Cast.ai vs Together AI: One Rents You GPUs, the Other Makes Yours Cheaper

Cast.ai reached a $1 billion valuation on 12 January 2026 by convincing companies to use less of the GPUs they already own. Together AI raised $800 million five months later, on 1 July 2026, on the opposite pitch: stop owning GPUs at all and rent inference by the token instead.

Both companies sit inside the same category on HokAI's own directory, generative AI infrastructure, and that is probably why you are reading a piece comparing them. But the pairing is close to a category error dressed up as a decision. Cast.ai is built for a team that already runs its own Kubernetes cluster and wants the GPU line item inside it to stop leaking money. Together AI is built for a team that would rather never provision a GPU at all.

Picking between them is really a question about whether you want to own compute or rent it. Answering it does not rule the other one out later.

Pick Together AI if you don't want to run infrastructure. Pick Cast.ai if you already do.

If your team has never provisioned a GPU node and has no plans to start, Together AI removes the decision entirely. Serverless inference on its published rate card runs from free for a small model like Ternary Bonsai 27B up to $15 per million output tokens for a frontier model like Kimi K3, with an on-demand H100 priced at $3.99 an hour, an H200 at $5.99 an hour and a B200 at $8.19 an hour.

Fine-tuning starts at $0.48 per million training tokens for models up to 16 billion parameters, rising to $1.50 for the 17 to 69 billion range and $2.90 above 70 billion, according to Together AI's own pricing page. You pay for what you use. Together AI absorbs whatever sits idle between your requests.

If your team already runs GPUs across AWS, GCP or Azure for training or inference and the bill keeps climbing anyway, Cast.ai does not replace that fleet. It automates the rightsizing, spot-instance scheduling and cross-cloud sourcing that keeps a cluster you already pay for from wasting the money you already spent.

The two are not mutually exclusive. A company can rent inference from Together AI for its product and run Cast.ai on the Kubernetes cluster behind everything else it operates.

The number that explains why Cast.ai exists at all

Average GPU utilization across Kubernetes clusters sits at 5%, according to Cast.ai's own 2026 State of Kubernetes Optimization report, broken down as 2% on Azure, 5% on AWS and 6% on Google Cloud. That means roughly 95% of provisioned GPU capacity is paid for and never touched by a running job.

That single number is the entire business case for a cost-automation layer. It is also why comparing Cast.ai to a rental service like Together AI misreads the product. Together AI's per-token and per-GPU-hour pricing already solves the idle-capacity problem, because idle capacity is Together AI's cost to absorb, not yours. Cast.ai solves the same problem for GPUs a team has already committed to owning.

What Cast.ai's GPU product actually does

Cast.ai markets its GPU offering as OMNI, which raises utilization through three mechanisms: time-slicing that runs 1 to 48 replicas on a single GPU depending on workload density, MIG partitioning that physically isolates slices of an A100, A30 or H100 into dedicated instances, and cross-cloud sourcing that finds capacity in a different region or provider when the primary one runs short.

Case studies on Cast.ai's own site cite Akamai reporting 40 to 70% in cloud savings and Yotpo cutting cloud costs by 40%, though neither figure is broken out as GPU-specific.

Cast.ai does not publish tiered pricing. Its pricing page states plainly that cost "depends on a few factors specific to your environment" and routes every prospect to a sales conversation. Third-party estimates put a starting Growth plan near $1,000 a month plus $5 per CPU, but that number comes from resellers, not from Cast.ai itself. Treat it as a starting point for a quote, not a real price.

The $1 billion valuation came from a January 2026 investment led by Pacific Alliance Ventures, the US venture arm of Korea's Shinsegae Group, on top of an earlier Series C led by G2 Venture Partners and SoftBank Vision Fund 2. That is a smaller raise than Together AI's, and it says something about the two markets: cost automation is a real business, but renting the actual GPUs commands the bigger check.

If a team is already committed to owning the infrastructure, the storage layer underneath the GPU workloads is a separate decision worth making deliberately. Archil mounts S3, GCS and other object stores as a real POSIX filesystem for agent and ML workloads, with a Kubernetes CSI driver for teams provisioning volumes the same way they provision compute.

What Together AI's rental model actually covers

Together AI's homepage currently advertises 2x faster inference and 60% lower cost through workload-specific optimization, plus a claimed 90% faster pre-training through its Kernel Collection, as verified live in September 2026. Customer references on the same page include Decagon, which reports a 6x cost reduction per conversation turn, and Vercept, which reports 11x faster inference. Together AI also recently launched a dedicated GPU cluster for Y Combinator startups that removes the traditional two-year reservation contract.

The $800 million Series C was led by Aramco Ventures, with Vista Equity Partners, General Catalyst and Nvidia among the other participants, and it lifted Together AI's valuation from $3.3 billion in early 2025 to $8.3 billion. HokAI covered the separate story of IBM's $240 million investment in Together AI in an earlier guide. The two rounds are not the same event.

