Michael Saylor violated his own forward guidance, bought back preferreds and compressed his dividend runway, and is now open to selling bitcoin. Jamie McAvity and I dig into the time bomb, bitcoin mining's generational buying window, and why the AI data center gold rush is sounder than 2021.
↓ Jump to the video and timestamps
I've been a hat tipper on Michael Saylor since Strategy started stacking. Not a fanboy. I had one of the most combative podcast episodes he's ever recorded, back when he was pushing the Bitcoin Mining Council and I was telling him to stop appeasing the ESG crowd. But on the accumulation strategy itself, I've tipped my hat. What he built was novel and, for a long time, defensible.
That changed. My antennae first peaked after the Q2 earnings call last year, when Saylor gave explicit forward guidance on ATM mechanics and then negged on it two weeks later. That was weird. Since then it's gotten worse: preferreds mismanaged, dividend runway compressed, a slide deck quietly adding "sell bitcoin and buy USD" as a fifth strategic pillar after years of "we never sell bitcoin." And now there are derivative vehicles stacking on top of the whole structure: Strive, DeFi stablecoins backed by MSTR.
The caveman read is simple. This seems too good to be true, and you can only financial engineer so far before you need actual cash flow.
I brought Jamie McAvity back on to talk about it. Jamie runs Cormint, one of the lowest-cost bitcoin mining operations in the country, and he's been working this intersection of mining, powered land, and AI compute longer than almost anyone I know. We covered the Saylor situation, the state of bitcoin mining right now (short answer: generational buying window), and why the AI data center gold rush is an order of magnitude bigger than 2021 but structurally different enough that it probably doesn't end the same way. Here's what we got into.
I pulled up mempool.space during our conversation and watched a negative 10% difficulty adjustment sitting about 31 hours out in real time. That tells you everything you need to know about where we are in the cycle.
Jamie's framing was sharp on this. The crowding of the bitcoin mining market over the last few years accelerated the transition toward GPU compute. Difficulty kept climbing in ways that were "unconstrained by economic reality for a long period of time," which made the economic case for new bitcoin mining investment look bleak. So capital rotated out. Now almost nobody is investing in new bitcoin mining capacity among the US public companies.
The market is weak and, in his words, unsexy. That's the buying signal. ASICs are being offloaded by miners pivoting to GPU compute. The difficulty is adjusting down. Jamie's view, and mine, is that ASICs are going to be a good buy throughout the summer, maybe through the end of the year. The new paradigm Jamie is describing is different from the old one.
You run ASICs for longer. You co-locate as close to energy production as possible, ideally right at the source. The fee market materializes because block times get volatile, driven by the cheapest energy inputs on the planet: solar and flare gas.
When the sun is hitting a region with heavy mining, blocks come fast, fees are low. When it's not, the backlog builds and fees spike. That volatility is the fee market actually working, and it's a healthy departure from the previous era where everything was running 95% of the time on a relatively flat fee schedule. The hosting industry that built up around the old paradigm, in Jamie's view, never should have existed. That era is over.
The comparison to the 2021 bitcoin mining frenzy is one I've been making for a while. After the Chinese mining ban, everybody and their mother decided they were a power and infrastructure expert overnight. Eyes lit up, capital flooded in, and the mining economics at that moment were extraordinary. Then FTX collapsed, bitcoin went to $17,000, and a lot of that capital got wiped out.
Jamie's case is that the AI data center gold rush is running on a structurally different chassis. You're not underwriting on spot commodity economics that can swing 90% in a year. You're signing 10-to-15-year leases with investment-grade counterparties, which you can then finance with 80% debt. The economics are locked in for a decade and a half. That doesn't mean there won't be a correction. Jamie is explicit that there will be bankruptcies from over-levered players, some leases will get broken, there will be litigation, but the core capital structure is far more resilient than bitcoin mining was in 2021.
The demand side is where my head spins. The agentic economy is maybe five or six months old depending on who you ask, and less than half a percent of the population has actually tried to implement any of it. Robotics hasn't even entered the picture yet. The regulatory barriers for full self-driving are beginning to ease. I can't fathom what the demand for compute looks like five years from now, and I mean that literally.
It's not a rhetorical hedge. In the ERCOT queue alone, there are 350 gigawatts of projects waiting on an 85-gigawatt system. I genuinely cannot model it.
A buddy at Luxor and I landed on a crude oil grades analogy in Austin last month. Different compute as different grades of crude, data centers as the refineries. Jamie thought the analogy was directionally okay but incomplete. His better comparison was the early days of GPU mining niche coins like Grin, where you had simultaneous price discovery in the token, difficulty climbing as new compute piled in, and software improvements that made existing hardware more productive all at once. The frontier AI labs are in a similar position: the quality differential between models a year apart is enormous, commodification hasn't happened yet, and nobody has a clean answer on whether the next massive training run is worth the capital.
Jamie and I have been talking about the grid for years. Since we last spoke, the political environment has done a 180. Pro-nuclear, pro-natural gas, grow-the-grid-at-all-costs is now the stated direction from the top. What surprised both of us is how strong and, in Jamie's word, baseless the opposition has been.
