Pinned Tweet
AWS and GCP are committing absolute robbery with their GPU prices
Introducing beam.cloud compute pools
> rent GPUs across 10+ regions
> H100s from 2.75/h
> 8xB300s available at this very moment
> Mount portable NVMe drives to any instance
You don't need to raise 100 million ymore - we're making compute available to everyone
Introducing beam.cloud compute pools
> rent GPUs across 10+ regions
> H100s from 2.75/h
> 8xB300s available at this very moment
> Mount portable NVMe drives to any instance
You don't need to raise 100 million ymore - we're making compute available to everyone
12
5
133
13,604
you don’t need much capital to sell 4 or 5 year GPU contracts
take a 20% deposit from the customer, hand it to an OEM or lender, and they finance the rest of the hardware without having to touch your balance sheet
spot is different because no one is putting up a deposit for it, so you’ve gotta finance it yourself
that’s why very few neoclouds can offer spot, even though it’s a way better business
but SpaceX doesn’t have this constraint
they can use all their money from Starlink and rocket launches to aggressively buy hardware and take on utilization risk themselves
take a 20% deposit from the customer, hand it to an OEM or lender, and they finance the rest of the hardware without having to touch your balance sheet
spot is different because no one is putting up a deposit for it, so you’ve gotta finance it yourself
that’s why very few neoclouds can offer spot, even though it’s a way better business
but SpaceX doesn’t have this constraint
they can use all their money from Starlink and rocket launches to aggressively buy hardware and take on utilization risk themselves
SpaceX is making a fortune renting GPUs on shorter terms. There’s huge demand for <1-3 year capacity, but almost no one else offers it at scale because they all finance on 4-5 year deals. Insane opportunity for whoever has the guts to do this next
3
0
17
1,382
very excited by the @TigrisData acquisition by @Dropbox - they helped us scale from early on
for context, basically every GPU cloud offers storage volumes, but you can’t move them between regions
if you use Runpod or Verda you’ll have a B300 in us-east and all your data, but then they’ll run out of capacity and your data is stuck in us-east
you’ve gotta move all your data to a new volume, which kinda sucks - if you’ve got 500Gi in a volume, it’s annoying to move that around
the reason very few clouds solve this is because distributed storage is a hard problem
at beam.cloud, we run across 14+ different regions, and basically all of our users have over 100Gi in storage
early on, we added multi-region storage powered by a distributed filesystem we built on top of Tigris + a caching layer
this lets us cache model weights and data near the servers running user workloads, and we don’t pay egress fees
everyone knows you need to be multi-region to scale GPU capacity, but far fewer people talk about how storage is what actually makes that GPU capacity available across regions
@ovaistariq and @HimankChaudhary are storage legends, and it's very smart move by Dropbox - excited to see where they go from here
for context, basically every GPU cloud offers storage volumes, but you can’t move them between regions
if you use Runpod or Verda you’ll have a B300 in us-east and all your data, but then they’ll run out of capacity and your data is stuck in us-east
you’ve gotta move all your data to a new volume, which kinda sucks - if you’ve got 500Gi in a volume, it’s annoying to move that around
the reason very few clouds solve this is because distributed storage is a hard problem
at beam.cloud, we run across 14+ different regions, and basically all of our users have over 100Gi in storage
early on, we added multi-region storage powered by a distributed filesystem we built on top of Tigris + a caching layer
this lets us cache model weights and data near the servers running user workloads, and we don’t pay egress fees
everyone knows you need to be multi-region to scale GPU capacity, but far fewer people talk about how storage is what actually makes that GPU capacity available across regions
@ovaistariq and @HimankChaudhary are storage legends, and it's very smart move by Dropbox - excited to see where they go from here
🚀 Dropbox plans to acquire Tigris, and everything you rely on today stays as it is: the product, the API, the pricing, the people.
Even more storage, free to move to any cloud.
