Slows down Dario Amodeis Is the call for a slower pace in AI now hampering the industry’s growth? The Anthropic CEO called for this over the weekend in his essay “We Must Pace the Frontier”—to deliberately slow the advancement of AI model capabilities— OpenAI CEO Sam Altman and xAI CEO Elon Musk publicly agreed within hours.

Amodei’s core argument: Recursive self-improvement and security incidents such as the OpenAI Hugging Face incident showed that the models’ capabilities are growing faster than the ability to control them. His three-step plan begins with independent auditors working directly within the companies. By this, he means not
halt to training, but rather more time for safety work to proceed in parallel with progress.

Why many see this as a brake on growth

Within the community, the call is being interpreted as a warning sign regarding the AI industry’s high growth rates. A slower pace in the release of new models could dampen demand for fresh computing power and thus weigh on the valuations of AI companies and Neoclouds. Coinciding with the announcement, market prices for Anthropic and OpenAI shares fell.

Why this misses the actual driver of growth

Amodei’s appeal targets the training of new models—and it is precisely this part of the AI value chain that is now the smaller and shrinking segment. According to Deloitte, by 2026, around two-thirds of global AI computing power will be directed toward inference (the ongoing operation of already trained models).

And it is precisely this demand for inference that is growing faster than ever before:

• $GOOGL (+2,04%)-Cloud order backlog: from $240 billion to $460 billion within one quarter
• Tokens processed by Google: 16 billion per minute, an increase of 60% compared to the previous quarter
• $MSFT (+0,7%): AI order backlog increased more than twelvefold
• GPU rental prices have grown from ~10 million/MW in early 2026 to as much as ~40–50 million/MW today

So the real bottleneck remains computing power itself, not the political will to slow things down. Microsoft and Google each have investment budgets of around $180–190 billion just to meet demand at all. Training pacing is unlikely to make much of a difference to this scarcity as long as current models are already so powerful for users that demand for their use continues to grow.

Overall , the appeal seems more like a shift in priorities—away from a pure race for capabilities and toward greater safety and control. In my opinion, it is wrong to believe that this will act as a brake on the entire AI value chain.

The largest and fastest-growing part of it—inference—remains virtually untouched by the debate.

This announcement could also come at a strategically opportune time: Training new frontier models costs a great deal of money (tens of billions per model). A deliberately slowed pace would alleviate precisely this cost burden and improve the profitability of Anthropic and OpenAI, just before both are set to make potential IPOs . This safety call—which, incidentally, protects their own balance sheets—comes at just the right time for the model providers.

What are your thoughts on this?

$IREN (+0,64%)$NBIS (-1,76%)$NVDA (+0,1%)$AMZN (+1,99%)$AMD (+2,37%)$WULF (+3,73%)$ORCL (-7,35%)$CSNDX (+0,98%)$AVGO (+0,52%)$SPCX (+1,95%)$PLTR (+0,75%)$META (+0,58%)$CRWV (-0,95%)$DELL (+11,62%)$BE (+6,74%)$MU (-0,14%)

Sources:

https://darioamodei.com/post/we-must-pace-the-frontier

https://www.deloitte.com/global/en/about/press-room/2026-tmt-predictions.html

OpenAI announced on September 3, released GPT-6 Astra , which shatters all previous AI benchmarks and can generate a playable first-person shooter map in 28 minutes.

Key metrics from OpenAI’s own release:

• FrontierMath Tier 4: 97.6%
• ARC-AGI-3: 99.9%
• OSWorld 2.0 (computer usage): 72.6% with about 47% less time per task than its predecessor, GPT-5.6 Sol, at 65.7%
• ExploitBench: 100% compared to 78.5% for Sol (previous model)
• Artificial Analysis Intelligence Index: 61.2 compared to 60.9 for Sol and 65.7 for Claude Fable 5.1

However, there’s a catch: Astra isn’t designed to be a better chatbot, but rather as computer operator:

It fills out online forms, updates customer records in a CRM system, organizes calendars, conducts research in the browser, creates documents, spreadsheets, and presentations based on templates, and tests finished websites for errors.

