Executive Summary
Today's sources point the same way. AI is quickly automating work that can be checked against a measurable result. Low-level infrastructure work is the first commercially important target, which weakens the lock-in that held up hardware and model pricing. Dwarkesh Patel now puts automated top-tier AI researchers at a 2028 median. Nate B. Jones reads Jensen Huang's exit from Nvidia as the loss of the industry's main dealmaker, at the same moment AI-assisted porting is softening CUDA. Matthew Berman's guest calls compute per problem a deciding competitive advantage and treats open weights as the counterweight to lab gatekeeping. For enterprise buyers this means switching costs are falling, supply coordination is getting worse, and agent compute is becoming a strategic input rather than an IT line item. Architecture and contracts should be built for portability now, while sellers still price as if lock-in holds.
What Changed
- Lab self-acceleration has numbers, though the labs supplied them. METR found Anthropic's progress on well-scoped tasks running at 2x last year's rate. Claude now leads 26% of Anthropic's R&D work, up from under 1% in February. OpenAI's heaviest researcher users went from about $1k to $7k+ per day in coding-agent spend over one summer.
- The CUDA lock-in has a measured crack. Anthropic says its Mythos model ported Claude prompt processing to AMD MI355X in about three days with almost no human help. The scope was narrow: one model, one stage, one chip, done by an unusually well-prepared lab.
- Nvidia's deal-making role has no clear owner. Colette Kress is interim CEO and stays CFO. More than $500B of preliminary outside-capital agreements and up to $105B of guarantees on the OpenAI-serving SB Energy site were priced partly on Huang's personal credibility.
- Export-control risk now hits buyers. With Huang gone, Nvidia loses its main voice against restrictions, and the AI Overwatch Act gains momentum.
Cross-Expert Synthesis
These three sources connect, and none of them says so. Patel's argument is that work with a measurable proxy can be optimized directly by RL and parallel agents. Kernel porting fits that exactly: correctness and throughput are checked automatically. The Mythos result is a small, commercial example of Patel's mechanism. Jones's "soft erosion" of CUDA is therefore not a one-off. It is the first visible case of a general effect: moats that rest on scarce engineering labor in checkable domains wear down at the speed of AI capability, not the speed of hiring. That applies to proprietary SDKs and migration tooling as much as to CUDA.
There is a real conflict on supply. Patel expects explosive demand for agent compute: per-researcher spend up 7x in one quarter, labs quadrupling compute every year, labor traded for compute. Jones expects diversification, margin compression and cheaper inference. Both can hold at once, with a falling price per token and a shrinking share of guaranteed capacity, especially if financing stalls and construction slips. Berman's guest supplies the link: if compute per problem decides who wins, access matters more than unit price.
The open-weights hedge has a weakness the sources leave unaddressed. If self-improvement compounds inside closed labs, the gap between frontier and open-weight models could widen, not narrow. Open weights then still serve as a floor for continuity and on-prem sovereignty, but stop being a substitute at the frontier.
Enterprise Implications
Hardware lock-in is now a negotiating position more than a technical fact. A buyer does not need AMD to match Nvidia across the board. Running part of the workload economically on AMD is enough to split orders. The same reasoning applies one layer up, to model providers.
Patel's argument for real-world domains has a direct private-AI consequence. If agents can absorb a Fortune 500 company's full email, Slack and document history in a day, the limit on enterprise AI is no longer model capability. It is how ready, permissioned and governed that institutional context is. The context store becomes the most valuable and most exposed asset, which is a strong argument for keeping it on infrastructure the client controls.
Agent economics will follow the labs with a lag. Plan for agent compute per expert employee to become a material budget line within 18–24 months, and for roles heavy in expertise and context to be contested by agents by 2028–2029 on Patel's timeline.
What To Do About It
- Hold Nvidia commitments for Q4 through Q2 until after the November 17 earnings call. Judge contracts on whether Nvidia can deliver the full package (supply, support and financing) over several years, not on list price.
- Run a bounded AMD or alternative-silicon pilot on a single inference stage. The aim is a credible alternative to negotiate with, not a migration.
- Put a model-abstraction layer in front of every production workflow and qualify one open-weight model per critical workload as a tested fallback.
- For non-US clients, map siting and customer-eligibility exposure under likely export-control legislation before signing hardware commitments.
- Make the institutional-context corpus a funded workstream: permissions, retention, lineage. It is what agents will consume, and what rivals or vendors would most like to hold.
Talk track: "If AI can port a production model to new silicon in three days, which clauses in your current contracts still assume switching is expensive?"
Risks and Blind Spots
Most of Patel's evidence comes from the labs themselves and from well-scoped tasks. A 2x gain on well-scoped work does not show that open-ended research can be automated, and he admits continual learning is unsolved. The porting claim comes from an IPO roadshow, where the incentive is to overstate. Jones's central bet, that no successor can match Huang's deal-making, is a forecast, and a strong permanent CEO announcement would weaken it. Berman's segment is a fragment with no data. None of the sources prices the downside where both theses come true at once: demand surges while financing-driven construction delays tighten supply, leaving enterprises paying less per token but unable to get capacity at any price.