Subscribe now

Zero Data Retention

FT recently dropped an article showing the spend split between different Anthropic models (indexed back to June). What was interesting about this chart (below) was that the recent growth across July wasn’t driven by Fable, but by Opus 5.

Many people asked questions about this. Why isn’t Fable 5 driving more adoption? Martin threw out a tweet that listed a number of potential rationales (I won’t summarize all of them, tweet is pasted below).

The overwhelming response to the question of “why has Fable 5 seen more lackluster adoption” was zero data retention (or ZDR). This topic has been front and center this summer. OpenAI wrote a blog post about it a few weeks ago. In my opinion, it’s becoming increasingly clear that users have a preference when it comes to data retention policies, and if you believe the majority of responses to Martin’s tweet, the labs have some decisions to make!

So what is Zero Data Retention? Typically, the standard API terms of service / arrangements has been the same for the last few years. The prompts / responses are held for some period of time (~30 days) to monitor for abuse. Abuse could be something like is the model used for hacking, or some other harmful intent. The labs want to know if the models were used this way so they could either do something about it, or use it as forensic evidence.

Of course, there are some models / customer contracts where nothing persists after the response is generated. No logs, no traces, no review queue. Nothing to leak, nothing for an employee to spy on, nothing for someone who steals employees credentials to see, nothing to subpoena, etc. Quick flag- zero doesn’t literally mean zero. Safety flags still exist even in this case (ie data that is flagged as harmful can be stored for much longer, but unflagged deleted immediately).

One distinction on this entire topic is policy vs architecture. A 30 day deletion policy in many ways is a “promise” (ie a policy). ZDR is more of an architecture. Policies / promises can be amended, terms of service can be updated, breaches can render it all irrelevant, etc. Architecture, on the other hand, has none of that. You can’t retroactively amend it. And we’ve seen a practical example demonstrating the policy vs architecture issue. In May 2025 a federal judge ordered OpenAI to preserve a lot of logs (I believe output logs). This included chats users had deleted. This was part of the New York Times copyright case. All of a sudden, the policy of “we delete after 30 days” became “we delete after 30 days unless someone says otherwise” (in this case the someone was a federal judge). For ZDR customers? There was nothing this judge’s order could do. There was nothing for OpenAI to preserve (for ZDR customers) because that data was deleted (architecturally) from the get go.

Ok circling back to Fable 5. When Anthropic launched it this summer (in June), they launched it with a 30 day retention policy - basically a 30 day retention of every prompt and output. AND everywhere the model is offered (their API, Bedrock, Foundry, Copilot, coding tools, etc). No ZDR exceptions, and existing ZDR agreements specifically did not apply to Fable 5 traffic. It’s easy to read that and think “huh, why, that seems weird.” But it’s actually reflects a very consistent viewpoint Anthropic has (which I don’t necessarily think is wrong!). The Fable series of models is by far the most capable model class ever created. And because it’s so powerful, it can be used in quite negative ways (cyber attacks, bio, etc). Anthropic would say the most serious misuse patterns only become visible (and thus truly block-able) when you can correlate across many interactions. On top of that, you can’t investigate something you never stored in the first place. If you take catastrophic risk seriously (which Anthropic clearly does), this data retention policy makes sense.

However, the rest of the market quickly reacted to this. Microsoft reportedly restricted its employees from using Fable 5. GitHub disabled it by default in CoPilot. And back to the chart (and tweet) I mentioned at the beginning of this post, the ramp data suggests the rest of the market did the same thing - routed traffic away from Fable 5. Of course, the rationale for the lower Fable 5 adoption is just a hypothesis. The cost of Fable 5 (relative to their other models or competing models) is clearly also a consideration for customers. Fable 5 is expensive, at $10 per million input tokens, roughly double GPT-5.6. Most workloads don't necessarily need the frontier (I've written before about the token bifurcation - the majority of tokens may flow to cheap, good-enough models while a thin slice of frontier tokens drives a majority the revenue). Opus 5 at half the price of Fable 5 always had the potential to cannibalize Fable.

However, the cost-only explanation (for the “lackluster” Fable 5 adoption) has some holes in it. Frontier models often complete the task in fewer tokens and fewer attempts. They’re more token efficient - expensive per token doesn’t mean the same thing as expensive per task (or per output). I do think (and hear very similarly) that the lower than expected usage of Fable 5 was more about the data retention policy (and remember, policy is different from architecture!). Ramp's own economist told the FT the retention requirement was a drag on adoption (in addition to price).

In the OpenAI piece I mentioned earlier, Offering Zero Data Retention for Frontier Models, they previewed something called Private Safety Processing. Customers can keep their data on infrastructure that they control (or sits encrypted under keys only the customer holds), and OpenAI runs automated misuse classifiers that scans for misuse patterns. OpenAI only receives an alert category and severity (not the actual data / prompt / response). I do think there’s some signal here. It’s not common to see a big launch-style blog post about retention policies…unless those policies are driving (or pushing away!) revenue. This is definitely a important topic, that will only become more important as the models get more capable.

So what does this all mean? Couple things come to mind. For one, the data flywheel inverted a bit. Some of the highest-value customers are now saying they’ll pay extra specifically to have their data “forgotten.” (so labs can’t learn from this to improve their products). The new improvement loop gets built from purpose-built data (RL environments, evals, opt-in partnerships), not from the data the customers spit out. Now the asset is that RL factory, not just the data traces.

One interesting thing for founders - your retention posture is now (in some ways) inherited from your vendor’s retention policies. “What does your model provider retain, and for how long?” is something I’d expect to show up on just about every security questionnaire you fill out. Companies using these models will want to get ZDR passthrough into your own contracts so you can pass that on. If not, architect for model portability, OR you’ll have to build that trust infrastructure yourself. There will definitely be demand for confidential inference, customer-held keys, audit tooling for AI traffic, etc.

For the labs: unilateral retention is appearing like a tax customers just aren’t willing to pay. Safety-through-retention either is something that starts to become an industry-wide standard, or it gets rebuilt cryptographically so retention isn’t required at all. My bet is the latter...

Two things I’ll end with. 1) the CISO is starting to become a lot more important constituent in the buying process. 2) For many years a strong pitch in software was “your data makes our product better.” The strongest pitch in AI might turn out to be something entirely different: “we never saw your data at all.”

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 4.4x

• Top 5 Median: 32.7x

• 10Y: 4.7%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 17.3x

• Mid Growth Median: 7.6x

• Low Growth Median: 3.9x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 4%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 35 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 22%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

What Could Go Right?

Whenever I’m thinking about making a private investment there’s (broadly) a couple main questions I ask myself:

1. What could go right

2. If things go right how big can it get

3. What could go wrong

The reality is the first two are kind of the same question, and the third question is completely irrelevant (in my opinion). I bring this up because so often the conversation around venture stage company investing is about the risks. Execution risk. Market risk. Competitive risk. Hiring risk. etc. What if the market isn’t big enough? What if they can’t execute on this vision? The deck is massively stacked against startups. The reality is only a miniscule fraction ever actually turn into big companies! And all the reasons to be hesitant are right! In a vacuum, the reasons to not invest in a startup almost always overwhelm the reasons to invest. And given the failure rate, the “bears” are usually right!

However - venture is a power law business. It’s a statement that’s so incredibly simple, and so incredibly viewed as “obvious,” yet at the same time it feels like most folk don’t actually invest or build a portfolio with this simple truth in mind (and I find myself drifting from it from time to time!).

If it goes right, how big could it get? That’s really the question to ask right now. The reason question three doesn’t really matter (in my opinion) is because of asymmetry. Being wrong comes with a 1x your money loss max downside. Being right comes with uncapped upside. An interesting sidenote here - a very large institutional LP shared some data on the distribution of outcomes in funds for top quartiles and bottom quartiles. Interestingly both the top quartile and bottom quartile funds had the same percentage of investments that returned <1x in a fund. 50%. I call this out to say - the percentage of deals you loose money on has no impact on top quartile vs bottom quartile funds.

The question of “how big could it get” also comes with extra weight currently. For the last ~15 years, the theoretical TAM ceiling on any software company was some flavor of software budgets. With many of the AI companies today, the TAM ceiling is no longer software budgets (but headcount / consulting budgets). Again, this is nothing revolutionary. But I’m constantly reminded of examples of this. And this concept of TAM ceilings rising with new technology leaps is nothing new. It’s one of the reasons the top decile outcomes keep growing and growing in size. When I got into venture, a $1b public company was considered an amazing outcome. Not anymore…BUT - oftentimes investing at the time it was hard to think of the top outcomes being much more than that, given this was the “ceiling” at the time (not an actual ceiling, but the general range of the top bucket of outcomes). Imagine passing on Snowflake in 2017 because “Teradata market cap is only $4b.” In moments of large tech disruption, the size of the prize grows rapidly!

It’s what can make investing today (and assuming today’s exit valuations will be tomorrow’s exit valuations) tricky. I’m certain the top decile outcomes in 2036 will be meaningfully higher than what they are today - but really investing with this frame of mind is hard. This isn’t to say every company will be a great company and it’s easy to invest right now. Quite the opposite. There’s more companies than ever getting started today (which makes being a VC extra fun). But the unfortunate truth remains, most won’t work out. The only thing that matters is finding the founders who can do the impossible! If things go right, how big could it get?

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 4.3x

• Top 5 Median: 30.5x

• 10Y: 4.7%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 17.2x

• Mid Growth Median: 6.8x

• Low Growth Median: 3.8x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Useful Life

A few weeks ago I wrote an article “The Second Life of a GPU.” In that post I discussed a (hypothetical) scenario where GPUs retained value past their (presumed) 4-5 year useful life. What impacts would this have on clouds / neoclouds or the broader compute complex? Could compute contracts be “re-contracted” after their useful life? All some version of the same question - is the useful life of these chips actually much longer than 4-5 years.

Well, this week on Coreweave’s earnins call we got a great nugget related to this.

“What we are seeing today is that the upside of recontracting is real as we remain largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs. So as our earlier generation fleets roll off their original contract, they offer the potential to deliver strong returns in the subsequent years. We are seeing this across our Ampere and Hopper fleet.

As an example, we recently signed an A100 contract that extends into 2029 at an attractive price. As a reminder, this SKU was introduced in 2020. Clusters of prior generations of architecture of our installed, energized production-grade compute already running at scale. They come with a proven ROI for customers.”

Then Jensen tweeted this:

Of course, both Coreweave and Nvidia have a vested interested in the useful life of chips elongating. However - I believe them. And we now have some concrete data to demonstrate these chips do have a useful life that (for now!) extends to 9 (and maybe more) years.

One one hand you can say this is just representative of the current moment we’re in. Compute constraints still exist everywhere, so why not take any form of compute, regardless of how old it is. While true, you still wouldn’t sign a contract for an A100 chip if you thought it was worthless. And for someone like Coreweave a recontracting deal like this is (probably) meaningfully profitable. The initial purchase (ie debt used to fund the purchase) probably was amortized over the useful life of the chip. My guess is that debt is now fully paid off, and incremental revenue from the re-contracting comes without a heavy financing (interest) cost.

I really believe we’ll see more and more of this - chips contracting well beyond their useful life. One reason - the types of inference requests will continue to spread out in complexity. We’ll start seeing the most complex inference workloads become even more complex. And these probably require the latest generation of chips. However, the long tail will also grow. And some of these requests will be incredibly trivial. Those won’t require the latest and greatest. There will be a lot of value created in a software layer that can understand the request coming in and route to the right chip in real time.

Another reason - I continue to believe that the “compute constraints” will last much much longer than people think. So far, the risk has repeatedly been to the upside!

Here’s an analogy I like. The market has been pricing GPUs like a melting ice cube. Turns out, they may have a freezer! And that freezer is an exponentially growing curve of inference (and heterogeneity of inference requests)

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 4.4x

• Top 5 Median: 36.1x

• 10Y: 4.6%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 18.1x

• Mid Growth Median: 6.7x

• Low Growth Median: 4.0x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 17%

• Median Gross Margin: 76%

• Median Operating Margin 3%

• Median FCF Margin: 20%

• Median Net Retention: 110%

• Median CAC Payback: 36 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 22%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Custom Tokens

I strongly believe there will be 3, very large, markets that emerge in the “model / token” market:

1. Frontier tokens (ie from the large labs like OpenAI and Anthropic)

2. Vanilla open weight tokens (Deepseet, Moonshot, etc)

3. Customized (RL / SFT) open weight tokens

So much talk recently makes these seem zero sum… And of course, they do take share from each other. If open weight models never existed closed models would take it all! BUT - zero sum assumes the size of the pie is constant…which it isn’t. The market is growing much faster than any one group can steal market share from the other. So everyone grows! This won’t happen for ever. Eventually the market will mature and we’ll settle on a more “stable” market share. But for now (and for longer than people will expect), hypergrowth persists!

I wanted to talk a bit today about the third bucket - customized open weight tokens. It’s a very common theme to hear these days - take an open weight model, customize it for your data / workflows, and then serve that specific model instead of a more general purpose frontier model. The hope is you get a cheaper and more specialized model. In many cases this is true! And I believe it. However, it’s not nearly as easy as anyone thinks.

There are broadly 3 main ways to customize a model.

1. Custom pre- training (the heaviest lift). Training a model from scratch on your own data

2. SFT (Supervised Fine Tuning). You show the model examples of what “good” looks like (prompt + response, in basic terms), and the model learns to imitate it

3. RL (reinforcement learning). Instead of showing the model examples of what good looks like, you let it try the task, grade the answers, and then let it learn from which answers scored well

Most people stick to the second and third options. Someone much smarter than me described 3 main pillars of this (lots of “lists of 3” in this post…):

1. Training infrastructure. This is all of the compute and distributed training that’s needed to run a training job. This includes many things like orchestrating GPUs, sharding models across them, managing the sampling loops for RL, keeping GPU utilization high, shuttling updated model weights from trainers to samplers after each training step, etc. It’s HIGHLY complex and takes real distributed systems engineers to make it work. Tinker from Thinking Machines makes this seamless.