Here’s what it means and what comes next: tigrisdata.com/blog/dropbox-p…
3
0
23
1,767
Why use Hetzner when you can use EC2
Why use Cloudflare when you can use CloudFront
Why use PlanetScale when you can use RDS
Why use Fly.io when you can use ECS
Why use Heroku when you can use Elastic Beanstalk
Why use Clerk when you can use Cognito
Why use Sentry when you can use CloudWatch
Why use Redis when you can use ElastiCache
Why use Cloudflare when you can use CloudFront
Why use PlanetScale when you can use RDS
Why use Fly.io when you can use ECS
Why use Heroku when you can use Elastic Beanstalk
Why use Clerk when you can use Cognito
Why use Sentry when you can use CloudWatch
Why use Redis when you can use ElastiCache
Rip all the sandbox startups
46
8
280
44,547
having a lot of fun using claude + blender + seedance
> Blender running in beam.cloud VM
> VM exposed via MCP for Claude
> Ask Claude to build scene in Blender
> Claude iterates on the scene based on each render
> Give final render to Seedance 2.5 to make into video
can't wa2
> Blender running in beam.cloud VM
> VM exposed via MCP for Claude
> Ask Claude to build scene in Blender
> Claude iterates on the scene based on each render
> Give final render to Seedance 2.5 to make into video
can't wa2
5
1
33
2,536
many cloud providers start by wrapping developer abstractions on top of AWS and GCP
since we started our business with GPUs, we needed to expand to niche suppliers early on and made our stack run on bare metal
very early on, we tried building on AWS and it was a nightmare
you couldn’t get GPUs, had to file quota requests, and it was 5x more expensive than neoclouds
then we started running on neoclouds, which helped a bit. you could get a lot more hardware for better pricing and with better DX
but the real unlock was going directly to niche suppliers and buying bare metal
instead of asking AWS for quota requests, we could find bare metal machines in smaller data centers and connect them to our control plane in 30 seconds
today, our cloud runs on thousands of servers across 14 regions, and it’s because we made this investment in distributed bare metal compute
building on hyperscalers is a natural starting point, but if you're serious about scaling efficiently you need to be at the lowest level of the stack
since we started our business with GPUs, we needed to expand to niche suppliers early on and made our stack run on bare metal
very early on, we tried building on AWS and it was a nightmare
you couldn’t get GPUs, had to file quota requests, and it was 5x more expensive than neoclouds
then we started running on neoclouds, which helped a bit. you could get a lot more hardware for better pricing and with better DX
but the real unlock was going directly to niche suppliers and buying bare metal
instead of asking AWS for quota requests, we could find bare metal machines in smaller data centers and connect them to our control plane in 30 seconds
today, our cloud runs on thousands of servers across 14 regions, and it’s because we made this investment in distributed bare metal compute
building on hyperscalers is a natural starting point, but if you're serious about scaling efficiently you need to be at the lowest level of the stack
6
1
33
2,512
Claude + Blender + RTX 4090 + 60 FPS
Introducing beam.cloud GPU Sandboxes
Give your agent a full cloud Ubuntu desktop with a GPU attached, streamed over WebRTC
4 vCPU, 8Gi RAM, 50Gi disk, $0.00045/second
Run Blender, SolidWorks, Unity, and any desktop software with one command
Introducing beam.cloud GPU Sandboxes
Give your agent a full cloud Ubuntu desktop with a GPU attached, streamed over WebRTC
4 vCPU, 8Gi RAM, 50Gi disk, $0.00045/second
Run Blender, SolidWorks, Unity, and any desktop software with one command
8
4
47
2,501
compute has always been a trade-off between cost and performance
most cloud providers hide this from you. you'll pick cheap or fast, and get rug pulled by an 8x CPU oversubscription
obviously, we decided not to do that.
