shoshin

Core: $nbis, $asts, $unh, $baba, $msft; Bio: $abvx, $rvmd, $cytk, $nktr

[shoshin @zen_tropy]

@HedgeDirty A carliment (with apologies to the owls) 😀

[Alexander @AlexfromBabylon]

$NBIS BREAKING: @mvcinvesting confirmed in a private Q&A during the Nebius Inflection event with management and other institutional analysts that the 1.2 GW DC in Penn, US is fully powered by SIGNED grid connected power to be deployed in a multi year period. That’s a major derisking in the physical datacenter buildout!

[shoshin @zen_tropy]

$nbis
It's hard to grasp the tsunami of compute demand that's surging towards us in the coming years w/physical ai . . .
"We also discovered a new type of "physical scaling": 8 robots exploring in parallel improves significantly faster than fewer ones."

Nebius sees it advancing, which is why, of course, they're building a dedicated robotics/physical-AI cloud stack w/$nvda, packaging workflow tools, funding ecosystem adoption programs. The goal is to develop a purpose-built, end-to-end cloud platform specifically for the robotics life cycle: simulation, synthetic data generation, model training, and real-world deployment.

https://t.co/h0p9Ehn3Y9

[Jim Fan @DrJimFan]

Today, we enable AutoResearch in the physical world for the first time! Introducing ENPIRE: we give 8 Codex agents a fleet of robots, an allocation of GPUs, and generous token budget. We set them free with a simple goal: solve the task as quickly as possible, keep the robots busy but stay safe, don't waste precious compute. Make no mistake.

Then humans step aside and our watch begins. The robot fleet starts to come alive: they learn to look for visual clues, reset the scene, practice novel skills, tinker with control stack, read papers online, debate, reflect, get stuck, and try again directly on the hardware. All we did is to give Codex an API to the world of atoms, and the rest is emergence.

ENPIRE is able to solve high-precision tasks like tying zip-ties, organizing fine pins, and installing GPUs all by itself. We also discovered a new type of "physical scaling": 8 robots exploring in parallel improves significantly faster than fewer ones.

A part of our NVIDIA GEAR lab now self-improves tirelessly over night. We just read the reports in the morning.

/goal: we all take a holiday and Jensen wouldn't even notice ;)

We will be open-sourcing everything, so you can host your self-running robot lab at home too! Deep dive in the thread:

[shoshin @zen_tropy]

@DrTomsLens Please—apple/samsung, sony/samsung, apple/ms, ford/gm . . . list goes on. It’s called co-opetition & there’s nothing special about it. Unless we’ve devolved so deeply into the scamosphere, we know longer expect people to honor their contracts.

[shoshin @zen_tropy]

$nbis, $crwv, $iren
Unlike saas, model margins compress with increased usage. . . Only thing “guaranteed is that exploding AI usage requires exponentially more infrastructure. That is why the most obvious beneficiaries of the AI boom remain the companies supplying the picks and shovels of the ecosystem.”

[Ricky Ho @rickyho_1989]

This chart from SemiAnalysis highlights one of the most misunderstood aspects of the AI industry today: revenue growth does not necessarily translate into profitability.

The market often assumes that a $20-200 monthly AI subscription is an extraordinarily high-margin software product. Historically that would be true. Traditional SaaS companies might generate 70-90% gross margins because serving an additional user costs almost nothing. Frontier AI is different. Every query consumes expensive GPUs, power, networking, and inference capacity.

The chart suggests that profitability is highly sensitive to user behavior. At low utilization levels, both OpenAI and Anthropic generate exceptional gross margins. However, as users increasingly rely on AI for coding, research, agent workflows, and long-context reasoning, margins deteriorate rapidly and can even turn deeply negative.

This is particularly important because AI adoption appears to be following a familiar pattern. Most users initially experiment lightly, but power users tend to expand usage dramatically over time. The more useful models become, the more inference they consume. Ironically, better products can create margin pressure rather than margin expansion.

The chart also helps explain why the industry is racing toward usage-based pricing, rate limits, premium tiers, and agent-specific subscriptions. Unlimited access sounds attractive from a marketing perspective, but frontier reasoning models can burn enormous amounts of compute. If a small percentage of users consume orders of magnitude more tokens than average, subscription economics quickly become challenging.

From an investment perspective, the biggest takeaway is that AI may not resemble traditional software economics. The winners may not be the companies with the highest subscription growth, but the companies that can continually reduce inference costs faster than user demand grows.

This is why every frontier lab is simultaneously pursuing two objectives: building smarter models and making them dramatically cheaper to run. Intelligence improvements alone are insufficient if inference costs grow faster than monetization.

The chart also reinforces why the infrastructure layer remains so attractive. Whether OpenAI, Anthropic, Google, xAI, DeepSeek, or someone else wins the model race, every additional query ultimately translates into demand for GPUs, HBM memory, networking equipment, power infrastructure, and data center capacity.

In many ways, this chart supports a view that AI is currently experiencing a version of Jevons Paradox. Costs per token are collapsing, yet total spending continues to rise because usage is growing even faster. The economic value is increasingly shifting toward those selling compute rather than those consuming it.

