Inference Price Wars, Sovereign GenAI Splits, and Enterprise AI Shifts to the Runtime Layer

OpenAI’s GPT-6 Cuts Force the Next Round of Inference Price Wars

OpenAI’s GPT-6 API price cuts reset the inference floor this week, with the smallest tier dropping to about $0.10 input and $0.50 output per 1M tokens, while Sol landed at $2 input and $10 output and Luna’s output price fell 58.3% versus GPT-5.6 Luna. Rivals moved fast: Anthropic launched Claude Opus 5.5 at $4 input and $20 output, a 20% cut from Opus 5, and xAI reportedly matched OpenAI’s new $2 input floor with Grok 4.7. Routine inference is turning into a margin-compression market, not a premium one.

That pressure is being reinforced by serving and hardware gains. AWS says quantization, speculative decoding, batching, and compilation can deliver up to 2x throughput and as much as 50% lower cost, while AWQ and GPTQ can shrink models by roughly 2-8x. Reporting also puts FP8 and FP4 deployments at roughly 20-40% throughput gains, and NVIDIA is extending low-precision execution into KV caches. With Menlo saying 76% of AI use cases are bought rather than built, Llama still the most adopted open-weight enterprise model, and MIT Sloan putting closed models at 87% higher run costs on average, model choice is becoming a metered policy decision by task, risk, and latency. That extends the routing layer from last week into governance and spend control, where practitioners now have to optimize not just which model to call, but when, why, and under what budget.

Sovereignty Rules Are Splitting GenAI Deployment

Microsoft’s regional commitments in the Middle East, Thales partnerships, sovereign stack efforts in Europe, the UK, and Nigeria, and the Pentagon’s blacklist of Anthropic show the next layer of the market taking shape: GenAI is fragmenting into governed deployment regimes. Hybrid infrastructure is becoming the default operating model, with public cloud handling training and bursty demand while sensitive or high-volume inference shifts to private, on-prem, or edge environments.

The economics and compliance case are converging. Deloitte says sustained cloud costs can exceed roughly 60%–70% of equivalent on-prem for predictable workloads, and IBM reports 61% of cloud leaders cite security or compliance as a reason to move specific workloads to private or on-prem. That makes the gating factor less about whether agents can perform and more about whether they can be deployed inside region-specific control regimes.

For operators, policy-aware architecture is now the next requirement after workflow KPI proof. For vendors and investors, the durable value is shifting further toward control planes, hybrid orchestration, and regional infrastructure partnerships rather than model differentiation alone.

Enterprise AI Moves Into the Governance and Runtime Layer

OpenAI’s launch on AWS Bedrock makes the next enterprise AI battleground explicit: not model quality alone, but the control plane around production agents. Enterprises can now access OpenAI frontier models, Codex on Bedrock, and Amazon Bedrock Managed Agents through Bedrock APIs, with AWS IAM access control, VPC and PrivateLink isolation, KMS encryption, CloudTrail logging, and AWS’s no-training policy for prompts and responses. AWS also says pricing matches OpenAI first-party rates, while keeping spend inside AWS commitments.

Databricks is pushing the same direction with Databricks Apps, Lakebase, Agent Bricks, and its new Unity Gateway CLI for connecting coding agents like Claude Code and Codex to governed tools, models, and spend controls. Ema’s $77 million Series B reinforces where capital is flowing: “AI employees” for HR, IT, and finance built around plan-execute-verify-approve workflows across 250+ business applications and 150+ models. VAST Data’s secure runtime launch points in the same direction from infrastructure. The strategic shift is clear: value is moving to vendors that can govern memory, routing, approvals, and execution, not just generate outputs.