Beam Cloud

What Beam Cloud actually is

Beam Cloud (beam.cloud, operated by Smartshare, Inc.) is a serverless compute platform built for AI workloads: inference endpoints, background jobs, agent sandboxes, and task queues, all deployed from Python or TypeScript with decorators rather than YAML. You define a container image and a function; Beam boots the container near your code, attaches GPUs if you asked for one, and bills by the millisecond while it runs. When traffic stops, capacity scales to zero and billing stops with it.

The company was founded in 2021 by Eli Mernit and Luke Lombardi and first shipped as Slai, an ML deployment tool; the current Beam product launched publicly as an open-source alternative to Modal, and it is Y Combinator-backed with Tiger Global, Charge Ventures, Hustle Fund, Soma Capital, and Alumni Ventures on the cap table. The part we care about most at ServerSpotter is architectural: the runtime, beta9, is open source under AGPL-3.0 and can be self-hosted or run in Bring Your Own Cloud mode on your own AWS, GCP, or Azure account. That makes Beam one of the few serverless GPU platforms where the vendor's retail price is a ceiling rather than a floor — if you hold committed-use discounts or cloud credits, the same API runs on infrastructure you already pay for.

The primitive set covers four shapes of work: web endpoints (HTTP-served model APIs), task queues (async jobs), sandboxes (isolated environments for untrusted or LLM-generated code, with gVisor isolation and optional Docker-in-Docker), and persistent Linux VMs for anything that does not fit a function. Storage volumes hold model weights and data, and memory/file-system snapshots plus GPU checkpoint restore are what make the cold-start story credible: a cached container with a restored GPU state comes back without a full boot.

Pricing: what you really pay

Beam's pricing has three layers, and mixing them up is the main way people misread their invoice. First, the subscription tiers: a Developer tier at $0 per month plus usage, and a Team tier at $89 per month plus usage. The Developer tier is not a toy trial — it includes $30 of compute credits every month, one seat, and concurrency for up to 5 GPU containers and 30 CPU containers. Team raises that to 50 GPU and 1,000 CPU containers, three seats (extra seats are $25 each), and longer log retention. There is no paid tier you must buy to run production work, which we appreciate; compare that with sandbox competitors that gate 24-hour sessions behind a monthly floor.

Second, usage pricing for serverless and sandbox work, metered by the second while code runs. The live pricing page shows serverless GPU tasks at $0.000192 per second for an RTX 4090, $0.000972 per second for an H100 PCIE, and $0.000303 per second for an RTX 5090. Sandbox CPU is billed at $0.0000375 per second per physical core (a physical core is a 2-vCPU equivalent), with RAM metered separately. Beam's own worked example: an endpoint on an RTX 4090 with 2 physical cores and 16 GiB RAM costs $1.77 per hour — $0.69 for the GPU, $0.76 for CPU, $0.32 for RAM — billed by the millisecond and only while the code is running.

Third, on-demand GPU containers billed per hour with published "from" rates: B200 SXM6 from $4.11, H200 SXM5 from $2.09, H100 PCIE from $1.83, A100 80GB SXM4 from $1.36, RTX PRO 6000 from $1.09, L40S from $0.76, RTX 5090 from $0.72, A6000 from $0.54, and RTX 4090 from $0.44 per hour. Those on-demand numbers are the ones worth comparing against the market: the H100 rate is well under half of what Modal's effective per-hour rate is by our reading of both pricing pages, and multi-GPU clusters (H100, H200, B200, B300 SXM, 8-GPU nodes with InfiniBand) are available when you outgrow single cards. Persistent VMs bill per second while running with a 50 GiB root disk included — the pricing calculator shows a 1 vCPU/1 GiB configuration at $0.0306 per hour ($4.90 per month at 160 hours of use).

Finally, Bring Your Own Cloud: Beam manages the orchestration and control plane on instances in your cloud account, you pay your cloud provider directly for hardware (or burn credits against it), and Beam charges a flat management fee of $0.019 per hour per vCPU and $0.009 per hour per GB of RAM. Storage volumes are included up to 1 TB; past that it is $0.021 per GB per month. One caveat on all of these numbers: Beam publishes "from" rates on the GPU line and the live page has moved prices at least once — an older comparison on Beam's own blog quotes H100 at $1.74 per hour where the current page says from $1.83 — so treat any single quote as a snapshot and confirm at order time.

Performance & infrastructure

The engineering claim that matters here is boot time. Beam's pitch is sub-second cold starts for cached images, achieved with a custom container runtime, a distributed storage layer that avoids re-downloading weights, memory snapshots, and GPU checkpoint restore — the ability to stop a sandbox and resume it with running processes intact. The vendor advertises restoration up to 35x faster than a traditional cold boot. We have not benchmarked Beam ourselves (our benchmark runs cover VPS plans, not serverless platforms), so treat those figures as vendor claims; the architecture behind them is at least the right shape for the problem, and the isolation model (gVisor, non-root containers) is the same family Modal uses.

