The networking gear is only a partial driver of the higher cost.
There is real scarcity for large, single-fabric clusters and they should be treated as a different product than 32-64 GPU increments (e.g. your standard inference P/D cell).
People tend to assume that the more GPUs you rent, the cheaper each GPU gets, but that's not really how the market works
Prices often bottom out around 32–64 GPUs, then rise as larger, interconnected clusters become harder to find and more expensive to operate.
At a certain size (~2,000 GPUs+), network designs even require an additional layer of switches, which can create massive incremental cost
Cluster size deserves a lot more attention in compute pricing negotiations
Ethan retweeted
Imagine if everyone had the tools to shape intelligence itself. That day is today.
Today we launch the Trajectory platform - the easiest and most intuitive tool to post-train models of your own. The foundation for continual learning.
With just a few agent prompts, you get access to a full creative suite to train, eval, and deploy trillion parameter models.
On top of that, we believe in a future where you don’t just shape your model once, it also keeps learning on its own from the world around it.
We’re releasing two research previews.
1. Continual grading scales your taste, so your model learns what good looks like from far more than you could ever grade by hand.
2. GS1, a new algorithm that lets models learn directly from what they do in production.
We’ve now raised 60M, with a new round led by @Sequoia at a 300M valuation to democratize the most powerful technology of today - training models - so that the lawyers, doctors, and engineers of the world can own their own intelligence.
You know what good looks like, now go build it. Own your own value function with Trajectory.
Ethan retweeted
If you announce a model's benchmarks without releasing it (weights, product, api, etc "later"), feels like almost a sure thing that someone releases a better model in your category before you get it out.
This is actually completely backwards, the largest customers (OAI/Ant/Meta) only take/highly prefer baremetal.
Managed k8s/slurm is great for small/medium labs. But even amongst these, you'd find a preference for baremetal.
The framing makes it sound like baremetal is easy; it's not.
to be a real neocloud you need to build software, not just supply power and GPUs
lots of people are buying sites and filling them with B300s
that’s cool, they’ll probably sell a few racks, but they’re gonna get blown out of the water by companies that build software
spot is fundamentally a software product. the platform is constantly shifting workloads around and cleaning up between jobs, and you’re solving many orchestration problems at scale
i’d be bullish on anyone doing spot or serverless like beam.cloud
and bearish on anyone buying power and GPUs and calling themselves a neocloud
> Argon is currently rolling out to a set of trusted cyber defenders in the Fairwind Program, with broader availability as soon as possible.
🙃
Jev nailed the product requirements:
- Ultrafast
- Dirt cheap
- Accurate enough
I feel like every "but we already have Jev" post I've seen has just 2 of those 3 but you need all of them.
The Jev moment happening now is similar in energy to the OpenClaw moment that happened in January, and also to the Opus 4.5 moment before that. Developers are scrambling to find use cases for a new hot thing, and it's a sudden blooming of creativity.
Models have typically NOT optimized for specific use cases and have gone the other direction, generalizing over all of them. This could be a very important moment for the whole AI ecosystem if this turns out to be the beginning of other model skews that make the market much more diverse, such as compaction, summarization, extraction and more, all of which can now be used to optimize harnesses and decouple them from provider lock-in.
We'll see what happens over the course of the next few weeks when the model labs optimize their small models more or try new ways of branding and packaging them.
Ethan retweeted
1/ We’re entering the era of agentic commerce, and we believe that world models are the missing engine.
Instead of browsing static product pages, agents will let you interact with products inside your own world before making decisions.
Today, @DecartAI we're releasing Anywear👇
🤖 Made with AI
Ethan retweeted
we made a chrome extension that lets you try on any clothes on any site in real time and its live today
anywear.decart.ai/free-sahas…
go break it😃
🤖 Made with AI
We've really enjoyed working with @grx_xce and the @Intelligence_ai team, they move fast and are super creative when it comes to designing evals!
Today, we're introducing @Intelligence_ai.
In 6 months, as a team of 10, we scaled from $5M to $60M ARR and 5.5M users across 190+ countries.
We raised a $7.9M seed, led by @IndexVentures with participation from @conviction, @A_StarVC, and @combinator to build DesignArena, a universal interface for accessing and evaluating the world's AI capabilities.
