Jen Willi

$IREN: Industry Signals from $CRWV, $NBIS Earnings IREN, CRWV, NBIS all had good earnings. I will go over key developments and explain new concepts like asset-lite revenue from both NBIS and Mirantis and how they differ. Contract Quality/Demand CRWV announced they were getting 20-25m/MW on 5 year deals. Strong proof of demand was that some of their deals were getting 100% prepayments on total contract value, not likely the 5 year deals. NBIS didn't breakout short-term deals vs 1-3 year deals vs 5 year deals but when average together they getting >25m/MW. They were also getting 100% prepayments on some deals covering both GPU and datacenter capex. Large prepayments doesn't reflect inability to access debt because there were a wide range of prepayments on deals of different lengths. Contract economics, IREN is doing a fantastic job with 5 year deals at 22m/MW and 90% GPU Capex Prepayments. Both of these attributes are on the same deal, while CRWV and NBIS do not match contract length to prepayments amount leaving more ambiguity. IREN's 40m/MW short term deals are in line with CRWV and NBIS short term deals. Prepayment wise, I really like NBIS getting 100% prepayment covering both GPU and datacenter capex and expect IREN to get some of those deals next quarter. CRWV has stated they are seeing 10x more quality demand than they can serve while NBIS sees that demand is exceptionally strong. IREN will have a fantastic time signing VR200 capacity in 2027. Asset Lite While talking about revenue, the Neoclouds are seeing singificant revenue associated with something called asset-lite. Funny but most IREN investors don't know this but Mirantis is asset-lite selling their orchestration platform to companies like Adobe, Metlife and Paypal. Now, there are major differences between how NBIS does asset-lite, how Mirantis does asset-lite and how AI platforms like FireworksAI, TogetherAI, Baseten, Modal do asset-lite. I will unpack it here. The AI platforms sell a broad range of managed services besides inference APIs so inference providers is a misnomer. Both selling inference runtimes or GPU orchestration software counts as asset-light which is why Mirantis is considered as asset-lite. The AI platforms are the easiest to understand: they buy bare metal capacity on multi-year deals from both hyperscalers and Neoclouds to serve their inference from fixed GPU cost. Nebius does asset-lite through providing the supply chain and architecture at the beginning and partnerships with revenue sharing at the end. In NBIS own words, "asset-light model. By that, we are bringing capacity online with the Nebius architecture. In this case, some revenues are recognized when we lock this capacity, the supply chain, and the architecture, and some of the revenues are recognized when the capacity actually comes online and being sold to customers" (6). Let me unpack what this means and why NBIS labeled this revenue is non-recurring. What many NBIS investors don't remember is that NBIS designs their own server and firmware and send the design out to ODMs like Foxconn to manufacture the servers. So the initial non-reccuring revenue is selling these NBIS designed servers to NBIS's asset lite partners like BoostRun. Then the recurring part is the revenue split between the partnership as the software platform runs. Now revenue split might be the wrong term and it the accounting works a different way, but you get the idea. The first question people might ask is: If NBIS's non-recurring revenue from asset-lite is 325m, what can we expect the recurring part to be? This depends on how the accounting works. We know the total revenue generated from 325m worth of servers is about 162.5m on a 2 year payback. Assuming a 20/20/60 split where 20% goes to BoostRun, 20% goes NBIS and 60% goes to payback the equipment cost, then NBIS will see about 32.5m as recurring. However, because NBIS owns the end customer, they are likely to first collect the full 162.5m annual revenue and then payback a split of the profits to Boostrun. This way they can count a larger recurring revenue towards their topline but the profit at the end of the day is the same, one method of accounting will just help the topline look better. The second question: how does the 325m initial payment look like in margins? Dell ISG (infrastructure server division) has margins between 10-14% which includes GPU, CPU, Networking, and Storage akin to the full architecture that NBIS is provider. However, Dell gets roughly 7-8% of that margin due to it's economies of scale which NBIS doesn't have. So I expect 5-7% hardware margins plus whatever NBIS charges as setup expert margin with total around 10-15%. There will be alot of haters that tell me that I understand this wrong and 325m is recurring once they renew. To these people: just look at the capacity of RunPod + BoostRun + Shadeform + Axe Compute (9) and tell me how they have the capacity to support 325m QUARTERLY revenue again? That's half of NBIS's quarterly revenue not including the asset-light portion. Does RunPod + BoostRun + Shadeform + Axe Compute revenue = 1/2 NBIS revenue? The combine market cap of those companies is between 1-4B or something like that, it's not even worth my time to look it up. Mirantis does not provide servers, any hardware, or architecture. They sell SaaS for mainly orchestration so the bad news is that they don't have big one time revenue bumps to pad revenue but the good thing is that their revenue is consistent. From IREN's 228m total revenue we know that 28m is BTC, 170m-180 is GPUs so by elimination Mirantis is 20-30m per 2 months. Note that Mirantis closed on August 4th so only 2 months counts towards IREN revenue. Thus for Q4, IREN can expect 30-45m asset-lite revenue. Now I will provide the fair full context: NBIS's asset-lite business is likely to grow faster