a16z dissects AI computing power: revenue doubles, stock price under pressure, capital expenditure astonishing

Original Title: Charts of the Week: Head In The Neoclouds

Original Author: Moses Sternstein, a16z

Editor’s Note: In the context of generative AI driving a new wave of computing power investment, discussions in the market about AI infrastructure are shifting from "Is there enough GPU?" to "Who can provide computing power sustainably?" As the consensus has emerged that model training, inference demand, and data center expansion are all on the rise, a more fundamental question begins to surface: Can the rapid growth in computing power demand truly translate into stable profits and cash flow?

In the "Charts of the Week" published by a16z New Media, author Moses Sternstein discusses the growth, valuation, and profit contradictions in the AI computing power market through the lens of new cloud companies like CoreWeave, Nebius, and Applied Digital, and further extends to horizontal SaaS, model routing, and talent competition in cutting-edge laboratories.

In this article, the author does not simply judge whether AI demand is strong, but rather breaks down the current AI transactions 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 in the AI industry is shifting from mere expansion to efficiency and returns.

First, there is the rediscovery of infrastructure value. In the past, land along railroads, natural gas pipelines, and cable networks served specific industries, only to be transformed into telecommunications and internet infrastructure later. Today, a similar asset revaluation is occurring again. Some new cloud companies that originally served cryptocurrency mining have operational experience in electricity, data centers, cooling systems, and high-density computing; after the explosion of AI demand, these capabilities quickly transformed into scarce computing power supply. The significance lies in the fact that competition for AI infrastructure does not start from scratch; early advantages often come from the recombination of old assets, energy resources, and engineering capabilities.

Second, there is the coexistence of high revenue growth and profit uncertainty. The early revenue growth rates of new cloud companies like CoreWeave once exceeded those of cloud giants like AWS during their startup phases, but the capital markets did not grant them the same level of recognition. The reason is that new clouds are not typical asset-light software businesses. GPU procurement, electricity access, data center construction, chip depreciation, and debt interest will rise in tandem with scale, even faster than revenue growth. This means that revenue expansion can only prove strong AI computing power demand, but cannot automatically prove that the business model has a sufficiently high capital return rate. What the market is truly waiting for is whether these companies can convert orders and revenue into sustainable free cash flow.

Third, the value of software is being re-differentiated according to AI impact. In the past, the market worried that generative AI would generally weaken the moats of SaaS companies, but Atlassian's performance shows that AI may also become a tool for increasing customer spending and product stickiness. Meanwhile, cybersecurity and observability software continue to receive valuation premiums because AI has expanded potential risks and increased enterprises' reliance on mature solutions. This means that the so-called "SaaS apocalypse" will not occur uniformly. Whether AI is a substitute product, drives down prices, or expands demand is becoming a new standard for software valuation differentiation.

Fourth, AI applications are shifting from "stacking tokens" to optimizing tokens. In the past, companies often tended to directly call the most powerful models or give engineering teams a budget to experiment; now, companies like Databricks are starting to use intelligent routing to match different models with varying prices and performance according to task difficulty, reducing costs while maintaining effectiveness. A decrease in the unit price of tokens does not necessarily mean a contraction in total AI spending: when unit costs decrease and application scenarios increase, the total consumption of tokens and overall market size may still continue to rise. Efficiency and demand are not mutually exclusive, but may form a mutually reinforcing cycle.

If this article can be compressed into a single judgment, it is: AI infrastructure has proven it can create high-speed growth, but the next stage of victory will depend on whether companies can convert that growth into higher capital efficiency. In this sense, the subjects discussed in this article are no longer just whether companies like CoreWeave can become the next generation of cloud giants, but whether the entire AI industry can transition from computing power expansion to sustainable business returns.

The following is the original content:

Under the "Neocloud"

In the early 20th century, the Southern Pacific Railroad Company had a large amount of idle construction rights on cleared land connecting cities and towns across the United States. The scope of railroad rights-of-way extended far beyond the tracks themselves, leaving many corridors available for development along the route.

Thus, this railroad company laid a communication network along the railway line, naming it the "Southern Pacific Railroad Internal Networking Telephony." By the 1970s, the company began commercializing this network, opening it up to a wider range of users.

Subsequently, two things happened simultaneously: on one hand, the monopoly pattern of the long-distance telephone market came to an end; on the other hand, fiber optic cables began to become commercially viable. The original communication corridors were transformed into fiber optic lines, and this network later became known by its English abbreviation "Sprint." The assets that once served the railroads thus became the backbone of the telecommunications revolution.

It was not just railroad companies that transformed existing physical networks into larger-scale commercial technology infrastructures.

