Alex A.C.

$NBIS Great visualization for all the doomers saying "but $NBIS has fallen 60% from peak, just take the L." Ok, dude. You do you. I'll hold and keep adding to this name for as long as demand outstrips supply and GPU prices keep rising. Yes, the Vineland delay was a hard hit, but look at why it happened. The site owner shut down the temporary power before the fuel cells were ready. That's a construction hiccup at a third-party site, and the Microsoft contract behind it is still in place. Management is also calling the $325M of asset-light revenue a one-time thing. Do you really think it can't happen again? Sometimes I feel like Nebius management stays conservative on purpose so they can beat their own guidance. They even called the $12–13B guide for 2027 conservative. And they're getting to break-even faster than you'd expect from a company spending this much. Customers now prepay more than 100% of the capex on new deals, the AI cloud business runs at a 53% EBITDA margin, and management is targeting pre-tax break-even by the end of 2027 while spending $40–45B. When the company breaks even, most of the upside will already be priced into the stock. I won’t wait until then, I’ll do like institutional players do, buy cheap and hold for as long as it makes sense.

소로스 펀드 매니지먼트(Soros Fund Management) 2026년 3분기(9월 30일 기준, 11월 13일 공시) 13F 보고서 $MU, $SKHY, $INTC, $META, $AMZN, $TSM, $GOOGL, $NVDA 매수 및 비중 확대 (Buy & Add) 메모리, CPU 반도체, AI 하드웨어 인프라, 클라우드 소프트웨어 종목을 집중적으로 매수 SK하이닉스 ADR (SK hynix ADR): 지난 7월 뉴욕 증시 상장 직후 분기에 곧바로 325,000주(평가액 $59.85M, 포트폴리오 비중 약 0.82%)를 신규 편입 마이크론 (Micron): 2분기에 보유량을 7.9배 늘린 데 이어, 3분기에도 22,422주에서 35,618주로 58.9% 추가 매수했습니다(평가액 약 $38M, 전체 비중의 약 0.5%) 인텔 (Intel): 기존 10만 주에서 283만 주로 약 28배 대폭 늘렸습니다. 3분기 말 평가액은 $340.3M으로 단숨에 포트폴리오 내 최대 보유 종목(1위) 메타 (Meta): 129,729주(+595.7%)를 추가 매수해 총 151,505주(평가액 $109.87M)로 비중을 크게 확대 마켓액세스 (MarketAxess): 488,804주(평가액 $80.12M)를 사들이며 3분기 신규 편입 종목 중 가장 큰 규모를 기록 셀레스티카 (Celestica): 셀레스티카를 125,000주(평가액 $45.2M) 신규 매수 데이터독 (Datadog): 249,085주(+2,578%)를 추가해 총 258,747주로 보유량을 대폭 늘림 상위 5대 보유 종목 (콜·풋 옵션 제외): 1위 인텔(4.63%), 2위 아마존(3.58%), 3위 TSMC(3.41%), 4위 알파벳(2.82%), 5위 엔비디아(2.67%)

Thanks @MichaelDell for the callout on the Silicon Data H100 Index! Rental rates have jumped hard. Our latest residual-value (DCF / going-concern) estimate for a connected H100 is still ~68% of original sale price. A powered, networked GPU generating income is not the same thing as a card sitting in a warehouse.

I was at Electronica this week. Some notes. 1. Everything is in shortage, and the consensus is that conditions will worsen. 2027 is expected to be a strong year for semiconductor industry revenue. Lead times are long, and customer supply chain teams are concerned. (The decline in semiconductor stocks does not match the mood at the event, most likely because equity markets are forward-looking.) 2. Robotics is clearly the next major theme. Hundreds of robotic demonstrations across the show. Many folks are now updating their profile as Physical AI experts on LinkedIn. 3. Agentic AI is moving into practical engineering tools. Siemens launched the HyperLynx Agentic Toolkit for natural-language-driven signal and power integrity verification, with claimed reductions from weeks to hours on complex high-speed designs. Other EDA vendors are advancing similar agentic workflows. 4. Skills and talent shortage is becoming a serious problem for Europe. 30% of semiconductor workforce in Europe is expected to retire by 2030. While new graduate input rate is less than 1%.

