazon, amd, intel, qualcomm, ibm and salesforce.
since then it put train on dgx cloud on the site, then its nim inference containers, then training cluster as a service on top of the lepton marketplace it bought in april 2025.
by mid 2025 nvidia owned the compute under hugging face and the containers on it. it didn't own hugging face.
hugging face is doing about $150m a year now, so at $12.9bn nvidia is paying 86x revenue.
-- the 2023 round was 129x revenue when the estimates were ~$35m for that year.
-- if they were near $100m when they said no, nvidia offered 70x revenue for a minority stake and is now paying 86x for the whole company.
-- so it's paying 23% more per dollar to own it outright than it offered for a stake with no say.
i don't think anyone else was bidding.
hugging face runs a router called inference providers.
-- every request goes to one of 18 partners, and by default it goes to whoever is fastest, meaning most tokens per second.
-- cerebras, groq and sambanova are on that list and none of them run nvidia chips, so when one of them is fastest for a model, that's where the request goes.
-- hugging face also writes optimum-amd and optimum-neuron, so the same models run on amd and trainium without cuda, and amd and amazon are on the cap table.
-- hugging face calls that zero vendor lock-in, and a $500m minority holder (what nvidia offered in late 2025) can't change a routing default.
chinese models were 41% of downloads on hugging face last year, more than american ones. there are more than 113,000 models built off qwen on the site, and qwen had passed a billion downloads by march. nvidia can't sell its top chips into china.
revenue went from about $100m a year in june to $150m in august and that's paid compute, storage and subscriptions.
i'd watch whether AMD, Amazon & Google take their 3x and leave, and whether the fastest default is still there once the deal goes through.
and whether qwen and deepseek keep releasing on a site owned by the company whose top chips they can't buy.
since then it put train on dgx cloud on the site, then its nim inference containers, then training cluster as a service on top of the lepton marketplace it bought in april 2025.
by mid 2025 nvidia owned the compute under hugging face and the containers on it. it didn't own hugging face.
hugging face is doing about $150m a year now, so at $12.9bn nvidia is paying 86x revenue.
-- the 2023 round was 129x revenue when the estimates were ~$35m for that year.
-- if they were near $100m when they said no, nvidia offered 70x revenue for a minority stake and is now paying 86x for the whole company.
-- so it's paying 23% more per dollar to own it outright than it offered for a stake with no say.
i don't think anyone else was bidding.
hugging face runs a router called inference providers.
-- every request goes to one of 18 partners, and by default it goes to whoever is fastest, meaning most tokens per second.
-- cerebras, groq and sambanova are on that list and none of them run nvidia chips, so when one of them is fastest for a model, that's where the request goes.
-- hugging face also writes optimum-amd and optimum-neuron, so the same models run on amd and trainium without cuda, and amd and amazon are on the cap table.
-- hugging face calls that zero vendor lock-in, and a $500m minority holder (what nvidia offered in late 2025) can't change a routing default.
chinese models were 41% of downloads on hugging face last year, more than american ones. there are more than 113,000 models built off qwen on the site, and qwen had passed a billion downloads by march. nvidia can't sell its top chips into china.
revenue went from about $100m a year in june to $150m in august and that's paid compute, storage and subscriptions.
i'd watch whether AMD, Amazon & Google take their 3x and leave, and whether the fastest default is still there once the deal goes through.
and whether qwen and deepseek keep releasing on a site owned by the company whose top chips they can't buy.
Verified
sandisk $SNDKB $SNDK put 25 dollars into its own stock for every dollar it put into a factory this year.
bought back nothing at all in either of the two years before, so it started the same year they got $59.8bn of flash reservations on the books.
and they don't even own the fabs, kioxia's JVs do.
bought back nothing at all in either of the two years before, so it started the same year they got $59.8bn of flash reservations on the books.
and they don't even own the fabs, kioxia's JVs do.
nvidia $NVDAB sorts its data centre sales into two piles, the big cloud companies and everyone else.
-- in may it said the cloud pile grew 115% and the other one 74%.
--in august it published the same quarter again and they're 93% and 91%.
the money never changed, $75.2mn either way. one customer worth $5.18mn got moved from one pile to the other in last year's figures.
https://x.com/nodiligence/status/2093645958777335981
-- in may it said the cloud pile grew 115% and the other one 74%.
--in august it published the same quarter again and they're 93% and 91%.
the money never changed, $75.2mn either way. one customer worth $5.18mn got moved from one pile to the other in last year's figures.
https://x.com/nodiligence/status/2093645958777335981
nvidia $NVDAB now carries $47.9bn of stock in private companies, up from $3.8bn a year ago. that stake is now about 3.4 times the value of the property and equipment nvidia owns.
most of the increase showed up in just two quarters: $14.1bn in the january quarter and another $20.1bn in the april quarter.
this is carrying value, not cash spent, and nvidia reported $15.9bn in “other non-operating” income in the april quarter, suggesting a meaningful portion of the jump is revaluation of positions it already had.
the filings don’t name any of the underlying holdings.
https://x.com/nodiligence/status/2094368992278565094
most of the increase showed up in just two quarters: $14.1bn in the january quarter and another $20.1bn in the april quarter.
this is carrying value, not cash spent, and nvidia reported $15.9bn in “other non-operating” income in the april quarter, suggesting a meaningful portion of the jump is revaluation of positions it already had.
the filings don’t name any of the underlying holdings.
https://x.com/nodiligence/status/2094368992278565094
Partly True
$SNDK put 25 dollars into its own stock for every dollar it put into a factory this year.
bought back nothing at all in either of the two years before, so it started the same year they got $59.8bn of flash reservations on the books.
and they don't even own the fabs, kioxia's JVs do.
follow me on X at @nodiligence.
bought back nothing at all in either of the two years before, so it started the same year they got $59.8bn of flash reservations on the books.
and they don't even own the fabs, kioxia's JVs do.
follow me on X at @nodiligence.
Log in to explore more content