Renting inference still means picking which of Together AI's 200-plus models to run, and price varies enormously by model size and provider. HokAI's model leaderboard ranks every GA model by blended price alongside benchmark scores, which is a faster way to make that call than reading a rate card model by model.

Not every team needs that much choice. Google AI Studio is the simpler version of renting: one vendor, Gemini models only, pay-as-you-go per token, no cluster to configure and no decision about which of 200 models to pick.

Where each one wins

Together AI wins for a team building a product feature on top of an open model where usage is spiky, unpredictable or still small. There is no cluster to provision, no idle capacity to manage and no commitment past this month's invoice.

Cast.ai wins for a team that has already made the infrastructure decision, usually for data residency, cost at scale, or a model that has to stay on hardware the company controls. In that world the GPU is a sunk cost the moment it is provisioned. The only lever left is utilization.

Neither wins for a team that has not yet decided whether it wants to own infrastructure at all. That decision comes first, and it is closer to a HokAI Smart Match question than a pricing comparison: run your constraints through it before pricing either vendor.

The obvious objection

The obvious objection is that this whole comparison is a category error: one company sells infrastructure automation, the other sells managed compute, and no buyer genuinely chooses between a FinOps tool and an inference API. That is true as a matter of product category.

It is also beside the point. HokAI's own traffic pairs these two names constantly, which means real people are typing this exact comparison into a search box without knowing yet that they are asking two different questions. Answering both at once, honestly, is the only way to actually help that reader.

For proof this pattern is not limited to language models, Tamarind Bio rents the same kind of managed compute for computational biology: over 200 structure-prediction and molecular-design models behind a no-code interface, with GPU provisioning handled entirely behind the scenes.

What switching would cost

Moving away from Together AI mostly means re-pointing API calls and re-testing latency against whatever replaces it. There is no infrastructure to unwind, because there was never any to build. Moving away from Cast.ai is a bigger commitment in the other direction. Its rightsizing and spot-scheduling policies are wired into the cluster's autoscaler, so leaving means either replacing that automation with an equivalent tool or going back to manual node management. Either path means someone re-learns why the utilization number was 5% in the first place.

Both companies are chasing the same 2026 shortage economics from opposite ends. Cast.ai squeezes more work out of GPUs that already exist. Together AI resells GPU time nobody else has bought yet. Expect more infrastructure products to blur that line rather than clarify it, not fewer.

Frequently asked questions

Are Cast.ai and Together AI direct competitors?

No. Cast.ai automates cost and utilization for Kubernetes GPU infrastructure a company already runs. Together AI rents managed GPU inference and fine-tuning by the token or GPU-hour, so its customers never provision hardware at all. A company can use both at once for different parts of its stack.

How much does Cast.ai cost?

Cast.ai does not publish tiered pricing and quotes each customer individually based on cluster size and cloud footprint. Third-party estimates put a starting Growth plan near $1,000 a month plus $5 per CPU, but that figure comes from resellers rather than Cast.ai own pricing page.

How much does Together AI cost?

Together AI serverless inference ranges from free for smaller open models to $15 per million output tokens for a frontier model like Kimi K3, according to its published rate card. On-demand GPU clusters run $3.99 an hour for an H100 up to $8.19 an hour for a B200, and fine-tuning starts at $0.48 per million training tokens for models up to 16 billion parameters.

Why does Cast.ai GPU optimization matter if a company already pays for its GPUs?

Because average GPU utilization across Kubernetes clusters is only 5 percent, according to Cast.ai own 2026 State of Kubernetes Optimization report, meaning roughly 95 percent of provisioned capacity goes unused. Cast.ai rightsizing, spot-scheduling and cross-cloud sourcing target that gap directly.

When should a team pick Together AI instead of running its own GPUs with Cast.ai?

When the team has not yet provisioned any GPU infrastructure and does not want to start, usage is spiky or still small, or the priority is shipping a feature rather than operating a cluster. Teams that already run GPUs for compliance or scale reasons are the ones Cast.ai is built for instead.

Covered in this guide

• Cast.ai: Cast.ai cuts Kubernetes cloud costs 50-70% using automated rightsizing and Spot instance management. Trusted by 2,100+ companies, reached $1B unicorn in 2026.
• Together AI: The AI Native Cloud: a full-stack platform for training, fine-tuning, and deploying open-source AI models
• Archil: This hosted file system turns S3 buckets into a POSIX disk with 2 vCPU serverless AI agent sandboxes, billed by the minute, free to start with no credit card.
• Google AI Studio: Free web-based IDE for building, testing, and deploying generative AI applications with Google's Gemini models
• Tamarind Bio: Runs AlphaFold, RFdiffusion, and other published biology models through one web app and API, used by 10,000+ scientists at biotechs and pharma companies.

Sources