2. Data. Proprietary data is the reason you want to customize a model in the first place. BUT - the data doesn’t come “training ready.” Someone has to convert the raw data (this could be support transcripts, agent traces, logs, etc) into a format a model can actually learn from. Today, this burden largely falls on the customer. There are companies like Applied Compute or Trajectory that help out with this part of the workflow. Or Handshake / Mercor / Micro1 who source data and deliver it “training ready.” This part of the workflow is hard to “productize.” Every companies data is different and is messy in it’s own way. For SFT you can usually just convert data to a JSON file, but the hard part then is the judgement - deciding what “good data” looks like (and then filtering out the bad data). For RL there isn’t a standard data format, but more of a “environment plus reward function” settup, which is even more bespoke / customized (and thus hard to productize)

3. Evals. This is how you ultimately measure whether the customized model is actually good at the workflow / task you want it to complete. For RL specifically, evals happen at each step (this is the reward function grading each attempt, so the eval runs inside the training loop)

Then you have some version of a “training recipe” that sits across each of these 3 pillars. You can think of this as more of a blueprint for the whole job. How you manage and sequence each step, which training method you use (SFT or a version of RL like GRPO, PPO, DAPO, etc or a combo of both SFT and then RL), which base model you start from, how you design the reward function and the hyperparameters (learning rate, batch size, etc), etc. The training recipe is the design work that connects all of the steps of “customization.” There’s definitely a gap in the market here. There are different products (or more commonly services) that help at each of the three pillars I listed above, but not a lot of solutions that stich everything together. A big reason for this is it’s hard to “productize” all of this. It’s naturally more bespoke for each company, and thus a better fit for services.

I’m quite excited to see the innovation that will come from this field in the next few years.

One implied point I’m making - it sounds nice to talk about “customizing models” but it’s actually very difficult. And what I laid out above is only one part of the difficulty… There are two more challenges! How you then ultimately serve and deploy these models, and then how you upgrade them over time (upgrade on new base model, or upgrade with new data, etc)

On challenge number 1 - serving the model. IF you’re able to get through all the steps I outlined above, you now have a customized model! But so what? What do you do now? How do you get it serving production traffic? A lot of the training platforms I listed above aren’r inference platforms (at least not yet! I expect that to evolve..). Some of them (like Tinker) have a sampling API you can hit directly, but I think of this more as a “test” vs anything that can serve production traffic / scale. So you can then either spin up your own inference stack on top of something like vLLM (again, very hard to do yourself), or hand the model to someone like Baseten / Fireworks to serve it (this is what vast majority of folks do, and both Baseten / Fireworks are getting more and more into the RL / SFT stages as well). Good news is serving a customized model on one of these inference clouds isn’t that hard. Generally the customized model keeps the base model’s architecture (whatever open weight model you started with), so all of the inference optimizations those platforms built for those open weight models transfer over to your customized model. The “downside” (if you can call it that) are the economics. You pay a tax for someone else removing complexity for you. This next part I’m a little squishier on (so someone from Baseten / Fireworks please correct me if I’m wrong!), but when those platforms serve “vanilla” open weight models they get to share GPUs with thousands of customers at a time serving that same model. Utilization of those GPUs are high, so per token costs can be lower. For the custom models - it’s yours alone. There is no sharing of GPUs or spreading out the costs. You need dedicated GPUs (a parallel might be single tenant vs multi tenant SaaS deployments?) This is largely (again, I think) why customized models make sense for high volume, repetitive, specific tasks, where the model can be running constantly and you’re not implicitly paying for idle GPUs.

Then there’s challenge number 2 - upgrading your model. This is definitely an underrated challenge… A customized model is a point in time snapshot, but unfortunately there are two “upgrade” cycles running at the same time, The first is your data. Your business changes, your customers ask for new things, you build new products, etc. The model you a few months back becomes stale quickly. So someone needs to be doing all of the customization steps I discussed in the first half of this post consistently (collect data, re-train, re-eval, re-deploy, etc). There is no “one time” cost to customize a model, it’s very much a recurring cost and recurring complexity.

The second upgrade cycle is new base models. Open weight models are improving so fast. Kimi K3, GLM 5.2, Deepseek v4 Flash. Whether it’s a new open weight model, or the next version of an existing one, there are CONSTANTLY new base models coming out. This means that you’ll spend time, energy and expenses customizing a model, only to go through it all again (in very short windows) when the next best open weight model comes out and you want to customize from a “stronger” base. The frontier labs will tell you it’s not worth it. Why go through all of this pain consistently vs just buying frontier tokens direct?

There’s certainly a lot of product gaps to make open weight model customization more efficient. But I believe we’ll get there. Baseten / Fireworks / Together and other inference clouds are attacking this market from their inference starting point. Thinking Machines is attacking this market from the training infrastructure side. Applied Compute / Trajectory are more “pure play” RL shops. Handshake / Mercor / Micro1 are attacking it from the data side. Neoclouds like Coreweave are entering this space. Evals players like Braintrust and Vals.ai attacking it from the evals starting point.

All of this to say, there are a lot of people going after this opportunity, all from very different starting points. I have no doubt we will see some massive platform companies emerge here shortly!

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 4.1x

• Top 5 Median: 31.8x

• 10Y: 4.7%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 19.3x

• Mid Growth Median: 5.8x

• Low Growth Median: 3.5x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 12%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 3%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 35 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 22%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

AWS CapEx ROI

The vast majority of earnings calls from public companies are pretty boring. But every now and then a CEO pulls back the curtain a bit and goes into a level of detail that I find fascinating. One of those calls was Jay Kreps from Confluent (on their Q1 ‘23 call) discussing the TCO advantages of using Confluent vs open source Kafka. On that call Jay went into a lot of detail about the differences between paying a cloud vendor vs hosting your own software. Jay’s one of the best infrastructure entrepreneur, so getting a peek into his brain like that was pretty cool.

Yesterday was another moment where I really geeked out on an earnings call. This time it was Andy Jassy from Amazon discussing the ROI on AI capex (which is clearly a very hot topic!). To set the stage, Amazon has guided to roughly $200b of CapEx in 2026. On the call yesterday they updated that guidance to $220b, and then Andy walked through why they believed the ROI was high on these capital investments.

First, he described two separate capital cycles that exist in this cumulative capex figure:

1. Data center

2. The equipment that goes into data centers (chips, servers, networking, etc)

On the data center component, he described these as 2 year projects. The spending starts 2 years before they can put any of the equipment into the data center. But once it’s up and running, it generates significant revenue on day 1, and continues to generate revenue for decades without another large up front expense (again, this is just for the data center capex build, not the equipment that goes in the data center). So think of the capital cycle as “decades” for the data center component.

On the equipment component, the cycles are much shorter. The lead time is months not years. It takes (in his words) 2 years to build a data center, so that spending starts 2 years before revenue hits. On the equipment side, he called out the purchasing hits “a few months” before they plug them in and data center starts generating revenue. This is important because as a company, they have strong visibility into the customer demand before making the purchase (again, because the lead time is months not years). If demand doesn’t exist, they don’t but the equipment. He called out a payback period of 3 years on the equipment, and a useful life of the equipment of “at least 5-6 years.” At the same time, most of their AI capacity is being contracted on 5 year terms. The important call out here - after hitting the payback period in ~3 years, the next 2-3 years print profits.

The last part to highlight is where these two capital cycles converge. A data center has a 30+ year useful life, while equipment has a 5-6 year useful life (again, according to Jassy). So each data center should see at least 5-6 "generations" of equipment cycle through it (the math there being 30 years divided by 5-6 years per equipment cycle). Every generation after the first has better economics than the one that preceded it, because the big upfront data center spend only happens once. The first generation of servers carries the data center build. Generations 2 through 6 just show up and print profits! (for the data center capex). That's why Andy was so willing to acknowledge the near term free cash flow headwinds. When you're building this many data centers all at once, and more importantly all ahead of monetization (with data centers having a 2 year lag), the cash outflows pile up before the revenue does. But his argument is that at some point revenue growth outpaces incremental capex growth, and the ROIC math takes over. What was cool was then Jassy compared this to a prior capex cycle - the first era of the cloud buildout. Andy claims AI margins are tracking core cloud margins at the same point in their respective evolution (or actually a little ahead). And the demand side isn't the constraint at all (which has been a very consistent message across all hyperscaler earnings calls). Even at $220b of capex they won't have enough capacity to meet 2026 demand, and he expects the same in 2027, and 2028 reservations are already coming. Then he dropped the mic at the end! Amazon used to think AWS could be a few hundred billion dollar revenue business. They now think it's a $1 trillion revenue business over time. Take with a grain of salt (every hyperscaler CEO is incentivized to tell this story right now), but it's quite a clear articulation of the AI capex bull case.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.8x

• Top 5 Median: 29.6x

• 10Y: 4.7%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 19.5x

• Mid Growth Median: 5.2x

• Low Growth Median: 3.0x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 12%

• Median LTM growth rate: 16%

• Median Gross Margin: 75%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 43 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Every week I’ll provide updates on the latest trends in cloud software companies. Follow along to stay up to date!

Subscribe now

The World Isn’t Zero Sum

I debated the title for this article a lot. The other option was “The World Isn’t Binary.” I’m still not exactly sure which title better describes this article, but I’ve found the principal of living with a positive sum mentality more impactful in my life so I went with that title. Anyway..on to the post!

There are no shortage of big debates right now in AI. Open vs closed models. Nvidia vs custom silicon. Will chips have value after 5 years. Will AI kill software. God model vs fleet of specialized models. Are we in an AI capex bubble or not. Model layer vs application layer. Will models commoditize. The list goes on! These were just a few that came to mind…

I won’t address all of these today, but the answer for most is “you’re not thinking big enough.” I’m clearly a hyper-optimistic person, I like to view the world through this lens! But I genuinely think these questions do a disservice to the moment we’re in. I don’t mean to invalidate the questions, they are smart questions, BUT in an early exponential moment I think it’s easy to “reduce” the moment to these simple questions that often seem zero sum in nature. When the size of the prize is expanding this quickly, the answer often turns out to be “everybody won.”

Let’s look at the current open vs closed model debate (or really the “Are Anthropic and OpenAI screwed as low cost open weight models take over” question). The TLDR of my perspective is that there will be both a massive market for closed frontier models AND open weight models. Does one come at the expense of the other? Of course. If open weight models didn’t exist, more dollars would accrue to the frontier labs. BUT - the total spend on models won’t stay static…If it did, sure the mix shift of frontier to open would shift towards more open over time (given the starting point today on frontier closed is such a high relative percentage). But, the topline isn’t static. It’s growing exponentially. So as more tokens start shifting to open weight models, the absolute dollar spend on frontier will still grow significantly (even exponentially). I hate to reduce it down to a cliche, but the size of the pie is growing much faster than mix shift is changing.

Same debate for Nvidia vs challengers or custom silicon. Will we have more custom ASICS, or new chips from upstarts like Etched, Positron, MatX, Fractile, d-Matrix, etc? Of course. Will the mix shift of today look less “Nvidia heavy” in the future? Probably. Nvidia is just such a massive percentage today. But do I think Nvidia will continue to grow meaningfully for a while? I do!

So the answer to both of these questions isn’t one winner one looser, it’s “many winners.” And I think in the early days of an exponential it’s easy to reduce questions down to “who wins and who looses” when the real answer is “both can win.”

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.3x

• Top 5 Median: 29.3x

• 10Y: 4.7%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 18.4x

• Mid Growth Median: 4.9x

• Low Growth Median: 2.7x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 12%

• Median LTM growth rate: 16%

• Median Gross Margin: 75%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 48 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Every week I’ll provide updates on the latest trends in cloud software companies. Follow along to stay up to date!

Subscribe now

Open Weights, Closed Prices?

There’s a new open weights model out this week many are talking about - Kimi K3. It’s not technically open weights YET, but on their launch blog they said" “The full model weights will be released by July 27, 2026.” At first pass, what stood out initially is the size of the model. K3 is a 2.8 trillion parameter MoE (16 of 896 experts active per token), which makes it the largest open model ever released (I think this is true, but haven’t fully fact checked everywhere). It’s ahead of DeepSeek V4-Pro at 1.6T (which I believe was the largest before K3). It has a 1M token context window, native multimodality, and seems to be built for agentic work. A lot of the launch materials / demos focused more on things like long-horizon coding, demos where it navigated massive repos, multi-hour autonomous runs, etc. On benchmarks (Moonshot's own numbers, so take with grain of salt, the technical report isn't out yet, and you never really know what the “details” behind how benchmarks were run are), K3 beats Opus 4.8 on the majority of published evals, (but still falls behind Fable 5 and GPT 5.6 Sol). My TLDR - Kimi K3 is arguably the best open model released, sitting roughly one generation behind the frontier.

The other thing that stood out to me is the pricing. Historically, open weight models have come out and been a fraction of the pricing of frontier models. Looking at GLM 5.2, that model was $1.40 per 1m input tokens and $4.40 per 1m output tokens. Using a 80% / 20% input / output ratio (which many benchmraks standardize to), the “blended” cost of GLM 5.2 is $2 per 1m tokens. This compared to Opus 4.8 at $9 and GPT 5.5 of $10.

