SPEAKER_1: Alright, so last episode the key insight was that rare earths represent a geopolitical and processing-chain bottleneck—not just a raw-material story. Now I want to swing back into AI, but from a completely different angle. Because after the chip suppliers and infrastructure names rerated, the question becomes: who actually owns and sells the compute? SPEAKER_2: Right, and that's where neoclouds enter the picture. McKinsey describes them as independent GPU-as-a-service providers that emerged specifically in response to global scarcity of high-end AI compute. They're not hyperscalers. They're not chip designers. They aggregate GPU capacity and sell access to it—often faster and cheaper than the big cloud platforms. SPEAKER_1: So not the large hyperscale clouds. Something different in structure. SPEAKER_2: Exactly. McKinsey's analysis notes neoclouds can price GPUs up to about 85% lower than large cloud providers. They offer more flexible contracts, faster provisioning, and specialized configurations. For a generative-AI startup that needs compute now without a multi-year hyperscaler commitment, that's a meaningful alternative. SPEAKER_1: Mm-hmm. So the business model is essentially: secure scarce GPU capacity, then resell access to customers who can't get it elsewhere or can't afford hyperscaler pricing. SPEAKER_2: That's the core of it. And the scarcity is real. J.P. Morgan's analysis describes a compute scarcity regime—orders for Nvidia GPUs reportedly totaling around one trillion dollars through 2027, major data-center markets running at roughly 1% vacancy, with more than 90% of capacity under construction already pre-leased. When supply is that tight, anyone who controls GPU capacity has pricing power. SPEAKER_1: Now, IREN is one of the named cases here. And most people know IREN as a Bitcoin miner. So walk through what actually changed. SPEAKER_2: IREN—formerly Iris Energy—pivoted around 2023 to 2024. Management started aggressively diversifying into HPC and AI cloud services rather than expanding Bitcoin mining capacity. By late 2025, the AI cloud fleet had grown to roughly 23,000 GPUs. And the company guided for more than 500 million dollars in annualized AI cloud run-rate revenue by end of Q1 2026. SPEAKER_1: Wait—so a Bitcoin miner becomes an AI cloud provider. What made that pivot actually viable? Because those sound like completely different businesses. SPEAKER_2: [short pause] Here's the counterintuitive part. Bitcoin mining campuses are already engineered for high power density and cooling. The physical infrastructure—power connections, cooling systems, high-density racks—translates directly to hosting liquid-cooled GPU clusters. IREN's Childress, Texas campus has power capacity increased to around 750 megawatts. IREN is also developing a dedicated 75-megawatt liquid-cooled AI and HPC data center in Texas, engineered to support roughly 200 kilowatts per rack and host Nvidia Blackwell-generation GPUs. SPEAKER_1: So the asset that made them a good Bitcoin miner—cheap, abundant power in a high-density facility—is exactly the asset that makes them a viable AI compute provider. SPEAKER_2: That's the key idea. And Reuters reported that acquiring or leasing space from Bitcoin miners with substantial power capacity can shorten the time to launch a new data center by roughly three and a half years. That time compression has real economic value when AI demand is outpacing new supply. IREN has secured approximately 2.75 gigawatts of grid-connected power capacity across West Texas—that's not just a data center asset. That's optionality. SPEAKER_1: Optionality is the right word. Because they can shift capacity between Bitcoin mining and AI cloud depending on which pays better at any given moment. SPEAKER_2: Exactly. Analysts covering IREN specifically highlight that the large power pipeline in West Texas and Canada gives the company an unusual ability to allocate between workloads based on relative economics. Think of it as turning power contracts into a tradable compute option. FY25 results showed quarterly revenue of 187.3 million dollars with triple-digit year-over-year growth—though Bitcoin mining still generated the majority of revenue at that point. SPEAKER_1: So the AI cloud segment is growing fast but still a minority of total revenue. That's worth flagging for someone tracking this as an investment signal. SPEAKER_2: It is. Multi-year AI cloud contracts cover around 11,000 of the 23,000 GPUs—representing roughly 225 million dollars of annual run-rate AI cloud revenue expected in operation by end of 2025. The contracted portion matters because it's the difference between compute-as-a-service revenue and speculative hardware ownership with no locked-in demand. Analysts note that incremental AI revenues can materially enhance data-center economics without requiring equivalent new capex. SPEAKER_1: Right—but what are the risks specific to this model? Because leverage, GPU depreciation, customer concentration—those aren't small concerns. SPEAKER_2: [emphasis] Several worth naming. GPU depreciation is real—Nvidia's roadmap moves fast, and hardware bought today may be undercut by next-generation chips within two to three years. Customer concentration is a risk if a small number of contracts represent most AI cloud revenue. Financing costs matter because building out GPU clusters requires significant upfront capital. And competition from large cloud providers is a real risk—if major hyperscalers undercut pricing, neoclouds can feel that pressure. Financial-sector research also notes that even older-generation