Pre-AI GPU Cloud (2010s)
Before the AI boom, GPU cloud existed but was fragmented. Hardware shops (Lambda), crypto miners (CoreWeave-as-Atlantic-Crypto), gaming-rental services, academic supercomputing — different industries with different incentives, all using GPUs, none integrated into a coherent category.
Hardware shops
Lambda Labs is the canonical pre-AI hardware shop. Founded 2012 selling pre-configured workstations and servers to AI researchers. There were others — SuperMicro variants, custom builders, smaller specialty vendors.
The hardware shop business sold capability rather than capacity. Customers bought the machine and ran it themselves.
GPU mining
Through the 2010s, cryptocurrency mining was a major GPU demand source. Bitcoin mining moved to ASICs by mid-decade, but Ethereum and other proof-of-work coins continued to use GPUs through 2022.
The mining ecosystem built:
- Operational expertise running GPU farms at scale.
- Datacenter infrastructure designed for GPU power density.
- Capital structures and financing relationships.
- Software stacks for distributed GPU operations.
Many neoclouds (CoreWeave, Crusoe, others) trace their operational DNA directly to crypto mining of this era.
Early cloud GPU offerings
The hyperscalers offered GPU instances starting in the early 2010s — AWS launched P-series GPU instances, Azure and GCP followed. These early offerings were expensive, limited in selection, and primarily used for visualization and scientific computing rather than ML.
Specialty providers — Vast (founded 2016), Paperspace, others — offered alternatives at lower prices, targeting researchers and indie users who couldn't afford hyperscaler rates.
Academic computing
Academic GPU computing existed at scale but mostly as institutional infrastructure — Oak Ridge's Titan, Argonne's Summit, university research clusters. These weren't customer-facing clouds; they were funded research compute.
The fragmented pre-history
The 2010s GPU-cloud landscape was fragmented. Each subspecialty (academic, mining, hardware, hyperscaler GPU instances) operated semi-independently. The strategic category called "neocloud" didn't yet exist because the unifying demand profile — AI training and inference at scale — hadn't yet emerged.
Takeaway
The pre-history shows the pieces that would assemble into the neocloud category. The catalyst came in 2022-2023. The next chapter examines that inflection.