From crypto mining farms to AI clouds: Why does a16z say the “new cloud” burns money as it grows?

Source: a16z New Media

Author: Moses Sternstein, a16z

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

Editor's note: In the context of generative AI driving a new round of computing power investment, market discussions on AI infrastructure,

There is a shift from “not having enough GPUs” to “who can provide computing power in a sustainable way”.

When model training, inference requirements, and data center expansion have become consensus,

A lower-level question is beginning to emerge: Can the rapid increase in computing power demand actually translate into stable profits and cash flow?

In “Charts of the Week” published by a16z New Media,

Author Moses Sternstein switched from new cloud companies such as CoreWeave, Nebius, and Applied Digital.

Discuss the growth, valuation, and profit conflicts of the AI computing power market, and extend further to horizontal SaaS, model routing, and cutting-edge lab talent competition.

In this article, the author did not simply judge whether AI demand is strong.

Instead, it breaks down current AI transactions into a set of lower level structural issues: how to reprice existing infrastructure,

Why is revenue growth not simultaneously improving market expectations, and why the AI industry's competitive focus is shifting from simple expansion to efficiency and return.

The first is the rediscovery of the value of infrastructure. In the past, land along railway lines,

Both gas pipelines and cable television networks served specific industries and were later transformed into telecommunications and internet infrastructure.

Today, a similar revaluation of assets happened again. Some new cloud companies originally served cryptocurrency mining.

It already has operating experience in electricity, computer rooms, cooling systems, and high-density computing;

When demand for AI exploded, these capabilities quickly transformed into scarce computing power supplies.

The point is that AI infrastructure competition doesn't start entirely from scratch; early advantages often come from a recombination of old assets, energy resources, and engineering capabilities.

Second, high revenue growth and profit uncertainty coexist.

The early revenue growth rate of new cloud companies such as CoreWeave once surpassed the initial stages of cloud giants such as AWS, but the capital market did not receive the same level of recognition.

The reason is that the new cloud is not a typical asset-light software business.

GPU procurement, power access, data center construction, chip depreciation, and debt interest will rise simultaneously with scale, or even faster than revenue.

This means that revenue expansion can only prove that AI computing power is in high demand, but it cannot automatically prove that the business model has a sufficiently high return on capital.

What the market is really waiting for is whether these companies can turn orders and revenue into sustainable free cash flow.

Third, the value of software is being re-differentiated according to the impact of AI.

In the past, the market feared that generative AI would generally weaken SaaS companies' moats, but Atlassian's performance shows that

AI could also be a tool to increase customer spend and product stickiness.

Meanwhile, cybersecurity and observability software continues to receive valuation premiums.

Because AI expands potential risks, it also increases the company's reliance on proven solutions.

This means that the so-called “end of SaaS” will not happen evenly.

Whether AI is an alternative product, lower prices, or expand demand, is becoming the new standard for software valuation differentiation.

Fourth, AI applications are shifting from “stacking tokens” to optimizing tokens.

In the past, companies often preferred to directly use the most capable models or give engineering teams a budget to test on their own;

Today, companies such as Databricks are starting to use intelligent routing,

Match models with different prices and performance according to task difficulty to reduce costs while maintaining effectiveness.

A decrease in the unit price of tokens does not necessarily mean a contraction in total AI expenditure: when unit costs decrease and application scenarios increase,

Total token consumption and overall market size are likely to continue to rise.

Efficiency and demand are not mutually exclusive, but may form a mutually reinforcing cycle.

If I were to reduce this article to one judgment, it would be: AI infrastructure has proven that it can generate rapid growth,

But the next phase of success or failure will depend on whether companies can transform growth into greater capital efficiency.

In this sense, the subject discussed in this article is not only whether CoreWeave can become the next generation of cloud giants,

It's about whether the entire AI industry can move from expanding computing power to sustainable commercial returns.

The following is the original content:

A “new cloud” above your head

At the beginning of the 20th century, the South Pacific Railroad held a large number of vacant construction rights on cleared land connecting cities and towns across the United States.

The scope of railway rights is much wider than the railroad tracks themselves, so there are still many corridors along the route that can be developed.

As a result, this railway company set up a communication network along the railway line.

It was named “Southern Pacific Railroad Internal Networking Telephony” (Southern Pacific Railroad Internal Networking Telephony).

By the 1970s, the company began commercializing this network and making it available to a wider range of users.

Subsequently, two things happened at the same time: on the one hand, the monopoly pattern in the long-distance telephone market came to an end;

On the other hand, fiber-optic cables are beginning to become commercially viable.

The original communication corridor was converted into an optical fiber line, and this network was later also known by its English abbreviation “Sprint.”

An asset that once served railways, became the backbone of the telecommunications revolution's infrastructure.

Railway companies are not the only ones transforming existing physical networks into larger commercial technology infrastructure.

In the 1980s, Williams Company (Williams Company) converted an idle gas pipeline into a fiber channel and founded WilTel.

