Crucible Capital

We’ve spent the last couple of months deep in financing and coordinating our own B300 cluster deployment. Suffice to say buying, financing, deploying and monetizing GPUs is a wild process, so we figured we’d outsource our findings with ideas on improving this market. DMs open if you’re working on any of this: 1. First important to address that compute is product led not asset led. Consumption experience is the product, so buying bare metal is only part of the equation of bring a cluster to market (also why 1GPUhr does not = 1GPUhr). Orchestration and networking are critical processes to maintain on an ongoing basis. H/t to the start ups doing this as part of the monetization (selling the compute via contracts or marketplace) process- @fluidstack , @sfcompute - we are interested in meeting all players in this space in addition to 1. DCIMs optimizing workloads and networking (investors in @Aravolta_X25 ) 2. applications that enable further homogeneity of compute 2. Financing compute is a tough business mechanically and financially. Mechanically, there are several operational processes to underwrite and numerous counterparties involved (your colo, your OEM, orchestration and networking support, your offtake customer, your financing partner, your insurance carrier, whoever you use for financial rails). On these... 3. In underwriting offtake partners, look through into credibility is critical. @SemiAnalysis_ is leading the charge with ClusterMax performance data on neocloud and marketplace performance which is non trivially important for underwriting your ultimate customer - but there is further room to expand on this (credit ratings to complement performace ratings, actual market DATA on traded $/gpu hour for spot and term vs. prop indices of hypothetical quotes) 4. Underwriting your colo: there’s no real market to compare pricing and performance of colocation datacenter providers. This is a largely OTC, relationship driven process of knowing who to call and what your monthly rent/kw should like as well as what upfront costs should look like. Further, colocation billing and payments systems are archaic. We invested in gnomos (@iamminesoc) for this reason. Someone build a marketplace for colocation! 5. Financing partners: we are fans of what @USD_AI_ is building, they’re one of the only lenders who don’t require sale leaseback for those seeking upfront depreciation benefits in the US, with a novel mechanism to secure liens onchain. More of this - compute economy capital flows belong onchain. 6. Procurement: working with a strong channel partner is essential, buying a cluster involves customization of your BOM for your use case, negotiation, and securing credit and delivery terms. Outside of networking and orchestration, there’s technical work required in designing a cluster fit for your customer or use case. Monetizing with multiple customers vs 1 customer or for training vs inference present different designs in your networking, memory and control planes. Tons to unpack here, probably in its own report. 7. We modeled returns for GPU investing exhaustively for different tax, off take, and financing scenarios. Ultimately, this is a modest income generating investment with attractive near term tax benefits in exchange for longer term tax liabilities. There is demand for this at scale to be clear and we are doing our part to fill it. But care around assumptions on taxes, utilization etc. is paramount in not getting carried away in return expectations. Given that this is a tricky financial puzzle to solve, we are looking to support companies that bring more tooling to this process, some ideas above but a ton of white space to run here, 8. TAX. Last but not least, tax can be the widowmaker or the deal maker here. You need to properly account for and optimize your upfront depreciation (active/passive - there are ways to do each), your recapture, your upfront sales or use tax, and your ongoing income and b&o. Region matters, entity structure matters, etc etc if you made it this far, DM with ideas/what you're working on, + a bonus meme for you:

@michielrauws2 @kellyjgreer @itsanhonour_ @Melt_Dem thanks for reading!

@kellyjgreer @itsanhonour_ @CrucibleCap @Melt_Dem Thanks for the write up, reads great on Kindle too!

come get your copies

@kellyjgreer @itsanhonour_ @Melt_Dem vibes

@kellyjgreer @doublezero @GigaEnergy_ @get_hydrahost drip 😍

who wants a printed copy?

to whom it may concern: we (@CrucibleCap) have a 20MW brownfield site w access to installed BTM generation w/in 4 months. grid connect w/in 12 months and all other fixings in place DM if you want to take on 20 MW of compute, or generators and turbines generally 😀

How to Build a Datacenter (for Dummies) Part I from @CrucibleCap and our friends @GigaEnergy_ @doublezero @get_hydrahost datacenter spend is going to 10x the prior largest US infra build in history this decade and still many ask us about how sites are built 🧵

the US power grid build: $230m in funding in 1935 vs: $3-4trn in expected datacenter infra spend this decade @CrucibleCap report on datacenter development w/ collab from @eden_ @doublezero @GigaEnergy_ @get_hydrahost dropping soon join us in NYC IRL Oct 2nd (link in reply)

beyond cool — @GEVernova is repurposing jet engines (run in reverse!) as power turbines to generate low-carbon energy for AI compute. 29 of them are going to @CrusoeEnergy to supply nearly 1 GW of electricity to Crusoe’s AI data centers.

