What this area covers
Training infrastructure encompasses the full hardware and systems stack that makes frontier AI possible: the chips that run matrix multiplications, the data centers that house them, the power contracts that feed them, the distributed training software that coordinates them, and the capital agreements that secure access to all of the above. It is the physical substrate on which scaling laws operate — and, increasingly, a primary competitive dimension in its own right.
Why it matters
The 2020 OpenAI scaling laws paper established an empirical foundation that has since driven hundreds of billions in capital allocation: model performance scales predictably with compute, data, and parameters along power-law curves. That finding transformed infrastructure from a cost center into a strategic asset. Labs that can train on more compute, faster, and more cheaply have a structural advantage in capability — which is why the investment numbers in this space have grown from billions to hundreds of billions in under a decade.
How it evolved
Phase 1 — Hyperscaler dependency (2019–2023). The era began with Microsoft's 2019 $1B investment in OpenAI and the establishment of Azure as OpenAI's exclusive cloud provider. Training on rented cloud GPUs was the norm; the infrastructure question was primarily "how many A100s can we reserve."
Phase 2 — Strategic lock-in and hardware co-development (2023–2025). Amazon's initial investment in Anthropic (up to $4B, announced late 2024) marked a qualitative shift: Anthropic engineers began writing low-level kernels and contributing to the AWS Neuron software stack, co-developing Trainium accelerators from the silicon up. This is not a cloud rental — it is vertical integration into the chip design process. The Stargate Project, announced January 2025, signaled that OpenAI was pursuing the same logic at even larger scale, targeting up to $500B in dedicated U.S. AI infrastructure.
Phase 3 — Multi-gigawatt buildout and supplier diversification (2025–2026). The current phase is characterized by commitments measured in gigawatts and tens of billions of dollars, and by aggressive diversification away from single-supplier dependency. OpenAI signed 10GW partnerships with both NVIDIA and Broadcom, a 6GW AMD deal, and a $100B+ AWS Trainium commitment — all within roughly twelve months. Anthropic assembled a parallel portfolio: 5GW of Amazon Trainium2–4 capacity, multi-gigawatt Google/Broadcom TPU agreements, 1GW of NVIDIA Grace Blackwell/Vera Rubin via Microsoft, and burst access to SpaceX's Colossus facility (220,000+ NVIDIA GPUs, over 300MW) within a month of signing.
Phase 4 — Custom silicon (2026–). The most recent development is OpenAI's entry into proprietary chip design. Jalapeño, co-developed with Broadcom and designed in nine months with AI-assisted design tooling, is OpenAI's first custom inference chip. Engineering samples are already running production models, with datacenter deployment planned by end of 2026. OpenAI and Broadcom have also announced a longer-horizon partnership targeting 10GW of OpenAI-designed accelerators by 2029. Custom silicon closes the last gap in vertical integration: owning the compute contract, the facility, and the chip itself.
The physical buildout
Compute commitments are materializing as concrete facilities. OpenAI broke ground on a 1GW Stargate data center in Michigan in June 2026. Anthropic committed $50B to purpose-built data centers in Texas and New York via Fluidstack, with sites coming online throughout 2026. Mistral AI, operating at a different scale, announced a 10MW inference data center in Les Ulis, France, scheduled for Q3 2026. The geographic pattern reflects both domestic AI policy pressures and the practical need to site facilities near power sources.
Geopolitical and security dimensions
The infrastructure layer has acquired geopolitical weight. Iranian drone strikes damaged at least three AWS data centers in Bahrain and the UAE in March 2026, disrupting cloud services across the region — the first known targeting of commercial cloud infrastructure during active conflict. The episode underscores that AI compute is now considered a strategic asset worth attacking, and that geographic concentration of data centers creates physical vulnerability. Anthropic's international infrastructure expansion for enterprise compliance needs, and the predominantly U.S.-sited nature of its new compute commitments, reflect awareness of this dynamic.
Efficiency research as a countervailing force
Not all the action is in raw scale. A cluster of research results is beginning to challenge the assumption that more compute is always the answer:
- Single-layer RL post-training: A 2026 arXiv paper found that training a single transformer layer — concentrated in the middle of the stack — can recover most or all of the gains from full-parameter RL post-training across seven model families and three RL algorithms. If this generalizes, it dramatically reduces the compute cost of alignment fine-tuning.
- Sparse attention: DeepSeek's V3.2-Exp introduced DeepSeek Sparse Attention (DSA), achieving efficiency gains during training and inference while matching the performance of its predecessor, accompanied by a 50%+ API price cut.
- MoE hyperparameter transfer: The Complete-muE framework enables near-optimal hyperparameter reuse across dense and Mixture-of-Experts architectures without costly re-tuning — reducing the overhead of scaling MoE models.
- Shannon Scaling Law: A theoretical framework modeling LLM training as noisy-channel information transmission introduces an SNR-based capacity limit that explains phenomena like catastrophic overtraining that classical power-law scaling laws cannot capture, potentially improving how labs allocate training compute.
- On-device MoE: MobileMoE demonstrates 2–4× fewer inference FLOPs than dense baselines on commodity smartphones, pointing toward a parallel efficiency frontier at the edge.
These results do not reduce infrastructure investment in the near term — labs are spending efficiency gains on more capability — but they suggest the compute-per-capability ratio is improving, and that the relationship between raw scale and model quality is more nuanced than early scaling laws implied.
Where it is heading
The structural trajectory is toward full-stack ownership: labs that control their chips, their facilities, their power, and their software stack will have the most predictable access to compute and the lowest marginal cost per training FLOP. OpenAI's Jalapeño and the Broadcom 10GW accelerator roadmap represent the leading edge of this. Anthropic's multi-provider portfolio is a hedge against any single supplier's constraints. The Stargate Michigan facility and Anthropic's Fluidstack data centers represent the physical manifestation of commitments made in 2024–2025.
The open question is whether efficiency research will eventually bend the curve enough to make the current scale of investment look excessive — or whether capability demands will continue to absorb every efficiency gain, keeping the infrastructure arms race at full intensity.