The company was later sold and eventually renamed WorldCom.

In the 90s, unidirectional coaxial cables installed for the cable TV business also experienced a round of large-scale and expensive upgrades.

It eventually became the infrastructure for Comcast and Charter to provide broadband internet services to consumers.

This has led to another type of enterprise: they also have ready-made infrastructure, and these assets are now being drastically remodeled and repriced.

To meet the needs of an emerging technology — this is a “neocloud” (neocloud) company.

In brief, Xinyun originally mostly engaged in energy- and computational-intensive cryptocurrency mining business, and then the AI wave arrived.

Suddenly, who has the right to use electricity and infrastructure,

and experience in building and managing high-intensity computing workloads — in CoreWeave's case,

Also included are plenty of GPUs — whoever is on one of the hottest tracks of the moment.

Of course, this isn't strictly a similar comparison.

However, if you look at the three largest listed Xinyun companies, their revenue growth rate is indeed remarkable.

We can only estimate the cloud business revenue of hyperscale cloud service providers in the early stages of their business.

But the general trend is already clear: the growth rate of new cloud companies is extremely fast, and significantly faster than the initial growth rate of the three major cloud service providers.

It should be clarified that in the overall computing power sales market, Xinyun is still only a relatively small participant.

They are still a long way from being hyperscale cloud service providers.

The quarterly revenue generated by hyperscale cloud service providers is several orders of magnitude higher than that of new cloud companies.

But at the same time, CoreWeave only used about 25 quarters,

It realized the $2.6 billion in revenue that AWS only reached in the 40th quarter after launch.

Once again: these businesses are growing really fast.

With such a high growth rate and riding the AI industry's ride, investors should be quite excited.

This is true to some extent, but the reality is a bit more complicated.

Although these companies generally performed well in the latest round of financial reports, over the past year, CoreWeave's stock price has dropped by about 16% cumulatively;

Only Nebius is closer to its previous high.

So while the story is still good overall, it's clearly less appealing to the biggest new cloud companies.

Recent market performance has been relatively lackluster, in part because many growth expectations may have been reflected in valuations.

For capital-intensive enterprises such as Xinyun, the market sales ratio is not the most appropriate valuation indicator, but it is still very intuitive to use to explain the problem.

The smaller, faster growing Nebius and Applied Digital have valuation premiums far higher than the much larger CoreWeave.

CoreWeave's revenue is still doubling, but it's no longer as fast as the 400% to 450% growth rate of the top companies.

If there is any real problem with the new cloud company, it is not growth, but long-term profitability.

New cloud companies need to continuously invest in chips, electricity, and physical infrastructure to scale up, and these costs are not low:

In CoreWeave, for example, its revenue growth was impressive, but its capital expenditure was even more impressive.

Other significant costs include chip depreciation — depreciation amounts to more than half of revenue

— and interest expenses that are still rising due to borrowing to build expensive infrastructure ahead of time.

This article is not intended to judge whether Xinyun will eventually succeed or whether its current stock price is reasonable.

Apart from the popularity of the topic itself, what I'm really trying to explain here is that Xinyun just happens to be a microcosm of the long and short tension in the entire AI deal.

On the one hand, they are on a scale far exceeding everyone's previous expectations,

Furthermore, the vertical market — the computing power market — continues to expand and has created a growth rate rare in history;

On the other hand, the cost of building such enterprises is also at a historically high level, requiring investment in huge fixed infrastructure that continues to depreciate.

A return to horizontal SaaS?

Here's a brief update on the changing market landscape of the “end of SaaS.”

The company that was the hardest hit by the previous software stock sell-off has performed quite well in the past month.

Over the past 30 trading days, horizontal software companies have had the highest performance among IGV Software ETF constituents — although since the date this set of data was collected,

They have taken back some of the gains.

Overall, the fundamentals of these companies remain solid.

Atlassian, in particular, has not declined as the market had previously anticipated under the impact of AI.

The productivity software company “both exceeded expectations” in terms of performance and guidance: cloud business revenue increased 31% year over year.

Revenue backlog orders are growing even faster. But perhaps the more critical sign is that AI is becoming a booster rather than a deterrent to business growth.

Atlassian says its AI assistant, Rovo, has been widely adopted; at the same time, customers using Rovo are spending nearly twice as fast as 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 is still slightly lower than other software categories.

With few exceptions, the expected market-sales ratio of horizontal SaaS companies, including Atlassian, is generally below the level corresponding to the “growth rate—valuation multiplier” trend line.

Once again, horizontal SaaS has only had a relatively good “month.”

If you want the market to believe that the “end of SaaS” has been cancelled, a month's performance alone is far from enough.

Of course, if your software business is in the field of cybersecurity or observability, that's a different story — for these companies,

The so-called “end of SaaS” never happened.

The cybersecurity sector continues to significantly outperform other categories in IGV software ETFs.

In this field, AI has instead become a tailwind factor: the market generally believes that AI has improved people's risk perception of cybersecurity threats.