The misinformation hitting AI data centers is the same playbook we watched get run against bitcoin mining: they'll use all the water, they'll drive up your electricity bills, local communities will be harmed. It's uninformed bullshit, and people are buying it.
My take on who benefits from that misinformation: China. They want to slow down American data center expansion because they want us buying AI tokens from them. Every protester using ChatGPT to make their anti-data-center flyer is doing the most ironic thing imaginable. Don't you want the most capable AI in the world built in a country where you still have rights and can push back against the state?
The alternative is an authoritarian surveillance state that deploys AI on its own citizens without restraint and without any mechanism for the population to resist. The opposition to American data center build-out is, at minimum, doing that country's work for it.
Texas and ERCOT stand apart from this. Lowest power prices, fastest-growing renewable sector, market-based power economy. Jamie's point about why high renewable penetration elsewhere causes price problems is worth sitting with: intermittent generation doesn't increase the capacity factor of a grid to meet peak demand. When thermal generators (the ones that run close to around the clock) have to absorb long stretches of suppressed pricing from excess renewables, their economic returns corrode and you lose the reliable baseload you need.
California is the cautionary tale. Texas built the model that works. The community economics in rural Texas are particularly good. Property tax from data center investment flows directly to school district budgets. If you're an electric cooperative, the adder on data center power goes back as a rebate to residential customers. Jamie sees genuine home runs for rural Texas communities that have available power and the right attitude.
This is the section I want to be precise about, because I am not calling for a short on MSTR and I'm not pretending to know exactly how the financial engineering unwinds. What I am saying is the caveman read is flashing.
My antennae peaked after the Q2 earnings call last year. Saylor gave specific forward guidance on the mechanics of their ATM and the guardrails they were putting on themselves for when they'd use it. Two weeks later they negged on that guidance. That was the first signal.
Since then, here's what I'm watching. He bought back the preferred convertible debt and reduced the dividend runway on the preferred instruments from roughly 18 months down to roughly 6 months, by Jamie's read. That's when the market started selling off. Then came the slide deck with a new fifth pillar: sell bitcoin and buy USD. After years of "we never sell bitcoin" and "always increasing sats per share," that's a reversal. Jack Mallers confronted him in Prague a couple days before we recorded this.
Ten-minute answer that was essentially a non-answer. That same day Saylor excluded Strike from a slide listing emerging bitcoin companies, which is, as I said on tape, more than a little pedantic given everything going on.
Jamie's message to Saylor was direct: you brilliantly financial engineered your way into a massive bitcoin position, and now you need to figure out how to create actual intrinsic value with that bitcoin. The jig is up on the financial engineering. His line that stuck with me: "The more that you deceive people from here on out, the more that this will be the high watermark on your reputation."
Mine is simpler: you can only financial engineer so far, but you need cash flow if you're running a business.
The thing that makes this more than just a single-company problem is the derivative stack. Strive, DeFi stablecoins backed by MSTR. You're building layers on top of a structure that is wobbling. If the deception at the base of that stack persists and something cracks, the contagion doesn't stay contained to MSTR equity.
One more dimension on the financial engineering angle. Saylor has been public about using AI to create these financial products. I've touched this technology firsthand and it's genuinely powerful. But it's also fallible, and it's sycophantic.
These are sophisticated calculators that can give you wrong answers and will tell you what you want to hear. How much of the product engineering here is a model being sycophantic back to a person who is already convinced the idea is genius? That's a real question. How much of the product engineering here reflects a model amplifying conviction rather than stress-testing it?
Jamie's core thesis on the competitive advantage side is one I think is underappreciated by the traditional infrastructure world.
Bitcoin mining built data centers oriented around one thing: convert electricity into bitcoin as cheaply as possible. Jamie says you can build a data center for $100,000 per megawatt including all materials and labor from a high-voltage input to the plug. You can run it at 2 cents per kilowatt hour. You can run it for free. And critically, the more you embrace flexible operations and invite downtime into your design philosophy, the lower your total cost of ownership. Fortune 100 companies built from the opposite starting point.
Zero downtime, always. When AWS goes down or Instagram goes down, the revenue loss is thousands of dollars per megawatt-hour equivalent: reputation damage, broken commerce streams, frustrated users. The design philosophy reflects that imperative.
AI compute is a different animal. It's commodity production. Marginal cost of production is what matters. Bitcoin mining data center developers who built and operated at the lowest possible cost are sitting on a genuine edge here.
Colossus is the proof of concept. Elon and crew built that thing in 122 days. When we got a look at how they did it, it was obvious: they thought like a scrappy bitcoin miner. Daisy-chaining gen sets, battery walls on the inside, no-downtime orthodoxy thrown out the window in favor of speed and cost.
Jamie's own best build was 110 days at $104,000 per megawatt. Those are his numbers for his own operation. Colossus confirmed the approach scales. His caveat is worth noting: if you need to order everything from scratch through the supply chain, 100 days isn't possible just due to lead times. Having inventory matters. But the philosophy is replicable, and it's where bitcoin miners have something real to teach the people who've been building data centers for Fortune 100 clients.
