Meta sells computing power, Palantir loudly lashes out in public, and Zhipu becomes a top Silicon Valley drawcard—the story of AI Capex needs a new way to be told.

Written by: Beatz

AI market corrected sharply again, this time because Meta said it might sell its excess AI computing power.

If this news had come out three years ago, probably no one would have found it strange. Cloud computing has always been a business of slicing up servers and selling them to others. Amazon, Microsoft, and Google have done it for years. New cloud providers like CoreWeave and Nebius are also on this path, turning NVIDIA chips into collateral for financing, and then turning that financing into more chips.

But when it comes to Meta, things take a different turn.

Meta hasn't understood computing power this way in the past. It buys chips, builds data centers, secures electricity and land for its own models, for its ad system, for its recommendation feeds, and for the superintelligence that Zuckerberg talks about getting closer. It's not a cloud provider. It didn't originally make money by renting out machines to others.

A company that used to say, "I need as many machines as possible because the future will consume them," now says, "If these machines are temporarily unused, we can sell them to others."

This doesn't directly prove an overcapacity in computing power, but it also can't be brushed aside lightly.

On the day of the market crash, Palantir CEO Alex Karp vented for nearly twenty minutes on a CNBC interview.

He was originally there to discuss the new partnership between Palantir and NVIDIA, but quickly pivoted to the token-based pricing models of OpenAI and Anthropic. He said CEOs privately complain that enterprise AI adoption is "paying for tokens that create no value, while handing over their own data." He even called the increasingly expensive model bills a "wealth tax" weighing down companies.

For the past two years, the discussion was about who dared to spend, who spent fast, and who could stack data centers first. Now the question is slowly changing. After the machines are bought, who can keep them running at full capacity all the time.

Meta's statement hasn't yet materialized into an official business. In public reports, Meta has an internal direction called Meta Compute, which might sell raw computing power, or, like Amazon Bedrock, place different models on its own infrastructure and sell them to developers. Zuckerberg previously mentioned at a shareholder meeting that external companies ask almost every week if they can buy Meta's API services or some compute, and they are willing to pay above Meta's cost.

He also added a remark at the time. They haven't done it yet because Meta feels it can still use that computing power itself.

If it can use it, renting it out is a choice. If it can't, renting it out is a painkiller for the balance sheet.

The hardest part to judge is right here. Meta might simply be creating a window in its construction cadence, selling off temporarily idle resources. Or it might be telling investors that the hundred-billion-dollar AI spending can't be sustained solely by a distant superintelligence; it needs to find a closer revenue line.

Both interpretations are plausible.

Demand hasn't disappeared, it's just starting to pick and choose

Capex is the core of the AI narrative, bar none. Like the liquidity flood of 2021, the expectation of capex is continuous growth, the flood keeps flowing, and all branches of the market go up together. As soon as Meta signaled it would sell computing power, many people's first reaction was that AI capex was about to collapse. Big companies finally admitted they bought too much; the semiconductor feast was over.

That's too simplistic.

Public data doesn't support such a clear-cut conclusion yet. AWS revenue grew 28% in Q1, reaching $37.6 billion, a rare fast growth in recent years. Google Cloud grew even faster in Q1, with revenue reaching $20 billion. Microsoft Azure is still growing at about 40%.

Amazon is still saying its capital expenditure this year could reach $200 billion. Alphabet raised its 2026 capital expenditure guidance to $180-190 billion. Meta itself raised its full-year capex to $125-145 billion.

These numbers don't look like a demand collapse.

They look more like a redistribution.

Cloud providers are in a different position than model providers. Cloud providers sell the road. As long as there's traffic on the road, no matter who built the car, they can collect tolls. OpenAI, Anthropic, enterprise customers, government clients, startups — they all end up on some data center, some chip, some network, and some electricity contract.

So the big three clouds can continue to be strong.

