Same Earnings Day, Opposite Stock Moves: An AI Capex Signal Checklist Agent Teams Can Borrow From Big Tech

Same day, opposite stock reaction

On October 28, 2026, Microsoft and Alphabet both reported earnings. Alphabet posted revenue of $127.6B (+25% YoY) with Google Cloud surging 103% to $30.7B, yet shares dropped nearly 8% after hours. Net income fell 25% to $26.2B on an $8.7B equity-investment loss, EPS of $2.12 missed the roughly $3.00 consensus, and quarterly capex jumped from $24.0B to $51.4B, pushing free cash flow to -$9.1B.

Microsoft moved the opposite direction the same day. Revenue hit $92.3B (+19%), Azure revenue was disclosed in dollars for the first time at $33.2B (+48%), operating income rose 21% to $45.9B, and GAAP net income climbed 32% to $36.6B. Capex came in at $39.4B — nearly double year over year, but below the roughly $50B the market expected — and shares rose.

The gap was signal design, not spend size

Both companies are pouring tens of billions into AI infrastructure. What split the reaction was how each company packaged that spending into a signal. Microsoft broke from reporting Azure growth as a percentage only and disclosed the absolute dollar figure for the first time, removing the suspicion that slowing growth was being hidden behind a ratio — and it extended data-center depreciation from 15 to 25 years, which mechanically lowered reported capex. Alphabet did the opposite: an $8.7B equity-investment loss buried its strongest growth signal (Cloud +103%) inside a single EPS line, and the capex spike had no new unit-economics disclosure to explain it.

A capex-signal checklist borrowed from Big Tech earnings for agent teams

Internal AI agent and automation budgets fall into the same trap. Reporting only that infrastructure spend went up gives leadership the same impression Alphabet's investors got: "the number got bigger." Before a budget increase, pin down cost-per-task and target auto-resolution rate as numbers — for example, infra cost per processed case under $0.08, auto-resolution rate above 70%, human-escalation rate under 15%. Without those three lines, there is no way to put capex growth and outcome improvement on the same table.

The failure pattern mirrors Alphabet's case almost exactly. Mixing one-off items — migrations, license switches, model-swap testing — into recurring cost in a monthly report leaves a "cost spiked last month" impression even after the number falls back, and budget owners start doubting the whole project. Go the other way and fail to separate one-off costs at all, and you miss the segment where cost is actually leaking.

The recovery strategy is to split the disclosure. Log recurring infra cost (tokens, inference, storage) and one-off cost (migration, testing, cutover) on separate lines from day one, and require that any quarter-over-quarter rise above 10% in per-task recurring cost be reported alongside whether auto-resolution improved. The rule that matters is blocking a cost increase with no matching outcome gain from rolling into the next quarter without explicit sign-off.

On the operational-checklist side, Microsoft's "disclose the dollar figure for the first time" move is worth copying. Adding absolute values, unit costs, and attribution to a metric that previously showed only a growth percentage is itself a signal of having nothing to hide. Log model ID, feature flag, tokens per task, completion rate, and cost attribution (which feature, which stage) as required fields, and put "cost per feature" and "unit-cost trend" on page one of the quarterly report instead of a raw capex total.

Microsoft's accounting change — stretching depreciation from 15 to 25 years to lower reported capex — cuts both ways. If the useful-life extension reflects a real change in asset life, disclosing it transparently is worth doing; if it is mainly a cosmetic assumption change, the same question returns next quarter. Internal reporting needs the same discipline: when an assumption changes, state the reason and the delta against the prior assumption in the same report.

The continuous-improvement loop only needs two questions asked every quarter. First, is per-task unit cost falling faster than capex is rising? Second, does recurring cost alone — stripped of one-off items — keep moving the project toward its target auto-resolution rate? If either answer is no, don't defer the decision to pause or shrink that feature until the next budget cycle.

Checkpoints to use now

A capex total, by itself, is neither a good signal nor a bad one. Alphabet doubled its cloud revenue and still lost the market on signal design; Microsoft nearly doubled capex and still won the market by disclosing dollar figures and explaining its assumption change. The same principle applies to internal AI project reporting — keep per-task unit cost, auto-resolution rate, and one-off-cost separation on the same table every quarter, and budget approval meetings run on numbers instead of impressions.

References

Alphabet Announces Third Quarter 2026 Results — SEC EDGAR (8-K Exhibit 99.1)

Microsoft Fiscal Year 2027 First Quarter Earnings Release — SEC EDGAR (8-K Exhibit 99.1)