Executive Summary
Today's sources agree on one thing: models are no longer the hard part to get. Leading-edge silicon is hard capped through 2030. GPU capacity is sold in rigid blocks that do not fit real demand. Open-weight models are close enough to frontier that swapping models often beats fine-tuning. The agent extensibility layer (MCP plus Skills) now sits under one neutral foundation. The scarce assets are physical supply, procurement skill, and the ability to verify what models and agents actually do. Altman's "discovery era" pitch, Goodfire's interpretability claims, and Modal's sandbox security posture all depend on verification. For enterprise clients, the advantage is moving away from model selection and toward capacity strategy and oversight infrastructure.
What Changed
- Agent standards consolidated. Anthropic donated the Agent Skills spec to the Linux Foundation's Agentic AI Foundation (AAIF), where MCP already sits. Both agent extensibility layers (tools and procedural knowledge) now have one vendor-neutral home, with OpenAI and Anthropic both participating.
- A credible supply ceiling. Intel's CTO puts the maximum growth in leading-edge wafer output before 2030 at roughly 2x, limited by ASML shipping under 100 EUV tools a year. AI compute demand has been growing about 3x a year. Memory is tighter than logic.
- Compute procurement became a finance job. Modal now runs a four-person procurement desk, multi-year prepayments across about 25 clouds, and a bare-metal team. Its business is turning three-year GPU blocks into on-demand slices.
- Interpretability moved toward operations. Goodfire reports that monitors reading internal model signals caught an agent substitution that monitors reading the transcript missed. The figures are secondhand and unverified.
Cross-Expert Synthesis
The supply story pulls in two directions. Intel describes a decade of structural scarcity in which hyperscalers and labs lock up whole fabs and everyone else takes the price offered. Bernhardsson refuses to own hardware because he sees tail risk of buying at the top if GPU prices collapse. Both can be right: wafer supply stays tight while rental pricing swings with demand, model efficiency, and the non-Nvidia accelerators he expects to gain traction in 2027. For buyers, the cost of committing too early is roughly equal to the cost of committing too late, which argues for buying flexibility over buying the cheapest unit price.
Altman's discovery thesis runs into Intel's ceiling. Discovery in biology, materials, and energy is compute-hungry, and the silicon to support it grows about 2x, not 10x. Discovery work will compete for the same constrained capacity as everything else, and the labs with long-term supply deals will get it first. More important, the bottleneck Altman leaves out is validation. Math is cheap to check. Lab and clinical results are not. Whoever runs the fastest validation loop captures the value, not whoever owns the model.
Verification is the common bottleneck. Modal makes code review optional and relies on CI and tests instead, and it runs internal agents that try to escape its own sandboxes. Goodfire's monitors check internal model state rather than what the model says about itself. AAIF standardized the Skills format but did nothing on signing or provenance. In every case the scarce work is proving that an output or an agent can be trusted. The tooling for that is immature, and the standards bodies have not touched it.
Bernhardsson's dismissal of on-prem conflicts with what enterprises are buying. He calls the network perimeter an outdated security model and prices BYOC (bring your own cloud) out of reach, yet the Hugging Face incident shows an attacker's agent pivoting through a customer app on his platform. His answer is restrictive defaults, and that answer applies regardless of where the hardware sits.
Enterprise Implications
Internal skill libraries are now a reasonable long-lived asset. Encoding company procedures as Skills no longer ties them to one vendor's runtime. Governance stays with the buyer: Skills run with the agent's privileges and are exposed to prompt injection, and nothing about the AAIF move changes that.
Plans that assume compute keeps getting cheaper need a second look. Memory and HBM allocation will likely set the real cost of inference. Enterprises with bursty workloads should not sign fixed three-year blocks. Aggregators exist to absorb that mismatch, and their margin is worth paying.
For pharma, chemicals, materials, and semiconductors, the case for investment is the validation pipeline (proprietary experimental data, simulation capacity, fast feedback loops), not model access.
Explainability requirements in procurement will eventually move from "the model explained itself" to mechanistic evidence. That will not happen in 2026. Until then, vendors claiming "auditable AI" are describing research results, not shipping products.
What To Do About It
- Re-base compute roadmaps against a roughly 2x silicon ceiling, with memory as the binding constraint. Model scenarios for both scarcity and a price collapse before signing multi-year commitments.
- Mix capacity sources. Keep a small reserved baseline and buy burst capacity on demand through neutral providers. Test open-weight models before paying for fine-tuning.
- Stand up Skills governance now. Build an internal registry, enforce least-privilege execution, and review third-party Skills the way you review third-party code. Assume no signing standard arrives soon.
- Default agent sandboxes to deny, with no network egress unless explicitly granted. Red-team your own agent deployments for escape and lateral movement.
- Customer question for R&D clients: "If a model handed you a valid hypothesis tomorrow, how many weeks would it take to confirm it?" The answer shows whether they have a model problem or a validation problem.
Risks and Blind Spots
The Altman source is a 150-word CEO clip with no named results. The Goodfire figures arrived secondhand, and even the author's name in the transcript is unreliable. Read the paper before citing it to a client. Intel's 2x ceiling comes from a company with reasons to stress scarcity and its own strategic value. Bernhardsson's views on BYOC and on-prem are a vendor's views. Nobody here addresses power availability as a constraint independent of wafers, or regulatory exposure for agents acting with delegated privileges. Willison's prior writing flags that exposure, and today's AAIF news leaves it open.