A few threads worth capturing that don't fit cleanly in the main sections above.
On AI and humanity: I tweeted the morning we recorded that I've never been more bullish on humanity. I don't buy the permanent underclass meme. We've been building a company brain at TFTC using agentic flows and I can see, touch, and feel what it does. It's real. Our survey data backs it up: roughly 60% of TFTC subscribers said they want more AI content. I think bitcoiners are actually better primed to receive this technology than most, because they've already done the mental work of accepting that a disruptive technology can fundamentally change the world even when powerful incumbents fight it.
On Satoshi's coins: Jamie and I are aligned. You can't touch them. His hope is that some well-resourced entity cracks the quantum problem and burns them as a flex, which would be genuinely sick. But I had a conversation in New York this week with a contact I won't name who made an argument I think a lot of people are completely overlooking.
His case was that Satoshi is alive, has a long-term plan, is waiting for the market cap to reach a threshold, and intends to deploy those coins as a philanthropist. We hold Satoshi up as a deity and assume the coins are either lost or untouchable forever. What if he's just patient? I'm not saying it's likely. I'm saying it's worth taking seriously as a third option rather than dismissing it.
On the bitcoin security budget: not panicking. Jamie has moderated his concerns somewhat and thinks the fee market will materialize when block time volatility gets real. My view is that making bitcoin useful enough on every layer (every protocol layer and every layer above it) creates demand for UTXOs.
Jevons Paradox applies here. The more efficient UTXO usage becomes and the more optionality we build around individual UTXOs, the more demand flows to them. Maybe that's blind confidence. But we're going to find a way.
The setup is as favorable as it's been in about five years. A negative difficulty adjustment is in progress, the major US mining companies are rotating capital toward GPU compute, and hardware is being offloaded. Prices on second-hand ASICs follow difficulty and miner sentiment, and both are pointing down right now. If you have access to cheap power and a long time horizon, the summer into year-end looks like a window.
The biggest structural difference is the revenue contract. Bitcoin mining economics in 2021 were underwritten on spot commodity prices that could swing 90% in a year. AI data center deals are being signed as 10-to-15-year leases with investment-grade counterparties, financeable at 80% debt. The capital structure doesn't blow up on a single price move. There will still be corrections and bankruptcies among over-levered players, but the base of the thesis is fundamentally more durable.
The proximate trigger, by Jamie McAvity's read, was Strategy buying back preferred convertible debt in a way that compressed the dividend runway on their outstanding preferred instruments from roughly 18 months to roughly 6 months. That spooked the market. Compounding that was a slide deck that introduced selling bitcoin and buying USD as a strategic option, a direct reversal of years of "we never sell bitcoin" messaging. Credibility on forward guidance had already eroded after ATM mechanics guidance was walked back within two weeks of the Q2 earnings call in 2024.
Strategy's mNAV (the premium of its market cap over the net asset value of its bitcoin holdings) has been used as a key marketing metric. Critics, including Jack Mallers at a recent Bitcoin conference, have argued that the way Saylor presents and uses mNAV is misleading to investors, particularly as the financial engineering has grown more complex and the gap between the metric's presentation and the underlying economic reality has widened. Saylor's extended non-answer when pressed on it publicly has not helped the situation.
ERCOT runs a deregulated, market-based power economy largely isolated from the national grid. That structure keeps prices responsive to supply and demand rather than rate-regulated cost recovery. Texas also has the fastest-growing renewable sector in the country, which adds generation capacity, and has historically maintained a strong investment environment for thermal generation (natural gas, coal, nuclear) that provides reliable baseload. The combination of competitive markets, diverse generation mix, and continued transmission investment is what other grids should be studying.
The bitcoin mining paradigm was built entirely around minimizing cost per unit of commodity produced. That means building at low capital cost per megawatt, operating at the lowest possible electricity rates, and treating downtime as a design feature rather than a failure mode. Traditional Fortune 100 data centers were designed around zero-downtime SLAs because outages cost those customers thousands of dollars per megawatt-hour equivalent in lost revenue. AI compute is commodity production: marginal cost is what matters. Bitcoin miners who built for cost rather than uptime are ahead of where the rest of the industry is heading.
The concern is that as the block subsidy continues to halve, transaction fees need to grow enough to keep miners economically incentivized to secure the network. The worry is that fees won't scale fast enough. My view is that making bitcoin useful enough across every layer (base layer, Lightning, every protocol built on top) creates organic demand for block space that drives fees. Jevons Paradox: make UTXO usage more efficient and you tend to get more UTXO usage, not less. Jamie has moderated his concerns on this somewhat and now thinks volatile block times driven by intermittent energy sources will create natural fee spikes that partially solve the problem. Neither of us is panicking.
Jamie McAvity is the founder and CEO of Cormint Data Systems, a Texas-based bitcoin miner and one of the lowest-cost producers in the country, operating a 130-megawatt facility in Fort Stockton. He founded Cormint in 2018 after a career in commodities and derivatives trading, and was recently appointed to Texas's Strategic Bitcoin Reserve Advisory Committee.