On capacity, the homepage publishes live utilization across three pools — workloads on customers' AWS accounts, GCP accounts, and Beam's own bursting capacity — which fits the BYOC design: your steady load can sit on your reserved instances while bursts spill onto Beam's fleet. The managed service's physical footprint is not publicly documented the way a VPS provider's is; where your containers run depends on the pool, and in BYOC mode it depends on your cloud regions. For teams with data-residency constraints, that BYOC mode is the answer: nothing leaves your VPC because the compute never leaves your account.

The ecosystem is thinner than a hyperscaler's but wider than most seed-stage infra startups: the beta9 runtime is on GitHub, the SDK is Python- and TypeScript-first, and named customers include Coca-Cola, Magellan AI, Shippabo, and Stratum according to the company's own launch materials. A public status page is linked from every page. What is missing on the reliability side is any published uptime SLA or independently audited availability record; the company is small and young, and that is a real consideration when an inference endpoint sits in your product path.

Where it falls short

Beam is a young product from a small team. That cuts both ways — pricing is aggressive and the founders answer tickets — but there is no SLA with credits, no long incident history to inspect, and no enterprise procurement track record. If your use case demands contractual availability guarantees, you are signing something custom or choosing a larger vendor.

Second, pricing complexity. Serverless per-second rates, on-demand hourly rates, sandbox per-core-second rates, BYOC management fees, and subscription tiers is a lot of surface area, and the per-core billing unit (a "core" is 2 vCPU) makes naive comparisons against vCPU-priced platforms wrong. For steady CPU-only workloads, serverless pricing is the wrong model and Beam says as much; the on-demand line exists for that reason.

Third, cold-start behavior is workload-dependent. The sub-second numbers apply to cached images and warm snapshots; the first invocation of a large uncached image still pays for pull and build time. The DB row we inherited claims cold starts "may exceed 5 seconds on first invocation" — we found no published measurement either way, so we treat both that figure and the 35x claim as unverified. Model-weight downloads on first run are the usual culprit regardless of provider.

Fourth, scope: this is a container platform, not a VPS provider. No per-instance SSH on the serverless line, no arbitrary kernel tuning, and the VM product is young. And regional choice on the managed fleet is limited compared with a global cloud; BYOC is the escape hatch, but it assumes you already run AWS, GCP, or Azure.

Who should pick this

Our pick for Beam is a team shipping AI features — inference APIs, agent sandboxes, batch jobs, fine-tuning or evaluation pipelines — with spiky or growing traffic and no desire to manage GPU fleets. Three profiles fit especially well. First, startups that want GPU economics without a GPU contract: the free Developer tier's $30 monthly credits cover real prototyping, and the RTX 4090 and A100 on-demand rates are the cheapest we have seen published for that hardware class. Second, agent platforms that execute untrusted code: the sandbox primitive with GPU support is a combination most competitors either lack or charge a multiple for. Third, cost-conscious scale-ups with AWS or GCP commitments: BYOC turns Beam's management fee into a thin layer over compute you already bought at discount, which is exactly where serverless pricing normally hurts.

If you need bare-metal root access, strict data residency without a cloud account, or a contractual SLA, look at the GPU cloud and serverless platforms categories for more established options. And if your workload is a plain website or always-on API, ordinary VPS plans will beat any per-second platform on cost.

Key takeaways

• Serverless GPU/CPU platform with millisecond billing, scale-to-zero, sandboxes, task queues, and persistent VMs — deployed from Python or TypeScript.
• Free Developer tier includes $30 of compute credits monthly; Team is $89/month plus usage; there is no mandatory production subscription.
• GPU pricing is the attraction: on-demand H100 PCIE from $1.83/hr, A100 80GB from $1.36/hr, RTX 4090 from $0.44/hr; serverless RTX 4090 at $0.000192/sec.
• The beta9 runtime is open source (AGPL-3.0) and self-hostable; BYOC runs the same API on your AWS/GCP/Azure account for a $0.019/vCPU-hr management fee.
• Young company, no published uptime SLA, and pricing with several billing units — verify current rates and your own cold-start behavior before committing.
• Best fit for AI inference, agent sandboxes, and spiky GPU workloads; wrong fit for always-on sites and bare-metal needs.

Sources

• https://beam.cloud/pricing
• http://beam.cloud/
• https://beam.cloud/about
• https://platform.tracxn.com/a/d/company/58ffb6b9e4b07ccdc564ac7c/beam
• https://www.ycombinator.com/launches/OuS-beam-open-source-modal-alternative
• https://apis.io/plans/beam-cloud/beam-cloud-plans-pricing
• https://beam.cloud/blog/modal-pricing-explained
• https://beam.cloud/blog/e2b-pricing-explained