than Mirantis. Mirantis has a older code base and likely needs a revamp of it's teams to see significant growth. Open Source Inference This is where CRWV and NBIS lead IREN but the competition here is fierce with Hyperscalers, FireworksAI, TogetherAI, Baseten, Modal and MoonshotAI and Deepseek serving their own modals. Among the Neoclouds, $CRWV is in the lead with over 250m of ARR (9). $NBIS is second with ARR approaching a multi-hundred-million dollar run-rate which a nice wayt to say almost 200m ARR (5). IREN currently has 0 managed inference revenue and I see this as a 2028 launch for them. However the bulk of all Neocloud's revenue is still bare metal and managed GPUs anyways with Anthropic max bidding bare metal at almost manage inference prices, I can wait until 2028 or even 2029 for an IREN launch of managed inference. Delays CRWV did the best here as they have learned from delays last year. They stated in their earnings call that there expect no political impact in Texas for their sites that are either operating or under construction. My ability to preemptively spot delays has been very accurate and I will monitor delays across the industry in 2027. I have a unique ability to spot delays due to my previous industry experience as a software engineer at Tesla for Megapack for which I have triaged many islanded site deployment (original name for BTM), my current experience as a networking engineer at Cisco, my wife's software engineering experience at Oracle Cloud, my mother-in-law experience with firmware on GPU servers at SMCI, close friends who work at AWS and GCP, and on the AI platform side, my first job at ArgoAI. I have industry experience in all parts of the stack besides AI model research and lower level semiconductors but have some knowledge as I did my Thesis on Federated Learning and TAed computer micro-architecture classes. I called out IREN's potentially deal issues on Oct 16 (2) but had really high expectations the delays would only be 1-2 months to Dec. However, I didn't account for lead times on the cables which would take another 1-2 months. I was able to, however, pinpointed the exact issue as network topology and that it was a misunderstanding of Infiniband specs by Microsoft's platform team. Together with @ilzmcfly, we were the first to warn about the Vineland delays (4). Unlike all the other in-accurate FUD regarding regulations and permitting in Vineland, I was able to accurately call out there there were going to be NO regulation or permitting problems once switched to BE fuel cells and the fuel cells would arrive on time and operating in minimal time (3, 4). I pinpointed the real issue at Vineland that the media and most of X still gets wrong to this day: the delay is in the behind the meter grid between the fuel cells and the datacenter (3). Behind the meter is still a mini-grid and the hardest part of the grid is not the long transmission lines but the power electronics that are need whether you behind or in front of the grid. The benefit of the grid is that the grid utilities have been working for 3-5 years slowly building it out. On July 8, I called out the exact timeline that at best we would be looking at a March 2027 delivery for Vineland (4) and you know what Arkady revealed at earnings today? "We now expect ARR to exceed $7 billion for the month of February, a shift of about two months, even before New Jersey ramps up to full capacity" (5). This means that February is before New Jersey ramps to full capacity. The over all timeline for Vineland is stated as H1 2027 (5). For IREN, I'm still expecting Mackenzie to be delivered Dec 2026 but Air Cooled Childress to be Jan-Feb 2027. NBIS has stated in earnings that "delivery dates at a number of colocation sites shifted into the first half of 2027, including the remaining New Jersey tranches" (5). Delays are common across the industry but some delays are longer than others and Nscale and Firmus will see "indefinite" delays. Profitability CRWV announced EBIT break even in 2028 while NBIS announced pre-tax break even by end of 2028. Once you put together all the context, those are actually similar targets. CRWV uses mainly debt so they pay heavy interest while NBIS main uses convertible bonds which is eventual dilution or cash-payback with acretion which are still costs but does not show up as interest. So EBIT breakeven for CRWV and EBT breakeven for NBIS are closer than the terminology suggests. Furthermore NBIS terminology end of 2027 is a runrate figure which means real break even is 2028 the same as CRWV's. IREN has split their financing strategy between both convertibles and asset backed debt. So using CRWV and NBIS terminology of backing out interest payments for debt based financing and backing out convertible dilution for convertible based financing to define break even profitability, I expect IREN to achieve EBIT/EBT profitability in 2028. IREN is not likely to announce until they have solidified the economics of the 200MW of VR200s at SW1. Once you figure that ORCL is also going to hit cashflow positive in H2 2028, you'll realize the top Neoclouds are running the similar schedule at different scales. SPCX hasn't officially announced either because they haven't finalized all their contracts like IREN but profitability seems very likely for them in 2028 as well. Price Action Even though all three earnings were good, they had different price reactions due to different expectations built into the stock price: IREN: 2x EV/2026-ARR (Jan/Feb 2027) CRWV: 3.3 EV/2026-ARR (Dec 2027) NBIS: 5x EV/2026-ARR (Feb 2027) For EV/2027-ARR, IREN at ~0.47x, NBIS at ~1.16x, CRWV at ~1.32x. All figures are year end ARR. 2027 figures are based on 17B for IREN, 30B for NBIS and 50B for CRWV from VR 30m/MW and adjusting for earlier-contracted capacity at lower rates like IREN-NVDA and NBIS-META.