Kimi K3 pricing is $3 per 1m input tokens, and $15 per 1m output tokens. Blended price of $5.40. This is starting to get MUCH closer to “frontier” pricing vs “open weights” pricing. I think there’s an interesting question here - is this the trend of future open weights models? Will their pricing start to converge to frontier pricing? What happens if the price gap between open weights and frontier models starts?

There’s also another part of the pricing discussion I’ve discussed earlier. How token efficient are the models? GLM 5.2 was much cheaper, but it was way less token efficient. So for the same prompt it would use more tokens than a frontier model (by 2-3x). So the “headline” price understated the actual price to serve. I haven’t seen an analysis yet on the token efficiency of Kimi K3, but there’s a good chance when you factor in token efficiency the pricing is even closer to an Opus 4.8 or GPT 5.5.

BUT - there’s another (maybe even bigger) pricing consideration with Kimi K3. Everything I just described is API pricing (which for a closed model is all that matters). But K3 is (about to be) open weights and historically, that meant the API price was just one option. If you don’t like the pricing from Kimi, you can simply download the weights and run the model yourself (or go to someone like Baseten).

With a closed model, you just have to pay whatever the lab charges, and the price includes their margin. You don’t have an alternative. With an open model, the free license itself was partially what made it “cheap”: you could download the weights and skip paying anyone's margin at all, or the inference providers could all serve the same model and compete on price / performance.

Maybe said another way, open weights were a cost lever and the “license” was the discount. At 2.8T parameters, that math on that “lever” doesn’t exist to the same extent (or at all). The model was trained at MXFP4, so the weights alone are ~1.4TB. Meaning 10+ H200s just to load it, before you touch the KV cache on a 1M context window (napkin math so don’t hold me to it). Moonshot is pretty up front about this as well. Their own launch blog recommends “supernode configurations with 64 or more accelerators.” The company releasing the weights is also saying that you need a 64+ GPU supernode to run them well. Which get’s into the next “hidden cost” - it’s as much a capex question as a self hosting one (and if you’re not self hosting, someone else like Baseten will, and they will have to run it on a larger compute footprint).

One force that historically factored into open model pricing (anyone can serve it) is much weaker when “anyone” means “anyone with a supernode and a serving stack tuned for a brand new attention architecture.” What we don’t know is what happens when they open the weights on July 27 - will 3rd party model servers undercut the $5.40 blended pricing? Will they be able to? We don’t really know the true margins on Kimi K3, so it’s harder to comment on the true pricing until we’ve had a bit more price discovery from those serving the models. A 2.8T model probably has a structurally higher serving floor than a 1T model no matter is serving it. The K2-era 10x discount existed (partially) because those models were both smaller AND served at thin margins. K3 only gives you the second one. For two years “open” and “cheap” were used interchangeably, but they were never the same thing. They were certainly correlated, because open models happened to be small enough that anyone could serve them. K3 is the first open model big enough to break that correlation. Which brings me back to the initial point - what is the true cost of open weights models, and how are they trending.

All of this to say - it’s very easy to make the big claim that “open weight models are going to kill the frontier! They’re way cheaper but just as performant!” However, there’s usually a lot more than meets the eye on these surface level arguments (however trendy they are to say on Twitter!)

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.7x

• Top 5 Median: 32.1x

• 10Y: 4.6%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 19.8x

• Mid Growth Median: 5.4x

• Low Growth Median: 2.9x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Own Your Weights

Alex Karp had a pretty spicy segment on CNBC last week talking about a lot of things (many of which took shots at the large AI labs). The part I want to focus on today was the concept he talked about of “owning your weights.” In summary: should enterprises “own” their models or rent them (ie buy from a large lab). Should you own your weights to give yourself more control, better cost management, maybe even better performance with with RL’s models, and ultimately remove a single point of failure (or reliance). This could come in the form of developing your own models, or RL’ing open weights models on your specific workflows / data. There’s a future version of this post that will focus more on “owning your harness” vs “renting a harness” from the large labs, but I’ll save that one for another time. At the end of the day enterprises want control.

The catch on the debate of “owning your weights” - a weight file is a melting ice cube that quickly becomes outdated. Taking ownership of model weights really just buys you a snapshot, or a frozen point-in-time artifact of the model. The quality of what you own doesn’t degrade in absolutes, but it does degrade in relative terms. The frontier models get better, the goalposts move, and what you have quickly becomes outdated. What you really want is not just a snapshot, but the training and RL pipeline engine to consistently improve, tweak, update, refine, evaluate and THEN own the weights of that assembly line output. So the broader point I want to make here is it’s not just the weights you want to own, but the data flywheel, RL infra and eval harness engine that creates the weights.

Saying this differently - the naive interpretation of “owning your weights” is simply grabbing an open weights model (like a GLM 5.2), deploying it, declaring model sovereignty, and calling it a day. I don’t think that will get folks as far as they’d like.

I do think RL starts to change the game, and I do expect to frontier labs to get more into this game. The frontier labs train on EVERYTHING. The training set these frontier labs trained on doesn’t include a companies specific workflows, edge cases, business definitions, KPIs, etc. If you take an open model and RL it against your actual workflows ( in your environments, with your reward signals, on top of your data topology, etc) you’re not just holding a static snapshot anymore. You have a “personalized” model, but also an engine to continue to “personalize.” And what I keep observing - a task-specific model tuned like this can a) beat a frontier model at the one thing it was trained for (ie have better performance), and b) come at a fraction of the inference cost (usually because it’s a smaller model). The frontier model is more of a generalist model carrying ALL weights. Your model only needs to do ONE job and be good and cost effective at it. The downside is you may pay a complexity task. Play this out, and any given enterprise could have thousands of smaller fine tuned models funning around that need to be governed, version controlled, audited, secured, etc. Someone is going to make a fortune playing that role for enterprises.

So back to what I said earlier - what people really mean when they say “own your weights” is so much more than owning the weights. It’s owning the engine and loop that builds the weights. So the real question for enterprises isn’t “frontier or open source.” It’s “which of my workloads justify building a loop?”

But of course, there are tradeoffs! Building these engines and loops requires a lot of technical expertise. Today there are two main options:

1. Use the frontier for everything. You have one (or a small number of model providers), require zero internal ML team, and you ride the capability curve for free - every time Anthropic or OpenAI ships, your product gets better and you did nothing. This will actually be quite appealing to many who don’t have the resources to do anything but this. But you’re probably going to pay a lot more (you’re a price taker with a single vendor), and you have a single dependency.

2. Own your weights and RL a fleet of smaller models. You can get task-specific models that are typically cheaper, have lower latency, and don’t come with a single point of failure (or dependency)...but you’re going to be running a larger internal ML org. And every frontier release resets the bar you have to beat to justify all of those internal resources you set up. It’s now a race! Can your loop compound faster than the frontier improves on your specific task?

For most tasks (most of the time) the answer is the complexity tax of owning your weights becomes too high (especially over time, this complexity tax compounds)... Which is why there’s generally a gap between what the headlines say and what enterprises actually do (some of that is because the headlines focus more on the bleeding edge enterprises). Everyone loves the idea of ownership and having model sovereignty. BUT - The spend data says enterprises keep writing bigger checks to the frontier labs every quarter. It’s a classic stated preference vs revealed preference conundrum. Most companies don’t have any kind of foundation of infrastructure or talent to run these loops.

What I get excited about is companies that can help close this gap. Take companies who don’t have the engineering resources, and give them tools to own their weights. Again, not just own a model, but own the engine to continually build / fine tune the weights.

Maybe it’s a new startup who does this, or maybe it’s someone like Databricks. Databricks has models, but not frontier models that compete with the latest OpenAI / Anthropic models. Someone who makes the bet that the durable asset here isn’t the model itself but the engine around it. The pitch for anyone here is pretty simple: bring your workload, we’ll help you build a task-specific model on your data, tuned to your tasks, running on infrastructure you control. You get the cost and control benefits of “owning your weights” without having the huge complexity cost and tax.

Now - This doesn’t mean at all that “Ah! The big labs are dead!” Quite the opposite… There will be a difference between who generates the most tokens and who generates the most revenue. There will ALWAYS be a meaningful portion of tasks that require frontier tokens. And those tasks / tokens will be the most expensive. So I think you could see a clear future where the percent of tokens generates really starts to skew open weights, but the closed weight models continue to generate more and mroe revenue as a) the size of the pie grows, and b) their share of “expensive” workloads grows and grows. I still believe the frontier will win huge amounts of revenue. The hard stuff, the reasoning-heavy work, the tasks still riding the capability curve - plus every company that can’t or won’t build the loop - that spend stays with the labs (AND comes with a lot of pricing power).

So should you own your weights? Not entirely the right question. The real question is should you own a loop - and for which tasks?

A follow up in the future I want to write about - simply “codifying” a workflow for a model to learn from is no easy task! I brushed this off in this post as something simple, but it’s far from simple to do.

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.5x

• Top 5 Median: 31.8x

• 10Y: 4.6%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 20.3x

• Mid Growth Median: 5.4x

• Low Growth Median: 2.8x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The End of Compute Scarcity? Not So Fast

I think there are two big questions hitting AI markets right now:

1. If SpaceX (via xAI) and Meta are all of a sudden turning around and renting out their compute capacity, are we really in a compute crunch or is the system starting to fill up with excess?

2. What does the rise of open source models and the end of tokenmaxing mean for Anthropic and OpenAI?

On point 1 - SpaceX famously created ~$2.32B of monthly revenue by selling (ie renting out) about 450k of GPUs to Anthropic, Google and Reflection. Details below.

Then this week news broke that Meta is planning to start a cloud computing business selling compute capacity. Looking at both of these trends (SpaceX and Meta selling capacity) you can build a bear narrative:

• There’s excess compute capacity that two of the largest buyers no longer need. This in turn translates to the end of “scarcity of compute”

• This is the canary in the coal mine for downward capex revisions for the hyperscalers

• Maybe demand simply isn’t there for what’s being built with AI (Zuckerberg also said this week AI agent tech is developing slower than anticipated)

I don’t think this bear thesis has legs for a few reasons. First, let’s look at some of the details of these deals. Meta hasn’t actually sold anything yet. On SpaceX (xAI), all of their deals are VERY short term in nature. Each side (both seller and buyer) have a 90 day “out” where at any point in the contract they can tap out and stop paying (or providing the compute) within 90 days. The price SpaceX was able to charge was also extraordinarily high. I believe SpaceX structured the deals this way because they plan on using the compute themselves in the future. And the buyers (Anthropic, Google, Reflection) were willing to pay such high prices because they couldn’t get the capacity anywhere else.

I also think you need to look at who’s selling the capacity (SpaceX and Meta). Dario had a great segment on a Dwarkesh podcast talking about the economics of compute buildouts. In summary, there’s broadly two ways to use compute: training and inference. And there’s a magic ratio of training / compute that economically makes sense for any given business (and that ratio will differ by company). Let’s look at either extreme. If you spend a lot of money on compute, and then use 100% of it for training, you won’t have any left to sell inference and make money. Or if you use 100% of it for inference you’ll have no capacity left for training and you’re R&D progress is kneecapped and stuck in time. Neither of these make sense, so there is some middle ground each comapny should deploy.

Looking at both SpaceX (xAI) and Meta. The unfortunate reality is their models were struggling…xAI basically lost their entire team and usage of their models fell off a cliff. They themselves had no demand for their own models, so they couldn’t even use their own compute to sell inference. This mean they had 100% of their compute on the shelf for training, but were also in process of rebuilding their entire team (I’m speaking in absolutes, but of course there’s more nuance than this, but I think it’s directionally accurate). And as I mentioned above, using 100% of your compute for training is a tough economic proposition. You’re going to spend a TON of money on the build, but then generate no revenue on the backend. Not a great trade. So what did they decide? Again, I’m speaking in conjecture here… They decided until they actually had models to sell, it made economic sense to rent out their capacity so they could bring money in and balance the cashflow equation. And the short term nature of these compute deals validates this - they want a 90 day out to take the capacity back when they’re ready for it. So on SpaceX selling compute, this felt like an idiosyncratic issue they themselves faced. AND the price they were willing to charge supports that scarcity still exists.

On Meta - we don’t exactly know what they’re plans are or what the deals could look like (so it’s hard to really comment), but we do know their own AI efforts have been struggling. Llama models have fallen behind meaningfully (especially compared to chinese open source models), and the team / morale issues they’re facing are well documented. This also could be a “we spent a ton of money on this compute, but aren’t generating enough revenue on it, let’s sell some capacity until we get our act together.” Again, we don’t actually know what their plans are so it’s hard to comment. I also think it’s entirely likely they see the business of AWS, Azure and GCP and just want to get into that game.

So in summary, I really don’t think there is any signal here on “scarcity ending” as from my vantage point there is no excess compute in the system. Anyone who’s willing to sell capacity (hyperscalers, neolabs, etc) finds buyers immediately. So not to be too glib about it, but I think these deals are nothingburgers on the “scarcity is ending” comments. AI laggards are just wanting to turn a cost center into a revenue generating opportunity.

Then we have point two. Does the rise of open source models and the end of token maxing spell the end of Anthropic and OpenAI’s reign?? No, I don’t think so, but the world is certainly evolving. Yippit data also supposedly points to no slow down in the labs revenue trajectory (but to be fair, if there’s going to be an effect it may just not have hit the P&L yet).

My main reaction to this “fear” is that we have to break down AI usage into two buckets

1. Who is generating the tokens

2. Who is generating the revenue

It would seem these two should be directly correlated, but I actually see a future emerging where 80% of the tokens go to “low cost budget” options (like cheaper smaller open source models), and 20% of the tokens go to the “frontier SOTA” tokens from the labs (where these tokens are way more expensive). In this scenario, I think the 20% will drive a significant portion of the revenue (because those tokens will be more expensive and for more mission critical work). To date, the “expensive” tokens have capture all of hte market share AND wallet share, but I see these bifurcating in the future. And this is a good and healthy evolution. But I deeply believe there will always be a market for frontier tokens, and even if that market is a smaller percentage of total tokens, it will be a big percentage of revenue.