We started shipping the Nvidia H100 in 2022. A year ago it rented for $1.96/hour. Today: $3.10. +58%, on a chip that was supposed to be obsolete by now. So much for GPU depreciation.

$SHEL CEO: Shell is using AI to anticipate refinery failures and sustain 99% utilization "And importantly, we are leveraging data and AI more so than we have ever done before, trying to anticipate where things are going to go wrong, intervening early and mobilizing folks more and more into that space. When we talk about performance discipline simplification, the work that the refining team, the manufacturing team has done, I think, has been stellar. I would say there's a lot more to do. We are not yet at where we need it to be. So there's more to still do. But having 99% utilization year-to-date, I think, speaks volumes of what the teams are looking at and how they're focusing on this."

Thanks for taking the time to work through the model this closely. I’ll go through all twelve because they connect: delivery timing affects revenue, contract mix affects funding, and both determine what existing shareholders ultimately retain. We’ll be coming out with refreshed models for $IREN , $NBIS and $CRWV soon, incorporating these earnings and the latest capacity, pricing and financing evidence. Where the rebuild is still underway, I’ll distinguish the October assumptions from what I can support today. 1. The gap between our model and Nebius’s guidance Your roughly 550 MW calculation works at our old blended yield. One adjustment: the 938 average billable MW supported $21.21B of owned-cloud revenue; $21.86B included other revenue. Against $12.5B, that leaves approximately $8.7B to reconcile. Our old model also depended heavily on short-duration sales. They contributed $9.55B of annual revenue and $25.32B of the $42.63B owned-cloud exit run rate. I would revisit acceptance timing and contract mix first. Stronger prices on new long-term agreements can coexist with lower revenue than our forecast if capacity arrives later or more of it goes into longer reservations instead of the higher-yielding short-duration pool. The CFO explicitly connected that allocation decision to financing conditions. I cannot honestly assign percentages of the gap to timing, capacity and price without completing that reconciliation. I also would not reaffirm $42.6B today. The refreshed model needs to establish the replacement moving forward. Your ~$23B exit calculation is a useful sensitivity, but management’s gigawatt statement concerns new capacity, rather than the entire closing fleet. Existing capacity still contributes, while new capacity needs acceptance and billing assumptions. December’s run rate and the following year’s recognized revenue also need separate treatment. 2. Q3 revenue excluding the asset-light item Your arithmetic is right: approximately $656M of group revenue, or $219M monthly. Cloud revenue excluding the item was approximately $649M, or $216M monthly. Neither calculation proves that everything remaining is recurring. I don’t think delayed additions alone fully explain the comparison with June. They explain missing growth, but an otherwise stable opening revenue base should not fall below its previous exit month simply because the next addition is late. There was a real Vineland power disruption: management said temporary generation wound down faster than expected while Bloom generation ramps. However, we lack a monthly bridge covering existing service revenue, utilization, credits and partnership recognition. I cannot confidently assign the shortfall to spot mix or conclude that June’s ARR was inflated. I would also avoid mechanically subtracting the entire $325M from September’s annualized figure. We have not established that all of it was recognized in September. 3. Receivables, deferred revenue and Q4 collections Yes, advance invoices awaiting collection are a plausible explanation for a substantial part of the movement. Nebius’s accounting policy permits deferred revenue when an advance payment becomes unconditionally receivable. Both receivables and deferred revenue can therefore rise before cash arrives. The cash-flow statement shows a $3.762B receivables outflow and $3.434B deferred-revenue inflow, a combined negative $328M contribution. Those net movements do not identify exactly which invoices remain unpaid. Collections could materially improve Q4 liquidity. I cannot put a reliable dollar amount on that improvement from the disclosed balances alone. I would separate advance receivables from ordinary service receivables and model collection dates explicitly. Collecting an opening receivable must not also appear as a separate new prepayment. 