2025-10-27 · 供应链分析

反驳IREN论点，指出NBIS全栈软件优势带来长期更高利润率。

中文翻译

观点很棒！但是…… 你的论点基于对公认会计准则(GAAP)与非公认会计准则(non-GAAP)的误解……这可能会让非会计专业人士和 X 上的其他人感到困惑。在对比硬件部门利润与全栈 AWS 利润率时，$IREN 报告的是非公认会计准则(non-GAAP)，而 $NBIS 报告的是公认会计准则(GAAP)。 92% 是一个非公认会计准则(non-GAAP)的硬件特定利润指标，它系统地排除了折旧、数据中心 overhead、网络费用、支持等。硬件利润率通常仅通过从收入中扣除直接电力成本来计算，忽略了真实公认会计准则(GAAP) 销售成本(COGS) 的组成部分。在报告时，销售成本(COGS) 的遗漏和会计原则很重要（例如，为什么加密公司如 $MSTR 的数字与像 ASU 2023-08 这样的加密报告如此不同，因此你不能将收入与加密持仓与 Robinhood 进行比较）。与此同时，$NBIS 71.2% 的毛利率是公认会计准则(GAAP)，并将托管费用(colocation fees) 作为 29% 收入成本的一部分。因此，当你查看 EBIT 利润率以获得全貌时，$NBIS (~30%) 要高得多，因为全栈驱动了盈利能力。当 $IREN 扩展到 5 亿美元 ARR 目标时，他们被迫纳入现实的折旧、数据中心成本、网络费用和运营。如果以一致的会计标准对两者进行标准化，$NBIS 的全栈模型在真实利润率 spread 上已经高出 15-25 个百分点。你说超大规模客户自带 Type-1 堆栈是对的。但是……我再次对其他 $IREN 帖子说过这一点，它们一直在回避这个点。再说一次，Nebius 的 Type-2 层不是面向客户的功能，而是优化内部运营支出(OPEX)。如果企业自带 Type 1 (K8s/Ray)，Nebius 的内部编排(Type 2) 仍然能提升其自身的利润率。这是运营支出(OPEX) 护城河，它不是客户使用的，而是关于你运行集群的效率。把它想象成 AWS 的 Nitro 虚拟机监控程序，但核心在于 AWS 如何驱动 >80% 的利用率和最小的人工运营。Nebius 的内部控制平面自动化了混合 SKU 的 GPU 调度、遥测和固件生命周期。每增加一个百分点的利用率，毛利率就扩张约 1-1.5 个百分点。Nebius 的内部编排降低了其运营成本基础并提高了利用率，无论客户是否自带堆栈，所以这不是非此即彼的关系。长期的利用率效率和自动化复利增长快于廉价电力/容量，$NBIS 的全栈在长期驱动更高的利润率，而其他方面临利润率压缩的恶性竞争。随着 $NBIS 扩展，其软件熟练度的增量成本趋近于零，结构性地推动利润率更高。相反，$IREN 的依赖模式意味着随着扩展，许可费（如果你采用你的数字，为收入的 3-8%）加上必要的、未量化的站点可靠性工程师(SRE) 运营成本将成为盈利能力的永久拖累，阻止公司实现 $NBIS 的结构性利润率上限。外包 Type 2 解决方案的许可成本很低，如你的 3-8% 数字，但当被视为 overhead 时，“20-30%” 的数字非常相关。回到 GAAP-non-GAAP，一旦 $IREN 在规模上被迫确认完整的公认会计准则(GAAP) 成本，与 $NBIS 相比不可避免的 15-30 个百分点的利润率差异将会显现。我同意你，短期的容量和基础设施是今天的一大优势，这就是为什么我仍然看好转型的加密货币矿工如 $IREN、$WULF、$CIFR，但在执行期间的软件护城河，在将收入转化为利润的长期来看，将大大优于裸金属基础设施（这还假设 $NBIS 无法扩展基础设施，而他们是可以的）。 $IREN 当前的优势是物理集成和廉价容量，它是资本支出(capex) 密集型的，并受制于电力市场收紧。$NBIS 的优势是运营支出(OPEX) 杠杆。$NBIS 更高的利润率是几年后基于全栈 AWS 式利润率的必然经济结果，而当其他人竞相触底时，利润率差异不会崩溃，而是随着规模扩大而扩大。廉价电力赢得下一个季度，高效的编排成为下一个 AWS。