To some extent we're already seeing something like this show up in enterprise data. Ramp's latest AI Index shows the top 1% of firms spending ~$7,500 per employee per month on AI while the median firm spends $11.38 (which is a huge ~680x gap). The spend is super concentrated in a small group of power users doing serious, mission-critical work, which is the group I’d imagine reaches for the frontier tokens. The tokens probably eventually spread out across cheap open source options, but the dollars aggregate / cluster at the top, which is a pattern I’d expect to hold.

And then there’s the end of tokenmaxing. This is ABSOLUTELY a trend. BUT - I think the tidal wave is too big. There will certainly be a flushing of wasteful spend that will provide a short term headwind to spending on OpenAI / Anthropic. BUT I think the macro trend of spending more on intelligence is so large it will overwhelm any optimizations. I do realize this paragraph is starting to sound a lot like 2021 when a prevailing attitude was “yes software stocks have high multiples, but they’re growing so fast that their growth will overwhelm the inevitable multiple compression.” (I probably even said something like that myself…). I’m saying something similar here - I think we’re so early in the AI S-Curve that optimizations along the way will look like small blips on the overall trend.

I think there’s a bigger risk to the large labs that I’ll write about next week. A broader trend of “own your weights” is emerging that provides an opportunity (and challenge) for the big labs.

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.5x

• Top 5 Median: 28.6x

• 10Y: 4.5%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 19.7x

• Mid Growth Median: 5.3x

• Low Growth Median: 2.8x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Time to Power

There’s a saying that will become more mainstream very soon. And that saying is “Time to Power.”

Time to Power is starting to matter more in the infrastructure buildout. There are of course bottlenecks everywhere, but access to power is a big one. Why does it matter so much? Let’s look at a hypothetical data center build. These oftentimes will cost billions, or tens of billions, and the majority is financed with debt.

As a debt provider the main question you’ll try and answer is “what is the likelihood the borrower will be able to pay back this loan, and under what timeline.” To answer that question, they’ll dig in on when the cash register will start ringing for the borrower (ie when the borrower will generate revenue). You could procure the chips, acquire the land, do everything necessary to build the data center - but if you can’t power it nothing matters. Without power tokens aren’t’ generated, and the borrower’s (ie builders) cash register never ringts.

This leads lenders to zero in on a key question - when will you have power flowing to your data center? Or said another way - what is your time to power? The faster the borrower can prove they’ll have access to power, the more likely they are to get financing for their broader project.

On top of this - any builder (think any neocloud) is not just trying to lock up the supply side of the equation (ie the data center, which requires financing), they also need to bring demand (ie customers or offtakers of their data centers). For any customer, they won’t want to commit time and resource to a net new neocloud site unless they believe that site will have access to power (on a timeline that makes sense). So time to power also matters for the end customer.

In this equation you’ve got financing on one side and customers on the other. Before either wants to commit a dollar they are both asking “what is the time to power.” The gating factor is time to power. Chips, land, and capital are all real constraints, but are solvable if you’re well capitalized. Power is trickier.

As I’ve dug in on this problem and question a lot has surprised me. I used to think the main problem behind time to power was generation - we just need to build more plants, manufacture more generators, etc. But this isn’t entirely the case. There’s actually TONS of generation being built (and power already being generated today that is looking for a “home”). The real problem is getting the power from where it’s made to where it’s consumed (ie data halls in a data center).

The US has roughly 1,200-1,400 GW of total installed and grid-connected capacity today (according to some Claude research). I focused my research on looking at everything - plants, panels, turbines, etc. The crazy figure is the amount of power that’s sitting in interconnection queues waiting to be connected to the grid (but stuck in that process). So what’s the figure of power that’s stuck in interconnect queues? It’s somewhere in the range of 2,300 to 2,600 GW. We have roughly two grids’ worth of power waiting in line! This doesn’t mean we have that much power built generating electrons today. This capacity in queues is often times projects. Projects that need to prove they have someone to sell to (more often than not the grid) so they can get financed and then built. So IF we made it easy to connect power, we’d generate a ton overnight (not literally overnight, but the generation potential is there). More capacity is stuck and trapped in the queue than the entire country currently has plugged in.

This trapped power is everywhere, and for all of these new data center build the faster they can “un-trap” this power the faster their time to power is.

The main point I’m making - we don’t have a generation problem, we have a transmission problem. The electrons exist (or are close to existing in the case of projects), they’re just not being “connected” fast enough. And it kind of gets worse. Most of what’s sitting in that queue never end up getting built - historically only something like 10-20% of queued capacity ever reaches commercial operation. The projects that don’t get there withdraw from the queue, usually because the wait drags on too long or the cost to connect gets way to expensive. Even in a record year like 2026 we’re adding maybe ~86 GW. Against a 2,500 GW line. So this queue ends up being a death march where most projects die...

This gets us to the actual bottleneck: transmission!

Most people want to talk about generation - the gas plants, the nuclear restarts, the giant solar farms, etc. But generation isn’t exactly the hard problem anymore (it still is, but not to the relative extent people imagine). The hard part is the wires. You have to move power from where it’s generated, across high-voltage transmission lines, down into a substation, and finally into the data hall where the GPUs sit. It sounds simple enough, but each one of these steps is its own multi-year project, that requires it’s own set of financing and approvals, and unfortunately they oftentimes happen in sequence vs in parallel.

PJM (the grid operator covering a big chunk of the mid-Atlantic and east) has some good data on this. Projects are spending something like 3 years just getting to an interconnection agreement... and then another ~4 years after that before they actually come online. You can “clear” the queue (by signing all the agreements, getting the environmental studies, etc) and still be years from energized, because the transmission upgrades and the substation and the transformers don’t exist yet (or just aren’t in place yet). Transformer lead times alone are now north of 160 weeks. Said another way, it could take 3 years just to get the equipment that steps voltage down so the power is usable.

When I said “time to power” matters to lenders at the beginning of this piece, this is where it all starts to matter. This many year gap is one of the main things lenders are underwriting. A site can have 1GW of power purchase agreements (PPAs) signed on paper but not get close to 1GW delivered for years. Unfortunately a contract is not delivered power, and a LOT can go wrong from contract signed to actual power delivered. A huge key to answering the “time to power” question is proving you’ve solved the transmission problem (not simply signing the PPA).

Then there’s the last mile. How do you get power delivered to the data halls to the GPUs? The grid delivers medium-voltage AC, but chips want DC (and not much of it, GPU cores run at right around a single volt). Getting from one to the other has traditionally meant a number of conversion steps along the way. Transformers, switchgear, rectifiers, the distribution gear - and every one of them looses a little energy along the way as heat. If you’re just looking at one rack this really doesn’t matter much. The lost power at each step is just a rounding error. BUT, at the scale of a gigawatt campus (or even a couple hundred mega watts) it starts to really add up and eat away at a real piece of your power budget that dissapears before it hits the GPUs. This problem is also getting harder, not easier, because rack density is going up meaningfully. Power per rack has gone from ~40kW in the H100 era to ~120kW with GB200, and it’s headed toward a full megawatt with Nvidia’s Rubin Ultra racks in 2027. At those densities the old approach just starts to break. Feeding a single 1MW rack the traditional way takes something like 200kg of copper busbar (according to Claude) - and a gigawatt data center would need on the order of 200,000kg of copper just in the rack busbars. This just starts to get crazy…

This is the whole reason Nvidia is pushing the industry toward 800 volt DC. Instead of flipping between AC and DC several times on the way in, you convert once at the facility, distribute high-voltage DC across the data center, and only step it down to chip-level voltage at the very end, inside the rack. All of the steps I mentioned above start to go away. Nvidia claims this new architecture strips out up to four conversion stages, cuts copper use ~45%, and takes ~30% off TCO (vendor numbers, so grain of salt - but I’m sure it’s directionally accurate). When power is the scarce input, the efficiency this last mile (and specifically how few times you flip between AC and DC), turns into a real competitive edge (especially on the cost side).

This is why we’re seeing whole industry rethink how power moves inside the building - pushing higher voltage closer to the rack, moving toward native DC distribution, redesigning the power path to remove as many conversion steps as possible, etc.

There is a lot standing between a neocloud and a running GPU. The grid operators (PJM, ERCOT, etc) were built to manage slow, predictable load growth. They weren’t set up to manage a single offtaker to ask for hundreds of MWs let along GWs at a time! These requests were the size of small cities historically! Then there’s the utilities (the ones running the studies and financing the upgrades), who are now seeing more data center demand in their queue than their all-time peak load. Then there’s the regulators (FERC, DOE, etc) - trying to reform a process that just was never designed for this and is always a step behind. And the bring-your-own-generation or behind-the-meter folks (Meta, Amazon, Google, xAI, Crusoe, etc) are all basically saying “the queue is broken so we’ll just build our own power on site to skip the line entirely.” Behind the meter is a real trend, with an interesting “takeaway.” The most sophisticated buyers in the market have decide it’s faster to become power producers than to wait for the grid..That’s kind of wild. To me it points to a much bigger problem - the entire electric grid and system was simply not set up to support the power demand of data centers and the intelligence build out. We need something new. A new DC grid. New transmission. New equipment at every step.

Bring it back to where I started the article. The lender wants to know when the cash register rings. The customer (the offtaker the neocloud will sell the capacity to) wants to know the site will actually run and get power. Both questions are really about power, and specifically how fast you can get it the whole way from the plant to the rack. Forever it’s seemed like constraints on building in tech were talent and capital. And both are abundant! You can hire anywhere and raise money anywhere. But now we have to also ask if the electrons will show up. The winners of this next build won’t just be the ones with the most chips or the cheapest capital (of course that matters). They’ll be the ones who figured out, early, that the scarcest input was never compute - it was time to power.

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 2.9x

• Top 5 Median: 26.7x

• 10Y: 4.4%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 16.6x

• Mid Growth Median: 4.6x

• Low Growth Median: 2.4x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Workflows are King

Workflows workflows workflows…I’ve probably written about this in the past (hard to keep track of everything I’ve written at this point, so I’m sure there will be some repetitive content!) but I wanted to circle back to it.

For SaaS companies, the conventional wisdom was that you developed a moat if you built a system of record - a platform that stored data. If you owned / controlled the data, you had a moat! Data has gravity. True, but that wasn’t the real moat. The real moat was the hundreds of workflows that grabbed data from that system of record and then got work done. Sometimes those workflows originated from the system of record platform itself. Sometimes the workflows originated elsewhere, and one stop of the workflow grabbed data from the system of record.

The real moat is owning the platform that the most workflows touched (and the platforms the most workflows touched was almost always the system of record). Swapping out the system of record was nearly impossible because it didn’t just mean porting the data to a new system of record, it meant rebuilding / verifying / testing / securing / etc ALL of the workflows. Many of these workflows were in the critical path or customer facing. The cost of going through this change management was nearly always greater than the value of swapping to a new system of record.

I think the same thing is repeating with AI, with a twist. In SaaS, the moat lived at the data layer. Workflows formed around to the system of record, and this is what made it so hard to rip out. With agents, we still have workflows, but those workflows start to become more dynamic vs static (for SaaS workflows). So now the moat / anchor moves up a layer. The moat isn’t a derivative of where the data sits (ie workflows touching data), it’s where the work gets orchestrated. This is similar to what I wrote about last week about the value of becoming the Clearinghouse.

So what would my take away be for founders? I think the instinct when you read this (the moat is now the orchestration) is to go try to build the orchestration layer from the start. I don’t think this should be the takeaway! At least that’s not how systems of record because what they did. Salesforce wasn’t the platform a thousand workflows touched from day 1. They started with something more narrow / niche, owned a single use case, got really good at that one thing, and then slowly expanded outward until more and more workflows were built around them.

I believe the agentic version will have the same playbook (albeit from a different starting point). Pick a single workflow (probably something that looks like it’s about to get commoditized), something that will become more important in the future but appears niche today (you want to be riding the right wave), and do this much better than anyone else. People will assume there’s no defensibility in this one niche (and why you’ll get underestimated).

Then use this one workflow / wedge as strategic real estate. Build other workflows around it that are adjacent. Then slowly build the orchestration layer around it all (the thing that manages, routes, and governs the agents doing the work). One of my favorite sayings is “startups aren’t static.” What you do today isn’t the only thing you’ll do. You’ll expand, go deeper in certain areas, etc. What’s important is that where you start is strategic real estate you can build around. Then you earn the right to “manage” (ie orchestrate) all of these workflows (ie agents)

The end state really is a new kind of system of record. In SaaS, the database held data and the moat formed around it in the form of workflows. In an agentic world, the “database” from the SaaS world becomes workflows. And the company that owns where those workflows get orchestrated I think will own the next moat.

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 2.9x

• Top 5 Median: 26.4x

• 10Y: 4.5%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 16.8x

• Mid Growth Median: 4.7x

• Low Growth Median: 2.5x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

[table]

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

Systems of Record Won the SaaS Era - Clearinghouses Will Win the Agents Era

Back in December I wrote about the fight to become the front door to the systems of record. In that post I wrote about distribution, who sits between the user and the data (and why sitting there is strategic). This post is an expansion of that post (and the 1 or 2 I wrote after about similar topics). What really should AI companies be racing towards? If systems of record won in the SaaS era (ie they had the durable moats), what’s the equivalent in the AI era? Of course the answer is still partially “the system of record”, where maybe you swap out “record” with something like “agents” or “work".” But let’s come up with something new :)

Let’s start by looking at the SaaS era, and what qualities created durable successful companies. In SaaS, one main goal just about every company aspired towards was to become a system of record. Store and govern critical data. Own the place where workflows kicked off (or where workflows touched). If you got there (Salesforce for customer data, Workday for employee data, NetSuite for financial data) you ended up with an incredibly deep moat. Everything integrated with the system of record. And when you had hundreds of independent workflows and solutions built on top of the system of record, even thinking about ripping out that foundational layer was impossible. Ripping out a system of record was so painful that customers would really avoid it at all costs... The change management was almost never worth it. If you think upgrading a piece of software to a new version breaks things, imaging ripping out the layer all your workflows touched.. Price increases, mediocre product velocity, difficult upgrade cycles... didn’t matter. The data was in there, the workflows were built around it, and the switching costs got harder and harder with every year that went by.

