Original Title: Charts of the Week: Head In The Neoclouds
Original Author: Moses Sternstein, a16z

Editor's note: Against the backdrop of Generative AI driving a new round of computational investment, the market's discussion on AI infrastructure is shifting from "Do we have enough GPUs" to "Who can provide compute power sustainably." As model training, inference demands, and data center expansion have become consensus, a more fundamental question is emerging: Can the rapid growth in compute power demand truly translate into stable profits and cash flow?

In a16z New Media's "Charts of the Week," author Moses Sternstein delves into the AI compute market's growth, valuation, and profit paradox through newcomers like CoreWeave, Nebius, and Applied Digital, further extending the discussion to horizontal SaaS, model routing, and frontline lab talent competition.

In this article, the author does not simply assess whether AI demand is strong but breaks down the current AI transaction into a set of more fundamental structural issues: how existing infrastructure is being repriced, why revenue growth has not improved market expectations in tandem, and why the competitive focus of the AI industry is shifting from mere expansion to efficiency and returns.

First is the rediscovery of infrastructure value. In the past, land along railroads, natural gas pipelines, and cable TV networks have all served specific industries, then repurposed as telecom and internet infrastructure. Today, a similar reassessment of assets is occurring. Some new cloud companies that once served cryptocurrency mining operations now possess operational experience in electricity, data centers, cooling systems, and high-density computing; after the AI demand surge, these capabilities swiftly transformed into scarce compute supply. The significance is that AI infrastructure competition does not start entirely from scratch; early advantages often stem from the recombination of old assets, energy resources, and engineering capabilities.

Second is the coexistence of high revenue growth and profit uncertainty. The early revenue growth of new cloud companies like CoreWeave once surpassed that of cloud giants like AWS in their early stages, but the capital markets did not accord them equal recognition. This is because new clouds are not typical asset-light software businesses. GPU procurement, power access, data center construction, chip depreciation, and debt interest will rise in sync with scale, even faster than revenue growth. This means that revenue expansion can only prove robust AI compute demand but cannot automatically prove that the business model has a sufficiently high capital return rate. What the market truly awaits is whether these companies can convert orders and revenue into sustainable free cash flow.

Third, software value is being redefined by the impact of AI. In the past, the market was concerned that generative AI would generally weaken the moat of SaaS companies, but Atlassian's performance illustrates that AI could also be a tool to increase customer spending and product stickiness. Meanwhile, cybersecurity and observability software continue to receive valuation premiums as AI expands potential risks and increases enterprise reliance on mature solutions. This means that the so-called "SaaS doomsday" will not occur uniformly. Whether AI is a product substitute, price depressor, or demand expander is becoming the new standard for software valuation divergence.

Fourth, AI applications are shifting from "Stack Tokens" to Optimized Tokens. In the past, enterprises often tended to directly invoke the most powerful models or allocate a budget to the engineering team for experimentation; today, companies like Databricks are starting to use intelligent routing to match models of different prices and performance based on task difficulty, maintaining effectiveness while reducing costs. A decrease in Token unit price does not necessarily mean a contraction in total AI spending: as unit costs decrease and use cases expand, Token consumption volume and overall market size may continue to rise. Efficiency and demand are not mutually exclusive but may form a mutually reinforcing cycle.

If this article were to be condensed into one judgment, it would be this: AI infrastructure has proven it can drive high-speed growth, but the next phase's outcome will depend on whether enterprises can transform growth into higher capital efficiency. In this sense, the objects of this discussion are no longer just whether companies like CoreWeave can become the next cloud giants, but whether the entire AI industry can transition from compute expansion to sustainable business returns.

The original text continues below:

Embracing the "New Cloud"

At the beginning of the 20th century, the Southern Pacific Railroad Company held extensive land grants on cleared land connecting cities and towns across the United States. The railroad's right-of-way was much wider than the tracks themselves, leaving many corridors available for development along the route.

As a result, the railroad company laid down a communication network alongside the railway and named it the "Southern Pacific Railroad Internal Networking Telephony System." By the 1970s, the company began commercializing this network, opening it up to a broader range of users.

Subsequently, two things happened simultaneously: the end of the long-distance telephone market monopoly and the commercial viability of fiber-optic cables. The existing communication corridors were transformed into fiber optic lines, and this network later became known by its English acronym, "Sprint." Assets once serving the railroad became the backbone of the telecommunications revolution.

The transformation of existing physical networks into larger-scale business technology infrastructure is not limited to railway companies.

In the 1980s, Williams Company repurposed idle natural gas pipelines into fiber optic conduits, establishing WilTel. The company was later sold and eventually rebranded as WorldCom. By the 1990s, the one-way coaxial cables laid for cable TV services underwent a large-scale, costly upgrade to become the infrastructure through which Comcast and Charter provide broadband internet services to consumers.

