So, what is your take exactly - that everyone could vibecode software stack above bare-metal for managed gpus or inference, or that only few could do that even with Nvidia software and Claude Code? I'm confused. Are AWS/Microsoft/Google selling bare-metal to Fireworks.ai mean that they couldn't sell tokens through their token factories and will lose that part of the software stack completely to it?
Finally! Someone who understands the full stack that AI Clouds operate in!
@NotASithLord here actually shows real industry knowledge in his answer to my question today: Why does $NBIS sells Bare Metal GPUs to Modal?
Modal owns zero GPUs or datacenters and have tremendous growth in managed services. FireworksAI, TogetherAI, Baseten, Modal all take this layer for Open Source.
Meanwhile, as of this time, Nebius is only selling managed services because anyone who is able to bring GPUs online can sell it any form they one. Hell, IREN can get contracts for managed services and they are still developing their platform.
Once GPUs stop becoming short supply, it will be Anthropic, OpenAI, Hyperscalers, FireworksAI, TogetherAI, Baseten, Modal that are front line for AI Platforms. Maybe Coreweave. Non-tech people don't know this.
Then bare metal operators will derive their profit from the efficiency of their operations plus margins from owning their own Colo.
Those who say vertical integration do not know that Nvidia does the design of hardware, foundational software, and networking and everyone else including Hyperscalers is just integrating it. Google do vertical integration for TPUs but it lags way behind Nvidia GPUs.
As long as compute supply is constrained, all bare metal providers like $IREN, $NBIS, $CRWV will be able to masquerade as managed services providers but none of them will be able to keep up with Anthropic, OpenAI, Hyperscalers, FireworksAI, TogetherAI, Baseten, Modal in this part of the stack.
If $NBIS was better than Modal at managed services, there will be absolutely no reason for $NBIS not to just put Nebius Cloud on top of their own bare metal instead of sell it to Modal. $NBIS experts who say "short-term pricing" do not realize that Modal is making margin on top of "short-term pricing", why can Nebius Cloud not do the same?
Then once it's all said and done, I'm not even sure TogetherAI, Baseten, and Modal can remain competitive in this part of the stack. DeepSeek just open source not just their model but their entire inference runtime. When Nvidia finishes their open source model, they will likely open source their fully optimized runtime for it too.
You’re confusing two different products:
• On-demand: pay as you go, without a long-term commitment or spot-style interruption risk. Nebius currently lists B200s at $8.50/GPU-hour.
• Spot: interruptible access to temporarily idle capacity. Nebius can stop those instances when it needs the resources for regular VMs. Prices fluctuate, and Nebius explicitly caps them below the corresponding on-demand rate. B200 spot pricing starts at $0.99/GPU-hour, so $1.30 is entirely plausible.
Nebius confirms that spot uses temporarily idle capacity. Whether that includes unused capacity reserved under long-term contracts hasn’t been publicly confirmed. If it does, spot instances can improve margins on capacity serving both on-demand and long-term customers.
This quoted post is unavailable.
1
1
237
My impression is that Jim read the leaderboard, but not the report. For a software specialist, criticizing a benchmark without first reading its scope, specification and methodology is embarrassing.
I actually read the report.
Both critiques make the same basic mistake: they treat ClusterMAX as a ranking of “the best AI clouds overall.” It explicitly isn’t. SemiAnalysis says “for the folks in the back: we are ranking managed clusters.”
1. “AWS/Azure trained the best models, therefore they should rank highest.”
This completely misses the scope. The report explicitly explains that frontier labs like OpenAI and Anthropic often do not want a managed cluster: at their scale they operate the scheduler/control plane themselves and co-design the infrastructure stack. So “Claude runs on AWS” tells you very little about AWS’s quality as a turnkey managed-cluster product for a smaller lab.
2. “Firmus only has 256 H200s, yet ranks above AWS.”
The cited 256-GPU figure is from a particular deployment, not a sensible measure of the entire company. More importantly, company size is not the ranking criterion. ClusterMAX tests things like orchestration, networking, storage, monitoring, reliability and support. It even documents plenty of problems with Firmus rather than presenting it as some flawless cloud.
3. “Oracle had a security incident, so it can’t rank above AWS/Azure.”
Again, this assumes ClusterMAX is some generic corporate-quality or cybersecurity leaderboard. It isn’t. The rating is based on the managed-cluster product and a specific test methodology. A separate security incident does not automatically invalidate the result.
4. “Vast has GPUs in people’s basements, yet ranks close to AWS.”