AWS even raised the price of an AI cloud service at the end of June, a service that lets customers pre-reserve GPUs. AWS increased the price of this service by about 20% starting in July. It had already raised it by about 15% in January. This is not the kind of move you see when demand is weak.

When supply is scarce, sellers raise prices.

But model companies may not all be so comfortable.

Model companies have more picky assets. Computing power doesn't generate revenue just sitting there. It needs to be constantly filled by smarter models, higher-frequency users, and more expensive enterprise workflows. Only when the model is good enough will users tolerate queues, limits, price hikes, and increasingly complex subscription tiers.

That's why Anthropic is seen as a different kind of company. Not because it's cheap, but because users are willing to entrust it with expensive tasks. Writing code, modifying systems, running long tasks, connecting to enterprise workflows — once these tasks truly enter the production environment, they consume far more tokens than casual chatting.

The problem for strong models is not enough machines.

The problem for weak models is no one cares about the machines.

Both problems are about computing power, but they are not the same thing.

xAI's line has a similar flavor. Grok hasn't formed a clear enterprise mind share like the strongest models, yet some computing power from Musk's system can flow to Anthropic. This move is more sobering than any slogan. Machines don't care about founders; they only care about who can keep them fully utilized.

The relationship between Google and Meta also shows things aren't simple. In June, there was news that Google restricted Meta's use of Gemini because the computing power Meta wanted to buy exceeded what Google could provide, even affecting some of Meta's internal AI projects. A company is considering selling computing power on one hand, while on the other, it can't buy enough top-tier model capabilities for some tasks.

This is not overcapacity in the traditional sense.

It's a mismatch. Because the bills are starting to get glaring.

Cloud providers can continue to raise prices because they sell certainty. Customers need GPUs they can secure for a period, a stable data center, and infrastructure that won't go down in the middle of the night.

But after enterprises get the computing power, the problem isn't over.

They still have to hand that bill to the CFO. The CFO won't ask how many tokens you used; he'll ask how much money those tokens saved the company, how much extra revenue they generated, and how many mistakes they reduced.

For enterprises, tokens become an electricity meter

That brings us back to Karp's interview at the beginning.

He described what many AI companies sell to enterprises as overselling. The day before the show, Palantir posted a nine-point statement on X about so-called AI sovereignty, specifically calling out the "tokenmaxxing" model. This term is hard to translate directly — it sounds harsh — but the idea is simple: treating token consumption as progress, burning money as usage, and the bill as productivity.

Karp put frontier labs like OpenAI and Anthropic on the table. His point isn't that enterprises shouldn't use the strongest models; it's that they shouldn't hand over their data, processes, and business judgment, and then pay an ever-growing bill based on consumption.

Palantir wants to sell something different. Not a universal chat box, not a single API, but putting data, approvals, permissions, operational rules, and AI into the same business system. What customers pay for isn't "how many times AI was used," but whether a production line, a risk control process, or a government task was actually transformed.

The people who really control the money in enterprises are waking up.

UBS recently spoke with enterprise IT executives, and one direction is clear. Many enterprises aren't stopping AI use; they're putting brakes on AI spending. About 60% of surveyed companies are cutting token costs and adding usage guardrails, especially those that have moved past the pilot phase and started integrating AI into daily processes.

This is also an interesting reversal.

Once AI goes from a toy to a tool, spending becomes harder. In the toy stage, bosses are willing to give budgets because everyone fears missing out. In the tool stage, the CFO asks who it saves labor hours for, who it helps sell more goods, and who it reduces risk for.

On this spreadsheet, tokens don't look like revenue.

They look more like an electricity meter.

You could say a fast-spinning meter means the factory is running. Or you could say the meter spins too fast while output hasn't increased, meaning the machine is problematic.

AI agents amplify this problem. A Codex study from OpenAI and several universities has some alarming data. In the first half of 2026, active Codex users grew more than fivefold; output tokens for some roles inside OpenAI skyrocketed — legal roles saw median monthly output tokens 13 times higher than November 2025, and research roles saw over 50 times higher.