Most evaluations try to simulate the real-world. We believe the real-world is the ultimate verifier.
People come to @DesignArena with a request. Models compete to fulfill the request, and users determine what works best for them.
Their live user behavior evaluates the models, improves how work is routed, and helps people access the right intelligence.
We've helped the world's leading frontier labs break the news on their SOTA capabilities.
What's the limit? Join us and find out.
Killer team, all the more relevant these days when labs are trying to get compute from any provider with capacity!
Today, SkyPilot is out of stealth.
Building custom intelligence is now existential. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
Frontier teams like @appliedcompute, @AbridgeHQ, @hippocraticai, @hcompany_ai, and @nubank already run on SkyPilot, with 10x faster time-to-intelligence and double-digit increase in GPU utilization.
AI teams today get compute anywhere they can. They then firefight compute fragmentation across providers. Researchers burn time on workload setup. Infra gets paged when GPUs go down. Frontier teams build slowly even on the fastest compute.
@skypilot_org turns your fragmented compute into one AI supercomputer, so you run frontier workloads faster. Many users manage 10,000+ GPUs across providers with SkyPilot. GPU hours consumption has grown 6x in the last 6 months.
1/ We're launching SkyPilot Platform — the AI compute platform for frontier AI teams to manage large GPU fleets and accelerate building custom intelligence.
Optimized for fleet management, team governance, and frontier workloads — pretraining, post-training, multi-cluster serving, and sandboxes. SkyPilot open source users can switch to the platform with a server URL change.
2/ We've raised over $20M led by @Lux_Capital (@breeves08), with participation from @AmplifyPartners (@dauber, @lennypruss), @coatuemgmt, @FoundationCap (@ashugarg , @JayaGup10), @RaceCapital, @thehousefund, and top operators like @alighodsi (CEO, Databricks), @JeffDean (Chief Scientist, Google), @rauchg (CEO, Vercel), @amasad (CEO, Replit), @ClemDelangue (CEO, @huggingface) and more.
We're hiring across Engineering and GTM to deliver the platform for the next decade of AI.
Above all, I'm excited to be building with the incredible team we've assembled, along with my cofounders Zhanghao @Michaelvll1, Romil @bromil101, Scott, and Ion @istoica05.
If you firefight AI compute, let's build.
ALT Today, SkyPilot is out of stealth. We help frontier AI teams build intelligence faster by removing their biggest bottleneck: AI compute fragmentation.
Ethan retweeted
This is a very good teaser but the details behind the scenes are even cooler!
For some intuition on CSM, Cerebras made everyone understand that if you can pool SRAM together on a big wafer you are going to go very fast. Etched can pool the SRAM without needing to keep the dies attached. This means they can push from 8x to 128x to 1000x+ chips all working together in ultra low-latency scale up domains, and then at the same time they can use LVI to pack the die with tons of flops for throughput. These are beautiful ideas and it’s great to see them coming to life!
Ethan retweeted
What if video editing had a “volume knob”?
We are excited to share new research from @DecartAI and @TelAvivUni, accepted to #SIGGRAPH2026. We found directions within video models that enable smooth control over edit strength, from subtle changes to full transformations!
Check out our paper & blogpost 👇
We're hiring in SF! Autoregressive diffusion models have their own unique challenges and our opt team paves the way for new experiences like Lucy.decart.ai
If this type of impact resonates with you, reach out to @hiba280700 with ref to your perf work!
stacked speaker lineup on perf ahead of GTC, come by if you're in SF!
Excited to have our chief scientist & co-founder Orian Leitersdorf speaking on realtime diffusion models — alongside @tedzadouri (FA4 first author) and @spectorb (ThunderKittens / @flappyairplanes) —
at a perf-focused evening, hosted by @Accel & @GPU_MODE.
Register now! luma.com/gpumodexdiffusion
Come chat with the team about realtime video at GTC!
We are hosting the first realtime AI event with @nvidia, @awscloud, @Deloitte, and @fal.
Meet the people building realtime AI and see what's shipping next from the lab.
🗓️ Tue, Mar 17 · 5–8 PM
📍 Fuze, San Jose
👉 Spots are limited, RSVP now → luma.com/burxctxy
