THE DALTARIS RADAR · October 5, 2026

The model is a commodity. Everything else is the business.

Salesforce admits the model doesn't matter. Anthropic measures the delegation gap. DeepInfra prints money on the pipes. Five stories on where value actually accrues in the agent era — and what practitioners should do about it.

#1

Dreamforce 2026: Salesforce bets the company on your data being correct

What happened: The 15 September keynote declared the "agentic enterprise": AIforce, a live interface layer (Claudeforce MCP server in open beta, Slackforce, Agentforce Coworker with 100,000 activations in 35 days); Koa, a CRM reasoning model built on NVIDIA's open Nemotron (pilot, enterprise GA winter 2026); seven job-ready agents; and a governance plane (Trusted Enterprise AI Harness, Salesforce Guardian) shipping alongside the agents. 7 billion agent work units and counting.

Why it matters: Salesforce gave up on the model — Gemini and Claude both run in its stack now. The pitch is grounding plus governance, which makes the agentic enterprise a pure data-quality bet (Forrester's read). And agents inherit the user's permissions, so your access-control gaps are now agent blast radius.

What to do: Before you celebrate what your agents can do, audit what they can read. The team that owns data correctness just became the team that owns agent safety — make sure it's the same team.

#2

Anthropic quantifies the delegation gap: 60% AI usage, 0–20% delegation

What happened: Anthropic's 2026 Agentic Coding Trends Report finds developers now use AI in roughly 60% of their work — but can fully delegate only 0–20% of tasks. Engineers delegate what's easy to verify or low-risk and pull agents back the moment a task turns design-heavy or ambiguous. Long-running agents are stretching from minutes to hours (one team ran a single 7-hour session across a 12.5M-line codebase), and multi-agent systems are predicted through 2026.

Why it matters: The bottleneck has moved off writing code and onto clarity about what to build. Delegation needs persistent context, testable outcomes, and explicit constraints: three things a prompt rarely provides, and all three are data/infra problems. A seven-hour agent run on a vague spec doesn't save time; it costs days. And with ~27% of AI-assisted work not existing before, AI isn't shrinking backlogs, it's expanding them — the new scarcity is judgment about what deserves to exist, grounded in evidence (five independent summaries converge).

What to do: Treat intent as infrastructure: persistent, versioned specs agents can execute and verify against, not prompts that vanish after the session. And before you delegate, audit the foundations — the 15-point Agent Readiness Audit (daltaris.com/audit/) covers data correctness, permissions, governance, evals, and operations.

#3

DeepInfra crosses $100M ARR: the model is electricity, the pipes are the business

What happened: Inference cloud DeepInfra announced (September 29) a $105M+ annualized run-rate — roughly 14x year over year — and 22 trillion tokens per week, tripled since May. It runs 200+ open-source models on one vertically integrated platform; a $107M Series B is funding footprint expansion including a first data center in Canada.

Why it matters: This is the lab-to-production shift in dollars. When 200 models are swappable on the same rails, the model is a commodity — consistent with the Dreamforce read — and the economics move to throughput, reliability, and cost per token.

What to do: Start treating inference spend like warehouse spend: instrumented, budgeted, and owned by someone. The model choice is becoming the least interesting decision in your stack; the data you feed it and the cost of running it are the whole game.

#4

OpenAI's DevDay: agents get their own computers, the frontier stays parked

What happened: At DevDay, OpenAI shipped Dots — agents with their own cloud computers — and GPT-6.1 Sol, a cheap high-volume model. Conspicuously absent: a new frontier model.

Why it matters: The lab is monetizing the present while the future is paused for safety. That gap — between lab caution and product speed — is the whole story of AI right now. Agents are a shipping product; the next capability jump is not on any committed timeline.

What to do: Build on what's shipping (agents, cheap inference), not on a frontier timeline nobody will commit to. If your roadmap assumes the next big model, rewrite it around what exists today.

#5

DeepSeek open-sources its AI stack for Huawei chips: the CUDA alternative takes shape

What happened: On September 29–30, DeepSeek open-sourced the Ascend versions of its AI software stack: TileLang for Huawei chips, plus five infrastructure libraries (DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect) — each mapping one-to-one to its NVIDIA versions. DeepSeek positions TileLang as easier to program than CUDA.

Why it matters: The April story was the model running on Huawei; this is the toolchain. Compilers, kernels, and communication libraries are the actual CUDA moat — chips without software are just silicon. If developers can build on Ascend without NVIDIA's software stack, the decoupling runs deeper than hardware. Watch the toolchain, not the leaderboard.

What to do: Map your single-vendor moats — not just chips, but the software layers (CUDA, proprietary APIs) your pipeline can't easily leave. Then ask how much of your ML stack assumes CUDA specifically. Optionality is the hedge.

SOURCES

✦ The Daltaris Radar lands weekly. Follow @Daltaris on X to get each issue first.