Another study puts it more bluntly. Agentic coding tasks can consume up to 1000 times more tokens than regular code chat and code reasoning. Token consumption for the same task can vary by 30 times across different runs.

That's the real foundation of today's computing power shortage.

Not that people are asking a few more questions from chatbots.

It's that software is starting to become a group of small workers that repeatedly read files, run commands, modify code, fail, retry, fail again, and retry again. They don't have lunch breaks, but every step consumes tokens.

When tokens become the electricity meter, whoever owns the power plant has power. But whoever wastes electricity will also be the first to be questioned.

When the bill gets thicker, cheaper models find their place

Once the CFO starts looking at this electricity meter, the next step is almost instinctive.

He'll ask: which tasks must use the strongest model, and which tasks only need a good-enough model?

That's when open-source models like GLM, Kimi, DeepSeek, and Qwen stop being just tech news. They become bargaining tools on enterprise procurement tables.

Even Marc Andreessen of top venture firm a16z said many AI practitioners now see Zhipu GLM-5.2 as among the first Chinese models that can match or even surpass leading US open models on most tasks. That judgment may not be the final verdict, but it gives enterprises one more option.

Coinbase provides a harder example. Brian Armstrong said the company switched its default AI model to open-source models like GLM 5.2 and Kimi 2.7, combined with model routing, caching, and context streamlining. Token usage is still growing exponentially, but AI spending has been cut by nearly half.

The destructive power of this statement is that enterprises can now break down model capabilities into separate purchases.

The hardest tasks still go to the most expensive models. Simple summaries, customer service, information extraction, templated code, internal knowledge base Q&A — those go to cheaper models and local deployments.

Open-source models don't have to win every battlefield.

They just need to convince procurement departments that not every kilowatt-hour has to be paid at a mansion's electricity rate.

At this point, Meta selling computing power is no longer an isolated news item.

It's the same story as Palantir ranting about tokens and Coinbase splitting open-source models: the AI spending chain is starting to be broken apart. The upstream sells certainty, the midstream sells results, and the downstream compresses unit prices. Every layer is still growing, but every layer is being asked: was the money worth it?

The hardest part isn't buying machines; it's keeping them busy

For the past two years, the easiest story to tell in AI has been scarcity.

Not enough GPUs, not enough electricity, not enough data centers, not enough engineers, not enough clouds to run models. This story was too smooth. When things are scarce, everyone instinctively rushes forward. Claim the spot first, sign the power contract, buy the chips, get the machines up.

During a resource grab, people don't count pennies.

Because the cost of being a step behind seems greater.

But Meta's news pushes another issue to the forefront. After you buy the machines, just because they're expensive doesn't automatically make them a good business. They need work every day, customers willing to pay, models to keep them fully utilized, and applications to turn costs into revenue.

That's utilization.

Utilization sounds cold, but it's actually brutal. It doesn't ask about your future; it asks whether your machine is running today. It doesn't care what you said at the press conference or whether you bought the most expensive GPUs. It only looks at one thing: did this money turn into continuous cash flow?

Cloud providers answer this question relatively easily. They sell infrastructure by nature. AWS, Google Cloud, and Azure sell roads, electricity, and server rooms. Customers who want to train models, run inference, or host applications all end up on some cloud.

So they can still be strong.

Strong model companies also have their answer. If the model is strong enough, users are willing to queue, enterprises are willing to integrate, and developers are willing to rewire their workflows around it. Then computing power isn't inventory; it's a bottleneck. The more machines, the more they can scale.

The hardest is the middle layer.

They have machines, stories, model teams, and big budgets. But the model hasn't reached the front, the product hasn't become a daily habit, and developers aren't willing to change their workflows for it. For this type of company, computing power goes from weapon to inventory after just one failed model release or one user migration.

Inventory isn't necessarily useless.

But inventory has to be discounted, rented out, or find new uses.