In the 1980s, Williams Company repurposed idle natural gas pipelines into fiber optic channels, establishing WilTel. The company was later sold and eventually renamed WorldCom. By the 1990s, the one-way coaxial cables laid for cable television businesses underwent a massive and costly upgrade, ultimately becoming the infrastructure for Comcast and Charter to provide broadband internet services to consumers.

This leads us to another type of company: they also possess ready-made infrastructure, and these assets are now being significantly transformed and repriced to meet the demands of an emerging technology—this is the "neocloud" company.

In summary, most neocloud companies originally engaged in energy and computing-intensive cryptocurrency mining, and then the AI wave arrived. Suddenly, those who own electricity usage rights, infrastructure, and experience in building and managing high-intensity computing loads—such as CoreWeave, which also includes a large number of GPUs—stand on one of the hottest tracks today.

Of course, this is not a strict comparison. But if we observe the three largest publicly listed neocloud companies, their revenue growth rates are indeed remarkable.

We can only estimate the early cloud business revenue of ultra-large-scale cloud service providers, but the general trend is already clear: neocloud companies are growing at an extremely fast pace, significantly faster than the growth rates of the three major cloud service providers during their startup phases.

It should be noted that in the entire computing power sales market, neocloud companies are still relatively small participants.

They still have a long way to go to reach the scale of ultra-large-scale cloud service providers.

The revenue generated by ultra-large-scale cloud service providers each quarter is several orders of magnitude higher than that of neocloud companies. However, at the same time, CoreWeave achieved $2.6 billion in revenue in just about 25 quarters, a milestone that AWS only reached in its 40th quarter after launch. Again, it is emphasized that the growth rate of these companies is indeed very fast.

With such high growth rates and riding the tailwind of the AI industry, investors should theoretically be quite excited. To some extent, this is indeed the case, but the reality is more complex.

Although these companies generally performed well in their recent earnings reports, CoreWeave's stock price has still fallen by about 16% over the past year; only Nebius is relatively close to its previous high.

Therefore, the overall story is still quite good, but for the largest neocloud companies, the appeal is clearly weaker.

The recent market performance has been relatively flat, partly because many growth expectations may have already been reflected in valuations.

For capital-intensive companies like neoclouds, the price-to-sales ratio is not the most suitable valuation metric, but it is still intuitive for illustrating the issue. Smaller, faster-growing Nebius and Applied Digital have valuation premiums far exceeding those of the much larger CoreWeave. CoreWeave's revenue is still doubling, but it is no longer on par with the 400% to 450% growth rates of the top companies.

If there is a real issue with neocloud companies, it is not growth, but long-term profitability. Neocloud companies need to continuously invest in chips, electricity, and physical infrastructure to scale up, and these costs are not low:

Taking CoreWeave as an example, its revenue growth is indeed quite impressive, but its capital expenditures are even more astonishing. Other significant costs include chip depreciation—depreciation amounts have already exceeded half of revenue—and interest expenses incurred from borrowing to build expensive infrastructure in advance, which are still on the rise.

This article does not intend to judge whether neocloud companies will ultimately succeed or whether their current stock prices are reasonable. Aside from the heat of the topic itself, what is truly meant to be illustrated here is that neocloud companies precisely constitute a microcosm of the tug-of-war in the entire AI transaction.

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

Return of Horizontal SaaS?

Next, let’s briefly update the changing market landscape of the "SaaS apocalypse." The company that suffered the most in the previous software stock sell-off has performed quite well in the past month.

In the past 30 trading days, horizontal software companies have ranked among the top performers in the IGV software ETF constituents—although they have already given back some of their gains since the data collection began.

Overall, the fundamentals of these companies remain robust. Especially Atlassian, which has not declined under the AI impact as the market previously expected.

This productivity software company achieved "double beats" in performance and guidance: cloud business revenue grew by 31% year-on-year, and the growth rate of revenue backlog orders was even higher. But perhaps the more critical signal is that AI is becoming a booster for business growth, rather than a hindrance. Atlassian stated that its AI assistant Rovo has been widely adopted; at the same time, customers using Rovo have spending growth rates nearly twice 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 remains slightly lower than that of other software categories.

With few exceptions, horizontal SaaS companies, including Atlassian, generally have expected price-to-sales ratios lower than the levels corresponding to the "growth rate—valuation multiple" trend line.

Again, it is emphasized that horizontal SaaS has only gone through a relatively good "month." To convince the market that the "SaaS apocalypse" has been canceled, a single month of performance is far from enough.

Of course, if your software business belongs to the cybersecurity or observability field, that’s another story—there has never been a "SaaS apocalypse" for these companies.