🚀 New project alpha alert: @pwrfinance (165 followers, account created 3 months ago) Description: "the currency of the compute economy backed by @cruciblecap + @montaukcc we are hiring" Visit web3alerts.app for more alpha :)

Meta’s new “tent” data centers ⛺️ have no diesel backup, just prefab modules and lightweight structures to get AI compute online fast. They will need workload management to maximize the utilization of every watt of power, esp on hot days. [insert @CrucibleCap portcos here!]

another big front of meter nuclear deal AWS buys 1.92 GW from Talen after Meta bought 1.1 GW from CEG both estimated at mid $80s/MWh, a 60% premium to PJM pricing while behind the meter regs drag along, hyperscalers are paying up to secure decades of nuke power more to come

Cembalest reiterated his bearish view on SMRs from his Heliocentrism piece on Odd Lots which was a great listen. he claims: 1. we've built no SMRs in the US to date 2. Chinese and Russian SMRs saw significant cost overruns and delays 3. unlikely this tech can be commoditized to bring costs down (like li-ion batteries, turbines, solar panels). huge respect for Cembalest and he's not wrong on 1 and 2. reactors have been a cost overrun nightmare historically. but it's worth unpacking why the future of domestic nuclear has a better shot. 1. positing "this is unlikely succeed" because we haven't done it yet is simply pessimistic. humanity is where it is today because we've done countless things we never did before. next 2. SMRs to date - built in China and Russia - breakdown of overruns: a. China's 300% overrun: China's model (HTR-PM or pebble bed module) was exotic to say the least. the design uses helium coolant and a graphite moderator which required extensive R&D. b. Russia's 400% overrun: the KLT-40S was again exotic, a floating plant built in a remote location and the project specifically had no cost control measures in place sum: both were extremely exotic FOAK builds and in Russia's case there was no enforcement in place to control costs. worth noting that China enjoyed a significant power surplus in the 2010s when building their SMR too. 3. most domestic reactors are more familiar, and while are still FOAK builds have a few things working in their favor: - the government is in our court. the strongest EO "DEPLOYING ADVANCED NUCLEAR REACTOR TECHNOLOGIES FOR NATIONAL SECURITY" actively seeks to reduce FOAK costs and support advancement as a national security imperative. -more standard designs - take Kairos, Oklo, GE Vernova. GEV's BWRX-300 uses traditional light water reactor technology. Kairos Power's fluoride salt-cooled high-temperature reactor and Oklo's fast neutron reactors, while advanced, build on historical experiments and operational reactors, such as the Molten Salt Reactor Experiment and Russia's BN-600 fast reactor, respectively. these should all be easier to build due to their technological familiarity vs. the far out Russian and Chinese experiments Cembalest references. yes we're doing something new. but we're grounded in research and we need to do it. renewables will take time to proliferate as Cembalest reminds us. turbines are on material backorder due to limited capacity expansion. we're not blind to the fact that reactor production is difficult, and we discuss it in the @CrucibleCap fuel report. but - SMRs have offtake agreements in place now and have all the tailwinds at their back (regulatory, capital, demand) to make this work. the history of domestic nuclear was egregiously prohibitive as discussed in the fuel report, but the opposite is the case now. where from here, all eyes on Kairos Hermes for first deployed by 2027 (so far 1 year delayed vs. 10+ years in China/Russia) bit.ly/nuclearfuelrep…

the @CrucibleCap x @standardnuclear fuel report is COMING real soon imagine what happens when @Melt_Dem and i collab with two nuclear physicists on a 30 page report we go off on how the US fu*ked itself and how we plan to fix it and we make funnies bc we cant help it

@huangkuan @Melt_Dem @_itsanhonour @kellyjgreer Should be working - keep the traffic coming

@Melt_Dem @CrucibleCap @_itsanhonour @kellyjgreer Your website is down now. Maybe this beautiful photo attracted too much traffic?

her name is @CrucibleCap it’s a cheesy photo but it makes me happy. you really *can* just do things.

LET THEM COOK congrats to our portfolio company @miranetwork sign up for their flagship app, Klok, and earn points when you ask anything and everything💡klokapp.ai

1/ @CrucibleCap is co-leading an investment into @nirvanalabsai with the prolific builders at @jump_ Nirvana is building next-gen infrastructure from silicon to network, reducing reliance on hyperscalers and dramatically lowering compute, storage, and bandwidth costs

nice little chat with @wsfoxley on @theMiningPod about why we @CrucibleCap are focused on aggregating the periphery of equipment backed finance rather than taking down gpu financing at scale

@KLorenCompany @itsanhonour_ industrial battery financing is and continues to be a graveyard

@itsanhonour_ @CrucibleCap Storage (buffering) needs to be central to this equation

yesterday, the entire Iberian grid simply switched OFF (Spain and Portugal) - thats like turning off TWO NYISO's (the grid that powers all of New York City and the surrounding state) ...

the @CrucibleCap thesis is simple: AI + grid software = intelligent, adaptive, resilient infrastructure. As the blackout showed, this isn’t optional anymore. If we don’t modernize the grid, it will break.