Here’s what users have already built with it in the first week:

Astra no longer explains programs like Blender or Unreal Engine—it operates them itself :

• Game map in 28 minutes: Riley Brown had Astra build a first-person shooter map via Codex with full computer access. The log shows 28 minutes and 16 seconds , 20 edited files and 80 automated tests passed —including weapon and audio timing as well as explosion effects.
• A playable world requiring no prior knowledge: Matt Shumer, neither a 3D artist nor a game developer, had a survival world generated in Unreal Engine and set up as a playable character.

The point here isn’t the graphics, but the duration: Tasks spanning hours and days that the model pursues on its own have been the biggest challenge so far.

Astra is also the first model that OpenAI has classified as “Critical” for cybersecurity:

It can find and exploit unknown security vulnerabilities without a human guiding every step. The released version therefore refuses to perform offensive tasks such as building exploits; this capability is only unlocked through the Daybreak defender program. Sam Altman spoke of a “new capability level” for the model.

The leap forward, then, lies in tasks that run for hours and require the management of multiple programs. In terms of pure cognitive ability, the gap is small—in the independent Artificial Analysis Index, Astra is practically on par with its predecessor and trails Claude Fable 5.1.

Those who just chat will hardly notice a difference. Those who delegate work certainly will.

What does this mean for the AI sector?

A model that runs for 28 minutes straight consumes many times more resources than a short chat session and drastically increases token usage. Agent-based work is thus primarily a source of demand for more computing power—and that ends up with $NVDA (+0,1%)the hyperscalers ($AMZN (+1,99%)$GOOGL (+2,04%)$MSFT (+0,7%)) and the neoclouds ($CRWV (-0,95%)$NBIS (-1,76%)$IREN (+0,64%)), which lease out this capacity.

Sources:

https://openai.com/index/gpt-6-astra/

https://openai.com/index/safety-overview-gpt-6-astra/

https://www.cnbc.com/2026/09/03/open-ai-astra-gpt-6-cyber.html

https://x.com/rileybrown/status/2095679352927056230

https://x.com/mattshumer_/status/2095596175705399482

$NBIS (-1,76%) CEO Arkady Volozh and CRO Marc Boroditsky just spoke at the Goldman Sachs Technology Conference.

Here are the key takeaways:

1) Demand continues to accelerate

Nebius made it clear that demand for AI infrastructure continues to far outstrip supply. Importantly, the visibility of future demand continues to extend.

A few months ago, management was still talking about visibility of about 18 months with demand remaining strong. Now Arkady says that this period has extended to 24 months or possibly even longer. Customers are already inquiring about capacity for Q1 and Q2 2028.

He also said that customers are currently requesting tens of thousands of Vera Rubin GPUs.

Marc emphasized this point, noting that demand is not only exceeding supply but is growing faster than supply.

2) Explanation of the Partnership Between Nebius and Palantir

The underlying thesis revolves around enterprise data.

Arkady believes that companies increasingly want to run open-weight models on their own proprietary data. These models are intended to be continuously improved within their respective enterprise domains. At the same time, companies want to retain control over the resulting data and the intelligence gained from it, rather than constantly feeding it back into external commercial models.

Palantir already has the enterprise software layer and the necessary tools to orchestrate such processes. What Palantir lacked was the underlying infrastructure.

This is exactly where Nebius comes in.

“We provide the entire stack, right down to our Token Factory. They take this stack, add their own tools and their customer base. For us, this is an excellent way to gain access to enterprises.”

This could give Nebius access to an enterprise customer base that would have taken considerably longer to build organically.

3) Enterprise demand appears to be reaching a tipping point

The Palantir deal is part of a larger trend.

Marc said Nebius has seen a “significant increase” in its pipeline with platform companies that themselves serve enterprise customers.

Some of these companies view Nebius as their first provider outside of the major hyperscalers. They are not just looking for training capacity, but for a partner that can support training, post-training, and inference across the entire AI lifecycle.

The evolution of customer segments is becoming increasingly clear:

AI Natives → Larger AI Natives → Digital Natives → Enterprise Platforms → Enterprises Themselves

Management continues to believe that enterprise customers will represent the largest portion of the market opportunity.

