Brandon Bailey on why bitcoin miners accidentally built the most valuable asset in the AI compute boom, which companies are executing the transition, and why the whole setup is net-positive for Bitcoin and the network.
↓ Jump to the video and timestamps
I've been in the bitcoin mining world long enough to watch the whole thing get built from scratch. The China mining ban, the gold rush to North America, the 2021 bull cycle where everybody was locking down land and power as fast as they could, the brutal bear market that followed. It was a steep learning curve and a lot of companies didn't survive it.
Here's the thing nobody who's currently panicking about AI "stealing" bitcoin's thunder is saying: everything those miners went through to acquire power is now the most valuable skill set in the entire AI infrastructure build-out. The hyperscalers showed up with trillion-dollar capex budgets and no idea how hard it actually is to get megawatts energized. The bitcoin miners figured that out the hard way. That means they're sitting on the scarcest asset in this whole boom.
I sat down with Brandon Bailey, who I've known for seven or eight years now from his time at Galaxy and beyond, and who just launched DI Metrics, a market intelligence platform built specifically to track this intersection of bitcoin mining and AI compute. It's one of the best conversations I've had on this theme, and I think Brandon is one of the sharpest analysts working this particular layer of the infrastructure build-out. Here's what we covered.
Brandon's framing of how this all started is worth laying out carefully, because the history explains everything that's happening now.
After the China mining ban, which took roughly half the network's hash rate offline almost overnight, there was a mad dash to reestablish bitcoin mining compute in North America. That's when you saw the first wave of publicly traded miners: Marathon, Riot, Clean Spark. The institutionalization of the industry was underway, and with it came a race to lock down land and power.
The economics of that moment made it all make sense. Bitcoin had come off lows that touched roughly $3,100 in the 2018 bear market and ran into the 2021 cycle. The margins on mining were extraordinary. Every company with access to capital was acquiring gigawatts of power and planning out 100, 200 megawatt data centers to run ASICs.
Then the cycle turned. The bear market hit. A lot of those power portfolios sat underutilized or got stressed. And then ChatGPT arrived.
What those miners didn't realize in 2021 was that the painful, expensive, time-consuming process of locking down land, negotiating interconnection agreements, surviving utility queues, and actually getting megawatts energized had given them something priceless. Energized power is the single biggest bottleneck in the entire AI compute build-out right now.
I watched this play out from the inside during the 2021 era, and I can tell you the learning curve was brutal. The AI world is now going through that same education by brute force, with hyperscalers throwing tens of billions of capex at a problem that can't be solved by money alone. You can't bully your way past a utility queue. The bitcoin miners already know that. The hyperscalers are just figuring it out.
This is layer zero of everything we're building. Energy infrastructure is the prerequisite for the agentic economy, for robotics, for all of it. And the bitcoin miners, almost accidentally, built the portfolio that everyone needs.
Bailey's comp analysis here is the clearest explanation I've heard of why this is such a compelling setup for investors.
Bitcoin mining companies today trade at roughly 4-6x EBITDA by his reckoning as an analyst in this space. Traditional data center operators like Digital Realty Trust and Equinix trade closer to 20-24x EBITDA. That's a massive discount applied to the exact same underlying asset, energized power capacity, purely because of how it's being used.
The reason for the gap is straightforward. Bitcoin mining revenue is volatile. You've got the halving cutting your economics in half every four years. You've got bitcoin price swings. You've got ASIC depreciation. The cash flow profile looks nothing like a data center running a 10-15 year lease with a Google or Amazon as the tenant.
That's the key insight. Once you swap out bitcoin mining as the off-taker and replace it with a hyperscaler lease, you've transformed the cash flow profile of the business. Guaranteed revenue from the most creditworthy tenants on earth, with debt financing available at favorable rates rather than constant equity dilution. Bailey's point is that you'd take that deal every single day if you could get it.
And the re-rating, when it happens, is violent. Core Scientific signed its first CoreWeave lease and the market moved dramatically, because the probability the market had assigned to that ever happening was essentially zero. Cypher, Hut 8, and Terawulf saw similar moves when they signed their first leases. Bailey estimates some of these companies ran 5-7x from their pre-lease lows to where they traded after signing.