The implication is straightforward: the AI leaders may eventually become extraordinarily profitable businesses, but that outcome is not guaranteed. What is guaranteed is that exploding AI usage requires exponentially more infrastructure. That is why the most obvious beneficiaries of the AI boom remain the companies supplying the picks and shovels of the ecosystem.

The question is no longer whether demand exists. The question is who ultimately captures the economic rents generated by that demand. Today, the evidence increasingly suggests that the infrastructure providers are capturing a larger share of those rents than the model providers themselves.

[shoshin @zen_tropy]

$abvx Doubled position, here. Case building market got spooked by misbegotten cancer signals in the safety data that weren't contextualized re: placebo arm px yrs. & disease pop background rates. Stellar specialist coverage on bioX + sell side coming round.
A risk, but, if true, possible future soc asset on sale . . .

https://t.co/VKFpIhwR9l

[Biotech_Jack @biotech_jack]

$ABVX Research update

[shoshin @zen_tropy]

$nbis Yet more evidence of the current pricing power availiable to infra/platforms that reserve capacity for short term, spot & recurring contracts. Here, it's Baseten confirming a pre-renewal price doubling on a B200 cluster (May notification for an October contract renewal)

🤔Didn't @daniel_koss just recently confirm that ~50% of nebius's projected arr will be coming from enterprises/ai natives and not LT hyper contracts . . .

https://t.co/hBYpef79x8

[Rittenhouse Research @RHouseResearch]

Baseten CEO @tuhinone tells Altimeter's @apoorv03 that one of Baseten's cloud providers has already indicated their B200 prices ($/GPU hour) are set to double when existing contracts expire and are up for renewal later this year.

"If you go out right now saying you want a thousand GPUs, truly.. people are talking about Q2 of next year. So 12 months out, maybe 15 months out.

We have a cluster.. in one of these clouds.. of B200s.. Our unit price right now is $2.63 an hour.. that's up for renewal in October. They came to us already in May and said $5.10 is the new price.. So double."

[shoshin @zen_tropy]

@BKad2005 CPU comments at 37:55 (“will be scaling Vera and Intel CPU’s”). Nothing on memory; thats just the other poster’s speculation

[shoshin @zen_tropy]

$nbis Your friendly reminder that Cohere, a Nebius customer, has special relationships with the Canadian & British governments . . .

[The Transcript @TheTranscript_]

Canadian PM Mark Carney: "The situation we’re in collectively right now with Mythos and Fable is something that can happen with overreliance on certain models. Nobody has done anything wrong in the situation. But we will have done something wrong if we just accept this, don’t take the lesson, don’t build out and diversify."

[Nebius Token Factory @nebiustf]

@ClementDelangue As we always say 🫡

[shoshin @zen_tropy]

@BKad2005 @BenBajarin Yeah, of course. But, at some point, the hypers who offer managed open model inference will look to consolidate and get competitive on price in that arena. Question is does the new enterprise fear of single model reg point of failure drive them there faster?

[Ben Bajarin @BenBajarin]

Workload routing to the right model for the right job & token cost is the future.

[shoshin @zen_tropy]

$nbis The strategic/structural imperatives for job-specific, open source model routing have been stacking up faster than can be cataloged . . .

To cost-optimization (tokenomics) & capability-maximization (leveraging both model & version-specific strengths, which are constantly in flux), enterprises must now add a new sovereign regulatory factor re: assessing their need for risk-mitigation in AI platform partners—a category which already included avoiding model & vendor lock-in/concentration to preserve negotiating leverage and bus. flexibility .

Sometimes the market moves toward you faster than you expected. These shifts are likely scaling earlier than they planned, but if anybody has a chance of meeting this accelerated moment, it’s the grounded, delivery-focused team at Nebius.

[Aaron Levie @levie]

The layer that can route to the best AI model for the particular job is going to increase in value substantially. There are at least 3 big reasons:

* Cost optimization: there are plenty of use cases where you need frontier intelligence for some tasks and something far cheaper for others. Even in the same task you may use frontier intelligence for planning and review of the work, but an OSS or cheaper model for the bulk of the workload. This is going to be standard across large buckets of work going forward.

* Capability maximization: despite the bitter lesson and models generally getting better in the same direction, there are still lots of differences between models. Some are better at tool use, others better at coding, and others again better at certain domains of knowledge work. The ability to route between these at different times is a huge advantage.

* Risk mitigation: while the Fable situation is somewhat of a black swan, it’s possible we’re heading toward a regulatory environment where governments may restrict models at different times based on their approval mechanisms or new things they discover. This means you’re going to want flexibility in being able to deploy workloads across different providers as a form of risk mitigation.

Ultimately, it’s going to increasingly be a a strategic advantage for the applied AI layer that they can effectively route between models. Will be very interesting to see how this evolves.

[shoshin @zen_tropy]

@kevinsxu @daniel_koss Had exactly the same thought from a distance. Ecosystem construction is a signature Arkady/Roman move; they’ve historically thought in terms of multi-decade infrastructure & developer network effects. Feels like the inaugural launch of an important series . . .