칩제조사들이 단백질 폴딩 및 AI 추론 모델을 관리형 마이크로서비스로 제공하여 깊은 ML 인프라 전문성이 부족한 기업들의 특화된 워크로드 배포를 단순화하고 있다.

For most of the last decade, predicting the three-dimensional shape of a protein complex was a capability only well-funded research organizations could deploy. That landscape is shifting rapidly. Open-source structure prediction models are now freely licensed for commercial use, and the largest chipmakers are packaging them as ready-to-deploy inference microservices available through cloud marketplaces. The Business Research Company projects the AI protein structure prediction market will grow from approximately $2.33 billion in 2026 to roughly $6.62 billion by 2030, representing a compound annual growth rate of approximately 29.8 percent. As the underlying models become easier to access, the competitive question for drug discovery is shifting to what capabilities companies add on top.

Structure prediction itself is rapidly standardizing. OpenFold3, an open-source biomolecular structure prediction model developed by the AlQuraishi Lab at Columbia University and the OpenFold Consortium, is available under the Apache 2.0 license for academic and commercial use. Unlike AlphaFold3, which carries restrictions on commercial deployment, OpenFold3 permits unrestricted commercial application. In March 2026, the Consortium released the complete training datasets, model weights, and training and inference code, allowing the community to inspect, retrain and extend the model rather than simply use it as a black box. The silicon vendors have followed: NVIDIA offers OpenFold3 as a downloadable microservice through its BioNeMo platform, while Advanced Micro Devices announced that Vultr would bring AMD Inference Microservices, built by AMD Silo AI and optimized for AMD Instinct GPUs, to the Vultr Marketplace. These containerized inference microservices allow biology teams to reach production-grade structure prediction endpoints through a marketplace transaction rather than undertaking a multi-year infrastructure project.

MindWalk Holdings Corp., a bio-native AI company, validated AMD Inference Microservices for OpenFold3 running on Vultr Kubernetes Engine and Vultr Cloud GPU powered by AMD Instinct MI325X GPUs. According to the company, a production-grade prediction environment deployed in minutes, and antibody-antigen inference time fell by approximately 5 times against MindWalk's prior environment. The company notes that these results reflect specific configurations evaluated and may not be representative of all workloads.

Rather than running a curated benchmark, MindWalk exercised production-oriented drug discovery workloads: single- and multi-chain proteins, protein-RNA and protein-DNA interactions, protein-ligand structures, and antibody-antigen complexes. The evaluation included the complete pipeline, encompassing CPU-intensive preprocessing and multiple sequence alignment alongside GPU-accelerated inference. "Antibody-antigen prediction is the workload we care most about, and it has always been the one where infrastructure got in the way," said Dr. Jennifer Bath, Chief Executive Officer and President of MindWalk. "Standing up a production environment in minutes, and seeing complexes of that size come back in seconds, changes how many hypotheses a team can afford to test in a week. That is the practical effect, and it is why we put real programs through this rather than a benchmark set chosen to look good." The evaluation also surfaced practical considerations around multi-GPU execution, inference controls and deployment workflows, which MindWalk says it returned to AMD to inform subsequent platform work.

MindWalk's validation demonstrates a model-agnostic approach: the company deployed, ran and measured a third-party open-source model on third-party infrastructure within its own discovery workflows and under its own governance. The company describes Vultr, AWS and AMD as the compute fabric beneath its platform, with high-compute work directed to whichever provider offers the best price-performance fit for a given workload. At the core of MindWalk's platform is HYFT Technology, a proprietary, function-aware biological representation refined over 20 years of curation, comprising 660 million patterns and 25 billion relationships. This representation underpins ReefIQ, the biological context layer for life sciences, and LensAI, the reasoning and application layer for target discovery, candidate diligence, portfolio decision support and agentic AI workflows. The company argues that value compounds in this representation layer rather than in any individual AI model running on top of it.

A standardized, Kubernetes-native path from open model to production workflow also shortens the distance between a research question and an answer within a regulated environment—an operational concern pharmaceutical CIOs frequently raise. AMD Inference Microservices for OpenFold3 are available through the Vultr Kubernetes Engine Marketplace as part of the AMD Enterprise AI Software stack.

The standardization of structure prediction has expanded the options for compute capacity. Advanced Micro Devices reported second-quarter 2026 revenue of $11.5 billion, up 50 percent year over year and a company record, with gross margin of 54 percent, operating income of $2.0 billion, net income of $2.3 billion and diluted earnings per share of $1.38. CoreWeave, NVIDIA, and Nebius Group have each reported rapid growth in AI infrastructure demand, widening the range of providers available for routing high-compute biology workloads.