From a cap rate perspective, his framework is that stabilized traditional data center operators like Digital Realty Trust are being valued at roughly a 5.5-6% cap rate. Many of these miners, even after signing leases, still trade at 8.5-9% cap rates because of execution risk. As companies deliver powered shells on time and construction risk goes to zero, that spread narrows toward parity. Each milestone closes the gap.
Bailey names the companies that have signed leases: Core Scientific, Hut 8, Cipher Mining, Galaxy, Terawulf, and Applied Digital. He gives special credit to Core Scientific as the pioneer, both for being first to sign and for being first to actually deliver energized buildings.
Core Scientific has roughly 590 megawatts of critical IT capacity leased with CoreWeave, and they've begun delivering on that capacity. That delivery is what matters most right now, because the market still assigns an execution discount to companies that have signed but haven't built.
And that execution challenge is real. I asked Bailey directly about this, because I think people underestimate how different building an AI compute facility is from building a bitcoin mining operation.
Bitcoin mining data centers were built with one philosophy: get it up fast and cheap. You wanted machines hashing as quickly as possible because difficulty was climbing and every day offline cost you. Cutting corners was sometimes rewarded. The "get creative, get it done" mentality was an asset.
AI compute facilities are the exact opposite. Uptime is everything. Redundant systems are non-negotiable. The cooling requirements, the power delivery architecture, and the intricacy of every component are orders of magnitude more complex than a bitcoin mining shed. The hyperscalers hand you a design spec and say build it to this standard exactly.
The question isn't "can you get creative." It's "can you execute to spec." That's a different discipline, a different team, and a different relationship with the operator.
Phase one is signing the lease. Phase two is delivering the building. Core Scientific has proven both. The others are working through phase two now. For investors, Bailey's point is that each successful delivery removes the construction risk discount from the stock. Tracking delivery milestones matters for that reason.
This is the part of the conversation that addresses the bubble question. The dot-com comparison doesn't hold, and Bailey agrees.
In 1999, a lot of what was being built had no revenue, no real demand, and no path to profitability. What's happening now is different. You're already seeing real revenue from Anthropic, real usage numbers from Google's Gemini, real enterprise adoption. The demand side of the equation is already showing up in earnings calls.
But more importantly, the reason the capex doesn't stop is game theory, and Bailey breaks it down at two levels.
First, the nation-state level. The United States and China are in a genuine arms race over AI leadership. The US has to foster an environment where this industry can flourish and make sure its builders have access to the raw inputs needed to win. That's a strategic imperative, not a discretionary investment.
Second, the hyperscaler level. Google, Amazon, Microsoft, Meta, Nvidia, SpaceX. These companies built massive cloud moats over the last decade and they all understand that AI threatens to upend those moats if they underinvest. None of them can afford to let a competitor end up with superior data center capacity. The logic of mutual assured investment keeps every one of them spending aggressively. Alphabet raised about $85 billion in June 2026 to fund its AI build-out, with Berkshire Hathaway anchoring $10 billion of the offering.
The thing you can't afford to do is underinvest and wake up three years from now watching a competitor run away with market share because they had more compute. So the money keeps coming.
I've also been running agentic AI infrastructure at TFTC for about two months now, using Claude Opus as the brain of our system, building out a company knowledge graph and semantic search so our agent has context before it burns tokens. The adoption curve for enterprise use is still very early. And then you layer in robotics, which hasn't even really started yet, and the token demand projections get almost incomprehensible.
I raised Jevons' paradox in the conversation. The worry is that as tokens get cheaper and more efficient, demand destruction follows. Bailey's read, which I share, is the opposite: as tokens get cheaper, consumption expands to fill the space, plus entirely new use cases open up that weren't economically viable before. You do not outrun demand for intelligence. Not in this environment.
This is the warning section, and it's important.