So what’s the equivalent prize in AI? I think it’s becoming The Clearinghouse for agents.

In financial markets, the clearinghouse sits between different parties that aren’t able to fully trust each other. The clearinghouse verifies / authorizes / settles trades, and ultimately keeps the receipt. Nobody really loves the clearinghouse, but it’s clear it has to exist for the ecosystem to transact.

Now think about where enterprise software is heading. Agents from tons of different vendors, acting autonomously, touching your most critical data, and even in the future spending real money. Some company has to sit in the middle of all that and decide: which agent is cleared to act? On what data? With what limits? And can you prove what happened after the fact? Whoever holds that seat holds incredibly “strategic real estate.” (and every founder I’ve worked with has probably heard me discuss strategic real estate over and over). That’s the clearinghouse.

This may sound counterintuitive, but owning the clearinghouse for agents (given agent companies themselves will want to be the clearinghouse) may create a deeper moat than the one systems of record had. A system of record controlled your data. It kind of controlled your workflows (but not always, oftentimes someone else controlled the workflows, but the data in the system of record was a critical part of the path). The Clearinghouse controls four things: memory (what your agents know), context (what they see and how it’s served), execution (what they’re allowed to do), and governance (who’s allowed to do what, plus the audit trail behind all of it). If migrating off a system of record was painful, migrating off the thing that holds your policies, your permissions, and your entire audit history is probably harder (especially when the agents start to handle more and more of the work). AND - I think these agent companies that become The Clearinghouse will start to look more and more like systems of record in their own right. Data in systems of record were oftentimes transactional data. Data in agent systems of records (ie Clearinghouses) will be agent traces, agent evals, agent telemetry data, agent A/B data, etc

And governance is obviously a really important part of the story here. Governance used to be more of a compliance checkbox at the end of the sales cycle. It was the thing the security team made you sit through after the deal was basically done and agreed to by the business folks. Now it’s what CIOs will focus on at the forefront of a deal (or at a minimum are thinking about from meeting #1). Once agents act autonomously, the buying question changes. It’s no longer “is the model good?” Every model is good (or good enough). The question becomes “can I see what every agent did, set policy on what it can touch, and prove it to my auditors?” The biggest companies are already showing the way... Databricks leads with evaluation and governance. KPMG just announced they’re wrapping Microsoft’s Agent 365 in their “Trusted AI” framework. These companies are selling clearance. Governance is some very strategic real estate on how you earn the clearinghouse seat (and like systems of record, what makes it nearly impossible to lose once you have it).

This is why everyone is racing. Microsoft (Agent 365 + Copilot woven through Windows and Office), Salesforce, Snowflake, Databricks aren’t fighting over model quality anymore (not that they ever really were…). They all see the same prize: the clearinghouse is the toll booth every agent action passes through.

And lots of people going after this! The data players (Snowflake, Databricks) believe they win from below (data gravity becomes clearing gravity, because the agents have to come to where the data lives). The OS and productivity players (Microsoft) believe they win from above (own the surface where employees actually kick off agents, and the clearing happens wherever the user is). The new crop of Agent native companies believe an entirely new layer emerges! All arguments are right about one thing: whoever wins gets to define the knowledge graph, the governance frameworks, and which workflows get automated first.

Which brings me to the founder question. If you’re building in this market, ask yourself honestly: am I on a path to becoming a clearinghouse, or am I building a feature that clears through someone else’s?

The good news - there are real paths to the clearinghouse seat, even for startups. The first is vertical: earn clearinghouse status in an industry the horizontal players won’t go deep on. Things like owning proprietary data, simplifying regulatory complexity, or workflow depth they can’t absorb from the outside. The second is a bit harder, but more impactful. Become the clearinghouse across clearinghouses. No enterprise is going to run only Microsoft’s agents. Or only Anthropic models. Someone has to govern the multi-vendor mess (the parallel was multi-cloud). Set policy across all of it, clear actions across all of it, hold the audit trail across all of it, etc. Become the “single pane of glass.” The incumbents can’t credibly be neutral here (Microsoft governing Salesforce’s agents? Good luck). That neutral seat is a super strategic position in software right now.

I’ll end with this. Becoming a system of record got you the moat in SaaS. Becoming The Clearinghouse for agents gets you that same moat in AI. The lock in just moves from your data to your permissions. The source-of-truth era is transitioning into the source-of-permission era. Founders should pick their path to the clearinghouse now, because in 18 months there won’t be room to be undecided.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.1x

• Top 5 Median: 26.8x

• 10Y: 4.5%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 16.7x

• Mid Growth Median: 4.8x

• Low Growth Median: 2.6x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 44 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Where Are the American Open Source Models?

I’ve been investing in open source companies for nearly my entire venture career. I love open source businesses, think they’re generally great for the ecosystem, and can also create a lot of commercial value (but this can be tricky!). We’ve seen all kinds of open source businesses become successful. Databases (Mongo, Clickhouse, etc), Data Infrastructure (Databricks, Confluent, etc) Developer Tools (Hashicorp, GitLab, etc). And many other categories.

The nuance lies in how you define “open source.” A lot of this comes down to what open source license the open source project uses. At a high level there are three main buckets. First is fully open and permissive (MIT, Apache 2.0, etc) - anyone can use, modify, and distribute the code however they want, including building commercial products on top of it. Second is copyleft (GPL, AGPL, etc) - the code is open, but if you distribute or build on it you have to open source your own code too. Third is source available - the code is publicly readable but the license restricts commercial use in specific ways. This is where you see licenses like BSL (Business Source License) and SSPL, which HashiCorp and MongoDB have both moved to at various points. The main point with source available licenses is that you can basically do everything you can in the permissive license, EXCEPT monetize the project (ie you can’t take the project, create a hosted version, and charge for it). The most “pure” opensource proponents would say only the fully open and permissive licenses are truly open source. Everything else isn’t.

The playbook for many of these open source companies was / is to first become the default in their field. This is the most important part. Spark became the default in large scale data processing, Kafka became the default for real time data streaming and event ingestion. Mongo for NoSQL databases. Etc. Each project became widespread and ubiquitous in their respective fields - they were / are the dominant platform for their use case. Then the commercial companies behind each project figured out a way to monetize. Sometimes this was through hosted offerings - where the company runs and manages the infrastructure for you, so you don't have to worry about deploying, scaling, or maintaining the project yourself. In other instances, proprietary features were gated behind a commercial offering (this could be better performance, enterprise features, etc). Regardless - there were playbooks to monetize, and successful examples of large commercial businesses being built on top of successful open source projects.

But this wasn’t always the case. For a long time, the conventional wisdom was that open source couldn't be a real business. If the code is free, who pays? And the skepticism was genuine - why would a company pay for something their engineers could just download and run themselves? Early open source companies had to fight hard to prove out the commercial model, and a lot of investors passed on these deals early on because the business logic wasn't obvious. "You're just going to give it away?" was a common reaction. And honestly, it wasn't an unreasonable question. The answer turned out to be: yes, you give the software away, but you charge for everything around it - the hosted version, the enterprise features, the support, the integrations, the proprietary engines, etc. The open source project served as distribution, and the commercial product was something different. And after a wave of successful open source commercial companies emerged, it became more obvious what was possible (commercially). And this was very important - investors (like myself) deeply believed successful open source projects could translate into successful commercial companies. This created an important incentive set - an incentive to fund open source companies (by investors) and an incentive to start commercial open source companies (by founders)

Let me tie this back to what’s happening in open source AI. Today, one of the biggest disappointments is a lack of a strong open source model ecosystem in the US. All of the frontier models are closed source from OpenAI / Anthropic. The leading opensource models are largely all from Chinese labs. DeepSeek being the most prominent example, but also Qwen from Alibaba, Kimi from Moonshot AI, and others. Jensen Huang has talked about the importance of having American leaders at every layer of the AI stack - chips, models, infrastructure, applications. The US is clearly winning at chips (Nvidia) and has strong incumbents at the application layer, but the open source model layer is a real gap. And that gap matters more than people realize. Open source models are how developers learn, experiment, and build. They're the distribution layer for the next generation of AI applications. If the dominant open source models all come from Chinese labs, that has long term implications for where developer communities form, where tooling gets built around, and ultimately where commercial ecosystems emerge.

So why hasn’t there been more of a blossoming US open source model ecosystem? I think it all comes down to incentives. In the same way investors had questions about the business model of open source companies many years ago, they have similar questions about open source labs today. How do you monetize an open source model? You could “host” it - which in the case of models probably means just means hosting an API endpoint people can hit for inference. BUT if you’re doing this you’re also going to directly compete with anyone else who can pick up your model and host an inference API endpoint. Baseten, Fireworks, all the hyperscalers, etc. And the question then becomes do you as the model provider have an advantage that lets you serve the model either more performantly or more cost efficiently (or both). The answer to this question is probably “maybe” but not by enough to win with high margins over the long term.

The next natural place is to move into fine tuning / eval feedback loops. Most folks want to take an open source model, fine tune it for their use case / data, observe / monitor / improve that fine tuned model (and rinse repeat). So the open source labs could move into this line of business and charge for it. But then you’re still going to compete with the Fireworks / Baseten / Hyperscalers who will also offer this.

And then there’s the question of what does it mean to be an open source model? There's actually a spectrum here too - and it maps pretty cleanly to the licensing buckets we talked about earlier. The MIT / Apache equivalent in the model world would be truly open weights with no commercial restrictions - download it, fine tune it, host it, build a business on top of it, do whatever you want. The source available equivalent would be something like Llama - the weights are publicly available, you can experiment and build, but there are commercial restrictions that kick in at scale. And then true open source, the equivalent of releasing the full source code, would mean releasing not just the weights but the training data, the training pipeline, the evals, everything needed to fully replicate the model from scratch. Almost nobody does this. So when we talk about the "open source model ecosystem," we're really mostly talking about open weights models with varying degrees of commercial restrictions. Which matters, because just like source available licenses changed the commercial dynamics for traditional open source companies, the restrictions baked into open weights models will shape where developer communities form, what gets built on top, and ultimately who captures the commercial value.

I think the main issue facing the open source model ecosystem is incentives. People (ie investors) just question the business feasibility of open source models. Today, most of the world is using the frontier models for everything. However, what’s really come into focus very recently is costs. It just doesn’t make sense to drive the Ferrari down the street to the grocery store…The Honda will do just fine. You don’t need the “expensive frontier” tokens for everything. I don’t know where the line is, but let’s just use the 80/20 rule. Maybe 80% of the use cases are generic enough where a smaller, less powerful, but cost efficient model will do fine. And for the 20% of use cases you’ll really need the frontier expensive tokens (i’m completely making up that ratio, just using an illustrative example).

So then - why shouldn’t open source dominate in this next part of the cycle?? When people shift usage from most expensive tokens to cheaper tokens, wouldn’t open source be the natural alternative? Yes - but where are the models?? The other baked in assumption is that the open source models are “good enough.” Said another way, the gap between the frontier closed source models and open source isn’t “that wide” (and more importantly is the trajectory widening or closing).

Right now all the best open source models Chinese models… And these models have largely gotten to where they are by distilling frontier models. But I think that was much easier when the models were more “basic” language models. Question / answer type models where you could create a data set by asking millions of questions to the frontier model. I think distillation is getting harder. The most frontier models are now being released with increasingly complex harnesses. Or the frontier models aren’t released via an API but are wrapped into a product themselves. Or look at Claude - playing around with Claude Code, I have workstreams that kick off 50+ agents in parallel. How do you distill a model with 50+ agents running in parallel?? It’s much harder than just getting millions of question / answer pairs… All of this makes it much harder to “distill.” And the cost to pre-train (without the head start of distilling) is massive. All of this to say - I find myself today probably on the side of the gap widening (not shrinking) between the frontier and open source.

So who will be able to raise massive amounts of money, to pre-train a model, that will maybe work, but might end up being too far behind the frontier to capture real market share, and even if you get all of the previous points right there are still questions about how you monetize over the long haul (the first half of this post). Everyone WANTS this to happen. But who’s going to fund it?

I go back to the early days of Neoclouds where Nvidia really propped up the ecosystem. They funded (massively) a class of companies who might not have gotten the funding elsewhere. Do we need King Jensen again?? Or do we need governments to fund it out of a sovereign interest?? I’m not sure. But I really hope the US open source ecosystem finds a way to thrive.

What I think it will take - some sort of research breakthrough that open source gets first. Let’s hope that happens! I also think (no inside info) at some point the closed source frontier labs will get back into open sourcing models. Maybe not the frontier. Maybe a slimmed down version of the frontier, or maybe open sourcing one model release prior to the frontier. This is the most hopeful scenario!

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.4x

• Top 5 Median: 28.9x

• 10Y: 4.5%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 16.3x

• Mid Growth Median: 5.0x

• Low Growth Median: 3.0x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 12%

• Median LTM growth rate: 16%

• Median Gross Margin: 76%

• Median Operating Margin 2%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 42 months

• Median S&M % Revenue: 34%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 13%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The Second Life of a GPU

Last week I wrote a post on the opportunity for Neoclouds. At the end I teased out an idea that these businesses could really surprise people if chips retained value after a 4-5 year useful life, and I wanted to unpack that a bit this week.