our new pricing lets you choose exactly which trade-offs t cheap compute? try a spot VM for $0.036/h
want faster and slightly less cheap? try a shared sandbox for $0.088
there's no generically "best" option
compute has always been a trade-off, and making that decision transparent to customers is one of our most important jobs as a cloud provider
most cloud providers hide this from you. you'll pick cheap or fast, and get rug pulled by an 8x CPU oversubscription
obviously, we decided not to do that.
our new pricing lets you choose exactly which trade-offs t cheap compute? try a spot VM for $0.036/h
want faster and slightly less cheap? try a shared sandbox for $0.088
there's no generically "best" option
compute has always been a trade-off, and making that decision transparent to customers is one of our most important jobs as a cloud provider
sandbox providers love to brag about offering cheap compute while selling 8x oversubscribed CPU instances
unfortunately, you don’t really know about this until trying the product
one thing we value in our providers is transparency - so instead of secretly overcommitting CPU, we’re introducing three new pricing tiers to beam.cloud today
𝗦𝗽𝗼𝘁 - dirt cheap compute that may be interrupted
𝗦𝗵𝗮𝗿𝗲𝗱 - shared vCPUs, non-interruptible
𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 - ultrafast gaming CPUs, no preemptions
we could have said “we’re the fastest” or “we’re the cheapest”
instead, we’re just giving you all the options - decide which price and performance fit your use case
3
0
10
1,265
sandbox providers love to brag about offering cheap compute while selling 8x oversubscribed CPU instances
unfortunately, you don’t really know about this until trying the product
one thing we value in our providers is transparency - so instead of secretly overcommitting CPU, we’re introducing three new pricing tiers to beam.cloud today
𝗦𝗽𝗼𝘁 - dirt cheap compute that may be interrupted
𝗦𝗵𝗮𝗿𝗲𝗱 - shared vCPUs, non-interruptible
𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 - ultrafast gaming CPUs, no preemptions
we could have said “we’re the fastest” or “we’re the cheapest”
instead, we’re just giving you all the options - decide which price and performance fit your use case
unfortunately, you don’t really know about this until trying the product
one thing we value in our providers is transparency - so instead of secretly overcommitting CPU, we’re introducing three new pricing tiers to beam.cloud today
𝗦𝗽𝗼𝘁 - dirt cheap compute that may be interrupted
𝗦𝗵𝗮𝗿𝗲𝗱 - shared vCPUs, non-interruptible
𝗣𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 - ultrafast gaming CPUs, no preemptions
we could have said “we’re the fastest” or “we’re the cheapest”
instead, we’re just giving you all the options - decide which price and performance fit your use case
📢 ANNOUNCING: New CPU pricing tiers
Excited to share new pricing for VMs!
You can now pay less for CPU work that can share vCPUs or tolerate interruptions. CPU functions, endpoints, sandboxes and VMs can run on one of three tiers:
Spot: $0.0063/vCPU-hour + $0.00135/GiB-hour
75–92% less than our previous rates. Run on spare capacity and may be interrupted with 30 seconds notice.
Shared: $0.011/vCPU-hour + $0.0055/GiB-hour
39–81% less than our previous rates. Shared vCPUs, not preemptible, up to 8 vCPU / 16 GiB.
Performance: $0.05/vCPU-hour + $0.01/GiB-hour
38% less than our previous sandbox rates. No CPU sharing and not preemptible.
These savings are for a 4 vCPU / 8 GiB workload and vary by product.
More details 👇
4
4
63
7,537
Excited to see beam.cloud recognized by @tryramp for being the #2 breakout growth company this month!
I wish I could share some elaborate strategy for how we got here, but it’s pretty simple: companies are looking for flexible compute platforms to run agents, GPUs, and inference
We’re by far the smallest company on this list, and I think it comes down to intensity and focus
We’ve had to say no to so many opportunities over the past few months in order to focus on our core business: building the best compute platform for the next generation of AI companies
And yes, we’re hiring!
