Executive Summary
The three sources share one point. Model choice now explains less of AI value than effort settings, routing and portability, while the top of the model market is getting more concentrated. Anthropic says an agent can now port its own models across chip families in days, which would make the hardware layer cheap to switch. If Berman is right, Anthropic's lead in model quality is growing faster than the benchmarks show. Jones shows that within a single model, a hidden setting can decide whether users find it useful or call it "stupid." Taken together, enterprise lock-in is moving up the stack from silicon to the model provider, and most of the value an enterprise can capture now depends on configuration and orchestration rather than on which vendor it picked.
What Changed
- Bringing up new hardware may be close to free. Anthropic claims Mythos ported Claude to AMD GPUs in three days and wrote nearly all the kernels itself. Earlier bring-ups took dozens of engineers months. This is an unverified roadshow claim, but it is concrete, and Anthropic's earlier incident postmortem, where some bugs hit only TPU deployments, supports the claim that it already runs on multiple chip families.
- Benchmarks are understating the gap at the top. Fable 5.2 leads Gemini 4 Carbon by about two points on Artificial Analysis. A practitioner with early access to Carbon kept going back to Fable anyway, and suspects the benchmarks are being gamed.
- The cheap tier caught up. Muse Ember matches Haiku 5.5 on the AA index at about a third of the cost per task. Ripple 65B runs on a MacBook or DGX Spark at $0.20/$0.60 per million tokens and scores 80.5% on Terminal Bench 2.1.
- Default settings hide capability. Paid ChatGPT seats ship at medium reasoning effort, the control sits behind a menu most users never open, and the model cannot report its own setting.
Cross-Expert Synthesis
Willison's account of the roadshow and Berman's flywheel thesis support each other but describe different moats. Anthropic's own pitch is about efficiency: physical capacity multiplied by algorithmic gains that compound, plus the freedom to buy whichever chips are cheapest. Berman's version is about demand: coding focus brings enterprise revenue, which brings usage data, which pays for the next model. Both say rivals are chasing a target that keeps moving away. Both also lean on recursive self-improvement, and neither offers evidence beyond the AMD anecdote. Treat that part of the story as pitch, not fact.
The real tension is about costs. Multi-chip portability and compounding efficiency should push per-token prices down. Berman reports the opposite at the top: Fable 5.2 is the most expensive model per task and burns through quota fast. The likely reconciliation is that Anthropic's cost advantage shows up as margin and supply security, not lower list prices, because a model without a close rival doesn't need to discount. Lower prices will come from below, from Carbon, Muse Ember and open-weights models.
Jones's small tip connects to Berman's routing advice. Both say spend should follow task difficulty: high reasoning effort or the frontier model for hard reasoning, cheaper settings or models for routine work. Effort level is routing inside one model; plan-with-frontier, execute-with-small is routing across models. Most enterprises do neither on purpose.
Enterprise Implications
Silicon matters less for strategy and models matter more. If agents can retarget kernels cheaply, CUDA's hold weakens for frontier labs first and, over time, for enterprise inference stacks. The dependency that remains is the frontier model itself, and that market looks more concentrated each quarter. That makes pricing power and supplier concentration the main planning risk for any workload that truly needs top-tier reasoning.
At the same time, credible US open-weights models make the private and on-prem option real for regulated workloads. The architecture that follows is two-tier: rent frontier reasoning for planning and judgment, and own the execution tier on local or sovereign hardware with small models.
Seat-based rollouts are probably underperforming for configuration reasons, not model reasons. Complaints that "the AI got dumber" should trigger a settings check before a vendor review.
What To Do About It
- Build task-level evals with effort and model recorded as variables. Index scores now mislead at the top and misreport performance at the bottom. Without your own evals you cannot tell a weak model from a misconfigured one.
- Set effort defaults by workload class. Analysis, code review and multimodal work get high effort; drafting and lookups stay at low or medium. Write it down, and ship a one-page guide to every seat holder.
- Design for planner/executor separation now. A frontier planner with Haiku-class or open-weights executors cuts spend and keeps the execution tier portable.
- Do not let architecture assume a chip vendor or a single model provider. Keep model calls behind an abstraction layer and keep a tested fallback model per workload.
- Talk track for buyers: "Your hardware is becoming interchangeable; your model dependency isn't. Which of your workloads actually need the frontier, and what happens to them if that one vendor raises prices 30%?"
Risks and Blind Spots
The AMD port, the compounding efficiency gains and the self-improvement claims all come from an IPO roadshow, which is the least neutral setting a company can speak in. Berman's view that Fable is far ahead rests on one person's hands-on use, and his claim that Fable 6 and 7 are already in progress is speculation. Agent-written kernels that run are not the same as kernels that are correct and fast in production; the TPU-only bugs are a reminder that every additional platform is another place for subtle failures. The Nvidia governance question, an affair with a possible Chinese-espionage angle now under independent investigation, is an unresolved supply-chain risk. A multi-chip strategy hedges against it for Anthropic, but most enterprises have no such hedge. Finally, Jones's UI details may change with plan and platform, but the underlying problem, defaults set for vendor cost rather than user outcome, will likely persist.