I am always baffled by how investors always confidently think that they know better than management Lord give me the confidence of a 24 year old broccoli hair pod boy telling a memory CEO that his business is cyclical, because he read it in a BofA primer last tuesday

Reading twitter investor's takes on this being a structural breaker to memory is so ... dunning kruger Like sure, no one thought to compress their KV cache until Google publicly released this technique when it was the single biggest pressure on their business !!! 😁

One of the biggest debates in the compute buildout: are these underlying businesses unprofitable and that there is no way to sustainably finance the investment cycle that we are undertaking. And I get it – we are spending a lot of money – likely $700B of capex in 2026 amongst the hyperscalers alone. This year, the hyperscalers will build about 20GW of incremental IT capacity. How will we pay for this? Built a framework to pressure test this – how the major players across multiple sectors monetize a GW of IT capacity. Before we get into things, want to quickly explain my methodology & frameworks - all this analysis is meant to be a useful framework to get people thinking. These are just my views – do your own work! Validate / reject my premises! We are all our own agents in this world. That said, unless otherwise noted, all metrics are based on publicly available information from latest calendar year (2025). The adjustments: CoreWeave operating margin of 15% is based on a discounted view on their publicly disclosed 20-30% LT margins / Oracle's 20% target LT margins on GPU cloud business. There is a lot of debate / discourse in the market on what the eventual margins there could be. For illustrative purposes on this chart, I've left them at 15% which largely assumes a shorter depreciation cycle than they have assumed, plus no incremental margins from software. This is akin to the early days of AWS - which in 2015 disclosed they were a 17% operating margin business, while creeping that balance up to 40% over 10 years (spoiler, did so with software!). OpenAI and Anthropic figures are based on publicly rumored figures of ARR and GW deployed and allocating a portion to "inference" vs "training" compute (I am assuming 60/40 for OAI, more training for Anthro). For instance - OpenAI has disclosed that their ARR at end of 2025 was roughly $20B which coincides with a power footprint of 1.9GW - a ratio of $10.5B ARR / GW. Part of that footprint is not revenue generating and is just training. Someone pushing back could say that is a feature, not a bug of a frontier AI lab – I don't agree, because it all depends on the slope of inference growth. Anthropic has had remarkably amazing utilization of their limited resources over the years. From my outside in work, contribution margins are awesome. Finally, note that for the chart, I use estimated GM% instead of OP%, this is for a visual framework and should leave an upper bound on the profitability of these businesses which are fundamentally "different software businesses than past ones". I included Snowflake & Salesforce Rev / GW just to add some context. After all, these are the 2000s and 2010s era companies that are powered by compute. I derived their figures by dividing AWS Rev / GW by product gross margins. For the chart, Snowflake OI uses Non-GAAP OP% from CY25 as they are negative on a GAAP basis. ----------------------------------- Ok with that out of the way, what is the point of the analysis I present? A couple of observations: 1/ A lot of this analysis is meaningless! Why? Because for Snowflake or Salesforce, this Revenue / GW is an output, not an input. They are simply not in the business of selling repackaged power – they are selling VALUE / utility. In the case of Snowflake, an infrastructure SW company, they are running a feature rich scaled cloud data warehouse at scale. This took a decade of R&D, refinement, continuous development – and is selling you a product that is reasonably hard to duplicate. But they cannot grow their business by just adding power. GW consumption is the output, not the input. The same holds true for Salesforce, or any other software company. Rather, these products are somewhat difficult to sell, due to the large contract values, duration of engagement, etc. Given their raw COGs are relatively low, a majority of their gross margin is invested into S&M to sell the product. 2/ Google and Meta are THE most profitable businesses from a pure monetization / GW. In fact, before the current datacenter investment cycle starting in 2022, these businesses' Revenue / GW were significantly higher. Critics say that the new AI business models are less profitable than their core ads business. And they are completely correct. In fact, Jensen always says this in his speeches – the truth is that in the world of retrieval based software, ads were the most profitable businesses known to mankind. 