That's what makes Meta's selling of computing power so glaring. It doesn't prove Meta's failure, nor does it prove that AI demand is gone. It just lets the market see for the first time that AI infrastructure can face the same problems as ordinary factories.

The factory is built. Where are the orders?

Computing power hasn't disappeared; it's just starting to become layered

So the best way to understand this is not "overcapacity."

That term is too crude.

A more accurate description is that computing power is becoming layered.

The top layer is still tight. The strongest models, the best clouds, the most stable GPU clusters are still fought over. AWS can raise prices because certainty itself has a price. Customers aren't just buying GPUs; they're buying a guarantee that a certain batch of machines will be available at a certain day and hour.

The middle layer is starting to get awkward. It might not be bad, but it's not scarce enough. It can run models, do inference, and sell to external customers. But customers will compare, negotiate, and ask why not use a cheaper model, why not use someone else's cloud, and why this batch of machines is worth its price.

The bottom layer will be gradually squeezed by open-source models and cost optimization. Enterprises won't always call the most expensive model for ordinary tasks. They will do routing, caching, compress context, and tier models.

Demand has grown up.

Children don't look at bills; adults do. AI will go through this process as it enters enterprises. In the pilot phase, everyone fears missing out; in the scaling phase, everyone starts counting.

After counting, the industry chain will no longer be as uniform as in the early days.

Some will continue to raise prices because they sell irreplaceable certainty. Some will switch to selling results because customers don't want to pay for consumption itself. Some will be forced to discount because good-enough alternatives appear. Some will rent out machines because leaving them idle looks worse than renting them cheap.

When these things happen simultaneously, the industry will look contradictory.

On one hand, computing power is scarce.

On the other hand, computing power is being rented out.

On one hand, token consumption is exploding.

On the other hand, enterprises are cutting AI spending.

On one hand, top models are getting stronger.

On the other hand, open-source models are getting cheaper.

They don't conflict. They simply indicate that AI has moved from a total volume story to a structural story.

The old railroad story will be told again

In the 19th century railroad bubble, railroads weren't fake.

The tracks were laid, goods really moved, cities really grew, and time was really shortened. Many valuable commercial networks did grow along those tracks.

But that didn't stop many railroad builders from losing money back then.

They lost not on direction. They lost on building too early, too much, in places without freight or passenger traffic, or borrowing too expensive money to build a line that would take too long to recoup.

The fiber optic bubble during the internet era is the same. Fiber optics weren't wrong. The whole world was later lifted by them. What was wrong was the set of ledgers that crammed decades of future demand into a few years of capital expenditure.

AI data centers will probably leave behind many useful things. GPUs will depreciate, power contracts will be renewed, data centers will upgrade equipment, and software will become more efficient at consuming computing power. What looks like outrageous token consumption today might be as ordinary as HD video traffic in a few years.

But assets have their own temperament.

They don't care if you believe in the future. They only care if someone comes to use them every day.

The signal from Meta selling computing power is stuck right at this point.

It's not the end of AI. It's not the end of semiconductors. It's more like when the capex narrative reaches its middle chapter, someone opens the door for the first time and lets outsiders see how many machines are in the warehouse.

Some machines will be consumed by top models.

Some machines will be rented by cloud customers.

Some machines will become cheaper in price wars.

And some machines will quietly wait for an application that hasn't arrived yet.

For the past two years, the market was willing to believe that all machines would eventually find their destiny. Now it starts asking: who will find it first, who won't, and who will find it but still won't make enough money?

Once this question comes out, the AI story changes.

It no longer belongs only to those who buy machines the fastest.

It belongs to those who can keep the machines running.

Original link

Click to learn about open positions at BlockBeats

Welcome to join the BlockBeats official community:

Telegram subscription group: https://t.me/theblockbeats

Telegram discussion group: https://t.me/BlockBeats_App

Twitter official account: https://twitter.com/BlockBeatsAsia