May 25, 2026

Inference Got Cheap. Attribution Didn't.

DeepSeek just made its flagship model discount permanent. Seventy-five percent off. The labs competing with it aren't going to hold the line; they never do. Frontier inference costs have been dropping for two years an...

Inference Got Cheap. Attribution Didn't.

Inference Got Cheap. Attribution Didn't.

DeepSeek just made its flagship model discount permanent. Seventy-five percent off. The labs competing with it aren't going to hold the line; they never do. Frontier inference costs have been dropping for two years and nothing in the market structure suggests that stops.

What that means practically: the cost of deploying an AI agent that talks to customers, recommends products, and drives purchasing decisions is approaching zero for any serious company. This is good for anyone building in the agent space. More deployments. Faster experimentation. Lower barriers for smaller teams.

What it doesn't change: the business infrastructure underneath those agents.


An agent that recommends a productivity tool to 10,000 users and drives 400 signups costs maybe $2 in inference at current rates, less by next quarter. The publisher who built the audience those users came from (the newsletter writer, the community moderator, the AI tool reviewer whose recommendation put their content in the agent's context) doesn't get paid anything, because there's no mechanism to pay them.

That's not a scaling problem. Cheap inference doesn't fix it. It makes it more visible.

The more agents there are, the more purchases happen through agents, and the more purchases happen without the attribution layer that would tell anyone who gets credit. Every one of those transactions is a publisher commission that doesn't exist yet, a merchant insight that never gets logged, a channel that has no visibility into what's actually driving its sales.


Practitioners who've been running agents in production have started naming this. The framework you chose barely matters. Model choice, orchestration layer, whether you went LangChain or bare-API: none of that is what kills production deployments. What kills them is the surrounding business infrastructure: billing, attribution, publisher relationships, commission logic. The AI part got commoditized first. The business part hasn't caught up.

This is the same pattern every technology infrastructure cycle produces. The core technology gets cheap fast. The layer above it (the one that makes the technology useful to a business rather than just to an engineer) takes longer, because it requires coordination across parties who aren't all aligned.

Payment infrastructure for the internet took a decade after TCP/IP was stable. The affiliate layer for mobile commerce took years after the smartphone was in everyone's pocket. The attribution layer for agent commerce is running the same playbook. The agents exist. The business infrastructure doesn't.


Here's what the gap looks like concretely.

A merchant running an affiliate program today has zero visibility into agent-initiated referrals. Their dashboard shows clicks, cookies, last-touch attribution. It doesn't show the AI advisor that recommended their product in a chat session, the MCP tool that surfaced their offer in an agent workflow, or the AI-assisted comparison that made the difference on a high-value purchase. Those events happened. They drove revenue. They're invisible.

A publisher who builds an audience around recommending software tools (a newsletter, a review site, an AI assistant with real user trust) has no mechanism to earn commissions when their content influences an agent recommendation. The agent read their output. The agent made the recommendation. The merchant got the sale. Nobody paid the publisher.

At fifty cents per thousand API calls, this problem scales fast. Cheaper inference means more agents. More agents means more agent-initiated purchases. More of that means a larger stack of unattributed revenue that nobody can see.


The attribution layer that solves this isn't cheaper to build because inference got cheaper. It requires different infrastructure: server-side event logging instead of browser cookies, signed attribution tokens that persist across agent sessions, programmatic payout rails that don't assume a human is initiating the commission review.

None of that is complicated. It just has to be built, and it has to work with how agents actually operate, not how human shoppers do.

The good news is that the pressure to build it is accelerating at the same rate as inference costs are dropping. Every month that more agents go into production is another month of unattributed revenue accumulating in publisher accounts that will never see it, and in merchant dashboards that have no idea what's driving their growth.

That pressure lands somewhere eventually. The question is whether the infrastructure exists when it does.

Syndicate Links is the attribution layer for agent-initiated commerce: MCP server, x402 token signing, programmatic publisher payouts. The docs are at syndicatelinks.co.


DeepSeek's permanent pricing announcement was reported by Bloomberg on May 23, 2026. The prior 75% discount had been in place since January 2026. Competing frontier labs have responded with their own pricing adjustments across Q1 and Q2 2026.