First - it’s important to go through some of the unit economics / business model of these Neoclouds to understand why the useful life of these chips matter. There’s largely three different types of “deals” different offtakers (ie labs, hyperscalers, AI natives, etc) make with these neoclouds. Bare metal, “managed kubernetes”, and “full cloud.” Bare metal is the most stripped-down offering. The neocloud delivers the physical GPUs, networking, and power, and the customer brings everything else (their own scheduler, orchestration, storage layer, software stack), essentially renting the raw iron. Managed Kubernetes is the middle ground. The neocloud handles the orchestration layer on top of the bare metal (so the customer doesn't have to babysit cluster management, node failures, networking config, etc), but the customer is still running their own workloads and software. Full cloud is the closest analog to what AWS / Azure / GCP offer. The neocloud bundles compute with a full suite of services (storage, databases, networking primitives, managed inference endpoints, observability, etc), and the customer is effectively buying a "cloud" experience rather than just chips. As you move higher up (from bare metal to full cloud) generally you can charge higher rates / hour because you’re offering more.

For this post, I’ll use more “bare metal” assumptions. From the Neoclouds perspective this generally means they’re signing a single offtake site - one customer takes the entire site. It’s not a “cloud” where many different customers (ie scaling AI native startups) all use / share the resources.

To understand the unit economics, let’s first dig into the costs. Neoclouds generally buy the hardware (vs lease, there’s tradeoffs to both). BUT - buying gives you access to the residual value of the chips (which is the point of this post), and generally a lower cost of capital. The other point to highlight is many of these businesses are structured as “sub projects” or “sub entities".” Think of these projects as an individual data center. There’s a parent co (the Neocloud), with a number of projects (ie data centers) sitting beneath the parent.

To finance each project, the sources generally include some combo of project level debt, project level equity (think co-investors who invest equity for return on their equity), and equity from the parent co.

The debt is generally tied to the single contract for the site. The debt provider is looking at counter party risk. What is the likelihood that customer is around for the life of the contract? If they stick around and pay, we (the debt provider get paid). If they go away, go out of business, break their contract, etc, then the Neocloud has to find a replacement customer or they risk not being able to pay back the debt. So it really matters who the end customer is. If that end customer (the offtake) has high credit worthiness, a good reputation, etc, the debt provider probably gives a better rate.

When I say the debt is tied to the contract, what I mean is the debt is amortized over the lifetime of the contact. If the single offtaker is signing a 4-5 year deal, the debt will have a 4-5 year term (ie paid off in 4-5 years).

Ok - so let’s bring this back to the original point of the article. What are the implications of chips having a longer useful life than 4-5 years. After the Neolabs have “completed” a contract with an initial offtaker (ie 4-5 year deal), the debt has been fully paid off. IF the chips have value after that period, the Neolab can “recontract” the site. It will certainly be at meaningfully lower price per hour per chip, BUT without an interest expense the profit margins can skyrocket. These recontracting deals will almost certainly be for inference vs training (folks will want to train on frontier chips), BUT if you believe inference will explode (which I do), there will be SOO much demand for inference compute in the future. And these recontracting deals could prove to be extremely profitable.

I’m writing this post much later in the evening than normal, and I think I’m starting to ramble / not write as clearly as I’d like…But the takeaway - I believe chips will have a longer useful life than the original contracts they’re being signed to support, and if they do we could see a meaningful lift to neolab profitability.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.2x

• Top 5 Median: 26.8x

• 10Y: 4.4%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 15.9x

• Mid Growth Median: 4.8x

• Low Growth Median: 2.6x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 0%

• Median FCF Margin: 21%

• Median Net Retention: 109%

• Median CAC Payback: 33 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 15%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The Neocloud Boom

** Caveat — This post is not a recommendation on current Neocloud stocks, or commentary on their valuation, but instead some personal observations on the market as a whole **

A Neocloud boom feels inevitable. Clicking out one layer, the data center infrastructure buildout feels like it could turn into one of the largest wealth creation moments ever in physical infrastructure. Now that I’ve spoken in absolutes like this, we can bookmark this post for later when we look back on “signs of the top” :)

Let me caveat this post with the fact that I’m very AGI pilled. Just about any estimate for “tokens consumed by X date” or model progress or data centers built or total demand I’m taking the over.

In all seriousness, the numbers are staggering. Rumors / reports peg Anthropic / OpenAI at ~3-3.5GW of capacity to end 2025. OpenAI has talked about getting to 30GW by 2030. Let’s assume Anthropic has similar plans. Just those two alone will bring on (or plan to bring on) ~55GW over the next ~4.5 years. Now let’s assume OpenAI / Anthropic represent 30-40% of the new builds over the coming 4.5 years (I asked Claude for its best guess on this figure, I have no more specific insights here other than Claude’s estimation). That’s >150GW of new capacity in the next 4.5 years alone.

Ballpark figures are ~$50b total costs to build 1GW. For a Blackwell dominated data center, that roughly breaks down into:

• ~70% for chips / compute.

  • 50% GPUs

  • 10% networking

  • 5% CPUs

  • 3% Storage

  • 2% assembly

• ~20% for data center shell + power infra (substations, transformers, switchgear, backup generators, hv interconnect, etc)

• ~5% cooling

• ~5% land, permitting, and other costs

So in total, ~$50b for a 1GW build, multiplied by 150GW = $7.5Tn total spend. Spread across 4.5 years that’s ~$1.7Tn / year, or ~5% of annual US GDP (~$32Tn / year). I asked Claude to come up with some historical precedents to put this in perspective:

• US Railroad building out in the 1880s was ~6% of GDP

• Telcom building in early 2000’s was ~1.2% of GDP

• The Manhattan Project was ~0.4% of GDP

So this “AI Buildout” is high, but not as high as the railroad buildout.

So that’s the costs - $7.5Tn over 4.5 years. Someone is going to make a ton of money financing all of that.

But how about the revenue / enterprise value creation serving all of it. This is where neoclouds come in. Here’s a quick scorecard on 3 public neoclouds today: Coreweave, Nebius and IREN

This is a totally made up multiple, but Coreweave / IREN are both at ~$90b of enterprise value per live GW. Nebius quite a bit higher (again, not a comment on valuation, just doing some fun math).

Let’s say the Coreweave / Nebius “math” holds for rest of industry and $90b of enterprise value is created for every deployed GW. If we’re going to bring ~150GW online in next 4.5 years, that equates to $13.5Tn of enterprise value creation…Again, this is all funny math…Lots of that build will be done by hyperscalers, a lot of it for their own internal consumption vs to sell to others. But the numbers are wild!

Even if you assume only ~20% is captured by Neoclouds, that’s over $2.5Tn of enterprise value created in the next 4.5 years. The numbers are so big, it can’t be served by only a few players. The only natural explanation is that we’ll see a boom in Neoclouds, hence the title of this post. In a few years I bet there will be a handful of very large Neoclouds, and then a long tail (including independently owned single site neoclouds, similar to how the IPP (independent power providers) market played out. Expect PE to play a big role!

Obviously you could say my initial assumption is BS - that we’re not going to bring online another 150GW by 2030. Either because you don’t believe AI will garner that demand (but I’ll point you to my initial caveat! I’m very AI pilled…). or you think the costs are just too high as a percent of GDP. But let’s dream the dream - this could be a MASSIVE opportunity…And as far as the venture communities go, there is still a lot of skepticism around neoclouds (I’d ask them, who else is going to build out the compute for AGI!). In many ways, Neoclouds are the perfect hedge on a venture portfolio :) If everything else goes to zero because AGI captured all the value, someone will have to build, deliver, and serve that compute :)

Ok last little bit here. I was done writing the post, but thought I couldn’t write a post about Neoclouds and not bring up SpaceX. SpaceX? How are they a Neocloud? I’m sure most of the readers of this blog have been plugged into the AI news cycle, but in case not, here’s a quick summary. xAI was a AI lab Elon founded. They created two very large data centers of their own. Colossus 1 (which is estimated at ~300MW), and Colossus 2 (which has total capacity of 2GW, but estimates peg it at ~500MW operational today). SpaceX then bought xAI. They then proceeded to “rent” capacity at Colossus 1 and Colossus 2 to Anthropic! In the SpaceX S-1, they gave some details on this commercial relationship.

According to the S-1, Anthropic is paying SpaceX $15b per year (!) for this capacity. They supposedly have all of the capacity in Colossus 1 (~300MW). Impossible to know how much of Colossus 2 they have, but let’s assume it’s ~40%? That would give them an incremental ~200MW. SpaceX/xAI has to keep some for itself…and Cursor has announced they have some of that capacity (which Elon as agreed to buy as well…are you keeping up with all the acquisitions??).

So that’s $15b / year for ~500MW. To put those numbers in perspective:

• SpaceX did $18.7b of revenue in the full year 2025. And they just “created” a new business line item of ~$15b in an instant! Insane. SpaceX was founded in 2002. It took them 23 years to get to an annual revenue figure of ~$19b. And with their new “neocloud” they generated almost that same amount. The opportunity for Neoclouds is massive!

• This math also assumes Anthropic is paying ~$30m / MW ($15,000M / 500MW). That would be really high…Most neocloud comps are closer to the $10-$12m / MW. Coreweave guided to ~$18.5b annualized run rate revenue to end 2026, and said they’re “on track to reach or exceed our target of more than 1.7GW by the end of 2026.” This math comes out to ~$11m / MW. If Anthropic was paying $11m / MW that would mean they have ~1.4GW of capacity from SpaceX (which isn’t possible).

The $30m / MW does feel high…BUT - the contract has a 90 day out. Anyone can cancel the deal with a 90 day warning. And Anthropic is VERY compute constrained. So might they stretch on a one time deal with an out? Maybe!

Either way - I’m bulled up on the opportunity for Neoclouds :) The revenue potential is clearly massive. Now they just need to prove they can make the business model work! On revenue multiples, CoreWeave trades at ~6x NTM rev, Nebius / IREN at ~10x. On NTM EBITDA, CoreWeave is trading ~10x, with IREN / Nebius ~20x.

I believe trillions of enterprise value will be created in the neocloud space in the next 5+ years. Go grab it! I may write a piece next week that discusses why these businesses could really explode if the chips retain value after years 4-5.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.2x

• Top 5 Median: 25.0x

• 10Y: 4.6%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 15.6x

• Mid Growth Median: 4.8x

• Low Growth Median: 2.6x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 1%

• Median FCF Margin: 20%

• Median Net Retention: 110%

• Median CAC Payback: 36 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 14%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The Real App Store Opportunity

Last October I wrote a piece saying OpenAI had their “App Store” moment after they released the Apps SDK. 7 months later that prediction doesn’t look great… I don’t think we’ve seen an explosion of custom ChatGPT apps. ChatGPT hasn’t turned into the “super app” yet. Hopefully one day they will!

I think there may be a separate “app store” moment happening with Anthropic. BUT - more of a B2B app store moment than B2C apps. Over the last few months I’ve seen massive adoption of “skills” in Claude. A skill is essentially an "onboarding doc" for an AI agent - a folder of instructions (often just a markdown file) that Claude pulls in only when the task calls for it. Anyone in a company can write one in an afternoon, which is why I think the distribution dynamic looks less like a consumer app store and more like internal tooling that spreads bottoms up. At our own firm we’ve seen a proliferation of skills being created and shared with the team.

What really changed thigs for me was the /skill-creator “skill.” It’s become so easy to start building skills (maybe skills are the new name for agents??) to automate real parts of my day to day. The biggest challenge I have with building skills is a creative one. I know there are 10x more ways I can be using / building skills, but I don’t always know where to start. So I discuss with my colleagues, with other friends in similar lines of work, to try and explore how other people are leveraging skills so I can do something similar.

The “app store” moment hasn’t happened yet, but I think we’re close. Today, we can share skills within our own organization. But we can’t yet share them externally. And there’s not great approval flows for making sure a built skill follows all security / compliance protocols. Anthropic has their own set of pre-built skills (including ones from partners like Notion, Figma, Atlassian, etc), but it’s managed by them (ie isn’t a marketplace). I met with a software vendor this week who has some of their own “skills” available out of the box (the interface is their own software UI, but they still called some of the automated workflows “skills.”)

All of this to say - I think the trend here is a set of skills people can build and then publish to some sort of marketplace. I think you can kind of do this today by exporting the markdown file and publishing on GitHub, but that feels janky. What people really want is to be able to find a skill, know it’s somehow been “blessed” by Anthropic (ie put through security reviews, deployed within Claude behind their governance, etc), and quickly customize it for their own set of integrations / permissions.

There's another angle here that I think founders building AI products need to internalize. For the last decade of SaaS, "we integrate with X" was table stakes. You couldn't sell into the enterprise without a long list of logos on your integrations page. I think "we have a skill for that" is about to become the equivalent over the next 12 months. If your customer's primary surface for getting work done is Claude (or any agent), and your product can't be invoked as a skill from inside that surface, you're functionally invisible. It doesn't matter how good your product is - if the agent doesn't know to reach for you, you're not in the workflow.

Everyone today has a “blank canvas” problem. They know there’s skills they could be building, but they don’t know where to start. A marketplace like experience of skills just makes too much sense not to happen! When will we see our first venture funded “skill”??