Bailey is explicit: a large portion of what's being marketed as interconnection pipeline will never be built. His estimate is that somewhere between 10-30% of the overall queue or interconnection requests will actually be permitted and constructed on a reasonable timeline. The rest is people who bought land, got into a queue, and are now selling the dream of future power without load studies, without real interconnection agreements, without actual line of sight to energization.
I've seen this exact movie before. In 2021 and 2022, after the China mining ban, there were people all over the country marketing power deals they didn't actually have. "We're in the ERCOT queue, we'll have 200 megawatts in 18 months." The queue was real. The timeline was not. A lot of capital got deployed against phantom power.
The AI build-out is running the same playbook at a scale that's an order of magnitude higher. The interconnection queue backlog is a documented, structural problem across every major grid in the country, with years of wait time and significant attrition between application and actual energization.
For investors, Bailey's point is that figuring out what pipeline is real versus what's being marketed is the core analytical challenge in this space. Companies with already-energized power don't carry that ambiguity, which is the foundational advantage bitcoin miners hold.
Here's where I want to push back on the narrative that AI is bad for bitcoin mining. It's not. Everything's bullish for bitcoin.
Bailey's medium-term thesis is that AI compute accelerates something miners have been talking about for years: bitcoin mining gravitating toward stranded energy. Grid-connected power is going to get allocated to AI compute, particularly at scale. That pushes bitcoin mining toward 1-5 megawatt stranded sites, distributed around the world, closer to the energy source.
That's actually better for the network. More geographic distribution, smaller operators, less concentration. Bitcoin mining has historically been too concentrated in too few jurisdictions. This pressure pushes it toward genuine decentralization.
There's also a potential hash price dynamic Bailey raises that's genuinely interesting. If grid-connected power gets absorbed by AI compute and miners can't quickly plug in ASICs when bitcoin economics turn favorable, you could end up with less total network power chasing the same block rewards. Hash price appreciates. Mining economics potentially enter a more durable, less crowded golden era.
He also flags the ASIC supply glut problem. Bitmain, MicroBT, and Bitdeer have been mass-producing machines to preserve foundry capacity regardless of demand. The result is an oversupply of ASICs sitting in warehouses. If AI compute absorbs grid-connected power and the addressable market for grid-connected bitcoin mining shrinks, the supply-demand balance for ASICs could normalize, which would make mining economics more sustainable over the long run.
And for countries outside the United States that aren't capturing the AI compute build-out, there's going to be a home for bitcoin mining. While I'd always prefer more American hash rate, a world where hash rate is genuinely distributed across multiple jurisdictions and energy sources is better for the network than the alternative. The China concentration lesson was painful. We shouldn't recreate it.
I want to spend some time on what Brandon actually built, because it's a good example of what these tools make possible right now.
DI Metrics started as a spreadsheet from his Galaxy days. He was manually tracking sites, power capacity, hash rate, all the granular information you need to actually invest in or analyze bitcoin mining companies. As those companies started transitioning to AI compute, the manual tracking got more complex and more time-consuming. Combing through SEC filings, investor presentations, press releases, trying to keep up with a rapidly expanding universe of companies and deals.
He asked whether the LLMs could automate that workflow. The answer turned out to be yes, even with no coding background whatsoever.
What he built is a market intelligence platform that goes granular in a way that Bloomberg or FactSet doesn't. Not just the balance sheet line items, but every site a company like Cipher Mining owns, the gross and leasable power capacity of each site, energization timelines, utility providers, lease rates, cap rate comparisons. The kind of information that previously required hours of manual research to compile.
The MCP connection is what makes it really useful. You plug DI Metrics into Claude or Codex, and instead of hours of manual work, you're asking natural language questions: give me every lease Cipher has signed and the rate, rank these companies by gross margin on their leases, show me the cap rate spread between Terawulf and Digital Realty Trust. That kind of analysis becomes a prompt.
I've been playing with it and it's genuinely impressive for anyone who's trying to stay on top of this infrastructure layer.
His point about the broader AI bottleneck trade is also worth emphasizing. Groups like SemiAnalysis and Fundstrat do excellent work on chips and semiconductors, but the data center layer, specifically these bitcoin miners converting capacity, is underanalyzed relative to its importance. DI Metrics is an attempt to fix that.
