The $100B market that doesn't exist yet, and the invisible tax it will charge

Last month’s agent infrastructure bill across my ventures came to roughly ₹2,47,000. Model inference was about 38% of it. Orchestration was 11%. Observability was 9%. Attestation, agent identity, and inter-agent communication combined came to under 3% — because those categories barely exist yet as paid line items, and most of what I use is home-grown.

Next to the total I wrote a question in the margin: how much of this bill do I pay in 2029?

The honest answer, if Chapter 9’s prediction on the agent dependency layer is right, is not much less — and in a completely different shape.

Infrastructure markets do not win by being cheap. They win by becoming invisible. And invisible infrastructure always taxes what runs on top of it.

PRED-003 — By December 2030, the combined market for agent orchestration, trust and attestation, agent identity, monitoring, and inter-agent communication will exceed $100 billion in annual revenue. This dependency layer will be larger than the agent application layer.

Confidence: 3 out of 5.

The chapter builds this prediction from a ratio argument. Combined revenue from internet infrastructure — CDN, DNS, SSL, cloud — exceeds $300 billion per year today. Infrastructure markets historically take five to seven years to materialise after the application layer emerges, then surpass it. The agent application layer is emerging now; the infrastructure layer is nascent. $100 billion by 2030 is the structural implication of that ratio, not a bottom-up extrapolation from 2026 revenue. Current agent infrastructure revenue is negligible. That is the point.

The ratio argument is solid. It is also incomplete. It tells you the total. It does not tell you which companies capture it.

Here is the part Chapter 9 does not answer, and the part I want this issue to answer: what shape does the tax take?

Every durable infrastructure layer in commercial history becomes a per-unit tax on the layer above. Visa and Mastercard charge 2–3% per swipe. AWS charges per GB-month of storage and per GB of egress. Cloudflare charges per request or per GB of traffic. Verisign charges per domain-year. Stripe charges per transaction. SSL certificate authorities charge per certificate-year. Notice the pattern. The rate moves; the shape does not. Shape is a bet on what the customer will grow into, not what the customer will pay today. Shape is harder to change than price.

Chapter 2 of the book identifies the five dependency-layer components. I am going to name my best bet on the most plausible tax shape for each one, and I will stand by the bet:

1. Orchestration platforms — per-agent, per-month. Like EC2 instances but for agent lifecycles. The orchestration platform meters as long as the agent runs. The customer’s bill grows as their agent fleet grows, which is exactly the curve the platform wants to ride.

2. Trust and attestation services — per-attestation-event. Like Verisign’s per-domain pricing or a certificate authority’s per-cert fee. The cost per event is a rounding error. The volume at enterprise scale — millions of attested actions per month — is the business.

3. Agent identity systems — per-agent, per-year, tiered. Like TLS certificate issuance, with a tier for “verified” vs “anonymous.” Identity is the category that moves from free to subscription the moment an enterprise has to attest to which agent did what.

4. Monitoring tools — per-event-ingested or per-metric-series. Like Datadog today, but with at least an order of magnitude more events, because agents generate events the way code does — continuously, not at human cadence.

5. Inter-agent communication protocols — per-message or per-byte. Like SMS APIs or, at the other end of the spectrum, SWIFT fees for financial messaging. The shape depends on whether the winning protocol positions itself as a messaging layer or a settlement layer.

Back to my own bill. In April 2026, four of those five components are either home-grown in my stack (attestation, identity, inter-agent comms) or bundled into an existing spend (monitoring, inside the cloud provider’s line items). PRED-003 is the prediction that those home-grown lines become real line items on every agent-powered venture’s bill by 2030 — with one of the five shapes above. The $100B total is the ratio implication. The shape is where the business is built.

Which means the bet at this layer is not about revenue. It is about shape. The companies that pick a shape matching the customer’s natural growth curve — per-agent, per-attestation, per-event — will compound. The companies that pick a shape matching the provider’s comfort — flat-rate, per-seat, per-admin — will be eaten by the per-unit players the second volume shows up.

The published falsification trigger:

If by December 2030, identifiable agent infrastructure market revenue is below $30 billion, or if agent application revenue clearly exceeds infrastructure revenue by more than 2x, this prediction is wrong.

That is the stated stake. The realistic failure mode is not “infrastructure does not grow.” It is bundling. If AWS and Azure absorb all five components into their existing line items — orchestration as an EC2 feature, attestation as an IAM extension, identity as a Cognito feature, monitoring as CloudWatch, comms as EventBridge — the dependency layer does not emerge as a distinct $100B market. It disappears into the hyperscaler revenue lines. The underlying thesis (infrastructure captures the value) is still correct, but PRED-003 as stated fails because no standalone market is measurable.

Bundling is the more dangerous failure mode, because it is the one where I look wrong while being right.

The other failure mode is open-source commoditisation. If the critical protocols — particularly attestation and inter-agent comms — settle into OSS commodities that no one can tax (the way Linux hollowed out proprietary operating system revenue), two of the five components evaporate as paid categories. The other three probably still hit the ratio, but the headline number tightens.

If you are building, buying, or pricing agent infrastructure at any layer, reply to this email with one thing: your per-unit pricing shape. Not the rate. Just the shape. Per-agent? Per-event? Per-seat? Per-inference? Per-byte? Flat-rate?

I will publish an anonymised distribution of shapes on atin-agarwal.com/predictions/pred-003-dependency-layer-market/, refreshed quarterly through 2030. If the distribution clusters around flat-rate seat pricing by Q4 2027, the invisible-tax thesis is wrong, and infrastructure has been captured by SaaS business models rather than platform business models. That is the quiet leading indicator for whether PRED-003 is on track or off.

If you are an infra founder: pick your shape before your pricing page goes live. The shape is harder to change than the rate, and your first ten enterprise contracts will lock you into it for a decade.

If you are an app founder buying agent infrastructure: read your contracts for the meter, not the total. The contract that meters per-inference or per-agent is the contract that owns you at scale. The contract that meters per-seat is the contract you will grow out of. Pick your dependency the way you would pick a cloud region — with full knowledge that it is expensive to move later.

If you are an investor at this layer: the thesis bet is about shape, not revenue. A $3M ARR company with the right shape compounds into a category winner. A $30M ARR company with the wrong shape is an acquisition at 3x revenue.

If you are a builder: the four components where I currently run home-grown — attestation, identity, inter-agent comms, and monitoring — are the four where someone else is going to build the business I would otherwise have to build myself. If you are looking for a startup idea with a structural tailwind, the list is right there.

This issue is drawn from Chapter 9 of The AI Agent Economy — 15 falsifiable predictions with dates, numbers, and explicit triggers for being proven wrong. Pre-order on Kindle — $9.99. Release July 1, 2026. atin-agarwal.com/books

Read the full PRED-003 entry on the public tracking page → atin-agarwal.com/predictions/pred-003-dependency-layer-market/

Previous issue: Issue 02 — Five job titles that don’t exist yet, but people are already doing the work → SUBSTACK-002 Next issue: Issue 04 — The AWS of agents will not be a model company → SUBSTACK-004

The five dependency-layer components are named as they appear in Chapter 2 and PRED-003: orchestration platforms, trust and attestation services, agent identity systems, monitoring tools, inter-agent communication protocols. No substitutions.