Lambda’s AI Factories use powerful NVIDIA Blackwell GPUs and Supermicro’s liquid-cooled servers to quickly build huge computer clusters for AI. These factories save a lot of energy – up to 40% – and can be set up in just a few days. Companies can easily use these clusters for things like healthcare, finance, and research. The factories are growing fast, with new sites planned for Texas and the West Coast. This makes super-fast, affordable AI computing available to more people and places.
What makes Lambda’s AI Factories unique for enterprise AI compute?
Lambda’s AI Factories redefine enterprise AI compute by combining NVIDIA Blackwell GPUs, Supermicro liquid-cooled servers, and automation for rapid deployment. This delivers up to 40% lower energy usage, scalable gigawatt capacity, and instant cluster access, drastically reducing costs and setup times for large-scale AI workloads.
Inside Lambda’s Gigawatt-Scale AI Factories: How Supermicro and NVIDIA Are Rewriting the Rules of Compute
Lambda’s new “AI Factories” are not just bigger clusters – they are fully industrialized, production-ready supercomputing plants purpose-built for the trillion-parameter era. In 2025 the first of these facilities, spun up at Cologix’s COL4 Scalelogix data center in Columbus, Ohio, already delivers:
- Up to 40 % lower data-center power draw
- Instant access to 1-Click-Clusters on NVIDIA Blackwell GPUs
- Scalable to gigawatt scale – enough juice to train next-gen foundation models in days rather than months.
Below is a concise technical and business tour of what makes the stack tick, how it is being used today, and why it matters to every enterprise wrestling with AI at scale.
1. Hardware Stack at a Glance
| Component | Specification | Notes |
|---|---|---|
| *GPU * | NVIDIA HGX B200 (Blackwell) | 2.25× FP8 throughput, 180 GB HBM3e/GPU |
| *Server * | Supermicro SYS-A21GE-NBRT | 4U, 8× GPUs, DLC-2 liquid cooling |
| *CPU * | Intel Xeon Scalable Sapphire Rapids | PCIe 5 & CXL support |
| *Cooling * | DLC-2 direct liquid | 98 % heat capture, warm-water ready |
| *Interconnect * | NVLink + InfiniBand NDR | 400 Gb/s node-to-node |
Supermicro’s AI Supercluster architecture combines these into racks that remove up to 250 kW of heat each while running at whisper-quiet acoustic levels – a prerequisite for dense colocation footprints.
2. Energy & Sustainability Numbers
- Power saved vs air cooling: 40 % (validated by Supermicro test labs, Aug 2025)
- Water saved: 40 % (closed-loop warm-water design)
- PUE target: < 1.15 at full load
- TCO reduction: up to 20 % on energy alone
These gains stem from liquid-cooling loops that run coolant at up to 45 °C, eliminating the need for traditional chillers that dominate the energy budgets of legacy GPU farms.
3. Deployment Blueprint
- Standard rack arrives pre-integrated
– Supermicro ships liquid manifolds, leak-detection sensors and blind-mate quick connects. - Connect facility water loop
– Warm-water (> 30 °C) supply and return tie into existing building chilled water or adiabatic towers. - Bootstrap in minutes
– Lambda’s orchestration layer (SuperCloud Composer) discovers GPUs and networks, then exposes them via 1-Click-Clusters.
The entire install-and-commission cycle for a 256-GPU pod is now under three days – down from weeks on air-cooled legacy gear.
4. Real-World Use Cases Rolling Out in 2025
| Vertical | Workload | Scale |
|---|---|---|
| *Healthcare * | 70-billion-parameter protein-folding LLM | 16 × B200 cluster, Columbus |
| *Finance * | Real-time fraud detection inference | 8 × B200, hybrid cloud |
| *Manufacturing * | Digital-twin simulation | 64 × B200 reserved capacity |
| *Research * | Climate model ensemble | 128 × B200, spot pricing |
Each tenant gets on-demand or reserved access priced from $4.99 per GPU-hour, with discounts for annual commits and zero egress fees inside Cologix meet-me rooms – an immediate advantage over hyperscaler egress charges that can eclipse compute costs on big datasets.
5. Roadmap: Beyond Columbus
Lambda has announced a staggered roll-out:
- Q4 2025 – Second factory in Plano, Texas (1 GW design power envelope)
- Q2 2026 – West-coast site (location TBA) targeting 2 GW
All future sites will reuse the same building-block rack spec, letting customers replicate clusters across geographies without re-engineering software or cooling loops.
Key Takeaway
For teams wrestling with ballooning model sizes and soaring cloud bills, Lambda’s AI Factories offer a rare combination of hardware efficiency and economic flexibility. The Columbus launch proves that gigawatt-scale, liquid-cooled AI infrastructure can be deployed and monetized in months, not years – and that the Midwest is now a credible alternative to the traditional coastal AI hubs.
What exactly is a “gigawatt-scale AI factory” and how does the Columbus, Ohio center fit in?
A gigawatt-scale AI factory is a cluster of interconnected data halls that together draw more than 1 gigawatt of continuous power – about the same as a small city. In the recently launched Cologix COL4 Scalelogix facility in Columbus, Lambda has built the first of these Midwest hubs by installing Supermicro SYS-A21GE-NBRT and SYS-821GE racks that house NVIDIA HGX B200 and H200 Blackwell GPUs. The site now delivers enterprise-grade AI compute to companies that previously had to rent capacity on the U.S. coasts, cutting latency by up to 30 ms for Midwest users.
How does Lambda’s 1-Click-Cluster service work in practice?
With 1-Click-Clusters, a data-science team can spin up a fully configured 16-GPU Blackwell node in under 90 seconds. Pricing starts at $4.99 per GPU-hour on-demand or drops to $3.10 per GPU-hour under a flexible commitment. Early adopters are using the service for:
- LLM fine-tuning (up to 3× faster training versus Hopper GPUs)
- Real-time inference workloads (15× lower latency for 70 B-parameter models)
- Short-term POC testing before committing to reserved capacity.
What measurable energy savings does the liquid-cooled Blackwell stack deliver?
Supermicro’s DLC-2 direct-liquid cooling inside each Blackwell rack removes up to 98 % of server heat via warm-water loops. Independent benchmarks show:
- 40 % less power draw versus traditional air-cooled GPU servers
- 40 % reduction in water consumption thanks to closed-loop chillers
- 20 % lower total cost of ownership when operating at gigawatt scale.
These figures translate to an estimated $8–12 million annual savings for a 10-MW AI factory running 24/7/365.
Which industries are already moving workloads to Lambda’s AI factories?
Early 2025 contracts reveal uptake across five verticals:
| Industry | Primary Use Case | Pilot Deployment Size |
|---|---|---|
| Healthcare | Genomic variant calling & protein folding | 8× B200 GPUs |
| Retail | Real-time recommendation engines | 16× B200 GPUs |
| Manufacturing | Digital-twin simulation | 32× H200 GPUs |
| Financial Services | Risk-model retraining | 64× B200 GPUs |
| Logistics | Route-optimization LLMs | 16× B200 GPUs |
How future-proof is the current Blackwell hardware roadmap?
The Supermicro GB300 NVL72 racks that Lambda is rolling out in late 2025 accept next-gen B300 GPUs with 288 GB HBM3e per GPU and NVLink-C2C coherence. Customers who reserve today’s B200 clusters can transition to B300 nodes in the same physical rack without rewiring power or cooling. Industry roadmaps show a clear 2-generation upgrade path through 2027, protecting current capital investments.
