The Twenty Minute VC · 3 Oct 2026 · From the week of 28 September
20VC: The Future of Datacentres: What You Need to Know | Why Everyone Gets GPU Depreciation and AI's Energy Costs Wrong | Who Really Makes Money From AI & Why Most Moats Don't Exist with Chase Lochmiller
These are notes on the conversation, checked against its transcript. The episode itself has the full discussion.
In brief
Harry Stebbings interviews Chase Lochmiller, co-founder and CEO of Crusoe, which builds and powers AI data centres and has raised a $3.9 billion Series F at a $30.9 billion valuation. They cover Crusoe's move from Bitcoin mining to AI, the real bottlenecks in data centre build-out (places to plug in GPUs, energy, skilled labour, long-lead electrical gear), and Lochmiller's case that fears about data centre water use and local energy prices are misinformation. Lochmiller explains Crusoe's economics: three products (data centres, GPUs, tokens), a portfolio of long take-or-pay GPU contracts and shorter, higher-margin managed services, and why he expects GPU useful life to run past the standard six-year depreciation. His central argument is that vertical integration, which he compares to an oil 'supermajor', gives Crusoe speed, cost visibility and a natural hedge as margin moves between layers of the AI stack. He also says he has come to believe that most moats are illusory and that speed of adaptation matters more.
For founders
- Lochmiller says vertical integration let Crusoe beat supply-chain bottlenecks: after committing to a one-year build where the next-closest bid was two and a half years, Crusoe found a medium-voltage power distribution centre had a 100-week lead time, and its in-house electrical manufacturing built one in 28 weeks.
- Crusoe runs a portfolio of compute contracts: roughly five-year take-or-pay rentals to credit-quality customers that pay back within the term, plus shorter-term rentals and managed services that earn higher margins but carry renewal risk.
- Lochmiller argues GPU depreciation fears are overstated: Hopper GPUs bought in 2023 now rent for more than when new, and he believes managed services that run on older chips will stretch useful life beyond the standard six years.
- He says he changed his mind in the last year and now holds that most moats are illusory and short-lived during fast technological change, so the real advantage is moving quickly and adapting.
- Lochmiller says Crusoe's apparent pivot from Bitcoin to AI was planned: Bitcoin monetised cheap stranded energy while the AI platform was built on the side, and ChatGPT's launch prompted a Bayesian shift of resources toward AI.
For revenue leaders
- Lochmiller describes blending contract types for a healthy overall margin: long take-or-pay contracts for predictable cash flow, plus shorter, higher-margin managed services that carry renewal risk.
- He reports that customers spend more money on closed frontier models but generate more tokens on open-source models, and that inference buyers weigh dollars per token alongside throughput and latency.
- Lochmiller says opposition to data centres is emotional and driven by misinformation, and that the industry has to tell its story with data on jobs, water use, energy prices and local tax revenue.
What was said 31, most useful first
Bitcoin miners cut about 98% of data centre cost by dropping reliability requirements, and Crusoe expected AI data centres to go through a similar redesign. Listen
Lochmiller traces Bitcoin mining from laptops to GPUs, leased co-location with five nines of reliability, and ASICs, until miners realised they did not need that much reliability. Low-cost, no-redundancy 'chicken coop' facilities removed roughly 98% of data centre cost and shortened payback periods. Seeing NVIDIA's roadmap of chip power rising from about 150–200 watts toward 300 and 600 watts per chip, he expected rising power density to change what AI data centres look like.
“they were able to eliminate something like 98 % of the total data center cost with these chicken coop type data centers.”
Crusoe built a 100-week-lead-time power component in-house in 28 weeks, which Lochmiller credits to having vertically integrated electrical manufacturing. Listen
For the first two buildings in Abilene, a little over 200 megawatts, Crusoe committed to a one-year build when the next-closest of 34 other developers bid two and a half years. Vendors quoted a 100-week lead time for medium-voltage power distribution centres. Crusoe's internal team sourced the components and made them in 28 weeks. Lochmiller describes the AI data centre supply chain as 'whack-a-mole', with the bottleneck shifting over time.
“When we went out and canvassed the market for vendors that were supplying these medium voltage power distribution centers, the lead time for this one component was a hundred weeks.”
A 140 MW Crusoe building in Abilene uses about as much water a year as roughly 10 single-family homes. Listen
He calls the claim that data centres use huge amounts of water false for modern AI factory designs. Crusoe uses water to cool GPUs in a closed loop: cold water enters the racks, goes to an outside chiller and the heat is exhausted. Most of the water consumed goes to staff bathrooms, handwashing and watering plants.