4. Financing the remaining capex and the share count Your residual is approximately right: $6–11.25B of 2027 capex sits outside the stated coverage range. The complete financing requirement also depends on opening cash, operating cash generation, debt service, fees and the liquidity reserve. At an illustrative $42.5B capex and 80% coverage, combined prepayments and asset-backed borrowing would be $34B. If advances supplied $18.5B, the implied borrowing component would be $15.5B. Microsoft/Meta-backed financing belongs inside that borrowing calculation. The roughly $11.6B of implied Q4 capex matters because it changes how much cash enters 2027. September cash cannot simply be carried forward untouched. My modeling order would be customer collections and operating cash, then borrowing supported by eligible contracts and assets, followed by the remaining corporate financing requirement. Converts and equity remain possible. The right mix depends on pricing, collateral, repayment schedules and dilution. Citi’s $3.5B facility also needs a refinancing or repayment assumption because it has a 364-day maturity. A replacement loan does not provide the full amount again as additional growth capital. The 7.6M remaining ATM shares would raise $1.14B at an illustrative $150 before fees. That is one tool, not a ceiling on potential dilution. Our old approximately 484M fully diluted year-end share count belongs to the old financing scenario; the refreshed number needs a rebuilt issuance and convertible schedule. 5. The 57.1% EBITDA margin Your implied Q4 calculation is reasonable. Applying approximately 40% full-year margin to $3.0–3.2B revenue implies roughly 35.5–36.2% in Q4, using the reported nine-month results. I would revisit our 57.1% assumption rather than carry it forward unchanged. Owned sites and better-priced contracts support improvement, but slower acceptance leaves costs running ahead of revenue. The one-off also prevents Q3 from serving as a clean recurring-margin starting point. Its associated costs were not separately disclosed, so subtracting $325M from EBITDA would not establish normalized profitability. I don’t have a completed replacement percentage yet. Management’s directional improvement comment supports the mechanism, but does not establish our old number. 6. IT MW versus facility MW I have not verified an explicit definition establishing that Nebius’s latest >$25M/MW quotation uses exactly the same denominator as our model. The letter identifies a weighted average of one-to-five-year deals, but that does not settle the electrical-load basis. I would keep that comparison qualified. Our model distinguishes facility power, equipped IT capacity, acceptance and billing. The 2,241.5 MW register and 1,613 accepted IT MW are different stages, not interchangeable totals. The old opening-year assumption was approximately 522 accepted IT MW. I do not yet have a defensible updated September/year-end pair. Dividing ARR by a blended price would produce a revenue-equivalent quantity, not prove physical customer-accepted capacity. 7. The four owned sites and New Jersey I would assess them separately: Lappeenranta: There is meaningful construction evidence. Fira confirmed work had begun, with first-phase completion scheduled for 2027. Caverion’s substation target is spring 2027. That supports a 2027 first contribution, while leaving the exact accepted MW and revenue timing open. It does not establish that all 310 MW will be operating. Mäntsälä II: The second 70 MW building has a specific substation contract, with completion targeted for autumn 2027. That supports a later-year contribution; commissioning and acceptance still follow power delivery. I would not give it a full year’s revenue. Alabama: Issued construction and electrical permits support real progress, but interior completion, commissioning and the legal challenge remain relevant. The October ruling narrowed the lawsuit while leaving the principal permit challenges alive for March. I regard 2027 operations as plausible, with unresolved execution and legal risk. Missouri: Construction has begun, but the campus ambition and generation permits do not establish a customer-ready 2027 phase size. I would retain a phased 2027 contribution as a scenario, with power delivery and acceptance controlling the amount. I have not independently verified every intervening construction or court milestone. For New Jersey, Q2 2027 is credible as management’s target because Bloom generation is already operating and the remaining task includes scaling it. It is not yet independently validated. I would model staged tranche acceptance, with a further-delay sensitivity, rather than assume all remaining Microsoft revenue begins on April 1. 8. Asset-light revenue and the $1.6B partner run rate It belongs to the same broad strategy, but the economic stream is different. The $325M reflected initial partnership enablement. Our approximately $1.6B exit run rate represented modeled continuing net partner economics. An upfront deployment milestone does not validate that recurring revenue assumption. More partnership transactions could produce additional recognition, but I would not annualize $325M or assume it repeats quarterly. Management’s initial 2027 guidance excludes potential asset-light and other partnership revenue. The refresh should separate deployment fees, ongoing service revenue and the partner capacity actually available for customers. 