AI Agent Budgets for 2027: What to Fund After the Q3 2026 Numbers

For the other companies, the period must not be silently substituted. As of early October 2026, Alphabet, Amazon and Meta calendar-Q3 CapEx figures were not yet in the evidence pack. Use Alphabet’s Q2 2026 earnings-call materials and Meta’s Q2 2026 results as earlier-period context only — not as substitutes for calendar-Q3 disclosures. Each company’s calendar-Q3 CapEx amount and any AI-specific allocation are to be verified.

For buyers, the practical change is a stricter distinction between supplier capacity and purchased value. For builders, it is the need to design for variable model prices, latency and availability without assuming that rising supplier CapEx will resolve workflow failures. Abundant, cheaper inference is a plausible scenario—not an observed outcome established by these materials.

Funding priorities for 2027 (eval, orchestration, human approval gates, infra/inference cost) vs what to cut

Fund evaluation first. Build task-specific test sets from permitted, representative work; include routine cases, exceptions and adversarial inputs. Score the complete outcome—correct action, justified refusal, escalation and downstream correction—not just the fluency of an answer. Reserve capacity for regression tests when prompts, models, tools or source data change. Without a baseline and repeatable measurement, a larger deployment buys exposure rather than evidence.

Fund orchestration and state management where workflows need them. Agents that use tools should have bounded permissions, explicit handoffs, recoverable state, retries and traceable records of tool calls and resulting actions. This is preferable to paying for elaborate multi-agent designs whose additional steps cannot be tied to better task completion. The relevant comparison is against a simpler workflow, including one that is deterministic or human-operated.

Fund human approval gates for consequential actions. Define which actions require review before execution, what information the reviewer sees and how rejection or reversal works. Approval is a control only if it is placed before the consequential step and reviewers can make an informed decision; a nominal sign-off after execution is not equivalent.

Fund inference-cost and infrastructure measurement, then optimize. Track cost per successfully completed task, including failed attempts, retries, tool calls, review time and exceptions. Test routing to smaller or less expensive models where quality permits, alongside caching and limits on unnecessary context. Budget for observability and supplier portability where they reduce operational risk. Any assumed saving from these techniques is to be verified in the target workflow.

Cut or defer broad compute reservations unsupported by measured demand; general-purpose model training without a specific advantage over available services; duplicate orchestration layers; and pilots that display impressive conversations but cannot demonstrate safe completed actions. Hyperscaler CapEx alone is not a reason to hoard capacity, nor proof that an enterprise should abandon every case for dedicated infrastructure. Those decisions require its own workload and contract evidence.

Kill criteria

Stop, redesign or contain an agent deployment when it cannot:

• beat an agreed baseline on task completion and error severity, using a representative evaluation set;
• enforce permissions and approval before consequential actions, with a workable rollback or remediation path;
• reconstruct what data, tools and actions produced an outcome, subject to applicable privacy controls;
• report the full cost per successful task, including human review and recovery, or stay within its approved cost envelope; or
• maintain acceptable results when a model, tool, data source or supplier changes.

The thresholds are organisation- and workflow-specific: their numeric values are to be verified before funding is released. A project should not pass because a supplier increased CapEx, or fail solely because another supplier did not disclose an AI-specific spending split.

Evidence pack (checklist)

A FY2027 decision file should keep the following together:

• Period-matched supplier evidence: company, fiscal versus calendar quarter, cash-flow or CapEx definition, direct investor-relations URL and disclosure date. Keep Microsoft’s March-ending FY2026 Q3 separate from calendar-Q3 2026 claims.
• Missing-data register: Alphabet, Amazon and Meta calendar-Q3 CapEx, AI-specific allocations and comparable definitions—to be verified against their eventual disclosures. Record that the supplied MMARW page extraction did not expose its body or figures.
• Workflow case: task boundaries, permitted data and actions, baseline method, evaluation set, severe-failure examples and owner of exceptions.
• Operating case: model and tool-call traces, approval records, latency, cost per successful task, review burden and recovery costs.
• Decision record: pre-agreed pass, redesign and stop thresholds; evidence of performance after changes; and contract assumptions tested against actual usage.

This pack separates what suppliers report from what an agent can deliver. The former informs procurement risk; the latter earns a place in the budget.

Sources

• MMARW, Hyperscaler Capex Q3 2026: What Microsoft, Alphabet and Meta Spending Tells Buyers of AI Services — buyer decision frame for CapEx vs rentable capacity.
• Microsoft Investor Relations, FY26 Q3 cash flows and FY26 Q3 performance — March-ending quarter (not calendar Q3).
• Alphabet, Q2 2026 Earnings Call materials — latest CapEx guidance/mix in the evidence pack; calendar Q3 to be verified on release.
• Meta, Reports Second Quarter 2026 Results — CapEx and guidance; calendar Q3 to be verified on release.
• Amazon IR quarterly results — verify calendar-Q3 CapEx when the release posts; do not invent a Q3 address.