Building a Global Cloud for AI Inference

Blackfuel’s first AMD-powered inference deployment points to a more open, outcome-driven future for AI infrastructure

With more than $250 million in contracted revenue under multi-year customer agreements, Blackfuel is building an inference-native neocloud that brings dedicated infrastructure and proprietary software together as one system. Its ambition is to turn compute capacity into reliable, reservable AI output at global scale.

Its first inference-only deployment in Barcelona will combine AMD Instinct™ MI355X GPUs with AMD ROCm™ open software, alongside Dell, NTT DATA and Digital Realty. This deployment is the starting point for a broader platform designed to optimize inference across accelerator architectures and hardware generations.

Blackfuel’s emergence from stealth reflects a broader shift as AI moves from experimentation into production. The company calls its platform an AI Token Grid: a network of architecture-specific accelerator clusters connected through a common software and commercial layer.

Global scale with local control

The opportunity is not simply to add more compute. It is to build an inference infrastructure market that gives customers local control while remaining connected to global innovation. For customers, that starts with choice over deployment location, models and operating environments, backed by defined performance and service commitments.

For enterprises, model developers and AI-native companies, that means more ways to access production infrastructure without defaulting to a single provider or operating model. Blackfuel plans to offer dedicated GPU clusters, managed inference and reserved serving capacity. Customers will be able to retain control of their compute environments or use Blackfuel to manage the serving layer against agreed workload and performance requirements.

The company’s approach also puts deployment location and capacity commitments closer to the customer’s control. Its Barcelona deployment will be hosted by Digital Realty and built on liquid-cooled Dell PowerRack systems featuring Dell PowerEdge XE9785L servers powered by AMD Instinct MI355X GPUs, with deployment and integration support from Dell and NTT DATA.

These details matter because they show how an international AI ecosystem can come together around production infrastructure. Compute, data-center capacity, storage, integration, software and commercial models all have to operate as one system. Blackfuel’s deployment brings those pieces together while giving customers another option for accessing locally operated AI capacity.

AI infrastructure is shifting from capacity to output

The next phase of AI infrastructure will be measured less by installed hardware and more by useful output. For inference, the practical questions are increasingly about how reliably a platform can serve models, meet latency and throughput targets, manage cost and continue improving as software and hardware evolve.

Blackfuel’s AI Token Grid reflects that shift. It is designed to turn dedicated compute into measurable, reservable AI output. Through its planned Token Platform, customers will be able to select models and reserve inference capacity against defined throughput, latency and data-location requirements without managing the underlying infrastructure.

Blackfuel is also introducing Logical GPUs, service units that allow customers to reserve a share of a model’s serving capacity for a fixed monthly fee rather than rent an entire multi-GPU deployment. Commitments can be aligned to a selected model, token quotas and applicable service levels. The underlying physical hardware and serving configuration can evolve while the customer’s agreed capacity and service levels remain consistent.

This model illustrates a broader change in how AI infrastructure creates value. Customers do not ultimately buy GPUs. They buy intelligence, throughput, responsiveness and business outcomes. Infrastructure providers are therefore moving beyond access to accelerators and building services around the output customers need.

Open software is becoming infrastructure strategy

As infrastructure becomes more service-oriented, the software layer becomes even more important. It determines which models can run, how quickly new tools can be adopted, how workloads are optimized and how much freedom customers retain over time.

For its AMD technology-based infrastructure, Blackfuel is building on AMD ROCm™ open software and open serving technologies including vLLM and SGLang. Its Token Platform plans to support selected models from Meta, Mistral, H Company, DeepSeek, Qwen, GLM, Moonshot AI’s Kimi, OpenAI (gpt-oss), MiniMax, Cohere, Reflection AI, NVIDIA (Nemotron), Liquid AI and Google (Gemma) at launch[JB4] , with coding assistants and agentic applications identified as priority workloads.

This is where an open software foundation creates strategic value. AI is evolving too quickly for infrastructure to be constrained to a closed set of models, frameworks or deployment choices. The AMD approach is grounded in open software and open standards, giving developers and operators greater flexibility to work with the models, frameworks and operational practices they use today while adapting as the market changes.

Open does not mean unoptimized. Blackfuel is combining open serving technologies with proprietary routing, allocation, metering and commercial-control software. The company is also developing an agent-driven process to build, benchmark and continuously optimize inference runtimes for each model, accelerator architecture and workload. Its goal is to lower serving costs and increase useful tokens per megawatt within defined quality and latency requirements.

Blackfuel as evidence of what comes next

The Barcelona deployment is important not only because it is designed to add new AMD-powered inference capacity in Europe. It demonstrates how a new generation of neocloud providers can combine open technology, locally operated infrastructure and flexible consumption models to turn compute into differentiated AI services.

Blackfuel is designing its platform to support multiple accelerator architectures and hardware generations, with optimized execution environments connected through a common service and commercial layer. That approach recognizes that infrastructure will continue to evolve. Customers need platforms that can optimize for today’s workloads without limiting tomorrow’s choices.

For AMD, this deployment shows how AMD Instinct GPUs and the ROCm open software stack can support providers building for production inference at scale. It also reinforces a larger point: openness is not simply a software philosophy. It is an infrastructure strategy that can expand customer choice, accelerate innovation and help providers adapt their services as models and workloads change.

Building the future of inference with choice

The industry has an opportunity to shape the next era of inference around openness, local control and access to production-ready capacity. Realizing that opportunity will require more than building clusters. It will require an ecosystem that can translate those clusters into useful AI output while giving customers meaningful choices across models, software, infrastructure locations and commercial approaches.

Blackfuel’s planned deployment in Barcelona illustrates that future. It shows what becomes possible when infrastructure, open software and a new service model come together around the needs of production AI. Blackfuel’s ambition is global: an AI Token Grid designed to make inference capacity accessible at scale while preserving local control and customer choice.

Learn more: Blackfuel: Launch announcement