The cybersecurity sector continues to significantly outperform other categories in the IGV software ETF. In this field, AI has become a tailwind: the market generally believes that AI has heightened awareness of cybersecurity threats, and no enterprise customer would rely on "vibe coding" to cobble together their own security solutions.

Whether this logic will ultimately hold true, of course, still requires time to test. But at least for now, the situation of traditional software companies is far from uniform.

Investors are highly focused on whether AI will bring benefits or cause erosion for each company, continuously revising their original judgments with each batch of new data—which is only natural.

Towards the Efficiency Frontier of Token Investment

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

Take Databricks as an example.

On the questions of "Which model should we use?" and "Which model is best?", Databricks has not adopted a winner-takes-all approach, nor has it simply given engineers a budget to decide how to spend it. It posed another question: "What if we develop a solution that automatically assigns the right tasks to the right models?"

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

Reportedly, Databricks' router can call more powerful, higher-priced models when necessary, while using weaker, lower-priced models when conditions allow, thereby "continuously reducing the average task cost by over 30%."

Overall, pursuing the "efficiency frontier" of token spending is hard not to be a good thing. This means that demand continues to grow, application scenarios are not only evolving at the performance frontier but also spreading to models that are not as top-tier. In the initial pessimistic narrative, these suboptimal models were originally thought to be quickly eliminated.

As we previously mentioned, efficiency improvements will expand the coverage of demand, which is precisely the dynamic the market hopes to see, similar to Jevons' Paradox.

Silicon Data's Token Price Strength Index shows that overall price strength is declining, especially as lower-priced open models capture a higher share in the continuously expanding market.

It is necessary to clarify a frequently misunderstood concept again: these indices measure the cost intensity of token spending, not absolute dollar amounts. It also depends on the quantity of token consumption and the comprehensive cost of tokens. This means that even if the price per token decreases, the total consumption of tokens and total spending amounts may still continue to rise.

What is truly important is that overall demand continues to grow, and the gradual movement towards efficiency frontiers in pricing and model selection will only further drive this growth. Notably, "AI demand" or "AI adoption" is not a single, homogeneous concept. There remains a significant gap between heavy users and other users. This clearly indicates that "always using the best model" may suit some companies, but certainly not all.

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

According to data from enterprise spending management platform Ramp, all companies are increasing AI spending, but the disparity between median spending and the top 10% of companies, as well as between the top 10% and the top 1% of companies, is extremely stark.

Ramp's data tends to be more biased towards tech companies, and this should be taken into account when interpreting it. However, the data shows that the top 10% of companies in spending have per capita AI spending about 50 times that of median companies.

This distribution is likely not coincidental. Companies that can unlock more value from AI spending are probably also the ones investing the most—though not every company is like this, at least a considerable portion of companies conforms to this pattern.

An analysis by Boston Consulting Group of 107 publicly listed companies found that companies in the top two quintiles of token usage have revenue growth rates significantly higher than those of other companies.

The core message here is that token demand and usage efficiency are mutually reinforcing: the more value enterprises gain, the more tokens they consume.

This process certainly involves a repeated trade-off between input and return, and R&D will always include some upfront costs. But for the vast majority of companies, indiscriminately "stacking tokens" has never been an effective strategy.

Therefore, it is clearly a good thing that companies will increasingly not need to adopt this approach in the future.

Talent Competition in Cutting-Edge Laboratories

New Media recently welcomed two outstanding team members to OpenAI, so we will conclude with a few interesting charts to look at the talent recruitment situation in cutting-edge AI laboratories.

Dario Amodei recently expressed concern that employees are placing money above mission. According to data from Levels.fyi, this concern may not be unfounded.

If we understand the data on the surface, Anthropic offers very high salaries to engineers, far exceeding those of engineers with comparable qualifications at companies like Google and Tesla.

It seems that being a member of the technical team is indeed a good thing.

Additionally, there is another set of data that is quite interesting.

According to data from Live Data Technologies—presented by Truist Securities—the talent sources for various laboratories have significant overlap but also clear differences:

Both companies have recruited a considerable number of talents from ultra-large-cap tech companies, but only OpenAI has hired from NVIDIA and Tesla, and both instances occurred in 2026.

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

Both laboratories have also recruited a significant number of employees from Salesforce and Stripe.

However, the overlap seems to stop there. Anthropic has recruited a considerable number of talents from SaaS companies, while OpenAI has almost none; OpenAI has recruited extensively from consumer internet, platform markets, and ad tech companies, while Anthropic has relatively few hires in these areas, except for Airbnb, Netflix, and Uber.

As for what these differences mean, we leave it to you to interpret.