Fractionalizing AI Infrastructure: Can Shared Compute and Data Democratize Innovation?

If today’s AI moats are built from compute, data, and tooling, then the most pro-competitive policy is to fractionalize those complementary assets—so more actors can reach scale without being owned by hyperscalers. That’s exactly what recent public programs attempt. The open question for leaders: Will these pools be big and usable enough to matter?

Why this matters: The economics are clear: when few firms control the complementary assets, value capture concentrates—even if knowledge diffuses. Azoulay, Krieger, and Nagaraj argue that policy that “fractionalizes or facilitates shared access to complementary assets” is one of the only realistic levers to avoid an incumbent-dominated outcome.

What’s changing now:

• U.S. NAIRR Pilot (National AI Research Resource): NAIRR is a federally‑led, public-private testbed providing researchers with access to compute, datasets, models, and user support, launched in January 2024. This pilot aims to expand through multi-agency partnerships and advance responsible AI goals set by the US government. Managed by the U.S. National Science Foundation, the list of partners already includes key government agencies (DARPA, NASA, NIH etc.) and leaders in AI (OpenAI, Microsoft, NVIDIA, Amazon Web Services, among others)

• Europe’s AI Factories (EuroHPC JU): EuroHPC JU is a network of regional “one-stop shops” that combine AI-optimized HPC, training, and expertise for startups, SMEs, and researchers. The initial seven sites selected in Dec 2024 include Finland, Germany, Greece, Italy, Luxembourg, Spain and Sweden. Six additional factory sites were chosen earlier this year. These factories will be set up in Austria, Bulgaria, France, Germany, Poland, and Slovenia. These factories are part of the wider package of policy measures, which includes the AI Act (legal framework on AI for developers and deployers), the AI Innovation Package, and the Coordinated Plan on AI.

• UK AI Research Resource (AIRR) & Isambard‑AI: The UK AIRR is a billion-dollar initiative enhancing sovereign compute capabilities, anchored by the University of Bristol’s Isambard‑AI supercomputer (21 “AI exaflops”)—now operational and offers public research access. Project partner companies include NVIDIA, HPE, Dell Technologies, and Intel.

• Canada’s Sovereign AI Compute Strategy: Canada announced a C$2 billion Sovereign AI Compute Strategy in December 2024 to build a national backbone for AI. This includes the AI Compute Access Fund (C$300 million, providing subsidized compute for SMEs and researchers), and the AI Sovereign Compute Infrastructure Program (C$705 million for Canadian-owned supercomputing). Canada also runs PAICE (Pan-Canadian AI Compute Environment), pooling resources across Mila, Amii, Vector, and other institutions to make compute capacity accessible to researchers. The approach emphasizes democratized access, SME support, and green energy-powered infrastructure as a differentiator from U.S. and European scale-first models.

• Open data commons. Non‑profits like Common Crawl and LAION continue to supply massive, open corpora for model training—vital, if imperfect, substitutes for proprietary data.

Reality check: Scale still bites. Meta alone aimed for ~340k+ H100 GPUs by the end of 2024, illustrating how private fleets dwarf most public pools. So, what does this mean for us?

• H100 GPUs are NVIDIA’s high-end AI accelerators (designed for training and inference of large models).

• Having 340,000+ of them means Meta is building compute infrastructure worth ~$10B USD in GPUs alone (not counting energy, cooling, networking).

• The private “hyperscalers” (Meta, Google, Microsoft, Amazon) are playing at a completely different level. Public compute programs (like NAIRR in the U.S. or PAICE in Canada) can’t possibly compete on raw scale with the hyperscalers.

Because they can’t match hyperscalers machine-for-machine, public initiatives should make their limited compute capacity go furthest by targeting where it creates the most social value. For example:

1.       Prioritized queues for basic research and safety:

• Let academics and safety researchers skip the line.

• These groups usually can’t afford GPU clusters, yet their work creates knowledge spillovers for everyone.

2.       Subsidized time for SMEs:

• Give small firms compute credits/grants so they can run training/fine-tuning experiments.

• Otherwise, only VC-backed or large-firm labs can play at the frontier.

3.       Turnkey MLOps so time-to-first-experiment is days—not months:

• MLOps = the tooling and infrastructure needed to get models from idea → training → deployment.

• If every SME or lab has to spend months setting up pipelines (data prep, logging, monitoring, cloud orchestration), then compute credits get wasted.

• Turnkey systems (like pre-built environments, APIs, and dev tools) reduce friction so that a researcher can go from idea to results in days.

• That’s critical: in fast-moving fields like AI, the cost of delay is as prohibitive as the cost of hardware.

Takeaways for leaders:

• Use the pools—but bring your stack. Treat NAIRR/AI Factories/AIRR as burst capacity — a place to test new models, validate early prototypes, or conduct safety/ethics research that benefits from neutral, academic oversight, not your only platform.

• Co-invest in data governance. Open data is only as credible as its provenance and red‑team hygiene; leaders should push Commons to publish structured documentation (source, filtering, IP checks).

• Design for portability. Make workloads redeployable across public pools and clouds; “exit options” are the best bargaining chip you have with vendors.