Kai Golden

capex enjoyer
@get_hydrahost

[goodalexander @goodalexander]

I should cause a credit event to get cheap GPUs

[Kai Golden @barestmetal]

@andrawesbahou @Melt_Dem Operative word is "banks."

I just got a partner a term sheet for a $300M line at 95% LTV with a 30 yo hedge fund that's never originated a compute deal before @ ~15%. This was at the end of a process with numerous private credit funds looking to lend at 80%LTV/~15% over non-IG.

[Kai Golden @barestmetal]

I finance brand new neoclouds for a living and am having no issues placing capital. Elon is universalizing his own problem, which is that investors constrained to the crowded rated markets are repositioning. Demand for GPUs expressed as financeable rental contracts outstrips supply by many orders of magnitude and private credit is all over it.

[Kai Golden @barestmetal]

The main question is how you create or preserve value when you do become a neocloud (aka hardware rental company)

You’re going to generate a ton of cash, but you’ll only get a terminal value on top of rental revenue potential if you can
a) generate value at the application layer, which few will OR
b) compete with your hardware suppliers à la Crusoe OR
c) own your data centers, turning lease opex into mortgage paydown and extending your modelable ability to generate cash

The ingredients of GPU rental businesses are necessary to keep any other AI business alive, but you’ll find investors looking at you more like a project finance company than a tech company

It’s what you do with the massive R&D budget your GPU rentals have given you that decides what kind of company you become

[Ariel Deschapell (b/acc) @NotASithLord]

Obvious when the biggest one already did, it’s called SpaceX.

Lots of companies that imagined they will train capable models or run capable inference services will end up “just” renting bare metal.

[wayne nelms @wayne_nelmz]

Compute capacity renewal rates are high and prices are increasing.

The status quo of a 6yr straight line depreciation schedule are being challenged.

There is simply not enough deployed compute in the world.

[Conor Moore @_ConorMoore]

With all of the focus on capex heavy businesses, VCs need to establish credit sleeves to stay relevant in competitive equity rounds

[Kai Golden @barestmetal]

*oceans are now battlefields font*:
data centers are now battlefields
https://t.co/cevMfiOOK2

[GeoInsider @InsiderGeo]

NEW: The Romanian Air Force has released footage from last night’s mission to neutralize two Russian Geran drones that entered Romanian airspace.

[WarMonitor🇺🇦🇬🇧 @WarMonitor3]

BREAKING: Ukraine has carried out the world’s first medical evacuation of a human via a drone from the battlefield the 30th mechanised brigade of Ukraine managed to save their injured comrade from a minefield.

Just absolutely insane.

[Kai Golden @barestmetal]

Middle strike is 600 km now. Our critical infrastructure is unprepared for a world where drone swarms make strategic bombing cheap and overwhelming.

[Defense of Ukraine @DefenceU]

Meet the Thor Hammer, a new Ukrainian middle-strike drone.

It carries a warhead of up to 50 kg and has a range of up to 600 km. Its navigation system combines inertial and satellite guidance with backup positioning, so it keeps finding its way when Russian EW tries to blind it.

Russian logistics hubs, fuel depots and command posts in the enemy’s rear are now within reach.

[Kai Golden @barestmetal]

@abcampbell A token is an economic output with value. More tokens per GPU/hour means more value per GPU/hour. The more money I can make per GPU/hour, the more I’m willing to pay for it.

[Kai Golden @barestmetal]

Value is leaving the hyperscaler/neocloud walled garden and reconsolidating closer to end customers. Capital is following suit.

It seems like AI compute lives in two worlds:
- Public neoclouds, which are in the business of standing up low IRR projects anchored by unsuable trillion dollar clients and whose debt has been long overpriced
- Private neoclouds and GPU companies invisibly serving the upper middle market, clipping high 30s and 40s returns serving the token factories and inference platforms the emerging AI economy runs on, which also happen to enjoy gorgeously above water unit economics

This is why Oracle can buy the wrong chips at the wrong price but still renews contracts at increasing prices. If my customers can create 10x more value per GPU/hour than before, so too creeps up the replacement value of a GPU/hour. In the age of inference (read: commercial viability of AI-native products), projects like Kimi, Deepseek, and Sail Research are inflationary, not deflationary.

This is also why CoreWeave debt backed by chips and MSFT contracts is cheaper than debt backed by chips and a corporate guarantee. Customer contracts are the true source of value and new return streams mean conventional clouds are no longer the only trade in town.

The space has reached escape velocity. The AI bubble isn't popping, but the hyperscaler/neo-hyperscaler bubble may be deflating. Capital is leaving crowded trades and reemerging where differentiated value is actually being created.

[zerohedge @zerohedge]

Data Center debt getting smoked today.

Coreweave unsecureds tag 1000 g-sprd bps for the first time ever. Oracle unsecureds trading like deep junk. Even Meta's Project Beignet about to hit record wide

[Kai Golden @barestmetal]

Ukraine is teaching Russia what we learned in the Gulf: concentrating critical workloads in conspicuous above-ground DCs is a bad idea

[Яндекс @yandex]

Всем привет, я Дима Пирогов, руковожу финтех-направлением Яндекса.

Сегодня утром был повреждён ещё один дата-центр Яндекса, дом наших сервисов. Из-за этого многие из вас до сих пор не могут проверить баланс, оплатить покупку или перевести свои деньги.

[OSINTRadar @osradar]

🇷🇺 Yandex Unable to Restore Its Data Centers Until at Least 2028 Due to Chip Shortages

[Kai Golden @barestmetal]

Solvable problems in human capital are one of many wedges the US can and should use to ensure LATAM powers American silicon and doesn’t become a forward operating base for China

@CarlMeacham and I get into the implications and urgency here https://t.co/pSFbWhBZh0

[Andres Oppenheimer @oppenheimera]

My article on Latin America's data centers for Americas Quarterly magazine: The region can become a global data center hub, but it won't be of much help unless it dramatically improves its education and worker training. https://t.co/dfKKycBDjw

[Kai Golden @barestmetal]

Price per FLOP is a function of value created and thus inflates at the pace of innovation.

Cheaper tokens -> more economic value/GPU/hour -> more expensive GPU/hours

Newer, faster servers accelerate development software efficiencies and bolster the value of all servers supporting inference, including old ones

The pie grows exponentially faster than hardware can come available. True scarcity is in properly configured equipment, not model improvements, and economic rewards will accrue accordingly.

[Nathan LeClaire @dotpem]

I can't emphasize enough how important this @a16z chart seems and how few have internalized it.

[Sci-Fi Archives @SciFiArchives]

The Art Deco details of the Hoover Dam are incredible. One of the most beautiful pieces of American infrastructure ever built.

[Fire Point @FirePointUA]

🔥 In response to Russian attacks on Ukrainian data centers, Ukrainian Defense Forces hit the "Yandex DC Kaluga" facility in Kaluga Oblast using FP-1 drones.

With a capacity of 63 MW, it is Yandex’s largest data center. The strike knocked several modules completely out of service, marking the second of Yandex’s five major data centers targeted by Ukraine.