AI Compute Expands, With Efficiency and Oversight in Focus

AI infrastructure decisions span three different problems: obtaining capacity, getting more work from existing hardware, and controlling operations. An $11.6 billion compute agreement and new FinOps controls for AI and agents put both investment and spending management in view.

More Capacity, Better Hardware Utilization

Anthropic has signed an $11.6 billion multi-year agreement with Akamai to expand its AI compute capabilities. Separately, SpaceXAI reportedly plans to add 660,000 AI GPUs this year, bringing its total to 1.44 million. Both developments concern access to substantial compute resources, but their commitments differ: Anthropic has secured an agreement, while SpaceXAI's additional GPUs remain a planned expansion. Neither announcement alone establishes what performance an application built on that capacity will deliver.

Modal's Quail addresses efficiency within existing hardware. The company reports a billion tokens per minute on a single GPU by combining a query planner with an inference engine. That result makes execution design relevant alongside hardware scale: the reported throughput comes from how the work is organized and executed. For engineers evaluating infrastructure, Quail offers a concrete efficiency result to consider alongside the much larger capacity commitments, without establishing that the same throughput applies to their own workloads.

Spending Controls Reach AI and Agents

Microsoft has introduced FinOps controls tailored to AI and agent operations, intended to improve visibility into and management of AI expenditure. This addresses a different layer of infrastructure planning from acquiring GPUs or increasing throughput. Organizations also need to understand what their AI operations cost. The controls may help teams connect resource use with project budgets, although the announcement does not establish how much spending they will save.

Incident Investigations Interrupt Model Training

OpenAI and Anthropic are investigating tens of thousands of incidents, with OpenAI pausing training. The pause makes operational oversight consequential for development schedules as well as deployed systems. It also belongs alongside the capacity announcements without being treated as their consequence: the reported investigations concern AI-system incidents, while the compute agreements and expansion plans concern infrastructure. The reporting supports the existence of the investigations and pause, rather than a conclusion about their eventual findings.