a16z's $1.1B Machine Age Fund and the cost of agentic AI

AI infrastructure is entering a new phase. The a16z Machine Age Fund makes the physical cost of agentic AI impossible to ignore. But the most important economic question for businesses is not only how to obtain more compute. It is how much operational work should become AI work in the first place.

On August 28, 2026, Andreessen Horowitz announced its 1.1 billion dollar Machine Age Fund for the physical infrastructure beneath AI: chips, memory, networking, storage, data centres, robotics and appliances. Its central observation is that as AI moves from chat to reasoning, coding and other knowledge work, both the amount of work and its token intensity increase by orders of magnitude.

Read the Machine Age Fund announcement

That is a supply-side thesis. More useful AI creates demand for more tokens, compute, memory, networking and power.

DataInbox addresses the same development from the demand side.

The cheapest AI operation is the one that doesn't need AI

An agent-heavy architecture asks a model to read sources, reconstruct the latest situation, interpret rules and decide what should happen every time work arrives. The model may be capable of doing all of that. Capability does not make it the right architecture.

Many business operations are deterministic:

• A status changes.
• An incoming event is validated against a schema.
• Two identifiers are mapped or merged.
• A policy condition is checked.
• A message is routed.
• A permitted state transition is applied.
• A change is stored as a diff.
• A current projection is updated.

These operations do not need open-ended reasoning. Turning each one into an agent call adds tokens, latency, variability and infrastructure demand without adding intelligence.

Put intelligence where intelligence is required

The sharper division of labour is:

LLMs for intelligence. DataInbox for operational state and execution.

Use an LLM when the task requires language understanding, ambiguity resolution, synthesis, judgement or generation. Use DataInbox to receive the business event, validate it, maintain current state, apply explicit rules, resolve permissions and provide the smallest useful projection to the model.

The operating sequence becomes:

1. Business event arrives.
2. DataInbox validates and processes deterministic changes.
3. The Inbox resolves current state, history, rules and permissions.
4. AI is called only when reasoning is required.
5. The model receives the permitted context for this task, not unrestricted access to every source.
6. The proposed result returns to the Inbox for validation, execution and evidence.

This does not make AI less important. It protects expensive intelligence from work that ordinary computation can perform more reliably.

Three architectures, one business assignment

This distinction should be measured rather than merely claimed. A useful benchmark gives three architectures the same business assignment:

1. Agent-heavy: an agent reads sources, reconstructs context and decides again for every event.
2. RAG plus agent: retrieval reduces the search space, while the model still interprets the retrieved material and reconstructs operational state.
3. DataInbox plus agent: events, state, rules and permissions are processed deterministically; the model receives a task-specific projection only when reasoning is needed.

The benchmark should report more than token price:

• Model calls per 10,000 business events.
• Input and output tokens.
• End-to-end latency.
• Deterministic operations completed without a model.
• Percentage of operations that genuinely required AI.
• Cost per 10,000 business events.
• Reproducibility, policy compliance and traceable evidence.

The central metric is not simply how cheaply a model call runs. It is what percentage of business operations needed a model call at all.

RAG helps, but retrieval is not operational state

Retrieval can reduce the amount of source material placed in a prompt. It does not automatically establish what is true now, which rule applies, whether an action is permitted or whether two records describe the same business object.

If the model must infer those answers again for each request, the organisation is repeatedly buying a probabilistic reconstruction of its own operation.

DataInbox keeps those answers attached to the business event and its current state. The AI receives a governed projection rather than a pile of source access.

The economic case becomes stronger as AI does more work

The Machine Age thesis is that increasingly capable AI will require a much larger physical infrastructure base. Our architectural inference is that businesses need an equally deliberate way to control demand.

DataInbox does not promise that every AI call can disappear. It makes the boundary measurable:

• Deterministic work remains deterministic.
• Current business state is maintained once instead of reconstructed repeatedly.
• Context is selected for the task instead of copied wholesale.
• AI is reserved for the steps where intelligence creates value.

That is why a business information and execution layer is not an optional optimisation when a company enables AI. It is what prevents every event, lookup, rule and state change from becoming another model workload.

The cheapest token is not a discounted token.

It is the token the operation never needed.

Explore the AI Economics Dashboard