“They use about the same water annually as about 10 single family homes.”
Energy prices typically fall, not rise, in communities where large data centres are built. Listen
His reasoning is that data centres catalyse investment in new generation capacity, and more megawatts get amortised over the same transmission and distribution infrastructure. He says Crusoe supports the industry bringing new power production online for large builds. He says falling prices are 'the opposite of the narratives that's being told'.
“with more megawatts being amortized over the same transmission and distribution infrastructure. And as a result, people's energy costs end up coming down.”
Crusoe manages GPU payback risk with a portfolio of contracts that trade margin against duration and renewal risk. Listen
Lochmiller treats the GPU hour as a near-commodity, noting that platforms are creating tradable compute futures. Crusoe rents capacity on roughly five-year contracts to credit-quality customers that pay back and cash-flow within the term. It also signs shorter-term contracts at higher margins that carry renewal risk, and sells managed inference and serverless fine-tuning on even shorter, higher-margin terms. The aim is a healthy blended margin across the mix.
“There are shorter term contracts that we'll do at higher margins. But they're riskier because at the end of the contract, it's like, is there a renewal?”
Hopper GPUs bought in 2023 now rent for more than they did when new, contrary to early fears that they would lose value after three years. Listen
When Crusoe made substantial Hopper purchases in 2023, the feedback was that the chips might not be valuable after year three, and early financings demanded very fast payback and heavy debt service coverage. Three years later, Hopper rental rates are higher than at launch. He says people underestimate developers' ingenuity in turning compute into valuable services.
“Here we are three years later, and the prices being charged for utilizing hoppers is higher than the rates that were being charged three years ago when they were brand new.”
Managed AI services that run on older chips will extend GPU useful life beyond the industry-standard six-year depreciation cycle. Listen
Crusoe depreciates GPUs over six years, which he calls the industry standard. Because Crusoe is 'long' energy, data centres and chips, it built managed services that hide the underlying chip from the customer. He says demand for cheaper intelligence on older, slower chips will persist, and he presents the longer cycle as a belief rather than something already proven.
“will ultimately, we believe, extend the depreciation cycle beyond six years.”
He changed his mind in the past year and now believes most moats are illusory, so speed and adaptability are what create advantage. Listen
He notes that VCs love to ask about long-term moats. He now thinks most moats are ephemeral, especially while model capabilities are advancing quickly. In his view, advantage comes from moving quickly and adapting to an ever-changing 'chessboard'.
“I think what I've changed my mind on is most moats are illusion. Most moats don't exist. Most of them are ephemeral”
He often finds wealthier founders make better entrepreneurs because they are not focused on protecting their downside. Listen
Stebbings frames this as a signal he looks for as an investor when judging potentially great entrepreneurs. His reasoning is that founders who have already made money focus on what is possible rather than on downside protection.
“often richer entrepreneurs are better entrepreneurs because they're not worried about downside protection and they just see what is possible.”
Crusoe's shift from Bitcoin to AI was planned from the start, with Bitcoin funding the business while the AI platform was built on the side. Listen
Lochmiller says the founding premise was that AI compute would be bottlenecked by energy. Crusoe put most of its early resources into using waste energy for Bitcoin mining, its best monetisation engine at the time, while building an AI platform that launched to paying customers in early 2022. He describes deciding when to move resources as Bayesian and probabilistic, and says ChatGPT's launch on November 30, 2022 shifted his view toward far greater AI compute demand.
“Bitcoin was the best monetization engine we had at the time, but we were building this AI platform from very, very early days.”
AI infrastructure does not need to sit in centralised hubs like Northern Virginia and will spread to wherever energy is cheap and plentiful. Listen
Most web applications run out of Northern Virginia, but Lochmiller says the time to serve a neural network is dominated by compute inside the data centre rather than network transit to it. That opens new geographies chosen for energy. Crusoe responded by investing in AI-specific data centre design and in sourcing energy.
“so much of the time to serve a neural network is actually the compute that's happening in the data center, not the time to get to the data center.”
Crusoe vertically integrated for availability, cost visibility and innovation rather than to stack margin. Listen
Lochmiller says electrical manufacturing is not an ultra-high-margin business, though margins have risen with shortages. The main benefits are on-time delivery, end-to-end visibility into true costs (he cites Elon Musk's 'idiot index', the end-product cost relative to raw materials), and freedom to design from first principles. The Abilene campus was designed as a one-gigawatt-scale computer running one cohesive workload on a single RDMA fabric.
