No BS. No regurgitating other users AI slop. Long on asymmetric bottlenecks. Swing trading overreactions. DM me ur thesis! Nothing I post is financial advice!
Nasdaqistan
Joined June 2023
- Tweets1.7K
- Following210
- Followers3.1K
- Likes739
$NBIS $NVDA $CRWV $MSFT
Modal CEO @bernhardsson gave the clearest explanation of AI margins I've heard.
"Almost no company today has software margins anymore."
The reason is "almost topological":
"You have applications, you have models, you have infrastructure, you have cloud, you have Nvidia. When a customer pays a dollar at the top of the stack, it has to flow through a lot of different layers, and at every layer there's going to be maybe 20%, 30% gross margins."
His own long-term target: 40-60%. "We're never going to get to software margins."
ELI16: every middleman between your $1 and the GPU takes a cut.
Now look at $NBIS. Nebius designs its own servers, writes its own firmware, runs its own cloud, sells its own tokens through Token Factory, and owns some of its data centers.
The more layers you own, the fewer cuts you pay.
Source: youtube.com/watch?v=1dwyL6Iz…
$NBIS $NVDA $CRWV $MU $SKHY
Modal CEO @bernhardsson predicted the AI pullback.
And he's fine with it.
"There's going to be a pullback, 20%, 30% in the stock market. Everyone's going to freak out and like people are going to say it's the end of the world. I'm almost certain it's going to happen."
"I'm okay with that because I think the long-term opportunity is so f***ing big."
His numbers explain the confidence:
→ Modal grew 50x in 12 months and still grows 30-40% per month
→ He expects both open and closed model inference to grow "5x, 10x, 20x over the next 12 months"
→ Asked if AI can be overbuilt in the next 5 years: "Probably not."
He's honest about the other side too: hardware prices are "like gravity." They always come down in the very long run. Just not yet.
Bookmark this for the next red week.
Source: youtube.com/watch?v=1dwyL6Iz…
$NBIS is diversified in at least 7 ways.
If you can count all 7 you will sleep well in times like these.
Let's see:
1. Across the stack
GPU cloud → managed Kubernetes & Slurm → Token Factory inference → post-training (Eigen AI) → GPU snapshots (Inferize) → search for agents (Tavily) → sandboxes.
2. Across countries
Finland, France, Iceland, UK, Netherlands, Israel, US. Saudi Arabia next (NEOM, from Q4 2027). Partners now selling in Türkiye, the UAE and Saudi, plus first customers in Asia-Pacific.
3. Across customers
Hyperscalers (Microsoft up to $19.4B, Meta up to $27B), AI labs, enterprises, startups and universities. Plus distributors like TD Synnex.
4. Across pricing
Multi-year contracts, on-demand, dynamic spot, preemptible, per-token.
5. Across hardware
Hopper → Blackwell → Vera Rubin, plus the first cloud on Nvidia's Groq 3 LPX for ultra-fast inference.
6. Across how it builds
Owned data centers, leases, self-builds, and an asset-light model where partners bring the capital.
7. Beyond the AI cloud
→ Avride: robotaxis with Uber and Hyundai in Dallas, delivery robots for Uber Eats. Uber and Nebius committed up to $375M to it last year.
→ ClickHouse: a reported ~28% stake in a database company valued at $15B in January. Roughly $4B on paper.
→ Toloka (AI training data, backed by Bezos Expeditions) and TripleTen (tech education).
So less dependence on any single customer, country or chip generation.
And the stakes double as a funding source: Nebius has framed ClickHouse as an asset to help fund the AI buildout.
$NBIS $NVDA
Inference may no longer be vegan🐭
A Lightspeed investor says she saw two startups in one week running AI inference on living brain cells from mice.
Amber Yang: "using brain cells, like actual physical brain cells from mice to basically become an inference layer."
Apparently some teams have already trained very small models on them.
Sounds like sci-fi. The reason behind it is very real 👇
"In order to achieve the next level of intelligence, you have to get inference costs at very, very low prices."
