capex enjoyer prev: founding director, hydra capital @get_hydrahost | investor, @incomedonediff
Joined October 2026
- Tweets53
- Following17
- Followers40
- Likes57
Kai Golden retweeted
watch how scarce compute becomes friends.
Anthropic's roughly $34.9B compute deal with Lambda can now be terminated by either company after Lambda failed to secure a facility lease by a set deadline.
It makes up about 85% of the $41.3B Anthropic committed to Lambda in deals signed in May, June and August, with terms running into 2034.
Anthropic only pays as capacity is delivered and accepted.
Neither side has sent a termination notice, and Anthropic says it does not currently intend to end the deal.
Otherwise, the Lambda contracts can only be canceled for an uncured material breach by Lambda or in other limited cases, and some tranche by tranche if Lambda misses delivery milestones.
Lambda is one of several providers, alongside Microsoft, SpaceXAI and Nscale, that give Anthropic access to $NVDA GPUs, including Grace Blackwell and Vera Rubin systems.
Kai Golden retweeted
the worst time to start a company is when everyone thinks it’s the best time to start a company.
people naively think raising $ is the hard part of building a company when it’s actually (1) hiring world class people + (2) competing for customers
both become harder for early stage startups in frothy markets.
all legendary companies were started at the beginning of market cycles, not after markets became frothy.
Kai Golden retweeted
The only thing worse than this is not having been outspoken about your crazy enough.
“Else, to-morrow a stranger will say with masterly good sense precisely what we have thought and felt all the time, and we shall be forced to take with shame our own opinion from another.” -Emerson
Kai Golden retweeted
human-in-the-loop
└ human-on-the-loop
└ human-near-the-loop
└ human-cc'd-on-the-loop
└ human-legally-liable-for-the-loop
Kai Golden retweeted
Cheaper tokens don't mean cheaper bills
Since May:
- Average token prices are down 67%
- Token spend is up more than 120%
Charts of the Week: a16z.news/p/charts-of-the-we…
The "Rack Spread" is unfolding, token prices are falling while GPU prices climb
Change since mid-July:
- B300 GPU rentals up 106%
- B200 up 66%
- H100 up 17%
- Average token prices down 58%
Charts of the Week: a16z.news/p/charts-of-the-we…
Where this whole thing is headed, half a yard of deployments later.
kaigolden.substack.com/p/the…
Echoing as the one that led the mentioned effort across more than $500m of speculative GPUs, the core thesis was that entities that can contribute creditworthiness beyond that of a customer contract can rip massive premiums for taking the merchant risk. And they have.
This also means that disaggregating the servicing layer from the CoreWeaves of the world basically unlocks fully funded projects alongside effective LTVs and more and better pools of subordinated capital, letting those with scarce servicing expertise and access to power service more supply and make more money than before. This presupposes, of course, financiers that are overtly long merchant risk/on demand (we have found them).
We have also observed that more dollars mandated to the space mean more leverage for those with physically scarce inputs (expertise-hours, sites, electricians, chip allocations).
Hydra Host has deployed several speculative clusters, independently financed by our partners who we manage the clusters for. Partners who have far more access to credit than startups.
A restructuring rather than a collapse is happening. Ultimately most clusters will be owned by investment institutions with the lowest cost of capital and that’s it.
We’ve anticipated this at Hydra and that’s why we’ve focused on getting many of these “near-zero” speculative clusters deployed.
Kai Golden retweeted
GPUs were supposed to depreciate like a new car off the lot. Today, 2020 legacy GPUs are still seeing their rental prices rise.
Real-time data for the compute economy, powered by Ornn.
Kai Golden retweeted
Yes, no kidding.
That’s why local GPUs will be sitting on distributed networks 24 hours a day doing training and inference.
sam: local inference is a terrible allocation of scarce resources
“a lot of people want local inference, and somehow they assume that the problem is solved now because we have open-source models, we're going to be able to run them on our computers, and so we're not going to need all the stuff in the data centre
but from a global allocation perspective, actually it's quite terrible to have everybody, like, have a bunch of the world's limited supply of DRAM sitting in their own computer and being used five percent of the time
you want maximum utilisation in the cloud if you're trying to deal with this very supply-constrained environment
somehow many people's intuition is the opposite of that”
👀
$IREN Mackenzie UPDATE:
Thermal imaging from 10/24 shows significant heat coming from data halls 4 and 2. All 8 segments of data hall 4 are showing enough heat to suggest commissioning. With fiber optics still being tied in, however, they are most likely not billing yet. As for hall 2, 6 of 8 segments are showing heat signatures and are most likely being commissioned, yet not billing since fiber optics has not been tied in.
Lots of work has been done since 9/23 footage was taken. I believe 3-4 halls could be online and billing before year-end.
1
129
