Loading
Carregando...Loading
Carregando...In the datacenter and Cloud provider market, you've likely heard the term hyperscalers. But in recent months, another term has become common when discussing AI at scale: neoclouds. While both serve a similar need, they were born with fundamentally different approaches.
Hyperscalers are the generalist clouds we know, with a vast catalog of services ranging from compute and databases to queues, security, and observability. Their strength lies in integration, robust governance, and global presence, offering managed AI products like Amazon Bedrock, Azure OpenAI, and Vertex AI.
Neoclouds were born with a specific purpose: to be the home of artificial intelligence. Their focus is GPU-as-a-Service. Instead of an extensive catalog, they offer simplified access to H100s, A100s, or B200s clusters, with high-speed interconnection, transparent pricing, and shorter queues to train and serve models. Examples of neocloud companies include CoreWeave, Lambda, Crusoe, and Voltage Park.
The main difference is in the approach. While hyperscalers offer a complete platform with MLOps, compliance, and dozens of regions, neoclouds specialize in providing raw computational firepower, whether bare metal or VMs, with simple APIs and total focus on network performance (InfiniBand/RDMA).
Andrej Karpathy (formerly at OpenAI and Tesla) is one of the most prominent AI researchers and engineers today. He recently launched an educational experimental project creating an LLM with power similar to GPT-4. This project is called nanochat.
The nanochat project demonstrates the complete lifecycle of a minimalist LLM, from pre-training to functional chat. In the guide, Karpathy suggests using an 8xH100 machine "from your favorite provider." His chosen company was Lambda, a neocloud. He did a quick 4-hour training (using 8xH100 GPUs) at a cost of about $100, and later ran a larger 33-hour training, spending approximately $800.
This example perfectly illustrates the value proposition of neoclouds: quick and frictionless access to cutting-edge hardware, with predictable hourly costs, allowing developers to experiment, fine-tune, and publish prototypes with agility.
The choice depends directly on your project's architecture. A neocloud is the ideal choice if you need cutting-edge GPUs (H100s, B200s, you choose) with short queues and want total control over the environment for intensive training, without navigating a sea of additional services.
On the other hand, a hyperscaler may be more appropriate if your application depends on an integrated ecosystem of managed services (data, messaging, security) and requires corporate governance. If the goal is to use ready-made AI products like Bedrock or Vertex AI, which already come with compliance and guardrails, the hyperscaler is the natural path.
It's worth noting that these boundaries are becoming increasingly porous, with hyperscalers integrating neocloud capacity to meet demand peaks.
The modern AI strategy is pragmatic and often hybrid. It's common to see teams doing training and fine-tuning on a neocloud to leverage performance and cost, then serving the model on a hyperscaler, where data, observability, and main application integrations already reside.
The key is portability. Using containers and neutral tools like PyTorch, Ray, and Kubernetes allows moving workloads between clouds as GPU availability, price, and latency change, avoiding lock-in.
Neoclouds deliver raw GPU performance, focusing on availability and simplicity for those who need to train models. The nanochat example shows that with the right tool, it's possible to go from idea to LLM prototype in a few hoursâan innovation cycle that defines the AI market's pace today.