And no enterprise customer would rely on “vibe coding” (vibe coding) to temporarily piece together their own security solution.

Whether this logic holds true will of course take time to test. However, at least for now, the situation of traditional software companies is far from uniform.

Investors are paying close attention to whether AI will bring benefits to every company or cause erosion.

And as each batch of new data continues to revise the original judgment — and rightly so.

Towards the cutting edge of efficiency in token investments

The market landscape around model usage, token consumption, and token spending management continues to evolve in a variety of interesting ways.

Take Databricks, for example.

On the questions of “which model should we use” and “which model is the best”

Databricks didn't take a winner-take-all approach, nor did they just give engineers a budget to decide how to spend it.

It asks 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,

However, it has developed a “smart router” (Smart Router), and the actual results are quite satisfactory.

Allegedly, Databricks routers can call more capable and more expensive models when necessary.

Use less capable and less expensive models when conditions permit, thereby “continuously reducing average task costs by more than 30%.”

Overall, it's hard not a good thing to go for the “efficiency frontier” of token spending.

This means that demand continues to grow, and application scenarios not only continue to evolve at the cutting edge of performance, but also spread to less capable models.

In the initial pessimistic narrative, these suboptimal models were thought to be quickly eliminated.

As we mentioned before, efficiency improvements will expand demand coverage; this is exactly what the market wants to see, similar to the Jevans paradox.

Silicon Data's Token Price Strength Index shows that the overall price intensity is declining.

In particular, as the cheaper open model takes up a higher share of the continuously expanding market.

Here it is necessary to once again clarify a concept that is often misunderstood: these indices measure the cost intensity of token spending,

Not an absolute dollar amount. It also depends on the number of tokens consumed and the total cost of tokens.

This means that even if the price of each token falls, the total token consumption and total expenditure amount are likely to continue to rise.

What is really important is that overall demand is still growing, while pricing and model choices are gradually moving to the cutting edge of efficiency.

It will only further drive this growth. Notably, “AI demand” or “AI adoption” is not a single, homogenous concept.

There is still a huge gap between heavy users and others.

This clearly shows that “always use the best model” may work for some businesses, but it certainly isn't for everyone.

Today, the market is rapidly developing more alternatives to choose from. Overall, this is a good thing.

According to data from enterprise expense management platform Ramp, all businesses are increasing AI spending,

But there is a huge gap between median spending and the top 10% of companies, and between the top 10% and the top 1%.

Ramp's data is generally more biased towards tech companies, and this needs to be taken into account when interpreting.

However, the data shows that the top 10% of companies spend about 50 times more on AI per capita than the median company.

It is likely that this distribution did not occur by chance. Businesses that can unlock more value from AI spending,

It's probably the company that invests the most — if not every company, at least a few companies fit this pattern.

Boston Consulting's analysis of 107 listed companies found that

Companies in the top two quintiles of token usage are growing significantly faster in revenue than other companies.

The core point I want to express here is that token demand and usage efficiency are mutually reinforcing: the more value an enterprise obtains, the more tokens it consumes.

Of course, there are repeated trade-offs between investment and return in this process, and R&D always includes some upfront costs.

However, for the vast majority of companies, uncontrolled “stacking tokens” has never been an effective strategy.

As a result, companies will increasingly need to do this in the future, which is obviously a good thing.

The battle for talent in cutting-edge laboratories

New Media recently had two amazing team members join OpenAI, so finally we'll use a few interesting charts,

Check out the talent recruitment for cutting-edge AI labs.

Dario Amodei recently stated that he is concerned that employees are putting money above their mission.

Judging by Levels.fyi's data, this concern is probably not unfounded.

If you understand it exactly according to the surface of the data, Anthropic offers engineers a very high salary.

Moreover, it is far superior to engineers with comparable qualifications from companies such as Google and Tesla.

It seems like being a member of the technical team is a great thing indeed.Also, there is an interesting set of data.

According to Live Data Technologies

——Compiled and presented by Truist Securities——The talent sources of various laboratories already overlap a lot,

There are also significant differences:

Both companies have recruited quite a few talents from technology companies with large market capitalization.

However, only OpenAI recruited people from Nvidia and Tesla, and they all happened in 2026.

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

Both labs have also hired significant numbers of employees from Salesforce and Stripe.

But the overlap between the two sides seems to have stopped there. Anthropic has recruited a lot of talent from SaaS companies, while OpenAI has almost none;

OpenAI, on the other hand, recruits a large number of people from the consumer internet, platform markets, and advertising technology companies.

Anthropic has relatively few hires in these areas, with the exception of Airbnb, Netflix, and Uber.

As to what these differences mean, I'll leave it up to you to interpret for yourself.

(This article does not constitute investment advice)

Twitter:https://twitter.com/BitpushNewsCN

Compare the TG exchange group:https://t.me/BitPushCommunity

Compare TG subscriptions:https://t.me/bitpush