Compute

What changed

Recent signals

• Anthropic announces access to substantial SpaceX compute capacity while raising product usage limits. A stated near-term capacity agreement is not independent confirmation that every GPU was delivered immediately.retrospective mapping
  supported inference
• Meta announces agreements involving existing nuclear generation and future reactor projects. Up to 6.6 GW by 2035 combines different projects and dates, not new operational capacity in January.retrospective mapping
  supported inference
• NVIDIA presents a six-chip AI infrastructure platform combining accelerators, CPUs and networking. Partner-system availability is planned for the second half of 2026; the announcement is not current capacity.retrospective mapping
  supported inference
• AWS announces GA of Trn3 UltraServers with Trainium3 chips and an expanded interconnect fabric. Advertised performance ratios are vendor claims and GA is not evidence of unconstrained global capacity.retrospective mapping
  supported inference
• Google says Ironwood TPU availability will expand over the coming weeks. This dated announcement alone does not verify the exact completed GA date.retrospective mapping
  supported inference
• OpenAI announces development of 4.5 GW of additional US data-center capacity with Oracle. Capacity under development is not already delivered electrical supply or operating compute.retrospective mapping
  supported inference
• Meta announces a twenty-year agreement for Clinton nuclear plant output beginning in 2027. Contracted existing generation and future uprates are not new operational power in 2025; later page updates are excluded.retrospective mapping
  supported inference
• Commerce directs non-enforcement and announces intended rescission of the AI Diffusion Rule. The announcement is not proof that the formal replacement regulation was already completed.retrospective mapping
  supported inference
• Google announces its seventh-generation TPU for inference workloads. Announcement and access requests do not establish general availability in April.retrospective mapping
  supported inference
• Large intended infrastructure investments and a new deferred-compliance export framework changed the planning environment, not current abundant capacity.retrospective mapping
  supported inference
• Trainium2 regional cloud GA added accessible accelerator options while new nuclear commitments addressed future infrastructure demand.retrospective mapping
  supported inference
• Longer model reasoning highlighted inference resource tradeoffs, while a signed nuclear PPA committed future data-center energy supply.retrospective mapping
  supported inference
• Previously controlled TPU v5p access became GA, and local AI-capable consumer PC hardware reached market availability.retrospective mapping
  supported inference
• Research compute access acquired a coordinated national pilot, while an AWS campus acquisition linked cloud expansion to staged nuclear-power supply; new Blackwell capacity remained future.retrospective mapping
  supported inference
• AI compute access is beginning to acquire a market-and-orchestration layer that can allocate burst capacity dynamically across heterogeneous infrastructure, creating an optimization path based on scheduling and utilization in addition to building more accelerator capacity.
  supported inference
• U.S. controls further tightened access to advanced computing chips and semiconductor manufacturing capability while broadening anti-circumvention and geographic scope.retrospective mapping
  supported inference
• Large AI training clusters and infrastructure-as-a-service training became explicit objects of AI-specific federal governance and reporting design.retrospective mapping
  supported inference
• Operational AI accelerator infrastructure diversified beyond NVIDIA-centered cloud offerings through AMD MI300X availability and Google's TPU v5p/AI Hypercomputer path.retrospective mapping
  supported inference
• H100-class AI acceleration became generally available through multiple major hyperscale cloud providers while remaining expensive, regional and concentrated.retrospective mapping
  supported inference
• Reflection’s pre-release open-weight signal advanced to a concrete early-access model debut while public weights and independent reproducibility remained pending.
  supported inference
• High-end AI supercomputing became more explicitly packaged as a commercial cloud service integrating accelerators, networking and AI software.retrospective mapping
  supported inference
• Open-weight models are increasingly being framed together with institution-controlled data and dedicated compute as a route to localized AI control, though Reflection's specific model remains pre-release.
  supported inference
• Geopolitical technology controls expanded from restrictions on specific accelerator exports toward formal constraints spanning advanced compute, supercomputer end uses, semiconductor manufacturing equipment and specified support for advanced chip production.retrospective mapping
  supported inference
• The next generation of transformer-focused accelerator infrastructure moved from announced architecture toward production and orderable systems.retrospective mapping
  supported inference
• Semiconductor capacity and access became objects of more direct state intervention through domestic manufacturing incentives and restrictions on access to advanced AI accelerators.retrospective mapping
  supported inference
• European AI-compute sovereignty is strengthening at the system-integration and manufacturing layers without implying full component or accelerator autonomy.
  supported inference
• Energy and carbon intensity became explicit metrics in AI infrastructure and large-model training announcements, but Q2 evidence does not establish electricity supply as a binding AI compute bottleneck.retrospective mapping
  supported inference
• Q2 research increased uncertainty about capability forecasting by documenting benchmark behaviors that appeared only above model-scale thresholds.retrospective mapping
  supported inference
• Frontier-scale accelerator infrastructure became more externally accessible through cloud preview, while scale, price-performance, interconnect and software utilization remained explicit constraints.retrospective mapping
  supported inference
• Research access to very large language models improved sharply, but the access bottleneck shifted toward licenses, approval, serving cost and the compute needed to reproduce or run them.retrospective mapping
  supported inference
• Frontier model scaling increasingly depended on distributed-systems software and high-bandwidth multi-pod networking, not only on adding accelerator chips.retrospective mapping
  supported inference
• By the end of Q1 2022, the clearest stack coupling ran through research, models, software, compute, chips, networks, manufacturing and governance; the selected evidence does not yet support persistent agents or agent permissioning as independent operational layers.retrospective mapping
  supported inference
• Model size alone was becoming an insufficient description of frontier progress; training-token allocation and inference cost emerged as important variables in the model-compute relationship.retrospective mapping
  supported inference
• Transformer workloads were becoming an explicit target of accelerator, interconnect and system architecture design rather than merely software running on generic compute.retrospective mapping
  supported inference
• Open and publicly accessible large-language-model ecosystems were becoming a material alternative route for research and application building, though at smaller scale than the largest closed models.retrospective mapping
  supported inference
• Dedicated AI supercomputing was becoming a distinct operational layer, with model progress increasingly dependent on large accelerator fleets, high-bandwidth networking and storage.retrospective mapping
  supported inference
• High-bandwidth memory remains a binding physical bottleneck in AI scaling, with demand above capacity and spillover tightening in broader memory markets.
  supported inference
• Large AI deployments are broadening beyond single-vendor accelerator stacks into integrated rack-scale systems where networking, cooling and open Ethernet architectures are strategic components.
  supported inference
• Data-center electricity and grid-upgrade cost allocation has become an explicit federal legislative issue in the United States, but no durable federal settlement has emerged.
  observed fact
• The AI stack is becoming more tightly coupled across models, agents, permissions, compute, chips, software, networks, energy and governance.
  supported inference
• Data-center electricity, water use and efficiency are becoming explicit objects of measurement and governance in the European Union.
  observed fact
• Agent safety is expanding from model behavior into runtime, hardware, identity and tool-access controls.
  supported inference
• China's AI stack is deepening vertically across chips, interconnects and developer software, with open-source tooling aimed at reducing dependence on Nvidia's CUDA ecosystem.
  supported inference
• Long-horizon AI and agent workloads are driving large multi-year commitments for cloud capacity, memory and supporting infrastructure.
  supported inference
• System-level scaling, interconnect design and software integration are being used to compensate for limits at the individual accelerator level.
  supported inference