4) xAI and $META (+0,58%) selling excess capacity

Goldman asked what would happen if companies like xAI and Meta built enormous amounts of infrastructure for their own use and later offered unused capacity on the market.

Arkady explained that this would largely consist of bare-metal capacity rather than a full AI cloud.

Someone would still need to take on this infrastructure, add the cloud and software layers, and then market it to end customers.

Nebius could potentially be one of those companies.

“If they could bring more capacity to the market, that would be good for the industry and good for us. We could take it on, repackage it, and resell it to end customers.”

5) The four major customers are already working on their next clusters

The four major customer wins announced last quarter were not one-time capacity deals.

Marc said that these customers already have additional needs and are discussing their next deployments with Nebius.

In some cases, they are asking for additional GB200 capacity, while all four customers are already discussing requirements for the Vera Rubin platform.

Furthermore, Nebius has other similarly large opportunities in the pipeline.

Important: Marc reiterated that these contracts were not won based on price, but rather on the reliability and performance that Nebius was able to demonstrate during the proof-of-concepts (POCs).

6) Older GPUs remain in demand for a specific reason

This is exactly the point I’ve been highlighting for quite some time.

Marc explained that older GPUs aren’t staying fully utilized simply because customers can’t get newer chips.

Many workloads—including RAG infrastructure, text prediction, and image generation—are actually more cost-effective on older hardware.

With these GPUs, customers already have a very clear understanding of the total cost of ownership (TCO), reliability, and performance.

“We actually have a list of new customers looking for chips from older generations as soon as they become available.”

This is exactly the dynamic I’ve been arguing:

The most powerful GPUs will increasingly be allocated to the most computationally intensive workloads, while older GPU generations remain economically viable when their performance is already more than sufficient for certain applications.

7) The Asset-Light Model Is Gaining Ground

According to Arkady, the asset-light strategy is designed to address two of Nebius’s most significant constraints:

How quickly Nebius can physically build out new capacity.

How much of this expansion Nebius can finance on its own.

The partners Nebius is in talks with are typically power utilities or data center companies that already have land, electricity, and, in some cases, access to affordable financing.

What they often lack is the expertise to move further up the AI infrastructure stack.

Nebius can contribute data center expertise, racks, the software stack, and customer demand to monetize this infrastructure.

Management says there is a long list of companies interested in this model.

Arkady describes each individual line in the internal pipeline as a major project in its own right.

Nebius is already working on several of these projects. The corresponding capacity is expected to come online starting in 2027.

In short: These statements paint a very positive picture for Nebius: strong and steadily rising AI demand, increasing enterprise activity, large follow-on orders, demand even for older GPUs, and an asset-light model that could accelerate capacity expansion.

What does an infrastructure provider gain by becoming the preferred partner of $PLTR (+0,75%) ?

Access:
$NBIS (-1,76%) It gains access to an existing corporate customer base without having to sell its own products—and does so precisely in markets where control over data and models is already a purchasing criterion.

Palantir provides the trust layer; Nebius provides the computing power.

Of particular interest are the planned modular data centers at locations with existing power connections: Those who already have a grid connection can deliver capacity faster than any new construction project.

Overall, this opens up a new source of demand for Nebius that would have been significantly more expensive to tap into on its own. The volume and timeline of the deal remain to be determined.

What’s in it for Palantir?

In the future, Palantir will be able to offer its customers sovereign AI, including computing power, from a single source, without having to build data centers itself or tie up capital for that purpose. In my view, the impact on revenue remains unclear—the company is still selling software, and without specific figures, it’s impossible to assess whether this will result in new customers or merely serve as an additional selling point.

What does it really take to outperform the World on an annual basis?

After a brief discussion—which wasn’t really a discussion at all—I came to the conclusion that four titles would be enough for me to hold my own against the World. Of course, that was just an excuse on my part to $RKLB (+1,49%) add a few more @Get_Rich_or_Die_Tryin

Well, I then decided to add 3 more titles.

$NBIS (-1,76%)
$PNG (-4,79%) And $ONDS (-0,24%)

Today, a little late, I invested 20k in the experiment.

I’m curious to see how the portfolio performs. I’ve also made it public on 212 in case anyone’s too lazy to enter the four stocks themselves.

BREAKING NEWS: Palantir $PLTR (+0,75%) appoints Nebius $NBIS (-1,76%) as its preferred Sovereign AI infrastructure partner

​👉 Key Highlights:

• ​Palantir and Nebius announce a strategic partnership for sovereign AI.
• ​Nebius will be Palantir’s preferred infrastructure partner for sovereign AI. Nebius’s computing and inference capabilities will be integrated into Palantir .
• ​Eligible customers will gain greater control over computing power, data , and models.
• ​Customers can deploy open-source models via the Nebius infrastructure.
• ​Models can be continuously updated with their own data .
• ​Integration combines the Nebius infrastructure with Palantir’s Sovereign AI Operating System (Sovereign AI OS).
• ​Companies will gain new computing capabilities for Palantir customers at an accelerated pace.
• ​The plans include modular data centers in locations where power supply is available.

​👉 Why this is important:

• ​Expands the offerings of Palantir in the field of sovereign AI for commercial enterprises.
• ​Could the additional demand for AI infrastructure and inference capabilities from Nebius .
• ​Gives companies more control over sensitive AI data and models.
• ​The partnership could facilitate the development of dedicated AI computing capacity .

$NVDA (+0,1%) Statements from the earnings call:

– First annual forecast in the company’s history: approximately 70% revenue growth in fiscal year 2028. Customer forecasts indicate a doubling —the 70% figure represents the maximum the supply chain can handle. The bottleneck is expected to persist at least until the end of fiscal year 2028.

– Revenue per gigawatt increases with each generation: $18 billion with Hopper, $25 billion with Grace Blackwell, $40 billion with Vera Rubin. Vera Rubin is expected to account for about 20% of data center revenue in the third quarter and to be the fastest product ramp-up in the company’s history.

– Half of the business comes from outside the hyperscalers: The ACIE segment generated $40 billion, up 138% compared to the previous year. Neocloud partners are expected to close the year with 8 GW of installed capacity, up from approximately 3 GW at the end of 2025.

– $AMZN (+1,99%) is expanding the partnership: an additional 2 million GPUs starting this quarter through the second quarter of fiscal year 2029. The five largest hyperscalers are expected to reach just under $800 billion in 2026, and around $1.3 trillion.

– Memory prices are squeezing margins: from 75% to 74% in the third quarter; the low point is expected to be 71–72% in the fourth quarter. Kress (CFO) describes the price trend as extreme and well above the company’s own expectations—citing the AI expansion itself as the cause.

– Kress (CFO) on circle financing: just under $50 billion has been invested in Frontier Labs, in addition to financing platforms involving six investors totaling more than $500 billion in third-party capital. “We do not issue loans.” These labs are expected to generate account for about a quarter of total revenue.

– China is left out: Less than 1% of data center revenue came from China in the second quarter; the outlook projects zero revenue from China.

$NBIS (-1,76%)$CRWV (-0,95%)$IREN (+0,64%)

$NVDA (+0,1%)

The server systems of Spitzenchips are affected Vera Rubin and Grace Blackwell, which will be delivered . The amount varies depending on the chip generation and memory configuration. The reason cited is the extremely high prices for RAM ($MU (-0,14%)$SKHY (+0,62%)$005930). As a result, part of the value creation in the AI stack is shifting downward—away from the computing chip and toward memory. Just how much of this NVIDIA itself is absorbing is likely to be reflected in the gross margin in the quarterly report on August 26 .

Overall, , this affects hyperscalers like Neoclouds and AI labs, which consequently face even higher costs —because Nvidia chips represent the largest cost factor in AI data centers.

Consequently, GPU rental prices will also continue to rise, which will have a neutral effect for these companies.

The extremely expensive financing, which continues to be weighed down by macroeconomic uncertainty. This price increase further adds fuel to the fire. 🔥

$GOOGL (+2,04%)$AMD (+2,37%)$INTC (+2,89%)$MSFT (+0,7%)$META (+0,58%)$IREN (+0,64%)$CRWV (-0,95%)$NBIS (-1,76%)$AVGO (+0,52%)

Source:

www.bloomberg.com/news/articles/2026-08-22/nvidia-customers-notified-about-ai-related-price-hikes-above-15