This brings us to another category of companies: those that also have existing infrastructure assets that are now undergoing significant transformation and repricing to meet the needs of an emerging technology—these are the "neocloud" companies.

In simple terms, neocloud companies primarily started out in energy- and compute-intensive cryptocurrency mining and then the AI wave arrived. Suddenly, those who have access to power, infrastructure, and expertise in building and managing high-intensity computing loads—such as in the case of CoreWeave, which includes a large number of GPUs—are now at the forefront of one of the hottest trends.

Of course, this is not a direct comparison. However, when looking at the three largest publicly traded neocloud companies, their revenue growth is indeed remarkable.

We can only estimate the early-stage cloud business revenue of the largest cloud providers, but the general trend is clear: neocloud companies are growing rapidly, and significantly faster than the early growth of the big three cloud providers.

It is worth noting that in the overall compute sales market, neocloud companies are still relatively small players.

They still have a long way to go to reach the scale of the major cloud services providers.

The revenue generated by the major cloud services providers each quarter is orders of magnitude higher than that of neocloud companies. However, CoreWeave, in about 25 quarters, achieved a $2.6 billion revenue milestone that AWS took around 40 quarters to reach after its launch. Once again, the growth rate of these companies is indeed very high.

Having such a high growth rate and riding the wave of the AI industry should theoretically excite investors. To some extent, it does, but the reality is more nuanced.

While these companies have generally performed well in their most recent earnings reports, CoreWeave's stock price has still cumulatively fallen by about 16% over the past year; only Nebius is closer to its previous highs.

Therefore, the overall story is still positive, but the appeal has evidently weakened somewhat for the largest of these new cloud companies.

The recent market performance has been relatively flat, partly due to many growth expectations possibly already being priced into valuations.

For capital-intensive businesses like these new cloud companies, the price-to-sales ratio is not the most suitable valuation metric, but is still a straightforward illustration. Nebius and Applied Digital, which are smaller in scale and have faster growth rates, have valuations far higher than the much larger CoreWeave. CoreWeave's revenue is still doubling, but it is not growing as fast as the top company with a growth rate of 400% to 450%.

If there is a real issue facing these new cloud companies, it is not growth, but long-term profitability. These companies need to continue investing in chips, power, and physical infrastructure to scale up, and these costs are substantial:

Take CoreWeave, for example, while its revenue growth is indeed significant, its capital expenditure is staggering. Other massive costs include chip depreciation—depreciation amount has exceeded half of the revenue—and interest expenses from debt taken on for the early construction of expensive infrastructure, which are still rising.

This article does not intend to pass judgment on whether these new cloud companies will ultimately succeed or if their current stock prices are justified. Apart from the hot topic itself, what is really being pointed out here is that these new cloud companies happen to represent a microcosm of the long versus short tug-of-war in the entire AI trading space.

On one hand, they are in a vertical market—the computing power market—that is far larger than anyone had previously expected and continues to expand, creating historically rare growth rates; on the other hand, the costs of building such enterprises are at historical highs, requiring significant investment and continuous depreciation of fixed infrastructure.

Horizontal SaaS Resurgence?

Below is a brief update on the ever-evolving market landscape of the "SaaS Doomsday." One company that was among the hardest hit during the previous software stock sell-off has shown quite a good performance in the past month.

Over the past 30 trading days, horizontal software companies have been among the top performers in the IGV Software ETF's constituents—although they have retraced some of their gains since the data collection date.

Overall, the fundamentals of these companies remain strong. Especially Atlassian, it did not decline as expected by the market under the impact of AI.

This productivity software company achieved a "double beat" in performance and guidance: cloud business revenue grew by 31% year-over-year, and the growth rate of revenue backlog orders was even higher. But perhaps the more crucial signal is that AI is becoming a growth accelerator rather than a hindrance. Atlassian stated that its AI assistant Rovo has been widely adopted; at the same time, customers using Rovo have seen their spending growth rate nearly double that of non-Rovo users.

This is good news for Atlassian, good news for Rovo, and good news for horizontal SaaS.

However, the overall valuation of horizontal SaaS still slightly lags behind other software categories.

With few exceptions, including horizontal SaaS companies such as Atlassian, their forward price-to-sales ratios are generally below the level corresponding to the "growth rate - valuation multiple" trend line.

Once again, horizontal SaaS has just passed a relatively decent "month." To make the market believe that the "SaaS doomsday" has been averted, the performance of just one month is far from enough.

Of course, if your software business is in the cybersecurity or observability realm, that's a different story – for these companies, the so-called "SaaS doomsday" has never occurred.

The cybersecurity sector continues to significantly outperform other categories in the IGV Software ETF. In this field, AI has actually become a tailwind: the market generally believes that AI has increased people's risk perception of cybersecurity threats, and no enterprise customer will rely on "vibe coding" to cobble together their own security solutions.