“it's not the main reason we're doing it is like stacking margin. It's more availability and getting availability, being able to deliver on time.”
In the 2026 inference and agent era, 'time to token' matters more than gigawatt-scale clusters. Listen
He describes 2026 as the era of scaling utilisation, agents and token consumption, which drives demand for inference. Serving models does not require a gigawatt-scale computer and can be done on much smaller clusters. The key metric becomes how fast a provider can go from nothing to delivering tokens.
“What matters is actually how quickly can I go from not having anything to being able to deliver tokens? So what is my time to token?”
Lochmiller names energy and skilled trade labour as the key data centre constraints, and says US power prices are globally competitive while US labour costs more than China's. Listen
He says the constraint shows up as a shortage of places to plug in GPUs, driven by available power and a finite supply of electricians, welders, plumbers and construction workers. Matching labour to the remote places where energy is available is a big challenge. Asked about China, he says China has subsidies but the US is globally competitive on power, while conceding that labour is far cheaper in China.
“There are a finite amount of skilled trade workers, electricians, welders, plumbers, construction workers in the United States.”
Data centre policy is hyper-local, and Crusoe favours factory-built infrastructure partly to reduce permitting friction. Listen
He calls regulation something to navigate and supports good policy, but says it slows development. Local communities care about job creation, energy costs, water, pollution, tax revenue and whether the data centre is a good citizen. Manufacturing more of the infrastructure rather than running everything as a giant construction project is one way Crusoe reduces that friction.
“That's actually one reason we also like the idea of manufacturing a lot more of this infrastructure as opposed to making everything a giant construction project.”
Crusoe's Abilene project will provide more than a third of local tax revenue and more than double the tax receipts going to schools. Listen
He says local businesses in Abilene report their best times ever from construction-related spending. He acknowledges the legitimate downsides as construction traffic, dust and noise, which he says are temporary, while permanent jobs and tax revenue remain to fund police, fire, roads and schools.
“For the school system, we're more than doubling the tax receipts that are going to the schools.”
The claim that 50% of planned data centres will never be built 'seems reasonable'. Listen
He declined to give a more precise figure. He lists what can derail projects: permits, entitlements, land owners who won't sell, large-load interconnection agreements with the local utility, and air permits for new generation.
“50 % seems reasonable.”
Lochmiller describes Crusoe as an 'AI supermajor' that sells data centres, GPUs and tokens, so vertical integration acts as a natural hedge when margin shifts between layers. Listen
He compares the AI stack to oil and gas: upstream, midstream and downstream. Exxon famously doesn't hedge its oil exposure because when crude prices fall, its downstream margins on gasoline and plastics rise. He expects margin in AI to move around across electrical, data centres, chips and services in a similar way.
“Well, they're sort of inherently hedged by being vertically integrated.”
Managed compute clusters are currently the highest-margin layer of the AI infrastructure stack because supply is so short. Listen
Asked which layer has the best margin today, Lochmiller names managed GPU clusters. He says this is true 'right this instant' and attributes it to a massive shortage of supply.
“Managed compute clusters is like incredibly high margin, like right this instant.”
Crusoe's GPU rental agreements are typically take-or-pay, and he does not see softening consumer AI demand as the first sign of trouble. Listen
Under take-or-pay, a customer pays for committed capacity, such as 100 megawatts, whether or not it uses it. Asked whether a crack in consumer demand, for example for ChatGPT, would be the first warning sign, Lochmiller says he doesn't think of it that way. His reason is that AI is transforming many areas of the economy rather than a single product.
“The GPU rental agreements are typically take or pay. Yes.”
Long build timelines force customers, including early-stage startups, to forecast compute demand years ahead, and Crusoe is building small modular data centres to offer just-in-time capacity. Listen
Crusoe builds its forecasts from customer conversations, but he calls longer forecasting horizons problematic for the industry. Startups are being asked to commit now to compute for 2028, which he calls 'an eternity'. Crusoe's answer is small, modular, factory-made AI data centres that cut time to delivery, especially for smaller clusters.
“earlier stage companies, earlier stage startups that are being asked to commit to compute in 2028.”
Lochmiller calls dollars per token a good but imperfect unit for the cost of intelligence, because not all tokens are equally valuable. Listen
He was responding to Crusoe investor Gavin Baker's claim that the lowest-cost producer of intelligence wins. Lochmiller says that alongside cost per token, buyers optimise for throughput (tokens per second for a given model use case) and latency (time to first token and time to last token).