On the AI implementation side, his personal learning curve mirrors mine. It started with "this changes everything," ran through a phase of "it's missing context, I have to prompt better," and came out the other side with a much clearer picture of what actually matters: giving the model the right context upfront, building systems that don't require constant iteration, and understanding where the models are genuinely strong versus where they'll hallucinate or miss.
My own experience at TFTC is that building persistent memory for your agent is the key advantage. We've been building our company knowledge graph and semantic search for about two months now. The agent pings that system first before burning tokens, gets the context it needs, and can one-shot things in our voice. Once you get that infrastructure in place, the productivity gains are real. The implementation details matter. Token maxing without intent is just burning money.
Bailey made a point I want to land on. Today's frontier models were trained on a fraction of the compute that's now being energized for the next generation of models. We're talking about the model improvements we've seen so far being trained on a tiny slice of what's coming online. If what we have today is already this capable, the step changes over the next two years are going to be mind-bending.
Bitcoin miners spent years navigating utility queues, negotiating interconnection agreements, and actually getting megawatts energized. That process is the primary bottleneck in the entire AI compute build-out right now. Hyperscalers can throw unlimited capex at the problem, but you can't brute-force your way past a utility queue. Miners who already have energized power are holding the scarcest asset in the build-out.
Bitcoin mining was built around speed and cost efficiency. Get the machines online fast, maximize the payback period, cutting corners was sometimes rewarded because hash rate was racing against you. AI compute data centers are the opposite: they require strict uptime, redundant systems, sophisticated cooling, and construction to a precise design spec handed over by the hyperscaler. It's a fundamentally different discipline, and not every miner can execute it.
By Bailey's analysis, the companies that have signed leases are Core Scientific, Hut 8, Cipher Mining, Galaxy, Terawulf, and Applied Digital. Core Scientific is the pioneer, the first to sign and the first to actually deliver energized buildings. The others are executing through construction and delivery now. Smaller miners are beginning to explore JV structures to capture the opportunity with less execution risk on their balance sheets.
I don't think the parallels are as clear as the doomers are claiming, and I said so on tape. The dot-com era was dominated by companies with no revenue and no real demand. The AI build-out already has real revenue, real usage, and enterprise adoption showing up in earnings. More importantly, the capex is driven by game theory at the nation-state and hyperscaler level. Neither the US government nor Google can afford to underinvest and lose ground, so the demand is structural regardless of whether any individual application works out.
Bailey's medium-term thesis is that bitcoin mining gravitates toward stranded energy sources, smaller distributed sites of 1-5 megawatts, as grid-connected capacity gets absorbed by AI compute. That's actually good for the Bitcoin network because it drives decentralization. There's also a potential hash price appreciation scenario if less grid-connected power is available for mining when bitcoin economics improve, because miners can't quickly plug in ASICs.
Bailey uses a cap rate framework. Stabilized traditional data center operators like Digital Realty Trust trade at roughly a 5.5-6% cap rate. Bitcoin miners transitioning to AI compute, even after signing leases, often still trade at 8.5-9% cap rates because the market is pricing in execution risk. As companies deliver powered shells on time, that construction risk discount goes to zero and the spread narrows toward parity with traditional data center comps.
DI Metrics is a market intelligence platform Brandon Bailey built for investors tracking the digital infrastructure space, specifically the bitcoin mining to AI compute transition. It tracks sites, power capacity, energization timelines, lease rates, and cap rate comparisons at a granular level that general financial data platforms don't provide. It connects via MCP to Claude or Codex so you can query the database in natural language. It's built for anyone doing serious analysis of this infrastructure layer.
Brandon Bailey is a Bitcoin mining and digital-infrastructure analyst and the founder of DI Metrics, a market-intelligence platform that tracks the bitcoin-mining-to-AI-compute transition: sites, power capacity, energization timelines, lease rates, and cap-rate comparisons, queryable in natural language through an MCP connection to Claude or Codex. He spent years at Galaxy covering the mining sector, and Marty has known him for seven or eight years.