GPUs retain strong rental markets, which provides some buffer, but the depreciation curve is still steep. SPEAKER_1: So the risks are real. But the structural case is that power access is as important as GPU ownership in this model—maybe more so. SPEAKER_2: That's the counterintuitive point worth landing. Power infrastructure takes much longer to build than data centers themselves—J.P. Morgan's analysis flags multi-year delays in securing electricity as a defining constraint. So a company that already holds gigawatts of grid-connected capacity isn't just an equipment owner. It's a scarce-resource holder. The market is increasingly pricing these sites as AI compute real estate, not just crypto operations. SPEAKER_1: The takeaway for someone following this chain: the bottleneck migrated again. From component suppliers—chips, foundries, memory—to the companies packaging and selling compute capacity to end customers. Neoclouds are the next link. SPEAKER_2: And that migration sets up the next episode directly. Because after commercial AI compute demand, the series moves to government-driven technology demand. Defense modernization is where procurement cycles, long-term contracts, and strategic hardware needs create their own rerating logic—and it follows a similar pattern of identifying who controls the scarce enabling layer. That's where the chain goes next. SPEAKER_1: Now, Nebius is another example of the neocloud model. Let's focus on how it aligns with the strategic logic of controlling scarce GPU capacity. SPEAKER_2: The emphasis should be on how Nebius exemplifies the neocloud model, focusing on its strategic approach to GPU capacity. Nebius, like IREN, leverages the neocloud model by controlling GPU capacity and offering flexible access to customers, bypassing hyperscaler constraints. SPEAKER_1: So two different paths to the same destination. One came from crypto mining, one from a legacy tech carve-out. But both are betting on the same structural gap. SPEAKER_2: Right. And that gap is real. McKinsey describes neoclouds as independent GPU-as-a-service providers that emerged specifically in response to global scarcity of high-end AI compute. The positioning is deliberate—complement the hyperscalers rather than compete directly. Focus on niche workloads, faster provisioning, and pricing that can run up to roughly 85% lower than large cloud providers. SPEAKER_1: Wait—85% lower? That's not a marginal discount. That's a different market entirely. SPEAKER_2: [short pause] It is. And that's why the customer base skews toward generative-AI startups that need compute now but can't absorb a multi-year hyperscaler commitment at hyperscaler prices. The neocloud fills that gap. Think of it like a flexible lease versus a long-term mortgage—same underlying asset, very different terms. SPEAKER_1: Mm-hmm. So for someone tracking this as an investment signal—what are the operating metrics that actually matter? Because GPU count alone doesn't tell the whole story. SPEAKER_2: Six things worth watching together. Installed GPU capacity. Megawatts of power access—and we've established why that's as important as the hardware itself. Utilization rate. Contracted revenue versus spot exposure. Customer quality and concentration. And financing costs, because building out GPU clusters requires significant upfront capital with a depreciation curve that moves fast. SPEAKER_1: That depreciation point is the one I keep coming back to. Nvidia's roadmap moves quickly. Hardware bought today could be undercut by next-generation chips within two to three years. SPEAKER_2: [emphasis] That's the pressure test. And it's why contracted revenue matters so much. Financial-sector research notes that even older-generation GPUs retain strong rental markets—neoclouds can offer lower-cost compute tiers on prior-generation hardware and still earn attractive returns. But the depreciation curve is steep, and uncontracted capacity sitting idle is a real risk. SPEAKER_1: So the difference between a viable neocloud and a speculative hardware bet is essentially whether the GPUs are earning revenue or waiting for a customer. SPEAKER_2: Exactly. Compute-as-a-service revenue with multi-year contracts is a fundamentally different business than owning GPUs and hoping demand shows up. For IREN, roughly 11,000 of 23,000 GPUs are covered by multi-year contracts—that's the contracted floor. The remaining capacity is where execution risk lives. SPEAKER_1: Now, the key idea from this episode—the bottleneck migrated again. Not from one component to another, but from suppliers to the companies packaging and selling compute capacity to end customers. SPEAKER_2: That's the right framing. And the market is increasingly pricing it that way. Analyst commentary describes the current environment as a compute scarcity regime—and in that regime, any operator who can bring GPU capacity online faster than hyperscalers captures a premium. Power access, existing infrastructure, and flexible contract terms are the scarce inputs now, not just the chips themselves. SPEAKER_1: The takeaway for someone following this chain: the rerating logic doesn't stop at component suppliers. It extends to whoever can aggregate and deliver compute to the end customer fastest. SPEAKER_2: And that sets up the next episode directly. After commercial AI compute demand, the series moves to government-driven technology demand—defense modernization. Procurement cycles, long-term contracts, strategic hardware needs. The same question applies: who controls the scarce enabling layer? That's where the chain goes next.