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.1x

• Top 5 Median: 23.4x

• 10Y: 4.5%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 14.7x

• Mid Growth Median: 4.4x

• Low Growth Median: 2.3x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 0%

• Median FCF Margin: 21%

• Median Net Retention: 109%

• Median CAC Payback: 33 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 15%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

When Machines Out-Eat Humans

Fun post this week. I want to write about my own (newly formed) definition of AGI. I think we'll hit AGI (or we can claim AGI) when as a society we decide the marginal unit of energy is better spent on a GPU (or whatever compute primitive exists at the time) than on a human. Said another way - when the energy consumed by compute becomes greater than the energy consumed by humans, we're making the implicit decision that we get higher utility out of sending energy to machines. All definitions of AGI are super wishy washy anyway, so why not through another into the mix! The reason I like this one is it's quite quantitative. I’ve run the math, and the answer is 2033 (as you’ll hear me describe later, it’s all a bit “funny math dragging assumptions to the right”) but that’s what makes it fun!

First - let’s start with some definitions. In today’s world people use power vs energy interchangeably, but they’re different. Power is a rate (how fast you’re using energy at any given time), while energy is a quantity (how much you’ve used over time). An oversimplified analogy - power is the speed your car is going (mph), energy is the total distance (miles) you’ve traveled. A car going 60mph for 2 hours has covered 120 miles. A data center pulling 100MW for 2 hours has consumed 200 MWh (mega watt hours).

From a unit standpoint, watts are the unit of power. My lightbulbs might be 9W (It needs 9W to turn on). I keep it on for 1 hour, and I’ve used 9Wh. Kilo / Mega / Giga / Tera are just scaling prefixes (1,000x, 1,000,000x, etc). When we hear power figures quoted for data centers we often describe them in mega watts or giga watts. For example, the Colossus 1 data center that xAI rented to Anthropic this week is quoted as 300MW. What does this mean? It means Colossus 1 needs transmission lines, substation(s), generation source(s) that can sustain 300 million watts of instantaneous draw. Think of this more as a physical infrastructure problem.

This data center won’t consume 300MW x 8,760 hours / year. Maybe it runs at 70-80% utilization (but AI training clusters are “running” constantly at higher utilizations).

The confusion most people have with power vs energy, is the “hour” component is not “division” it’s “multiplication.” It’s not MW per hour, it’s MW-hours (and yes, I just wrote a “it’s not X it’s Y sentence…tried to edit it out so it doesn’t sound like AI slop, but the it’s not X it’s Y just made too much sense). Super quick math to make this concrete - let’s say my house consumes 1KW for 10 minutes, then 1.3KW for 50 minutes. To calculate the total energy used:

• From 0 to 10 minutes: drawing 1 kW for 10/60 of an hour = 0.167 kWh

• From 10 to 60 minutes: drawing 1.3 kW for 50/60 of an hour = 1.083 kWh

• Total energy over the first hour ≈ 1.25 kWh

Let’s also use the 300MW Colossus site, running at 80% utilization(ie running 80% of the time)

• 300MW x 8,760 hours / year x 80% utilization = 2,102,400 MWh

• This translates to 2,102 GWh, or 2.1TWh

• So the energy consumption over a year is 2.1TWh

I walked through that lead in so that I could then bring it back to my new definition of AGI! When will total energy consumed by machines be greater than humans. To answer that, let’s first try to get a baseline on how much energy is consumed by humans to sustain human life! A very meta question…Obviously there’s no easy answer, but I used Claude to help frame it (ie guesstimate it). Focusing on just US.

Here was Claude’s answer: 27,600 TWh of total US energy consumption. This is how Claude broke it down:

Narrow definition:

• Residential energy use: 5,600 TWh

• Food systems (ag + processing + transport+ retail): 2,900 TWh

• Healthcare delivery: 590 TWh

• Water + sewage treatment: 290 TWh

• Total = 9,400 TWh

Moderate Definition. Everything in the narrow definition plus:

• Personal transportation: 9,400 TWh

• Commercial buildings: 4,100 TWh

• consumer goods manufacturing: 2,000 TWh

• Total = 20,500 TWh

Most Aggressive Definition

• According to the US Energy Information Administration (EIA) the US consumed 94.2 quads of energy in 2024. Converting that to TWh = 27,600 TWh

• This is all the estimate for all of the energy the US economy consumes, in any form, measured at the source. Every barrel of oil burned in a car. Every cubic foot of natural gas burned in a furnace. Every ton of coal or pound of uranium fed into a power plant. Every BTU of solar or wind generated. etc. A figure you may see quoted if you ask Claude how much electricity is consumed annually in the US is closer to 4,000 TWh (but that is just looking at the electric grid > offtake figure). Most energy is never electricity to begin with (and that’s the gap). This is the difference between the 4,000 TWh figure and the ~28,000 TWh figure

• The gap between the “moderate” definition and “aggressive” definition is largely military, freight, etc

So in summary, we can look at annual human energy consumption in the US as

• 9,400 TWh in most conservative case

• 20,500 TWh in the moderate case

• 27,600 TWh in the most aggressive case

What’s a good assumption for total annual data center energy consumption? I asked Claude, and the answer was ~300TWh. Here’s how the math breaks down

• Roughly 40GW of total data center capacity (this felt slightly high, but the Claude estimate was based on all hyperscalers, all labs and neolabs, and all neoclouds in the US)

• If you assume 80% utilization, then:

• This equates to 280TWh of data center energy consumption

So today, the data center energy consumption is 280TWh. Earlier, I said the “moderate” estimate for annual energy consumption to support humans was 20,500TWh. This means today, data centers consume 1% of the energy humans do. Long way to go to AGI :)

So when will we hit the 50% threshold?? Let’s do some more “funny math” :). Let’s say annual population growth is 0.5%, end users consume 1-2% more energy per year, but we get 1-1.5% energy efficiency gains / year. This leads us to ~0.5-1% annual growth in energy consumption, or 21,440 TWh in 2030.

Now, let’s assume OpenAI is going from 2GW in 2025 to 30GW in 2030. And let’s assume the rest of the industry grows at the same rate (maybe aggressive, but just for fun let’s make that assumption). This would get us to 12,600 TWh in 2030 used for data centers. The lines are getting closer! 12,600 TWh for data centers vs 21,440 TWh for human consumption = almost 40%!

Ok final assumptions :) Let’s assume data center growth continues at the same rate in 2030 and beyond as it’s projected to from 2025 to 2030. And let’s assume human consumption continues at ~0.75% / year beyond 2030. When do the lines cross?? The answer is 2033!! This would continue an exponential, which is very hard to do…But if we went from 40GW today to 600GW in 2030, at the same compounding annual growth rate we’d be at ~3TW in 2033. Doing all the funny-math converting that to TWh gets us to ~22,000TWh by data centers in 2033. Extrapolating the human consumption at 0.75% / year gets us to ~22,000 TWh in 2033. The lines have crossed!

Of course, this is all assumption upon assumption upon assumption, and I’m no energy expert at all, but thought it’d be fun to run through this exercise.

AGI by 2033!

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.2x

• Top 5 Median: 21.1x

• 10Y: 4.4%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 12.6x

• Mid Growth Median: 5.2x

• Low Growth Median: 2.5x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 1%

• Median FCF Margin: 20%

• Median Net Retention: 110%

• Median CAC Payback: 36 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 14%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The Death of Per-Seat Pricing?

All three Hyperscalers (Amazon, Google, Microsoft) reported earnings this week.

There was one quote from this earnings cycle that I think will get a lot less attention than it deserves. It came from Satya:

“The basic transformation of any per-user business of ours - whether it is productivity, coding, or security - will become a per-user and usage business. That is the best way to think about it.”

Big statement! The per-seat licensing model is the foundation that the entire modern SaaS industry was built on. It’s how so many IT budgets are structured. It’s how every renewal conversation goes. It’s how every comp plan is designed. Did Microsoft just say the seat is dying? Or at least, being repurposed? This isn’t the first time I’ve written about this topic...Nearly 2 years ago I wrote a similarly titled article, “Is seat based pricing dead“?

Amy Hood went further. She framed the new model as “a licensed business plus a consumption business applied far more broadly than I think people have thought about.” And later, Satya put a really nice frame on it - “the seat-based pricing is just entitlement to some consumption…there are some base usage rights that get bundled in or packaged into seats.” Said another way - the seat remains, but it becomes a packaging mechanism for prepaid consumption. Beyond a certain level, you’re paying per token, per agent action, per outcome.

They also announced a larger pricing change of their own! They moved GitHub Copilot to a usage-based model effective June 1. ~60% of their Dynamics 365 customer service customers are already buying usage-based credits. The Copilot credit-consumption offer was up nearly 2x QoQ. They’re already living in the new model.

There’s actually a tell buried in the bookings line. Hood called out that D365 bookings growth was impacted by weaker renewals “as customers balance spend between the traditional per-seat and the emerging seats-plus-consumption model.” The transition is already showing up as a drag on a legacy bookings metric. Customers are pausing on their old seat renewals because they’re trying to figure out the new pricing. It’s happening right now.

Why does this matter for founders building software companies?

A few reasons. First - if Microsoft is publicly committing to this shift, every other software company gets some air cover to do the same. The market will tolerate the transition pain because the largest player in the world is wearing the same lumps. If you’ve been afraid to introduce a usage-based component to your SaaS pricing, watching Microsoft go first is permission.

Second - the metric set you’ve been using to evaluate your business will need to change. ARR is a clean metric for a per-seat business. It is a much messier metric for a “per-seat + consumption” business. NRR will likely become more volatile (consumption ebbs and flows). Bookings will be lumpier. The “rule of 40” framework, the public market revenue multiple, the way you set sales quotas - all of it gets harder. Investors are going to have to learn a new language, and it’ll probably take some time to work its way through the system.

Third - the agent economy is fundamentally a consumption economy. An agent doesn’t buy a seat. It does a task. It uses tokens. It calls tools. If you’re building a software company today, and your business model assumes you’re going to charge a flat fee per human user, you’re solving for a world that’s about to get smaller. The world that’s growing is “per outcome, per task, per token.” Get there before you have to. Just about every startup I’ve worked with that went through some sort of seat > usage business model transition regretted not doing it sooner…

The thing I keep coming back to is - Microsoft is the least incentivized of any company in the world to do this. They have the largest installed base of seat-based subscriptions on the planet. Office, Windows, Teams, Dynamics. Every dollar they “convert” from seat to consumption introduces volatility into a model their investors have been comfortable with for two decades. The fact that they’re still leaning into it tells you everything about where they think this is going.

Seat-based pricing is becoming more of a wrapper than a product. The product is the work that gets done.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.2x

• Top 5 Median: 18.4x

• 10Y: 4.4%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 10.7x

• Mid Growth Median: 5.0x

• Low Growth Median: 2.4x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 0%

• Median FCF Margin: 21%

• Median Net Retention: 110%

• Median CAC Payback: 33 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 15%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

The Impending Employment Explosion

I’m a perpetual optimist. It’s hard for me to see the world through any other lens! Sometimes I’m naive, but overall I think it’s a better way to live (and, generally, hard to bet against humanity’s resilience).

One of the larger debates surrounding AI relates to the impact it will have on employment. One side (booo, the pessimists!) argues we’ll see a collapse in employment as AI takes everyone’s jobs. The other side argues some form of “Jevon’s paradox” - with massive positive economic benefits. Given my intro, I think it’s clear what side of that debate I fall on!

There are many ways I’ve framed this in the past (I think I’ve even written about it in a prior week’s edition). However, I heard Jensen recently articulate it much more eloquently (surprise surprise, he’s better at framing this than me!). The crux of his argument (which I wholeheartedly agree with), is that the framing of “AI will take jobs” is wrong. In order to actually debate this, you have to separate the job, from the task. What does this mean? Let’s use an example. The job of a software engineer is to build high quality software. A task of a software engineer is writing code in an IDE. AI may in fact automate one task of a job (ie automating writing code which will no longer be done manually), but it won’t replace the job - building high quality software. Because there is more to the job than just one of its tasks. In fact, I’d argue AI will have the opposite impact! AI will make the task much easier, leading to an explosion of people doing the job! This isn’t anything novel (it’s a common claim to say AI tooling will lower the barrier to creating software, leading to lots more software created!). But I hadn’t heard it framed the way Jensen did - separating the job from the task.

What’s the common pushback to this? Well what if the one task represents a significant portion of the job! Saying software engineers are “software builders” is just putting lipstick on a pig! They’re really just code writers!” I get the pushback, but I think it misses something important. The hard part of being a software engineer was never the typing. It was never the syntax. It's understanding the problem deeply enough to know what to build. It's knowing how to architect a system, how to make the right tradeoffs, how to debug the thing when it breaks in some way nobody anticipated. The code is the output (not the job). And look, maybe the skeptic is right that writing code is 60-70% of the time spent today. Fine! But that just proves the point even more. If you can compress 70% of the time, you don't get rid of the person… you get that person shipping 3-4x more. And what does every company in the world want? More software. More tools. More internal apps. More automation. More everything. The demand for software is nowhere close to saturated. We've been supply constrained, not demand constrained. So when you make it much easier to supply software… you don't get fewer engineers. You get more. The nature of the software engineer job will change (probably a lot!). But the job itself? It gets bigger, not smaller.

Here’s a hypothetical example I like to use. Let’s say we existed in a world before cars were invented. And if you wanted to travel from your home to the grocery store you used a horse drawn carriage. Let’s call the person operating the carriage the chauffeur. Then, let’s assume some technological innovation happened EXTREMELY quickly, and all of a sudden cars were hitting the road (the key part of the analogy here is the part in all caps - this change happened QUICKLY, just like AI has). A pretty pessimistic take may be “Ah! This is terrible! Think of all the chauffer’s who will loose their job!” But it’s pretty easy to see the world through a different lens. What it means to be a chauffer stays the same (you help people get from point A to point B). The task of a chauffer was operating a horse drawn carriage (and maintaining it, buying parts for it, etc). Now, the task of a chauffer changes! Their job is still helping people get from point A to point B. BUT the task changes - now they need to learn how to operate a vehicle (how to drive a manual transmission car) vs how to use the reins to help horses navigate.

