FluidStack is raising a $1B round at an $18B valuation, months after being valued at $7.5B. The company aggregates underutilised GPU capacity from data centers globally and provides it to AI workloads at prices that undercut AWS and GCP. That valuation trajectory reflects a real demand signal: engineering teams building serious AI products are finding that hyperscale cloud GPU pricing does not work at the scale their products require. The emergence of specialised AI compute providers — FluidStack, CoreWeave, Lambda Labs, and others — creates a genuine infrastructure decision that most engineering teams have not had to make before. This article frames that decision clearly, with a practical framework for when specialised AI compute makes sense versus when you should stay on the hyperscalers.
What We'll Cover
Why Alternative AI Compute Is Emerging Now
Three structural factors are driving the growth of specialised AI compute providers simultaneously. First, GPU supply has grown faster than hyperscaler data center capacity can absorb, creating a market for providers that can aggregate distributed GPU capacity outside traditional cloud facilities. Second, AI training and inference workloads have very different compute profiles from general-purpose cloud workloads: high GPU utilisation, long-running jobs, predictable batch patterns, and specific memory bandwidth requirements. Hyperscalers optimised their platforms for variable, latency-sensitive, multi-tenant compute — not for the sustained, GPU-heavy, predictable patterns of AI. Specialised providers are purpose-built for AI's compute profile. Third, price competition is intensifying. AWS and GCP GPU instance pricing reflects their infrastructure investment cost, support overhead, and margin requirements. Specialised providers with lower operational overhead can offer equivalent hardware at 40-70% lower cost for the right workload types.
Hyperscale Cloud vs Specialised AI Compute
The choice is not binary — most sophisticated AI engineering teams use both, routing workloads to the most appropriate provider. Understanding the genuine advantages of each is the prerequisite for making good routing decisions:
- Stay on hyperscale for: workloads with variable demand that benefit from auto-scaling; workloads needing deep integration with cloud-native services; regulated workloads requiring enterprise compliance certifications; real-time inference APIs where SLA guarantees matter; and workloads where your team lacks the operational capacity to manage a multi-provider infrastructure
- Evaluate specialised providers for: sustained high-volume AI training runs; batch inference jobs with predictable load patterns and latency tolerance; fine-tuning runs on large models where cost per GPU-hour is the primary variable; and situations where hyperscale GPU availability is a bottleneck
The Infrastructure Decision Tree for AI Workloads
When evaluating whether to use a specialised AI compute provider, work through these questions in order:
- Is the workload latency-sensitive? — Real-time inference APIs serving user requests need hyperscale SLAs and global edge presence. Batch workloads do not. Route latency-sensitive workloads to hyperscale.
- Does the workload require cloud-native service integration? — Training pipelines deeply integrated with S3, RDS, or Lambda need to rebuild those integrations when migrating. Calculate the migration cost before comparing prices.
- What is the sustained GPU utilisation rate? — Specialised providers offer the best economics for workloads with high, sustained GPU utilisation. If your workload runs GPUs at 20-30% utilisation, solve the utilisation problem first (batching, scheduling, model serving optimisation) before provider selection matters.
- How much operational overhead can your team absorb? — Specialised providers offer less managed infrastructure. You may need to manage your own Kubernetes, monitoring, and networking. This real cost must be weighed against per-GPU-hour savings.
What This Means for Engineering Teams
The practical implication is that AI infrastructure decisions now require the same rigour as architectural decisions. Defaulting to AWS for all AI workloads leaves money on the table; moving to specialised providers without proper evaluation creates operational risk. Engineering teams need people who understand both the AI workload characteristics and the infrastructure trade-offs well enough to make these decisions confidently. Our Cloud & DevOps engineers have experience designing multi-provider AI infrastructure, including cost modeling, portability architecture, and operational tooling. If you are evaluating whether a specialised AI compute provider makes sense for your workloads, a technology roadmap engagement will give you a clear answer before you commit to any infrastructure change.
Frequently Asked Questions
What is FluidStack and how does it work?
FluidStack aggregates underutilised GPU capacity from data centers globally and provides it to AI workloads through a unified API. Rather than building its own data centers, it acts as an infrastructure broker — matching available GPU capacity with AI workload demand. This model allows pricing significantly below hyperscale cloud GPU rates for compatible workloads.
How much cheaper is specialised AI compute compared to AWS?
For equivalent hardware (H100, A100), specialised providers typically offer 40-70% lower per-GPU-hour pricing than hyperscale cloud for sustained workloads. The actual saving depends on your utilisation rate, workload duration, and how much managed infrastructure you need. Short-duration, low-utilisation workloads may see less benefit.
What are the reliability risks of specialised AI compute providers?
Specialised providers generally have fewer redundancy guarantees, less mature SLA commitments, and smaller support organisations than hyperscalers. For training workloads that can tolerate interruption and restart from checkpoints, this risk is manageable. For production inference APIs serving user requests, the reliability risk usually outweighs the cost savings.
How do you migrate AI training workloads from AWS to a specialised provider?
Containerise your training code, switch from S3 to an S3-compatible object store (most specialised providers support this), set up your own monitoring and alerting, and implement checkpoint-based fault tolerance. For most ML training code, the actual compute migration is straightforward — the integration decoupling is the effort.
Should startups use specialised AI compute from the beginning?
Not necessarily. Early-stage teams benefit from hyperscale cloud's managed services and lower operational overhead. As your AI workload matures and compute costs become a meaningful line item, evaluating specialised providers for specific workload types makes sense. The crossover point is typically when AI compute cost exceeds $10-20K/month and you have a team member capable of managing the operational overhead.