90% contribution margin, with hardly any need for any S&M, with a baked in 20% growth algorithm per year by increasing the efficacy of the ads. But the truth is at some point, this algo slows down - scale slows down. In the same way software businesses could not grow their business by building power, these businesses could not either, there was a natural rate of adoption on these businesses, tweaked over the year with ad load and engagement. 3/ Infrastructure providers largely monetize at ~8-12B / GW and are the closest to the underlying hardware. I have a whole post in my drafts on this (still working!)... The thing I want to call out on the hyperscalers / neoclouds is that the core rental business of hardware usually starts out at ~10-15% operating margin. You can trace this back to the early days of AWS (which I may add, also was criticized as hugely money losing before they showed the world how profitable it was). Everyone thinks of these businesses as 40% EBIT businesses, which they largely are, but that was built over the years by selling software attached to their hardware. The core EBIT margins of just the hardware without adding value services is usually around 10-20%. Core cloud ARR / GW is closer to $12B / GW - you can derive this from AWS disclosures on power. The new accelerated compute infrastructure is around $10B of ARR / GW which is consistent from OCI, CRWV, and Nvidia. The way they all move to higher OP margin is attaching software to it at significantly higher blended gross margins -- the same way the hyperscalers built this during the 2010s. 4/ The model providers. Most controversial / interesting in this post. But perhaps the most applicable here. I am reminded most of the mid 2010s of Uber / Airbnb / Netflix and people / media claiming that these businesses would never make money. But it's all about the unit economics. If you can make 50-70% gross margins, then you can choose to allocate those GM dollars in a few ways. You gain significant operating leverage at scale. And my guess is gross margins likely move higher (another discussion for another time). But of note, VS the past generation of companies, the research compute budget is the significant outlier. This will likely be further concentrated at a certain time - continue to decrease as a % of the company budget, and more inference innovation techniques will be pushed - most of the benefits to consumers, while incremental ~3-5% GM gains will be kept per year... One of the great realizations in this exercise is that there are many ways of balancing a business to make money. In the case of software, they are hugely efficient / profitable from a "GW" perspective - and as a result, invest all their earnings into S&M to sell their product, which leads to a OP margin that is relatively low. For the hyperscalers, their gross margins are notably lower than their SaaS counterparts, but because their business is so large and have a high degree of trust with their customers, they are able to attach a considerable amount of their first party software while spending considerably less than their SaaS peers to sell that incremental $, yielding significantly higher operating margins. The internet providers are both hugely profitable, and need to invest little in their business, so really grew bloated over the years, investing in frivolous things and innovation grinding to a halt... until AI came along. Now they have a great target to invest in, with likely ways to enhance their core as well. In the case of AI – in the past few months, we have just crossed the uncanny valley of "model usefulness". They have largely gone from moderately useful chatbots / research tools, to very useful autonomous & agentic. Therefore, the name of the game will quickly shift to inference throughput & latency optimizations. As long as we are riding this S-curve, more compute = more revenues = more operating leverage for the model providers. And we are just starting... On this latest Nvidia earnings call, Jensen was asked how the hyperscalers will pay for their investments. He replied: "I am confident in their cash flow growing... in this new world of AI, compute is revenues... I am certain at this point that we are at the inflection point, we've reached the inflection point and we're generating profitable tokens that are productive for customers and profitable for the cloud service providers." For me, this switch flipped in the middle of 2025 - and really took off in late last year. Opus 4.5 and GPT 5 were tremendously valuable models, that were incredibly useful. We're seeing it now from the testimonies of the likes of Karpathy, etc. But anyone paying close attention to this knows / feels like everything has changed. Inference & usage is in take off mode & these profitable tokens are at the core of it all. These views are my own – not a view of Altimeter. Do your own research & look forward to discussing!