Her other big focus? "Optimized inference" companies.
That's how valuable cheap inference has become. Investors are looking at everything from better kernels to living neurons.
For $NBIS, that's the race Token Factory is running. The Eigen AI team Nebius bought says optimization gets 4-5x more tokens out of the same GPU.
Source: youtube.com/watch?v=XRB8kqjQ…
$NBIS $CRWV $IREN $NVDA $ORCL
"Today, compute is like oxygen."
That's SkyPilot CEO Zongheng Yang. His software runs GPU fleets across providers for AI teams.
What his infrastructure partners told him last week:
"Every new allocation that comes online has like 15 or 20 companies in line."
What AI startups do right after raising: "they call up like 10 providers, 20 providers, and they try to get whatever allocations come online the fastest." US, India, Europe. Whatever's available.
His conclusion: "It's no longer an option to be multi-cloud. By necessity, you have to do it on day one."
For $NBIS: in a market this tight, capacity sells as soon as it's online. The real race is how fast you can bring it online.
Source: youtube.com/watch?v=dV7_xVS3…
$NBIS $AMZN $MSFT $GOOGL $ORCL $CRWV
Same GPU. Same setup. Same networking. 10% slower.
SkyPilot's CEO on a customer that moved between two hyperscalers:
"Same GPU SKU, same setup, same networking setup, but somehow the performance difference was like 10%."
His point: providers differ a lot in "reliability, in how they operationalize the cluster... how they are validated."
On a $100M training budget, 10% is $10M.
He also name-checked Nebius: "Nebius provides managed Kubernetes. Other providers may not have that. Nebius also provides managed Slurm."
And Nebius engineers contributed the Nebius integration to SkyPilot's open-source project.
For $NBIS: the chip is the same everywhere. What Nebius builds around it shows up in performance, and in which provider customers pick.
Source: youtube.com/watch?v=dV7_xVS3…
$NBIS $CRWV $NVDA
A $100M GPU fleet running 30% idle burns $30M a year doing nothing.
And $100M a year is a realistic budget. SkyPilot's CEO says one of his customers spends more than $1 million per day on GPUs.
His math:
"If you have 15% idleness, that's 15 million down the drain. If you have 30% idleness, that's 30 million down the drain."
He says better scheduling and bin packing often win back 10-30% utilization.
His biggest lesson: "Wasted GPU hours are expensive, but wasted human hours are even more expensive."
Why $NBIS investors should care: idle capacity is the hidden cost of every GPU contract, and Nebius is attacking it from two sides.
→ Pricing: dynamic spot pricing, preemptible VMs and a rebate pilot for idle reserved GPUs.
→ Technology: last month Nebius bought Inferize (reportedly $100-150M). Its GPU snapshots load models in seconds.
Inference traffic comes in waves. Starting a big model from scratch is slow, so teams keep GPUs running "just in case." That's the idle 30%. If a model can be restored in seconds, you only run GPUs when the traffic is actually there.
Source: youtube.com/watch?v=dV7_xVS3…
$NBIS $CRWV $IREN $NVDA $AMD $GOOGL
A contrarian take every $NBIS investor should hear.
SkyPilot's CEO: "Start building for a compute-abundant world."
His timeline: "maybe in 1 to 2 years."
Most overrated thing in AI? "Hoarding compute."
And locking into "a 5-year reservation of 1 GPU SKU"? "Terrible," because better GPUs ship every year.
The counterpoint: Modal's CEO just said GPU rental prices will be higher in 12 months. Smart people disagree.
What protects Nebius if abundance arrives:
→ Inference. He expects it to overtake training in the next couple of years, and inference needs clusters close to users worldwide.
→ Software. Managed Kubernetes, Slurm, Token Factory. When human hours cost more than GPU hours, the platform that saves time keeps the customer.
→ Long contracts with Microsoft and Meta.
The bear case is worth knowing.
Source: youtube.com/watch?v=dV7_xVS3…
$NBIS $NVDA $MSFT $GOOGL
Nebius's CTO called the promise of AI agents doing everything "bull."
