Executive Summary
Today's sources agree that the hard limit on enterprise AI is no longer model capability. Performance now comes from fitting the model to its environment, whether that environment is silicon, a business process, or a permission boundary, and the gains do not carry over to environments that were never fitted. Huawei and DeepSeek are the clearest hardware example: a frontier‑competitive software stack running on chips roughly five years behind Nvidia. Tenex shows the same pattern inside enterprises: 90% have rolled out an assistant but only 5% run non‑engineering agents in production. Altman puts OpenAI's moat in integration, data, and infrastructure scale rather than model lead. For Fred's clients, this means advisory value has moved from model selection to integration architecture: access control, process definition, owned context, and portability.
What Changed
SemiAnalysis has produced the first strong public evidence that a non‑Nvidia stack (chip, compiler, scale‑out fabric) can reach frontier‑level software efficiency. Stock Ascend 950PR kernels reach 99.7% of the compute ceiling on BF16 GEMM, HBM bandwidth utilization matches H100 at scale, and DeepSeek now maintains Ascend ports of six core kernel repos. Absolute performance still sits between A100 and MI300X. The gains also do not travel: there is no DeepGEMM on x86, ARM dependencies are hardcoded, and there is no NPU+x86 CI. When upstream DeepEP maintainers declined the Ascend port, it moved under DeepSeek's own GitHub org. That is a forked supply chain in progress.
On the demand side, Tenex's figures from roughly 100 enterprise accounts give a hard funnel: 90% have rolled out an LLM, 60% have done real training, 50% have working coding agents, and 5% have non‑engineering agents in production. Output per engineer has roughly doubled since December, and token spend runs 1x to 2x salary.
Cross-Expert Synthesis
Co‑design is the source of advantage, and it does not generalize. DeepSeek engineers tune kernels, KV‑cache layouts, and scheduling against one supernode configuration. Tenex tunes agents against one standardized process and keeps the "why" in a git‑tracked harness. Altman argues agents waste compute parsing interfaces built for humans, and that this justifies purpose‑built hardware. All three point the same way: the performance gain lives in the fit, and the fit is local. That explains why Tenex is bearish on generic vertical agent products and why Ascend is not a CUDA drop‑in.
Cost per task is falling; total spend is not. Altman expects Jevons‑style demand growth as cost per task falls. Tenex sees the same effect from below: heavy token use ("token‑maxing") is harder to sustain because per‑task costs keep dropping. These views fit together. Unit costs collapse, consumption expands, and budgets set as fixed compute allocations break. SemiAnalysis adds that, given China's cheaper power, TCO and performance per watt will settle long‑run hardware competitiveness more than benchmark FLOPs will. The economics are moving to cost per outcome.
The two sources disagree on the value of the frontier. Altman uses unreleased math proofs as evidence of an internal lead. Tenex doubts a model 10x smarter is worth paying for on most economically valuable work. Both can be true: frontier gains matter for discovery workloads, while enterprise value is gated by diffusion. For most of Fred's clients, the Tenex view should drive spending and the Altman view should drive watch lists.
Verification is the new form of labor. Altman says heavy AI users are busier, not less busy, and admits sycophancy persists even with strict custom instructions. Tenex runs Codex as a reviewer and has humans redirect background agents. Labor shifts toward direction and checking, and a model that agrees with everything weakens the checking step.
Enterprise Implications
The 45‑point gap between coding agents (50%) and non‑engineering production agents (5%) is the addressable market. It is not a model problem. It is two unsolved infrastructure problems: per‑user data permissioning across every assistant, and processes too inconsistent to evaluate. Neither is solved by any frontier lab's roadmap.
Sovereign or non‑Nvidia inference is now a credible discussion for clients exposed to China or under data‑residency requirements, but only for DeepSeek‑shaped workloads on DeepSeek‑shaped hardware. Any proposal that assumes general Ascend readiness on x86 servers is wrong today.
Systems of record (Salesforce, SAP) are safe because rip‑out risk is too high. Mid‑tier SaaS and thin vertical agents are exposed. Software portfolio reviews should sort vendors by data or compliance moat, not by feature set.
What To Do About It
- Lead with permissioning. Design an access gateway (identity‑scoped retrieval, audit, revocation) before any broad agent rollout. This is the CISO's top pain point, and few vendors have productized it.
- Require process standardization before an agent build. No defined workflow means no evals and no production agent. Make this an explicit gating stage.
- Budget per outcome, not per seat or GPU‑hour. Model token spend as a ratio to the labor it supports, and plan to recalibrate quarterly as unit costs fall.
- Own the context layer. Keep decision logs, plans, and eval sets in the client's repositories, model‑agnostic and portable. That is the asset that survives vendor churn.
- Customer question: "Of the agents you have deployed, how many touch a non‑engineering process in production, and who approved what data each can read?" Most clients will answer "none" and "nobody," which makes the engagement scope obvious.
Risks and Blind Spots
Tenex's funnel comes from one services firm's pipeline, which is likely biased toward early adopters. The true production rate may be below 5%. Altman's capability and disease‑treatment claims are unverifiable and serve his commercial interest. Treat them as positioning. SemiAnalysis's five‑year hardware gap and its TCO thesis lack the promised TCO comparison. Watch for it before advising on any Ascend decision.
The structural risk is dependency. Paine's account of World War I coalitions has no AI content, but it supplies a planning heuristic: when partners cannot act alone, nobody can exit, even after aims diverge. Multi‑party AI programs (lab, data platform, integrator, business units) run on different success metrics. Clients without independent capability, meaning a portable model interface, owned evals, and a second vendor path, will find they cannot leave a program that has drifted from their objectives. Set exit criteria before signing, not at month 18.