And what happens in this world? We have way more chauffer’s because transportation has instantly become way more efficient (ie less productivity waste). At the same time, entire new industries are born servicing the automotive industry! Auto mechanics, oil and gas, auto manufacturers, etc.

Yes - the task of operating a horse drawn carriage, or manufacturing the carriage itself went away. But the job of a chauffer saw a spike in demand (leading to more employment), and a trickle down of job creation in adjacent industries.

I’m so convinced we’ll see the same thing happen with AI, and an explosion of employment is coming. And we do have some precedent to support this! ATMs came out in the late 1960’s. However, from that point in time to the mid 1980’s the number of bank tellers in the US doubled. Tellers went from cash-in / cash-out machines (the task) to relationship bankers who cross-sold financial products (the job). The task got automated, but the job got more valuable. At the same time, the task got way cheaper! According to Claude: “Before ATMs, you needed ~21 tellers to staff a branch. After ATMs hit saturation, that dropped to ~13. So the cost per branch went down meaningfully.” Cheaper branches lead to more branches. And more branches led to more tellers (in aggregate). I asked Claude to chart the number of bank tellers in the US. What do you see, number of bank tellers exploding at the same time the ATM came out)

You could make a similar argument around lawyer growth in the 70’s, 80’s and 90’s. The job of a lawyer is to protect and advance their client's interests, the some of the tasks include researching case law and drafting documents. Before the PC, a huge chunk of a lawyer’s time was spent on tasks that technology was about to automate. Legal research used to mean physically going to a law library, pulling case books off shelves, reading through indices, cross-referencing citations by hand. That could eat days or weeks. Then Westlaw (1975) and LexisNexis (late 70s) came along and compressed that to hours. Later, full-text search made it even faster.

Document drafting was similar. Contracts, briefs, motions - all typed up by hand or dictated to a secretary who typed them. Every revision meant retyping the whole thing. Word processors (and later PCs with WordPerfect / Word) made it so one lawyer could produce 5-10x the document output. So did things that automated the “tasks” of lawyers (PC, internet, LexisNexis, word processors, etc) lead to a reduction in the number of jobs held by lawyers? Here’s a chart from Claude:

I feel like there’s so many other examples I could use. Radiologists (this is an example Jensen uses). In 2016 Geoffrey Hinton (Godfather of Deep Learning!) predicted radiologists would go away because of AI. What happened? The job of a radiologist is patient care. The task is reading and interpreting scans. AI got really good at the task. But the job didn't shrink (it grew!). Because the demand for imaging exploded (more types of scans, more conditions screened for, aging population), and radiologists' job of providing patient care evolved toward more complex interpretive work, interventional procedures, and clinical consultation

The through line is the same: automating a task is not the same as automating a job. And more often than not, it leads to more of the job, not less. I expect AI to be no different.

Quarterly Reports Summary

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.0x

• Top 5 Median: 18.4x

• 10Y: 4.3%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 10.6x

• Mid Growth Median: 4.9x

• Low Growth Median: 2.2x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 0%

• Median FCF Margin: 21%

• Median Net Retention: 109%

• Median CAC Payback: 33 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 14%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment

Subscribe now

I have two topics I want to write about today:

1. Are we in an “over earning” period in AI

2. Follow up to last weeks post on the importance of the harness

Rising Tide, Hidden Risk

I’ve been thinking a lot recently about comments Satya made a few years back. If we rewind the clock to mid / late 2022, the biggest thing on software companies & investors’ mind was “when will the optimizations end.” The ZIRP period of 2020-2021 created a buying frenzy - no one was thinking about costs (when it came to cloud / cloud software spend), they were only thinking about growth and capturing more market share (oversimplification, but you get the main point I’m making). At the end of the day, the market was providing cheap and abundant capital (for public and private companies), and investors (for both public and private companies) were rewarding growth (ie placing the most emphasis on growth when determining valuation multiples). Once ZIRP ended and interest rates went up, capital got more expensive, investors started rewarding different behavior (efficient growth), and ultimately every CFO took a look at their P&L and realized they had a TON of unnecessary spend. Redundant spend, inefficient spend, or completely unnecessary spend was everywhere. After a couple years of “reckless spending” the tide turned, and every CFO was looking for “optimizations.” Way to reduce spend, consolidate spend, and in general reduce the wasteful spend. This was pain lasted quite some time! And this brings me to comments Satya made over and over again on earnings calls - cloud consumption trends tend to be quite cyclical. You have period of rapid expansionary spending, followed by periods of rapid optimizations - and this cycle repeats.

When I look at where we are in the AI spending cycle, we’re very clearly in a period of “rapid expansionary spending.” No one is really thinking about the ROI of the spend when budgeting. So many of the companies I work with say some version of “I don’t want to cap how much our employees use these tools, I want their creativity flowing, I want them trying out everything, and ultimately I want them becoming experts in these tools.” A common conversation went something like “our spend on AI tools (more often than not Claude) doubled from Jan to Feb, tripled from Feb to March, and is trending way up in April. And I don’t care if it 10x from here.” Then of course on top of this you have the “tokenmaxing” companies. Meta was in the news on this recently - the idea is “as an employee I need to show I’m using these AI tools a lot so I can show my boss I’m “AI Fluent" so I don’t get laid off. I want my token consumption to be super high, so I’m just going to have background agents running constantly, doing nothing, to give my boss the perception I’m using these tools a lot, even though they’re all doing empty tasks.” Clearly this kind of excess spend can’t last forever…

The common thread? No one is thinking about the costs. Just the potential.

I don’t know when the tide will turn, or what will cause it to turn. But at some point people WILL start looking at their AI spend and start thinking about “optimizations.” Should I be using smaller cheaper models vs larger ones? Should I be segmenting different use cases for different models? Should I run a particular background agent once per day vs once every hour? the list goes on.

In 2022, what caused the tide to turn was interest rates. Now? Interest rates won’t (I don’t think) be the reason the tide will turn. What most likely happens is the spend on AI tools balloons SO much, that companies have no choice but to think about the costs, and how to optimize. And when this happens, it could be painful for a number of AI companies experiencing hyper growth currently. There are just too many markets, that have too many competitors, that are all growing at ridiculous rates.

The thread I’m pulling at - a lot of companies in AI today are “over earning” just like many cloud companies were “over earning” in 2020 / 2021.

Just to call it out - I think the total token consumption will SKYROCKET over the next 10 years regardless of when this period of optimization happens. But today, everyone seems to be benefiting. When the tide shifts, the dispersion of who will truly benefit will shift, and a smaller subset of companies will benefit disproportionately.

Call out for founders - really introspecting your own customer / revenue / growth to understand “am I growing as part of the rising tide or am I differentiated” will be quite important…Many companies in the 2021 period got this wrong - they didn’t realize they were just catching the rising tide. They over hired, over fundraised, etc. And the unwind was painful.

Topic #2: Is the Harness Really the Moat?

In summary, last week I wrote a post that argues that the code and orchestration surrounding an AI model (the "harness," maybe formerly known as a "wrapper") matters a ton - maybe even more than the model itself? I’m back to write a quick follow up to this post.

I do think the harness matters a ton, and will continue to matter. BUT - it’s far from the only thing that matters. And we’re aready starting to see the evolution of some of the most successful application AI companies. (1) Start as a thin wrapper, (2) develop a complicated harness, (3) post train your own models, (4) lean into pre training your own models.

The vast vast majority of the market is still in phase (2). But we’re already seeing companies like Cursor (and other early AI winners) move into 3 and 4 (some form of developing their own model).

There are many reasons why AI companies may want to have their own model.

First - cost. Costs balloon when using someone else’s model. Your fixed costs will be higher developing your own, but variable costs may be lower. Second - capability. A purpose built model for your application and your domain could perform better than a general model. Third - control of your own destiny. No one wants to be beholden to someone else for the most critical layer of their infra…

And fourth - you may have no choice! There’s a very real world where the large labs (OpenAI / Anthropic) decide to keep their most powerful and current versions of their model for themselves, powering their own applications. And the models they release via API to customers may be “one version prior” of a model (ie one before the latest and greatest) or a distilled version of the current powerful model.

Why would they do this? Two main reasons I can think of. The labs are getting more and more into the application space. Why would they want their competitors (ie Cursor competing with Claude Code from Anthropic or Codex from OpenAI) to have access to the latest and greatest? Why not save the latest and greatest for their own product, and make their competitors use an older, less powerful version? I could certainly help the labs win more business.

Second (and this is all purely speculation), as these models get more and more powerful (ie Mythos speculation from Anthropic), the labs may view them more and more as a safety / security threat. Maybe they will want to be the only ones who “release” them to the world wrapped around a product? They could view themselves as the only ones truly capable of “protecting” users from the uncapped potential of the models. Said another way, they might trust themselves more to build applications around the powerful models, but not your average developer..

End of the day, this is all speculation. It may happen, it may not. But if you’re a breakout AI company, do you want to leave it to chance? Probably not. Developing your own model carries too many benefits, and could protect you in too many “downside cases” that I think we’ll see more and more companies move from the harness as the first moat, to developing their own model as the second moat.

The takeaway for founders I work with - are you building your own research / labs team? if not, maybe time to think about doing it…

Top 10 EV / NTM Revenue Multiples

Top 10 Weekly Share Price Movement

Update on Multiples

SaaS businesses are generally valued on a multiple of their revenue - in most cases the projected revenue for the next 12 months. Revenue multiples are a shorthand valuation framework. Given most software companies are not profitable, or not generating meaningful FCF, it’s the only metric to compare the entire industry against. Even a DCF is riddled with long term assumptions. The promise of SaaS is that growth in the early years leads to profits in the mature years. Multiples shown below are calculated by taking the Enterprise Value (market cap + debt - cash) / NTM revenue.

Overall Stats:

• Overall Median: 3.2x

• Top 5 Median: 17.2x

• 10Y: 4.3%

Bucketed by Growth. In the buckets below I consider high growth >22% projected NTM growth, mid growth 15%-22% and low growth <15%. I had to adjusted the cut off for “high growth.” If 22% feels a bit arbitrary, it’s because it is…I just picked a cutoff where there were ~10 companies that fit into the high growth bucket so the sample size was more statistically significant

• High Growth Median: 10.5x

• Mid Growth Median: 5.0x

• Low Growth Median: 2.4x

EV / NTM Rev / NTM Growth

The below chart shows the EV / NTM revenue multiple divided by NTM consensus growth expectations. So a company trading at 20x NTM revenue that is projected to grow 100% would be trading at 0.2x. The goal of this graph is to show how relatively cheap / expensive each stock is relative to its growth expectations.

EV / NTM FCF

The line chart shows the median of all companies with a FCF multiple >0x and <100x. I created this subset to show companies where FCF is a relevant valuation metric.

Companies with negative NTM FCF are not listed on the chart

Scatter Plot of EV / NTM Rev Multiple vs NTM Rev Growth

How correlated is growth to valuation multiple?

Operating Metrics

• Median NTM growth rate: 13%

• Median LTM growth rate: 15%

• Median Gross Margin: 76%

• Median Operating Margin 0%

• Median FCF Margin: 21%

• Median Net Retention: 109%

• Median CAC Payback: 33 months

• Median S&M % Revenue: 35%

• Median R&D % Revenue: 23%

• Median G&A % Revenue: 14%

Comps Output

Rule of 40 shows rev growth + FCF margin (both LTM and NTM for growth + margins). FCF calculated as Cash Flow from Operations - Capital Expenditures

GM Adjusted Payback is calculated as: (Previous Q S&M) / (Net New ARR in Q x Gross Margin) x 12. It shows the number of months it takes for a SaaS business to pay back its fully burdened CAC on a gross profit basis. Most public companies don’t report net new ARR, so I’m taking an implied ARR metric (quarterly subscription revenue x 4). Net new ARR is simply the ARR of the current quarter, minus the ARR of the previous quarter. Companies that do not disclose subscription rev have been left out of the analysis and are listed as NA.

Sources used in this post include Bloomberg, Pitchbook and company filings

The information presented in this newsletter is the opinion of the author and does not necessarily reflect the view of any other person or entity, including Altimeter Capital Management, LP (”Altimeter”). The information provided is believed to be from reliable sources but no liability is accepted for any inaccuracies. This is for information purposes and should not be construed as an investment recommendation. Past performance is no guarantee of future performance. Altimeter is an investment adviser registered with the U.S. Securities and Exchange Commission. Registration does not imply a certain level of skill or training. Altimeter and its clients trade in public securities and have made and/or may make investments in or investment decisions relating to the companies referenced herein. The views expressed herein are those of the author and not of Altimeter or its clients, which reserve the right to make investment decisions or engage in trading activity that would be (or could be construed as) consistent and/or inconsistent with the views expressed herein.

This post and the information presented are intended for informational purposes only. The views expressed herein are the author’s alone and do not constitute an offer to sell, or a recommendation to purchase, or a solicitation of an offer to buy, any security, nor a recommendation for any investment product or service. While certain information contained herein has been obtained from sources believed to be reliable, neither the author nor any of his employers or their affiliates have independently verified this information, and its accuracy and completeness cannot be guaranteed. Accordingly, no representation or warranty, express or implied, is made as to, and no reliance should be placed on, the fairness, accuracy, timeliness or completeness of this information. The author and all employers and their affiliated persons assume no liability for this information and no obligation to update the information or analysis contained herein in the future.

Subscribe now

Share Clouded Judgement

Leave a comment