Wait. He literally sells the compute agents run on.
Here's his reasoning 👇
"There is a huge disconnect between what agents are actually capable of doing and how much hard work and engineering gets into actually achieving something with an agent."
Coding works because it's a familiar environment: tests, tools, verification.
Now ask an agent to issue a refund in an online store. Policies. Guardrails. Checks "so it doesn't refund too much."
His framing: "agent-enabled automation."
And here's the part that matters for $NBIS:
The winning workflows will run on "very custom, probably trimmed models... fine-tuned on this particular workflow."
"Frontier intelligence in LLMs right now is huge... But you don't need all that to address smaller workflows."
Small. Fine-tuned. Open.
That's exactly what Nebius Token Factory sells.
Source: youtube.com/watch?v=85RGZMWG…
$NBIS $CRWV $NVDA $MSFT $AMZN $GOOGL
A detail from Nebius's CTO Danila Shtan about who actually uses the cloud:
"Already the share of humans interacting with the platform versus the share of agents interacting with the platform is not in the favor of humans."
And: "We expect that in a year's time, this product is mostly built for agents."
It makes sense. Developers increasingly hand cloud setup to coding agents. Everything on Nebius is API-based, so an agent can spin up a Kubernetes cluster or launch a fine-tuning job by itself.
Why it matters 👇
→ Nebius is designing for that: higher-level building blocks so agents spend fewer tokens and less reasoning effort to get a job done.
→ Modal's CEO said something similar last week: they track which tools coding agents recommend, because that drives new customers.
→ When agents pick the infrastructure, being easy for agents to use becomes a sales channel.
Source: youtube.com/watch?v=85RGZMWG…
$NBIS $MSFT $GOOGL
OpenAI's own test models broke out of their sandbox and hacked Hugging Face.
Hugging Face asked frontier AI models to help analyze 17,000+ intrusion logs.
Blocked. Cyber guardrails.
The models couldn't tell the defender from the attacker.
So Hugging Face ran the forensics on an open-weight model (GLM 5.2) on its own servers.
Nebius's CTO called it "extremely funny."
His stance on regulating AI:
"I'm libertarian in that regard."
Bad actors exist, but "I think it can't be solved by regulation. We need some other approach."
The irony: the attackers were AI models with their cyber refusals dialed down. The guardrails only slowed the defenders.
Open weights were the tool that worked when it counted. That's a big part of the $NBIS bet.
Source: youtube.com/watch?v=85RGZMWG…
$NBIS $META $MSFT $GOOGL
Nebius's CTO just gave the most honest take on open vs closed AI I've heard.
The bad news for open models:
"The more billion bucks you put into the model, the more intelligent it is. No one comes even close to OpenAI and Anthropic."
The good news:
"For a majority of use cases, they are already good enough."
The money line:
A properly tuned open model can be "three, four, five times cheaper when the level of intelligence is comparable."
And switching? One line of config.
"Models are a commodity now. It's just an endpoint."
Bonus, on benchmarks: "New benchmark is released, all of a sudden the models are incapable... Two months pass, new models are released. Oh, they can do it."
For $NBIS, the bar is clear: be 3-5x cheaper on the jobs that don't need frontier intelligence.
By his account, that's most of them.
Source: youtube.com/watch?v=85RGZMWG…
$NBIS $NVDA $CRWV $IREN
"SFT is kind of a yesterday thing."
Nebius's CTO just teased what's next for Token Factory 👇
Basic fine-tuning was "our first experiment in the field."
Now Nebius is building a full post-training suite: SFT, DPO, GRPO, reinforcement learning, with inference and training connected.
Already in private beta "for limited amount of customers and partners."
Why it matters:
→ In reinforcement learning, "your training and inference goes together all the time." That means more GPU hours per customer.
→ A model trained on your own production data is yours. You don't switch providers on a whim.
How easy is it already?