Same category error. ClusterMAX requests a 32-GPU cluster with high-bandwidth networking, shared storage, monitoring and Slurm/Kubernetes. A random RTX 5090 marketplace listing in someone’s house is not the thing being benchmarked.
5. And the tiers aren’t a distance metric.
“Vast is one tier below AWS” does not mean “a bedroom GPU is almost as good as AWS.” Participation Ribbon is explicitly the category for providers that, in SemiAnalysis’s words, do the “bare minimum.”
6. Why is AWS only Bronze then?
Because SemiAnalysis actually tested the managed-cluster experience and found provisioning problems, storage issues, EFA complexity/latency and broken remediation paths. Their conclusion is explicit: “Managed clusters are not AWS’s priority.”
There are fair criticisms of ClusterMAX: the tier aggregation is somewhat opaque, and the Vast rating in particular could use much better documentation of exactly what configuration was evaluated.
But neither of these posts really makes those criticisms. Jim mostly takes the leaderboard, assumes it means “best AI cloud/company/infrastructure,” and then ridicule contradictions that disappear once you read the methodology.
I assume he didn’t even read the excerpt about Iren (I won’t cite it). I’m not sure how can you provide a good technical analysis on the ai space when you completely ignore the report of the largest analytical house in the ai domain. And even if we assume that the ranking is paid Iren could use its money better to pay for it rather than sponsor a sports team. Semianalysis is projected to have 100m 2026 revenues and certainly affect the mindshare of possible Iren customers more that the sports team does.
Issues with ClusterMax 3.0
1. $ORCL had a huge security breach today (1) and is somehow ranked above AWS and Azure.
2. Firmus AI has a grand total 32 racks / 256 GPU H200s (2) generating 32m ARR and is a better AI cloud than AWS.
3. Cursor’s Composer Model was trained on FireworksAI (3), ChatGPT was trained on Azure, and Claude trained on AWS. None of these are apparently top AI Clouds. None of the top AI models are from “top AI clouds”. Hmm…
ClusterMax is clearly pay to rank ad.
(1) teapawt.org/rdd147/status/21026490…
(2) firmus.co/newsroom/ai-singap… - site taken down but you can get the archived info from ChatGPT
(3) cursor.com/blog/composer-2-t… (list FireworksAI and Colfax as collaborators)
You are not paying much, Nebius RnD budget is tiny compared to its revenue (67m out of 527m expenses in q1 2026), it's just derisking. If fireworks/databricks/Palantir can get great margins while buying bare-metal, Nebius/Coreweave will get even better margins on the same services. And Nebius will just kill the software that won't pay off. Buying just Fireworks (Palantir and Databricks are more diversified) may be a risk because they sell very compute-intensive services while relying on third-parties for bare-metal, but not for Nebius. As for the risk Iren also competes with hyperscalers in the medium-term on bare-metal, in 3-5 years hyperscalers will only buy colocation, so Iren is secured as a colocation provider only. As a short-term (6-12 month) trade Iren may be more lucrative because it's revenue bump should be steeper. Every company that can deliver gpus should be ok before the supply/demand imbalance eases. Roman Chernin said in his interview with @daniel_koss that some companies come to Nebius' inference platform because they don't get enough capacity on other platforms, so their software revenue will expand while the capacity is constrained. The moment gpus scarcity is even temporarily gone you don't want to be in either Nebius or Iren irrespective of their long-term potential, so I don't understand all that talk about the long-term.
$PLTR: An End to End AI Platform for Enterprise
Certain segments of FinX think the only companies that have a chance of running AI inference is HS + $CRWV, $NBIS are have a completely myopic view outside of stock tickers.
Here you see a real engineer from Palantir explain that they provide “the tools to evaluate, tune prompts, change architecture, and run all combinations across model providers and families”.
“This Intelligent Engine allows customers to run Nemotron models in sovereign environments with our critical infrastructure products….AIP, Ontology, Foundry, and Apollo. It will also enable customers to fine-tune these open-weight models for their domain-specific workflows, all securely inside their controlled perimeter.”
$ORCL has an AI Platform but $PLTR buys $ORCL bare metal. $PLTR may buy bare metal from $IREN or $NBIS but it will be bare metal.
This is why I invest in $IREN: the AI Software Platform side of Neoclouds has a lot of strong players. Less is better, I don’t want to be paying that much for a company to compete against Hyperscalers, Frontier Labs, Palantirs, Databricks, FireworksAIs of the world. For Neoclouds, I prefer as much bare metal as possible because that’s what companies like Palantir will be buying.
2
2
4,708




