OpenAI unveiled GPT-5.6 Sol last week: a multimodal model optimized for Cerebras hardware, achieving 750 tokens per second for enterprise agent workflows. The number is impressive. Cerebras' wafer-scale processors are purpose-built for this kind of throughput, and 750 tokens per second is fast enough that inference latency is no longer the bottleneck in even the most demanding agentic workflows.
In the same week, Anthropic committed $200 billion to Google Cloud for five-year TPU access.
I have been writing separately about each of these announcements, but I want to bring them together because together they form the clearest argument I have encountered for why sovereign inference is not a philosophical position — it is a practical necessity for organizations that want to control their AI economics over any meaningful time horizon.
The Speed Story and What It Actually Means
750 tokens per second on Cerebras hardware is a real achievement. It is approximately 10-15x what a well-optimized self-hosted inference setup on commodity AMD or NVIDIA hardware delivers today for a model of comparable quality. For use cases where inference speed is the constraint — real-time customer interactions, low-latency agentic loops, interactive code generation — this gap is significant.
But I want to examine what the speed story is actually selling, because there is a framing choice embedded in leading with the token-per-second number.
The organizations that need 750 tokens per second are organizations with high-volume, latency-sensitive inference workloads. They exist. They are real. For those organizations, Cerebras-hosted frontier models may be the right infrastructure choice, and the speed justifies the dependency.
The organizations that do not need 750 tokens per second — which is most organizations deploying AI for internal workflows, document processing, code assistance, analytical tasks — are being asked to make infrastructure decisions with this headline in their peripheral vision, as if the speed requirement applies to them. It often does not.
A self-hosted inference setup running GPT-OSS 20B or Qwen3 Coder 30B on a three-node AMD cluster delivers 50-80 tokens per second. For a Kaigents workflow that processes one engineering task at a time, with a human-in-the-loop checkpoint at each quality gate, 60 tokens per second means the model responds in under two seconds for most task steps. The agent is not waiting on inference. The human is waiting on the human's own review. 750 tokens per second changes nothing about the total task throughput in this workflow.
The 750 tokens per second headline matters for a specific use case profile. Know whether your use cases match that profile before letting the number anchor your infrastructure thinking.
The Dependency Story and What It Actually Means
I have covered the Anthropic Google commitment separately, but let me restate the core point in the context of the speed story.
The $200 billion commitment creates a structural dependency: Anthropic's API pricing is downstream of Google's TPU pricing for the next five years. This is not speculation. It is the arithmetic of infrastructure commitments. When a company commits $200 billion to one provider's infrastructure, the marginal cost of their product is partly determined by the terms they negotiated and the trajectory of that provider's pricing.
For enterprise buyers, this creates a compounding problem with the speed story. If you build workflows on frontier model APIs — Claude, GPT-5.6, Gemini — because of the speed and capability advantages, you are simultaneously accepting the dependency structure those APIs represent. The speed is real. So is the dependency.
The organizations that are most exposed are those that have built the deepest frontier model integrations — not in terms of technical sophistication, but in terms of economic dependency. If 70% of your AI-driven product value is delivered through an API whose cost structure you do not control, and that cost structure is upstream of a $200 billion hyperscaler commitment, you have a concentration risk that is not fully visible in your current budget.
The Sovereign Inference Middle Path
The argument I want to make is not "avoid frontier models." It is "understand the use case split and build sovereign capacity for the portion that does not require frontier capabilities."
Frontier models for frontier tasks. Genuinely novel reasoning, complex multi-step analysis, safety-critical response generation, tasks where the quality difference between a frontier model and an open-weight model is material — these use cases belong on frontier model APIs. Accept the dependency. Budget for the cost. Monitor the pricing trajectory.
Sovereign inference for the rest. Document processing, structured data extraction, code generation for known patterns, classification, summarization, agentic task execution with defined scope — for these tasks, open-weight models on sovereign infrastructure are competitive today, and the quality gap is narrowing. Running them on infrastructure you control eliminates the API cost, removes the dependency on any provider's pricing decisions, and keeps the derived intelligence (the embeddings, the processed outputs, the interaction logs) in your perimeter.
My sovereign stack — AMD Ryzen AI hardware, Kubernetes orchestration, Lemonade Server for inference — runs GPT-OSS 20B and Qwen3 Coder 30B. These are not frontier models by the headline metrics. They are production-grade models for the task profile I described above. The inference cost is electricity and hardware depreciation. Neither is subject to a five-year contract renegotiation.
The Arithmetic That Matters
Let me put this in concrete terms. If you are spending $50,000/year on Claude API for a mixed workload — some fraction that genuinely requires Claude's capabilities, some fraction that open-weight models could handle — and the sovereign infrastructure to run that open-weight tier costs $15,000/year (hardware amortized over three years plus electricity), the math of the split is:
If 40% of your workload is frontier-capable and 60% is sovereign-capable, you have reduced your Claude exposure from $50,000/year to $20,000/year, added $15,000/year in sovereign infrastructure, and net-reduced your AI infrastructure cost by $15,000/year — while also reducing your concentration risk by 60%.
These are illustrative numbers. Your actual split depends on your specific workload profile. But the structure of the analysis is the right structure. Know the split. Price both paths. Make an active choice about where the dependency is acceptable and where it is not.
What the Week's Announcements Together Signal
750 tokens per second and $200 billion in the same week are not contradictory signals. They are complementary.
The speed announcement says: frontier model inference is getting dramatically faster, and for latency-sensitive use cases, the cloud-based frontier model API path is genuinely compelling.
The dependency announcement says: the economic structure of that cloud-based frontier model API path is becoming more concentrated in single-provider infrastructure, and the cost trajectory of that concentration is not something you control.
Together: use frontier models where they matter, build sovereign capacity everywhere else, know the difference between the two, and make the infrastructure choice with your eyes open about both the benefits and the dependencies.
The case for sovereign inference has never been stronger — not because frontier models are bad, but because the dependency structure they represent is becoming more concentrated, more legible, and more consequential at exactly the moment when sovereign alternatives are becoming more capable, more accessible, and easier to operate.
The window for building sovereign capacity before you need it is now. After you need it is too late to build it without disruption.