I wish I could share some elaborate strategy for how we got here, but it’s pretty simple: companies are looking for flexible compute platforms to run agents, GPUs, and inference
We’re by far the smallest company on this list, and I think it comes down to intensity and focus
We’ve had to say no to so many opportunities over the past few months in order to focus on our core business: building the best compute platform for the next generation of AI companies
And yes, we’re hiring!
NEW: SPACEXAI is the fastest growing software company according to Ramp spend data.
I find the following:
- SpaceXAI's advantage in coding + its new consumer offering could accelerate business adoption of its models, much like what powered Anthropic's rise.
- Continued growth in ultiple winners in AI dictation (Wispr Flow and Granola both take a spot on our list).
5
5
37
2,660
shocked by how many people are selling $50M B300 clusters with zero space or power lined up
this isn’t the same as Nebius or CoreWeave signing a deal and using the prepay to place their order with Nvidia
I’m talking about people with zero experience in the industry - real estate guys, convenience store owners, crypto bros - signing 128 node B300 clusters and then racing to figure out how they’re going to deploy them
if everything goes according to plan, they’ll figure it out at the last minute. the networking gear gets delivered on time, the nodes go online, and everyone is happy
but it’s getting weird out there. every week there’s a new bottleneck in the supply chain putting any RFS date at risk, and people are underwriting large clusters that vastly exceed the cash on their balance sheet
one of the best buying opportunities for GPUs in the next 12 months is going to come from companies that defaulted on their loans or went bankrupt flying too close to the sun
this isn’t the same as Nebius or CoreWeave signing a deal and using the prepay to place their order with Nvidia
I’m talking about people with zero experience in the industry - real estate guys, convenience store owners, crypto bros - signing 128 node B300 clusters and then racing to figure out how they’re going to deploy them
if everything goes according to plan, they’ll figure it out at the last minute. the networking gear gets delivered on time, the nodes go online, and everyone is happy
but it’s getting weird out there. every week there’s a new bottleneck in the supply chain putting any RFS date at risk, and people are underwriting large clusters that vastly exceed the cash on their balance sheet
one of the best buying opportunities for GPUs in the next 12 months is going to come from companies that defaulted on their loans or went bankrupt flying too close to the sun
40
13
310
53,488
the timeline makes it seem like everyone is crushing it rn
this month alone I’ve seen like five companies I was 100% sure were killing it on X quietly shut down or get acquihired
just a reminder that everyone is dealing with their own shit and no one has any idea what’s really going on
this month alone I’ve seen like five companies I was 100% sure were killing it on X quietly shut down or get acquihired
just a reminder that everyone is dealing with their own shit and no one has any idea what’s really going on
4
1
34
2,230
people underestimate how much GPU networking and storage actually costs
networking easily adds an extra 10% to the cost of a cluster
external storage, like WEKA or VAST, is an extra $1M+ to setup - plus a monthly recurring fee
one of the many reasons compute markets are still early is because GPU clusters are nuanced
many people see a $5 B300 and think it's a fungible commodity, but these details are everything
networking easily adds an extra 10% to the cost of a cluster
external storage, like WEKA or VAST, is an extra $1M+ to setup - plus a monthly recurring fee
one of the many reasons compute markets are still early is because GPU clusters are nuanced
many people see a $5 B300 and think it's a fungible commodity, but these details are everything
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
1
1
40
2,672
Stop letting Claude deploy to your AWS account
Introducing beam.cloud stacks
> Deploy any GitHub or Dockerfile to the cloud
> Spin up Postgres, Redis, MongoDB
> 100% serverless and scale to zero
> Your credentials never enter the sandbox
You don’t need to waste tokens on Teour agent can host anything in seconds
Introducing beam.cloud stacks
> Deploy any GitHub or Dockerfile to the cloud
> Spin up Postgres, Redis, MongoDB
> 100% serverless and scale to zero
> Your credentials never enter the sandbox
You don’t need to waste tokens on Teour agent can host anything in seconds
1
0
15
1,555