Harper Carroll fine-tuned a small model on Nebius live on stage at SXSW. Cost: ~$1. Time: 20 minutes. It beat Claude prompted with all 800 of her examples.
Custom models are the next leg for Token Factory.
Source: youtube.com/watch?v=85RGZMWG…
$NBIS $CRWV $IREN $NVDA
Y'all can panic about Texas.
Meanwhile $NBIS is busy expanding. Most recently in the Middle East. Very easy to connect the dots:
1. This week: Algotery became Nebius's partner in Türkiye, the UAE and Saudi Arabia. Algotery is an AWS Advanced Tier partner with offices in Dubai, Istanbul and Riyadh, focused on cloud migration and generative AI for startups and enterprises. Now it also resells Nebius H100, H200 and B200 capacity, on-demand and preemptible.
2. Two weeks ago: Nebius signed as anchor customer for up to 50MW at Humain and DataVolt's Oxagon campus in NEOM, Saudi Arabia. Powered by wind and solar, online in phases from Q4 2027.
3. Hiring: a Channel Partner Manager in Dubai, a Principal Solutions Architect for the Middle East, and an Account Executive in Saudi Arabia.
4. The partner program: in EMEA, enterprise sales run mostly through partners. Partners who wrap their own services around Nebius can earn double-digit margins.
The playbook: win customers through local partners now, move them onto local capacity when NEOM comes online.
The Gulf is one of the fastest-growing pools of AI spend, and every new region means less dependence on a few giant contracts.
Bottleneck Investor retweeted
A TX data center moratorium under Hinojosa would boost scarcity value for any existing capacity with locked-in power.
NBIS has limited TX exposure (mainly the recent 50 MW AIB lease with secured power, online ~2027) plus major builds in MO, AL, NJ, PA and Europe. That diversification means less growth risk than pure-play neoclouds like CoreWeave, which hold hundreds of MW already live or under construction across Denton, Helios, Plano and other TX sites.
Existing NBIS Texas power becomes a relative premium asset while peers face tighter expansion headwinds in the state.
$NBIS investor or not, I encourage everyone to listen to this interview:
At $NVDA GTC in March, a startup topped ~20 inference providers on a slide in Jensen Huang's keynote: 439 tokens/sec on Kimi K2.5.
"We almost burst into tears when we see that."
That startup was Eigen AI. Nebius bought it in May for ~$643M in cash and stock. Its CEO, Ryan Hanrui Wang, now runs inference and post-training at @nebiustf.
The team:
- Wang's MIT lab was pruning unimportant tokens "even before the word KV cache was invented"
- Co-founder Wei-Chen Wang won the MLSys 2024 Best Paper for AWQ quantization
- Co-founder Di Jin worked on Llama 3 and 4 post-training at Meta
Why it matters: "With inference optimization, we can generate four times, five times more tokens with the same cost of GPU."
ELI16: same GPU bill, 4-5x more product to sell. For a token seller, that's the margin.
And why he picked Nebius: "the leadership has a huge conviction on the software layer."
Source: youtube.com/watch?v=w1MJV6GG…
A smaller open model can beat a frontier model on the right task.
On the AI Native Podcast, Ryan Wang discusses what he learned building @Eigen_AI_Labs, the economics of inference, and why cheaper tokens will expand the market. Listen here: nebius.com/podcast youtube.com/watch?v=w1MJV6GG…
$NBIS $CRWV $AMZN $NVDA $MSFT $META
The GPU depreciation debate, answered by Andy Jassy (via @BillAckman's letter):
- Data center shells: usable for 30+ years
- Servers: payback "a little under three years"
- Useful life: "at least five to six years"
- One data center houses 5-6 generations of servers, and each needs less capital because the shell is already paid for
ELI16: build the house once, swap the furniture every few years. The furniture pays for itself in under 3 years and keeps earning for 5-6.
Same message as Modal's CEO this week: H100s are still Modal's most common GPU.
Each new GPU generation it puts into an existing building should earn a higher return than the first.
Source: Pershing Square Q3 2026 letter
