NurungJi🫎

We are officially 24 hours away from the Based TGE! Tomorrow, March 30, marks a massive milestone for our community.

오늘 많이 바쁠 예정. 일단 로켓랩 쪽에서 악재(뉴트론) 하나 터졌고, 그린란드 쪽 딜도 빨리 해결되서 방향성 잡는 구독글도 정리하고. 우주 산업 쪽에서 '발사주권' 관련 이벤트도 하나 살펴봐야할 듯 휴가 끝. 로켓슬라 풀가동 준비 중.

🔥 The Hidden 48B Alpha in the $MSFT - $IREN Deal 1⃣ The market is missing a key detail The market views the Microsoft x IREN deal as a static, 5‑year rental of GB300 GPUs. This view is dangerously simplistic. It ignores the single most critical engineering feature in the contract: -- "Flexible rack densities (130–200 kW)." Why demand 200 kW capability for a 130 kW rack in a fixed deal? The answer lies in what I call "The Empty Hall Paradox" and Double Optionality. 2⃣ The "Empty Hall Paradox" – Shell vs. Racks Think of a data center as two layers: •🧱 The Shell (concrete, power feeds, cooling loops) •🎛️ The Racks (GPUs) — swappable every 2–3 years By over‑provisioning Layer 1, they unlock a massive "hidden upgrade" for Layer 2. Imagine a 50 MW IT-load Horizon hall: • Scenario A (the Microsoft deal today): Racks run at 130 kW. You need 385 racks to fill the 50 MW power budget. The room is full. (50,000 kW / 130 kW ≈ 385 racks.) • Scenario B – The Trap: You upgrade per‑rack density to 200 kW but keep the total power cap at 50 MW. You hit 50 MW with only ~250 racks. (50,000 kW / 200 kW ≈ 250 racks.) Result: ~135 rack positions sit empty. About 35% of your expensive data‑center shell is wasted. -- Side note – Rack density, PUE & headroom Rack density is how many kW of IT you can safely run in one rack (130 vs 200 kW). PUE is total site power divided by IT power. Building a hall for 200 kW racks but only running ~130 kW today means the power and cooling plant has headroom to handle hotter future GPUs, short‑term load variation and firmware TDP bumps without tripping breakers. Early on that can make PUE look slightly worse (same overhead, less IT), but it’s exactly what lets you later densify, raise IT load and improve the effective PUE and $/MW over the life of the site. 3⃣Why Microsoft wanted 200 kW racks – The "Rubin Option" Microsoft didn’t demand 200 kW capability to leave the building empty. They demanded it to unlock a training‑cluster upgrade – call it the Rubin Option. By designing the shell to scale both rack density (up to 200 kW) and total power (up to 77 MW), they avoid the trap. They can later swap in next‑gen Rubin racks, run them hot, and keep the hall full. 4⃣ What the image shows – the three panels The image above visualizes this arbitrage. • In Panel 1 (blue): Microsoft is just renting the GB300 racks: 385 racks @ 130 kW = 50 MW. Shell full. • In Panel 2 (orange): density rises to 200 kW but the power cap stays at 50 MW. Only ~250 racks are populated; ~35% of the shell is stranded. • In Panel 3 (red): both density and total IT power rise: 385 racks @ 200 kW = 77 MW. Same hall, +54% IT power and a path to ≈5× compute vs Phase 1 once Rubin‑class racks arrive. Microsoft isn’t just renting the blue racks (Panel 1). They are also buying the option to realize Panel 3. When NVIDIA’s Rubin arrives (~2027/28), they can flip the switch: • Same floor plan. • Same cooling loop. • +54% more raw IT power utilized. 5⃣The math of the Rubin upgrade Rubin is projected to be roughly 3.3× faster for training than GB300 Blackwell at similar per‑rack power. If Microsoft exercises its option to: • upgrade to Rubin (≈3.3× speed), and • maximize hall density (1.54× capacity, from 50MW -> 77MW) then the total uplift becomes: -- Total Training Uplift ≈ 3.3 × 1.54 ≈ 5.0× The facility they are moving into today is effectively a stealth 5× training supercluster in waiting. 6⃣Double Optionality – Not just a win for Microsoft This isn’t only a win for Microsoft. It’s a strategic weapon for IREN. If Microsoft swaps to Rubin in Year 3, what happens to the ~76,000 GB300 GPUs? They don’t go in the trash. IREN suddenly has several optionality plays: IREN’s "Recycle & Redeploy" Playbook (examples): 🥶The Canada Retrofit – IREN is already testing GB‑class liquid cooling at Prince George. They could shift the "used" GB300s to Canada for their own higher‑margin AI cloud. It may be one reason they tested liquid cooling on a small B200 batch there and structured those leases with return options after 2–3 years. ~76k GPUs also rhymes with earlier Canada GPU numbers. 🔁 The Dell Buyback – Similar to the B200 deals, Dell often includes buyback or trade‑in clauses. IREN could potentially swap GB300s for credit on new Rubin systems. 🌵 Sweetwater / Horizon Cloud Expansion / IaaS – Move GB300s to the massive 2 GW Texas expansion for lower‑tier inference workloads, or into Horizons 5–10 for IREN’s own cloud or another partner. … The point isn’t which scenario is most likely. It’s that a lot of optionality exists once the GB300s are freed up. 7⃣ How does this change the economics for $IREN? This setup could trigger a "Revenue / Margin Supercycle" for $IREN. Building a 200 kW‑ready shell is heavy upfront — roughly $15M–$16M per MW of IT for the first 200 MW. But when Microsoft flips the switch and densifies a hall from 50 MW → 77 MW, the math inverts: • The shell and core infrastructure are already paid for. • The effective cost per MW drops to $9M–$10M/MW. • That’s a ~40% reduction in CAPEX per MW of IT. Then add the Rubin Multiplier: -- 3.3× (Rubin) × 1.54× (density) ≈ 5× compute. If that capacity is fully sold, site‑level revenue theoretically scales toward that 5× ceiling versus Phase 1, while infra costs per MW fall. If you scaled revenue linearly with compute, the 9.7B headline for Horizons 1–4 could theoretically point to ≈48.5B of revenue potential after a full Rubin + density upgrade. Realistically, Rubin contracts and pricing won’t be perfectly linear to raw compute, and IT load might not be pushed all the way to 77 MW in every hall. But even if revenue "only" doubled to ~20B, the embedded upside would still be enormous. The key is that the capacity and the optionality are already baked into the data‑center build. 8⃣Conclusion – The "130–200 kW" is the Key The "130–200 kW" flexible rack density clause is the hinge of the entire $9.7B deal. - 🔵 For Microsoft: It’s a construction-free growth option. They can turn a 100 MW GB300 training island into something approaching a 5× Rubin-era supercluster simply by swapping racks and raising the dial. - 🟢 For IREN: It’s a de-risked path to becoming a major AI-infrastructure operator with a pipeline of high-value recycled GPUs and a much lower effective $/MW over time. This is not a pure rental. It’s a symbiotic hedge against one of the fastest hardware cycles in history — with the optionality literally poured into the concrete.