shoshin retweeted
Fluidstack is closing at $40bb but $NBIS has a $50bb mkt cap + Avride and Clickhouse!?!
$nbis Rich technical discussion w/nebius vp of product as a prelude to upcoming GTC. Interesting tidbits re: training still brunt of workloads; traditional co's now joining ai/dig. natives on platform; future will be serving continuous lifecycle loop of training/post-trainging-inference . . .
Closes with a nebius recruiting pitch: "We need a lot of people; please come work witrh us."
h/t to @anthonybs for posting.
youtu.be/wnYMHktPIzQ?si=UYjr…
$nbis, $pltr The all-you-can-eat builder buffet is now open & growing. Own your intelligence.
nebius.com/blog/posts/three-…
$nbis, $nvda, $pltr I'd opt for a material deepening of support over an acqui-hire, but either way directionally bullish for enterprise uptake of hosted ow inference, given the risks associated (rightly or wrongly) w/chinese models.
Rock solid, long term backing for a competitive (& continually scaling/improving) western ow ecosystem improves the model supply available to nebius, while its palantir relationship offers a specific route for deploying those models into controlled, domain-specific enterprise workflows . . .
We’re going to see a whole lot more of this kind of connect-the-extant-dot “discovery” . . .
holy fucking shit.
90+ opus 5.5 agents spent three days doing actual materials research and surfaced two candidates for a class of room-temperature magnetic semiconductor researchers have been searching for.
and one of them already fucking existed. humans made it. studied it. published papers on it. it was literally synthesized in 1999.
then 27 years later a swarm of opus 5.5 agents comes along and goes “wait, this might have the exact properties everyone has been looking for.”
there must be an unbelievable amount of science hiding in plain sight simply because no human has connected the right papers, calculations and ideas yet.
$nbis
Reflection releases their new ow (SMoE) model, Bean. Taking aim out of the gate @ increasingly cost-sensitive enterprises, who need an alternative/supplement to frontier & chinese models. Specifically focsed on efficient inference, and "more intelligence per token."
reflection.ai/blog/introduci…
$nbis
Reflection is a key Nebius customer. Arkady & co. marshalling their compute to supplement (and risk-manage) their portfolio of ow models—and getting paid while doing it. . . .
$nktr Off treatment data in on alopecia that suggests a durable immune system rebalancing. 75% keep hair re growth after 4 months off tx; 63% after 6 months off. 10 % grow new hair.
$NKTR In alopecia areata patients, SALT Score ≤20 response maintained in 75% at 4 months off treatment and in 63% at six months off treatment.
Hair regrowth continued for 6 months off treatment with 19% of patients achieving SALT Score ≤10, an improvement from 7% at the end of 52 weeks of treatment
🤔 Interesting . . . OpenAI opens a B2B marketplace essentially for ow inference . . .
xitter.cf/8teAPi/status/21050146…
shoshin retweeted
Woah, huge bullish signal for open models that OpenAI feels a pressure to serve users with these models. It's not a world where OpenAI alone can build all the models it's customers need!
Baseten joined the @OpenAI B2B marketplace today as one of the first open-model inference providers.
The best AI companies are already running a mix of closed and open models at huge scale. We're excited to give OpenAI enterprise customers the ability to use open models powered by Baseten natively within Codex and through the Responses API.
👀 $nbis "preemtible compute now reprices algorithmically every 15 minutes"
xitter.cf/tanayj/status/21046639…
Nebius ran a B200 auction and found it was pricing 15% too low, per CRO Marc Boroditsky
- pipeline pricing was 15-20% below what buyers would pay
- preemptible compute now reprices algorithmically every 15 minutes
- they have 4-6 customers lined up per unit of compute
(via @theinformation podcast)
$ABVX More UEG abstracts: Long term (up to 7 yrs) Ph2 OLM data showing endo remission maintenance/ increases despite dose descalation & no new safety sigs
programme.ueg.eu/2026/week/d…
$ABVX 👀 UEG sub group data out. Look @ the placebo Δ's for px's refractory to other tx's, including JAK's. Clin remissions across the board for up to 4 prior tx failures.
xitter.cf/lcbnuoH/status/2104042…
$ABVX UEG new maintenance data
programme.ueg.eu/2026/week/d…
34-42% overall CR in AT-IR
24-39% endoscopic remission in AT-IR
27-33% CR in 4+ AT failures or Jak failure
= 9b MC, right?
$nbis Marc B. FChat yesterday @ Fellows Forum. Nuggets:
Vision: "Our vision is ultimately to become the first AI-native hyperscaler. We do not think of ourselves as a neocloud. We think of ourselves as the next AWS, Azure, or GCP. We believe the path in front of us can produce $100 billion of revenue in the relatively near future."
Demand: “For every GPU, we have a customer ready to sign. We have 4 customers ready to sign for every GPU; we’re saying no 75% of the time.”
Visibility: Sees “reasonable clarity” for about 18 months at current demand rates . . . "customer willingness to sign up for tens of thousands of GPUs in the middle of next year is very high."
Org. Growth: Says company is "approaching 2,000 employees and expects to have a footprint of 20 data centers by year-end, more than double where we were last year. Our core financial performance is growing at a 5x rate . . ."
Inference to Intelligence Platfom: Reports inference solutions are the fastest-growing part of the business and quickly on a path to becoming a billion-dollar business, "we are confident inference becomes the lion’s share of the market opportunity we pursue."
But also observes training does not disappear, "what we cannot yet anticipate is how customers will take a model into a commercial cycle and then have to come back and retrain it." Points out that clients are taking open-source models & post-training them with their own data. If this becomes a recurring, dynamic cycle, Nebius (having inference & training on the same platform, potentially sharing/optimizing the same compute) is "uniquely positioned to create that virtuous cycle."
All told, describes compute, training/post-training capabilities, inference, etc., as discrete capacities that people are cobbling together but that can be knitted together to make a powerful intelligence platform: "My confidence is based on success & momentum in the market we'll be selling an intelligence platform as opposed to just selling inferencing."
Co-opetition w/ Hyperscalers: Describes fantastic relationships/contracts w/ HS's but also reports regularly winning business from HS's "because we meet the AI builder where they are . . . we increasingly give them capabilities to build models, post-train models, and deliver inference in a uniform, cohesive fashion that supports their business objectives. Hyperscalers face a dual challenge: they must manage their existing businesses. They cannot simply say no to large-scale enterprise agreements and global agreements with digital-native customers. They face an innovator’s dilemma in allocating resources and capital expenditure. We get to focus all our resources on one problem set around AI. With nearly 1,000 engineers, I would guess our investment is deeper and wider, and that can continue to propel us forward.
With hyperscalers, the real question is how agile they remain—or, more importantly, how agile we remain. Can we keep pushing the frontier of the capabilities we deliver? My confidence is very high."
Developer/ Partner Ecosystem: Says nowhere near the size yet of HS ecosystem , but " We are building a community around the company. We are early, but I anticipate that in the not-too-distant future we will report a builder community of more than one million.
youtube.com/live/iLhMOdYWapI…