“Not all tokens are created equal, so it's not like every single token has the same, you know, unit of value, but I think it is a good indicator.”
Cheap, fast inference depends on keeping GPUs busy, mainly through KV cache management across memory tiers. Listen
He calls the GPU the most valuable and expensive thing in the data centre, so idle time is 'money burning'. The KV cache lets previously seen tokens be looked up rather than recomputed, but it can outgrow on-chip HBM. Moving data efficiently across GPU HBM, system DRAM, NVMe drives and the cluster's object storage is what lets a provider perform well on latency and throughput.
“the GPU is actually the most valuable thing in the entire data center. It is like the most expensive thing. So it's the thing you want to make sure you're keeping busy.”
From Crusoe's managed inference vantage point, more money goes to closed frontier models but more tokens are generated on open-source models. Listen
He thinks open source will be very important, partly for data sovereignty because companies want to own their models and intelligence. He also thinks there is a lot of untapped private data that could produce large performance gains.
“I do think that people are spending more money on closed source frontier models than they are on open source, but they are generating more tokens on open source than closed source.”
Lochmiller expects a mix of custom company models and closed frontier models, with neither approach dominating. Listen
Asked whether enterprises training their own models on private data will eat into frontier labs' business, he calls it possible but says there will always be demand for the frontier. Private data can be added to closed frontier models or to self-owned open-source models. He points to Cognition post-training its own model and Harvey possibly doing similar in legal, and notes the frontier labs are also pursuing domain-specific knowledge.
“I don't think it's going to be like one approach is going to be the dominant approach.”
Lochmiller sees more room to collaborate than compete with specialised players such as Fireworks across the AI infrastructure stack. Listen
Asked whether Fireworks competes with Crusoe's managed inference, he says Crusoe competes at multiple layers and that Fireworks does a great job on certain workloads and user experience. He uses the oil analogy: Exxon competes with other upstream drillers but sells pipeline capacity to those same competitors.
“I just view there are so many more ways for us to work with folks across the AI infrastructure stack than compete with them.”
Crusoe is probably better off as a public company because of its capital needs, but has not decided when to list. Listen
Asked whether Crusoe will be public by the end of 2028, he says he isn't sure. He points to the large amounts of capital needed for data centres and GPU clusters and the access to scaled capital that public markets provide.
“we do think that ultimately the company is probably better off in the public markets. It's just a matter of like, when does that make sense for us?”
Crusoe makes 'think like a mountaineer' a core value, meaning contingency plans, endurance and a strict safety culture. Listen
Lochmiller says mountaineering requires a plan A, plan B and plan C for weather changes, gear failure, a sick partner or avalanches, plus a willingness to endure a long, hard expedition. He applies this to a business operating in both the physical and digital worlds. He extends it to safety: 'safety first, second, and third' on large construction sites and protection against digital attacks on the platform.
“there's this saying in mountaineering that getting up to the top is optional, getting down is mandatory.”
Starting Crusoe after earlier financial success let him take bigger risks, which can lead to bigger outcomes. Listen
Lochmiller had worked in quantitative finance and had a reasonable amount of financial success before founding Crusoe after climbing Everest in 2018. He says knowing he could cover rent and feed his family even if the company went to zero was psychologically empowering. That security made him willing to take bigger swings.
“it's empowering to take bigger swings, empowering to take bigger risks, which can ultimately lead to bigger outcomes.”
People are emotional about data centres because they are the physical face of AI job fears, and the industry needs to make its case with data. Listen
He says data centres have become a polarising issue ahead of the US midterms because people worry about losing jobs to AI. He argues that so far AI has created a large number of skilled-trade jobs and driven US reindustrialisation. He says the industry must 'storytell effectively' using facts on jobs, energy costs, water and tax revenue.
“the biggest aspect is people are very emotional about data centers.”
Lochmiller protects family time by taking red-eyes, staying home most weekends and being unreachable from 7 to 8 a.m. when he is home. Listen
He says he often takes worse flights and red-eyes so he doesn't miss his kids' bedtime. He spends weekends coaching soccer or helping with swimming. When he is home in the Bay Area, everyone at the company knows he takes no meetings from 7 to 8 a.m. while he makes breakfast and gets his kids ready for school.
“So if I'm home in the Bay Area, everybody knows that I'm unreachable from 7 to 8 a .m.”