英文原文

Great points! But... You did base your thesis off misunderstanding gaap vs. non-gaap... which could be confusing to non-accountants and other people on X. $IREN reported non-GAAP and $NBIS GAAP when you look at hardware-segment profit to full-stack AWS margins. The 92% percent was non-GAAP, hardware-specific profit metric that systematically excludes depreciation, data-center overhead, network expenses, support, etc. Hardware Profit Margin is typically calculated solely by deducting direct electricity costs from revenue, ignoring the components of true COGS GAAP. And when you do reporting, COGS omissions and accounting principles matter (eg. why crypto firms have like $MSTR wildly different numbers from crypto reporting like ASU 2023-08, so you don’t compare revenue with crypto holdings to Robinhood). Meanwhile $NBIS 71.2% gross margins was GAAP and includes colocation fees as part of 29% cost of revenue. So when you look into EBIT margins for the full picture, $NBIS (~30%) is vastly higher because full stack drives profitability. When $IREN scales to the $500M ARR target, they’re forced to incorporate realistic deprecation, data center costs, network expenses, and operations. If you normalize both under consistent accounting, $NBIS’s full-stack model already operates 15–25 points higher in true margin spread. You’re right that hyperscale clients bring their own Type-1 stack. BUT... I’ve said this again to other $IREN posts that keep side-stepping this point. Once again, Nebius’s Type-2 layer isn't for customer-facing features, it’s optimizing internal OPEX. If enterprises bring their own Type 1 (K8s/Ray), Nebius's internal orchestration (Type 2) still drives its own margins up. That's OPEX moat, it's not what customers use, it's about how efficiently you run the fleet. Think of it like AWS’s Nitro hypervisor but central to how AWS drives >80% utilization and minimal human ops. Nebius’s in-house control plane automates GPU scheduling, telemetry, and firmware lifecycles across mixed SKUs. Every percentage point of utilization is ~1-1.5 points of gross margin expansion. Nebius's in-house orchestration lowers their operating cost base and increases utilization, regardless of whether customers bring their own stack, so it’s not quite one or the other. Long-term utilization efficiency and automation compound faster than cheap power/capacity and $NBIS full stack drives higher margins long term, while others face margin compression on a race to the ground. As $NBIS scales, the incremental cost of their software proficiency approaches zero, structurally driving margins higher. Conversely, $IREN's dependency model means that as they scale, the licensing fees (3–8% of revenue if you we go with your figure) plus the necessary, unquantified SRE operational costs will remain a permanent drag on profitability, preventing the firm from achieving $NBIS's structural margin ceiling longer on. Licensing costs for outsourced Type 2 solutions are low like your 3-8% figure, but the "20–30%" figure is highly relevant when viewed as overhead. .Going back to GAAP-non-gaap, once $IREN is forced to recognize full GAAP costs at scale, the inevitable 15–30 point margin delta against $NBIS will materialize. I’d agree with you that short-term capacity and infrastructure is a great advantage today, which is why I’m still bullish crypto miners pivoting like $IREN, $WULF, $CIFR, but the software moat during execution, would vastly outperform bare mental infrastructure in the longer term when it comes to converting revenue to profit (and this is pretending $NBIS can’t scale up infra, which they can). $IREN current advantage is physical integration and cheap capacity, it’s capex-heavy, and subject to power-market tightening. $NBIS advantage is OPEX leverage. $NBIS higher margins is an economic inevitability in a years time when rooted in full stack AWS-like margins, while margin delta isn't collapsing as others race to the bottom, it's widening as scale hits. Cheap power wins next quarter, efficient orchestration becomes the next AWS.

在 X 上查看原推 ↗