9. Raising long- and medium-term pricing Yes, the new evidence supports revisiting those inputs upward for new bookings and renewals. Existing signed contracts retain their own economics. However, a weighted average across one-to-five-year deals does not establish the same price for every five-year contract, hardware generation or customer. CoreWeave’s $20–25M range provides another useful comparison, subject to consistent power and service definitions. Rubin could sustain prices well above $25M if its useful output per MW and customer demand support them. I would not assume that outcome solely because it is newer hardware. Delivered performance, utilization, contract duration, customer advances and competing supply determine the realized return. The refresh needs both a sustained-scarcity case and a normalization case. 10. IREN’s delivery, financing and Texas exposure There is still risk of slipping beyond February. Customer acceptance follows equipment delivery, installation, networking and testing. A customer-requested design change explains part of the revised schedule, but does not eliminate remaining execution risk. The grace periods are also tranche-specific. IREN’s annual filing describes windows extending from mid-Q4 2026 into the beginning of Q2 2027. That is contractual protection around delivery, not a promise of revenue by January or February. On funding, the company identifies $15B of remaining FY27 capex, $9B of cash and committed financing/prepayments, and $6B of targeted additional funding. “Fully funded” applies to the initial approximately 0.3 GW program supporting >$4B ARR, subject to funding conditions. It does not cover every subsequent expansion. I would expect a combination of GPU financing, customer advances and financing against the owned data centers, potentially supplemented by corporate capital. Equity remains a possibility. I would not label the entire $6B an equity requirement or count prospective ABS proceeds as already secured. FY27 also ends in June, so it must be separated from calendar-2027 spending. Sweetwater 2 is exposed to the Texas process. Conditional Batch Zero inclusion does not settle every permit or energization approval. TCEQ’s pause extends to infrastructure supporting data centers, and December 10 is the targeted report date, not an automatic reopening date. The incoming governor’s stated approach increases policy uncertainty, but the election itself does not determine the legal treatment of each phase. I would apply additional timing risk to unfinished approvals rather than assume either a blanket exemption or a blanket shutdown. 11. CoreWeave’s prepayments and renewals They are material to the plan. The qualification on unsecured debt matters: management does not intend to issue additional parent-level unsecured debt at current yields through 2027. Secured borrowing and other financing remain central. Of approximately $9B expected in near-term customer payments, more than $3B relates to recontracting existing infrastructure. That is attractive because the equipment is already installed. But the $9B is cash received ahead of service, not immediate profit. Management described approaching $4B of incremental annualized revenue potential from 2027 renewals at current pricing. The nearly $10B figure includes 2028 renewals too. Neither is automatically revenue earned during 2027. The path to positive adjusted net income in 1H28 also needs deployment, operating leverage and financing costs to cooperate. I would stress slower collections and weaker renewal pricing separately. Reduced new hardware spending on renewed capacity helps, but power, hosting, maintenance and other obligations continue. 12. Huang’s departure and Nvidia’s commitments I would distinguish existing agreements from future discretionary support. Nvidia confirmed Huang’s departure and appointed Colette Kress interim CEO. I have not found an announced cancellation of these agreements accompanying that transition. The $36B figure is Nvidia’s disclosed aggregate AI-cloud commitment as of July 26. It is not $36B earmarked for these three companies, nor an unconditional revenue guarantee. Nvidia says those commitments decline as capacity is used by third parties or Nvidia itself. CoreWeave’s disclosed residual-capacity arrangement runs through April 13, 2032, subject to delivery, availability and termination terms. It covers qualifying unsold capacity, rather than guaranteeing every future expansion or financing obligation. My immediate concern would be changes to future investments, new backstops and capital-allocation policy. I would retain existing disclosed contracts in the models subject to their terms, while testing less generous future ecosystem support.

There have been no order cuts for Nvidia, according to four senior executives in the upstream semiconductor supply chain. Despite rumors suggesting order cuts resulting from Jensen's scandal, orders remain unchanged for 2027 and mass production orders are looking ahead to 2028. An insider said the company expects orders to remain strong for the next 2-3 years. Another semiconductor industry insider stated that Nvidia will not abandon the global AI market simply because Jensen Huang resigned, and Nvidia's long-term competitiveness will continue. One industry insider pointed out that after Jensen Huang's resignation, the competitive landscape between GPUs and ASICs may partially shift towards ASICs, but the overall AI industry will not experience a bubble. $NVDA $GOOGL $MSFT $META $ORCL $AMZN