AWS Plans 2 Million NVIDIA GPUs by 2028 as AI Infrastructure Becomes Multiyear Bet

AWS commits to 2 million NVIDIA GPUs in 2027–2028

Amazon Web Services announced plans to deploy 2 million additional NVIDIA GPUs across its global infrastructure during 2027 and 2028, including Blackwell Ultra, Rubin, and Rubin Ultra systems. The commitment extends AWS's partnership with NVIDIA into CPUs, networking, AI factories, open models, data processing, and robotics.

The announcement is a capacity signal, not a price cut. Enterprises planning large AI deployments gain more leverage to negotiate multiyear AWS capacity commitments, reserved infrastructure, and model-hosting agreements. The delivery horizon means near-term GPU shortages and regional availability constraints remain unresolved. The risk for buyers is tighter coupling to NVIDIA's software and hardware stack as alternatives from AMD, Google's TPU business, and custom cloud silicon continue to lag in enterprise adoption.

The scale of the commitment reflects competition with Microsoft Azure and Google Cloud, both competing for the same scarce accelerator capacity and large training and inference workloads. NVIDIA remains the dominant enterprise accelerator supplier, and this partnership reinforces that position.

Crusoe closes $3.9 billion Series F for AI data centers

Crusoe, an AI data-center and infrastructure provider, announced an initial closing of a $3.9 billion Series F co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners. The size of the round indicates infrastructure providers are competing not only on GPU access but also on power procurement, data-center construction, and vertically integrated operations.

Crusoe competes with GPU-cloud operators including CoreWeave, Lambda, Nebius, and the hyperscalers. More capital could allow Crusoe to add capacity and compete for enterprise workloads that cannot obtain sufficient hyperscaler capacity. Buyers should examine delivery schedules, GPU generations, regional locations, power contracts, service-level commitments, and exit rights rather than treating the financing as proof of available capacity. The reported funding is concrete; comparable customer-count or utilization data was not provided.

Sovereign AI infrastructure gains traction in Canada and Australia

HIVE Digital Technologies' BUZZ High Performance Computing partnered with ProCogia to provide dedicated GPU capacity in Canadian data centers for LLM hosting, model training, agentic systems, and vertical AI applications. BUZZ cited a fourfold expansion of liquid-cooled capacity in British Columbia, a planned 320-megawatt sovereign-AI build in the Greater Toronto Area, a $220 million GPU contract with Bell AI Fabric for Cohere, and a $350 million AI-cloud-services agreement with an investment-grade enterprise customer.

Canadian organizations in regulated sectors may gain another option for keeping data and model workloads within Canada. The disclosed figures are commitments and infrastructure plans—not evidence of broad enterprise adoption. Buyers should verify commissioning dates, available GPU types, uptime history, security certifications, and whether capacity is actually reserved.

NVIDIA and Australian partners are planning land, power, and shell capacity for a buildout of up to 2 gigawatts by 2027, centered on NVIDIA DSX AI factories. Australian enterprises and government customers could eventually gain more local access to NVIDIA-based training and inference capacity, potentially reducing cross-border data-transfer and sovereignty concerns. The announcement describes planned power and site capacity, not currently available production capacity or published workload pricing.

Snorkel AI raises $350 million for data-centric tooling

Snorkel AI disclosed a $350 million Series E led by Insight and S32, with participation from GV and other investors, at a reported $3.5 billion valuation. The company develops data-centric tools for building and managing AI training datasets. Snorkel competes with data-labeling, synthetic-data, evaluation, and machine-learning-data platforms including Scale AI, Labelbox, and Databricks.

The round signals continued investor interest in the data layer of AI infrastructure, not only GPUs and data centers. For enterprises, the practical buying question is whether data-preparation and labeling platforms reduce the cost and time required to adapt models to proprietary data. The funding itself does not establish customer counts, benchmark gains, or return on investment. Procurement teams should request measurable evidence such as labeling-throughput improvements, model-quality uplift, annotation error rates, and integration costs.

What to watch: Capacity commitments versus operating reality

The week's developments point to three shifts. First, capacity is becoming a multiyear procurement issue. AWS's planned 2 million GPUs and Crusoe's $3.9 billion financing show providers committing capital years ahead of demand rather than relying only on on-demand cloud supply. Second, sovereignty and locality are becoming infrastructure differentiators. The Canadian and Australian projects emphasize regional power, data residency, and domestic compute rather than lowest-cost global capacity. Third, buyers should distinguish commitments from operating capacity.

The announcements provide substantial figures—GPU counts, gigawatts, financing, and contract values—but generally do not provide current utilization, enterprise customer counts, or production-readiness dates. Enterprises should negotiate contracted capacity with delivery penalties, verify power availability and cooling infrastructure, and build multisource procurement strategies to avoid single-vendor lock-in. The infrastructure being announced today will not arrive until 2027 or 2028. Budget and planning decisions made now determine whether your organization has access to the capacity it needs when it becomes available.