Whether this logic will ultimately hold up still needs to be tested over time. But at least for now, the situation of traditional software companies is by no means uniform.

Investors are highly focused on whether AI will bring benefits to each company, or if it will cause erosion, continuously adjusting their original assessments with each new batch of data – and this is only natural.

Moving Toward the Efficiency Frontier of Token Deployment

The market landscape around model usage, token consumption, and token expenditure management continues to evolve in various interesting ways.

Take Databricks as an example.

When it comes to questions like "Which model should we use" and "Which model is the best," Databricks didn't take a one-size-fits-all approach or just give engineers a budget to decide how to spend it. It posed a different question: "What if we develop a solution that automatically assigns the right tasks to the right model?"

Databricks is certainly not the only company to do this, but it has developed an "Smart Router," and the actual results have been quite satisfactory.

According to reports, Databricks' router can call upon a more powerful and higher-priced model when necessary, and use a less capable and lower-priced model when conditions allow, thereby "continuously reducing the average task cost by over 30%."

Overall, pursuing the "efficiency frontier" of Token expenditure is hard to overlook as a positive development. This means that demand continues to grow, with use cases evolving not only at the performance frontier but also spreading to models that are not at the cutting edge. In the initial pessimistic narrative, these suboptimal models were expected to be quickly phased out.

As mentioned earlier, efficiency gains will expand the coverage of demand, a dynamic similar to Jevons' Paradox that the market hopes to see.

Token Price Strength Index from Silicon Data shows that overall price strength is decreasing, especially as lower-priced open models take up a larger share in the expanding market.

It is important to clarify once again a concept that is often misunderstood: these indices measure the cost strength of Token expenditure, not the absolute dollar amount. It depends on both the quantity of Tokens consumed and the comprehensive cost of Tokens. This means that even as the price per Token declines, the total Token consumption and total expenditure may continue to rise.

What truly matters is that overall demand is still growing, and pricing and model selection are gradually moving towards the efficiency frontier, further driving this growth. Particularly noteworthy is the fact that "AI demand" or "AI adoption" is not a single, homogenized concept. There is still a huge gap between heavy users and other users. This clearly indicates that "always using the best model" may be suitable for some enterprises, but certainly not for all.

Today, the market is rapidly forming more alternative options. Overall, this is a good thing.

According to the enterprise spend management platform Ramp's data, all companies are increasing their AI expenditure, but the gap between the median spend of all companies and the top 10% of companies, as well as between the top 10% and the top 1% of companies, is extremely wide.

Ramp's data tends to skew more towards tech companies, which should be taken into account when interpreting. However, the data shows that for companies ranking in the top 10% of expenditure, their per capita AI expenditure is approximately 50 times that of median companies.

This distribution is likely not coincidental. Companies that can unlock more value from AI expenditure are probably also the ones investing the most – even though not every company does, at least a significant portion seems to follow this pattern.

Boston Consulting Group's analysis of 107 public companies found that companies in the top two quintiles of Token usage had revenue growth rates significantly faster than other companies.

The core point here is that Token demand and utilization efficiency are mutually reinforcing: the more value a company gains, the more Tokens it consumes.

Of course, this process involves a constant trade-off between investment and return, and R&D always carries some upfront costs. However, for the vast majority of companies, indiscriminately "stacking Tokens" has never been an effective strategy.

Therefore, it is clearly a positive development that companies are increasingly moving away from this approach in the future.

Talent Battle in Frontline Labs

New Media recently had two outstanding team members join OpenAI, so finally, let's use a few interesting charts to look at the talent recruitment situation in cutting-edge AI labs.

Dario Amodei recently expressed concern that employees are prioritizing money over mission. Looking at data from Levels.fyi, this concern may not be unfounded.

If we purely interpret the data on the surface, Anthropic offers very high salaries to engineers, significantly higher than similarly seasoned engineers at companies like Google, Tesla, and others.

It seems that becoming a part of a tech team is indeed a good thing.

In addition, there is another set of data that is also quite interesting.

According to data from Live Data Technologies presented by Truist Securities, the talent sources of various labs show both significant overlap and distinct differences:

Both companies have recruited talent from large-cap tech firms, but only OpenAI has hired from Nvidia and Tesla, and both incidents occurred in 2026.

Databricks, Snowflake (recently), Palantir, and DeepMind are also common talent sources for the two companies.

Both labs have also recruited a considerable number of employees from Salesforce and Stripe.

However, the overlap between the two seems to end there. Anthropic has recruited many talents from SaaS companies, while OpenAI has done so sparingly; OpenAI has extensively hired from consumer internet, platform markets, and ad tech companies, whereas Anthropic has relatively fewer hires in these areas, except for Airbnb, Netflix, and Uber.

As for what these differences mean, that is left for each individual to interpret.

[Original Article Link]

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia