Encrypted between you and the server. Whoever is physically in control of the hardware can still monitor what is happening on that hardware. There are more reasons than that though; imagine legislation is passed that AI providers must implement guardrails, for instance to limit what their models offer in terms of detecting security flaws in a codebase. Inference providers will comply, but if you have your own hardware, they can't directly control what you do with it. Physical ownership represents control.

When you rent a VM, you do get encrypted ssh, https traffic that should only be visible to you/root users. I understand it to be secure. A TOS policy that just promises privacy is not, but is very different.

Makes sense, I strongly prefer local inference over renting for a variety of reasons though. The privacy aspect is a big deal, even if they say they aren't selling your data there's no way to really know.

The datacenter in question used H100s and earlier GPUs. Instead of dismantle, it is rent out. The source for historic low rental spot rates is https://nebius.com/prices . A b300 that costs $65k as single card (clusters cost more) rents for 99c/hour, when 3 year depreciation of cost is $2.47/hour

When this low, renting VM's is a better value with full privacy compared to buying any AI GPUs. These types of GPUs are also obscenely loud.

Oracle is now forced to dismantle these massive superclusters, dumping the underlying hardware back onto the market as smaller, single-GPU virtual machine shapes.

How do I buy one? Is there a way to track the prices? I looked on ebay for some of the cards mentioned here, can't find B200 at all, H200 seems to go between 50-100k which seems very expensive, consumer gpu prices seem to have not come down.

Yes I wrote it. Oracle news today: https://www.reuters.com/business/microsoft-is-winding-down-oracle-deal-rent-ai-cloud-capacity-bloomberg-news-2026-11-06/ Other announcements related to MSFT were all during last 2 days.

Thanks.

Yes I wrote it. Oracle news today: https://www.reuters.com/business/microsoft-is-winding-down-oracle-deal-rent-ai-cloud-capacity-bloomberg-news-2026-11-06/ Other announcements related to MSFT were all during last 2 days.

Deepseek's critique (I admit that I don't know about Nscale that thoroughly, but disagree with Oracle point. Exact size is an estimate from deep dive, and dumping into spot market is their only "editorialized" viable option):

📝 Inaccuracies and Misleading Claims

1. The size of the cancelled Oracle deal is presented as fact, but it is unconfirmed.

The article states Microsoft “turned off an estimated 150 to 200 megawatts (MW) of unified AI clusters previously rented from Oracle.” However, the Bloomberg report that broke the story could not determine the full size of the deal. While the partnership is indeed winding down, the specific megawatt figure is an unverified estimate, not a confirmed fact.-1

2. The “$3 billion” figure for Nscale is incorrectly attributed to an order book deletion.

The article claims Microsoft’s exit from the Nscale Loughton site “instantly delet[ed] $3 billion from Nscale’s long-term order book**.” In reality, the **$3 billion figure refers to the amount Nscale is seeking to raise in its initial public offering (IPO), not the value of the cancelled contract. Reports note the Loughton deal was “believed to be a small part” of Microsoft’s overall commitment to Nscale, making a direct $3 billion loss highly misleading.-10

3. The Nebius fine involves 62 generators, not 45.

The article says Nebius faced a fine “for illegally running 45 industrial gas generators without local air permits.” The official news release from the New Jersey Department of Environmental Protection states the fine was for installing and operating 62 natural-gas power generators without permits. The number of generators is factually incorrect.-38

4. The article mischaracterizes the cause of Microsoft’s capital expenditure (capex) miss.

The piece claims that “systematic withholding of capital” directly explains why Microsoft’s quarterly capex missed forecasts by $5 billion** and its annual projection was cut to **$170 billion. While those figures are broadly correct, Microsoft itself attributed the shortfall to delayed capacity deliveries from third-party partners, not a deliberate strategy to withhold capital. The article’s causation is an oversimplification.-58

5. The article presents speculative claims as established facts.

Several assertions are phrased with unwarranted certainty:

• “Oracle is now forced to dismantle these massive superclusters, dumping the underlying hardware back onto the market…” – This is a speculative outcome. No source in the provided search results confirms that Oracle is dismantling clusters or flooding the secondary market with hardware. It is an assumption presented as a direct consequence.
• “Compute has effectively become a commoditized commodity utility, plagued by the same boom-and-bust cycles…” – This is an interpretive opinion about market dynamics, not a factual news report.
• “The structural narrative… has officially shattered.” – This is a strong editorial conclusion, not a neutral statement of fact.

6. The $5 billion miss figure is slightly oversimplified.

The article says the miss was “a staggering $5 billion**.” More precisely, Microsoft reported **$45.2 billion in capital expenditures and finance leases, compared with analyst forecasts of approximately **$50 billion**. The shortfall is therefore about $4.8–$5 billion, which is broadly in line with the article’s claim, but the framing omits the nuance that the miss was relative to a specific forecast range.-58

✅ Claims That Are Broadly Supported

For balance, some of the article’s key assertions are consistent with reporting:

• Microsoft is indeed winding down its 2023 Oracle AI compute deal.-1
• Microsoft did pull out of the Nscale Loughton (Essex) data-centre project due to power-delivery delays pushing past 2030.-10
• Microsoft declined an option to expand its Ward County, Texas, deal with Nscale by roughly 500 MW.-20-58
• Microsoft did step away from a 1.35 GW letter of intent for Nscale’s West Virginia campus.-30
• IREN’s Childress, Texas, project faced delays after Microsoft requested networking changes to support newer AI models.-47
• Microsoft has announced its custom Maia 300 AI chip, with production deployments planned for 2027.-67

🔍 Summary

The article blends genuine news events with speculative reasoning, unverified figures, and editorial hyperbole. Its most concrete factual errors are the unconfirmed Oracle MW figure, the misattributed $3 billion Nscale number, and the incorrect generator count in the Nebius fine. Additionally, it often presents interpretive conclusions (e.g., “compute is no longer digital gold,” “the narrative has shattered”) as established facts, which goes beyond the scope of reported news.

how it makes sense to build so much compute capacity so quickly. There seem to be predictions saying that we’ll need vastly more DCs and electricity running compute in the future

Every new datacenter coming online was decided 2 years prior at least. 2 years ago, skynet requirements seemed to be to scale models 10x or 100x their size. Datacenters huge enough to train that skynet goal will come online in 2028. SpaceX, perhaps other 2 labs, have centers big enough for 10T parameter models (about 5x the size of largest published frontier sizes). These are inherently 5x more expensive and 5x slower than frontier for inference, and meant to be trained with 5x the data at 5x the cost = 25x total training budget.

This is a speculative value proposition that I doubt to be a smart jump in model size. But 3x larger training capacity datacenters are scheduled for 2028, which is skynet dependency rather than following the economic efficiency (smaller/smarter) trend that is dominating currently instead.

Path that was laid out years ago is that skynet will buy every token, and so compute will always be scarce. US must pick AI as the last battle on China, because there are none left, and entire financial industry must go all in on skynet future. Financial industry did, but they are tapped out. Political environment is not to give government a $1T token budget for skynet at this time. AI revenue growth is plateuing, and GPU pricing is tanked.

The .com bubble was less desperate, or dependent on a government corruption backstop for skynet. But as everyone who participated in the gold rush, they didn't fully evaluate the consequences of competition for shovels.

Anthropic or OpenAI will almost certainly get bailed out. There's no chance that the US will throw in the towel and allow China to lead this tech uncontested.

What's the source of the article? Did you write it?

Make hay while the sun shines, they needed to inflate investment and revenue as much as they corks, before everyone realised it's a scam

I'm not that familiar with the details of what's going on in this industry, and this post is probably the most I've read about the actual mechanics of it.

What I don't understand is how it makes sense to build so much compute capacity so quickly. There seem to be predictions saying that we'll need vastly more DCs and electricity running compute in the future, but isn't that the same as:

I predict that within 100 years computers will be twice as powerful, 10 000 times larger, and so expensive that only the five richest kings of Europe will own them

Or possibly like when GPU bitcoin mining peaked, but now with orders of magnitude more money behind it?

Oracle is circling the drain, here's hoping Ellison is flushed away by it.

I expect Google, Nvidia, even Intel to survive through this, Anthropic might still for a while but days at the purported 1T dollar valuation are numbered.

🙄

No lmao we do this to run from our software optimization problems. It's not gonna get better seeing the way people use LLMs to write code.

Are we getting better at packing bits into the same bandwidth or just increasing bandwidth?

If you gave your email to Claude, then it must not have been private in the first place.

No

It is based on OpenHarmony which is open source under the Apache License https://gitee.com/openharmony/build/blob/master/LICENSE

Proprietary in the same way android is as far as I know. In that both have an open source foundation (AOSP and OpenHarmony respectively) but the final product built on top is proprietary.

Do you say the same of Microsoft? Or Apple? They both have their own OS. Huawei has 114,000 employees in Research and Development. Microsoft has 77,000. Apple has 166,000 in TOTAL and do not disclose how many are in R&D.

Huawei used AOSP, and Linux as a foundation for earlier versions of Harmony OS, but have since switched over to an entirely bespoke HongMeng kernel.

I'd personally rather it be FOSS but I'm still glad to have any sort of competition in the OS space. It should be normalized that all programs should just have to compile themselves into different versions for all OSes. The more OSes other than Windows we have the more companies will want this. I think Unity game engine already does this where it gives you a Linux, Mac, and Windows executable by default.

But it being closed source makes sense for its usecase too. The Chinese Government want this to be used in government there, just like Windows is in the US, and so they want the code locked down.

If I had to choose what OS to use my order of which ones I'd prefer probably goes:

GNU/Linux

FreeBSD

HarmonyOS

MacOS

Windows

I think the most important thing is that people realize Huawei is not making some Linux distro. They're not a small or medium sized company. They're a peer to Apple and Microsoft. An absolute behemoth that can support their own entirely in house software stack+ hardware ecosystem. That's the real story here.

I wouldn't be surprised if they also viaolating the GPL

ARR: anualized revenue run rate. An "unofficial accounting number" with no standard base period. By bragging about $65B(illion. T is trillion) in July, the start of their third fiscal quarter (Q3), and whole quarter showing no revenue growth, it gives a bad impression for the company.

IPO: Joining the stock market day. Valuation is number of total shares * price per share at open. S1 is the paperwork they have to fill prior to that day. They get to file their homework in draft stages until the SEC (government) accepts it as complete. In previous drafts of their paperwork, they had left a line with October revenue was ______ . They did not fill in that blank on Nov 2 when they did disclose a lot of info. Certain reason is that the number looks bad.

rest of the acronyms/words are company and names of models released.

Why not ask LLMs for help? /s