Actionable insights — Microsoft: The Open-Weight Hedge

1. Decompose a headline backlog by its anchor customer before crediting the visibility

1. Take the reported remaining performance obligation (RPO) / backlog and compare it to annual revenue — that ratio is the raw visibility claim.
2. Ask management (or read the call) for the figure excluding the largest counterparty. A backlog whose growth survives that exclusion is a genuinely broadening business; one that doesn't is a single-customer receivable in disguise.
3. Then look at the sequential composition: who added commitments this quarter? Concentration that stops worsening at the margin is a different risk profile from concentration that keeps deepening.
4. Cross-check against capacity commentary — a backlog is only bankable if the supply to deliver it is arriving.

• Whether the ex-anchor growth rate is disclosed again next quarter; renewed sequential additions from a single lab; and any gap between backlog growth and the capacity being brought online.

2. Diarise accounting changes that move real spending out of the reported CapEx line

1. Read the guidance section for depreciation-life extensions and lease-classification changes — both flatter reported results without changing a dollar of cash committed.
2. A useful-life extension slows depreciation, lifting reported profit; a shift from finance leases to operating leases removes commitments from the CapEx figure entirely.
3. Note the effective date and build your own bridge: from that period on, compare capacity/commitments (leases signed, megawatts, RPO) rather than year-over-year CapEx.
4. Treat the change as neutral to cash but negative to comparability — and be sceptical of any subsequent "CapEx growth is moderating" narrative built on the new basis.

• The first FY27 print where CapEx growth appears to decelerate; disclosure of operating-lease commitments in the footnotes; depreciation expense growing slower than the asset base.

3. Locate the moat one layer above the commoditizing technology

1. Assume the headline technology gets better and cheaper for everyone — then ask what in the stack does not compress: proprietary data, distribution, governance, or the ability to turn an answer into an action.
2. Quantify the data asset in operational terms (events processed, customers, years of history) — that's the piece a competitor cannot buy at any price this year.
3. Look for a router/orchestration design that mixes cheap first-party capability with expensive third-party capability only where required — the cost curve, not the benchmark score, is the strategic result.
4. Apply the same test to the marketplace: if the platform wraps every model with one set of APIs, governance and billing, its economics are indifferent to which model wins.

• The escalation ratio drifting (more traffic to the expensive third-party model means the first-party cost curve isn't holding); first-party silicon efficiency claims (40% better perf/watt on Maia 200) showing up in cloud gross margin.

4. Read a coalition's signatory list — and its absentees — as a positioning map

1. When an industry letter or alliance forms around a policy question, list who signed, who joined late, and who stayed out.
2. Map each name to the business model the policy protects or threatens — the alignment is usually economic, not philosophical.
3. Treat the holdout's stated alternative as the real forecast of where regulation may land, since a dissenter has to be specific.
4. Then translate the plausible outcomes into a company-level sensitivity: which revenue line breaks if the restrictive version passes?

• Whether legislation targets distillation narrowly or open weights broadly; chip-control expansions; and whether the holdout's conditions (mandatory safety testing) appear in draft rules.

5. Price "sovereignty" as a deployment feature, not a slogan

1. For any enterprise-AI vendor, check which deployment topologies it actually supports: public cloud, cloud-connected, and fully disconnected/air-gapped.
2. The disconnected option is what unlocks regulated buyers (banks, governments, healthcare, defence) who cannot route queries to a foreign-hosted API — it is a licence to bid, not a marketing line.
3. Check who supplies the model and who supplies the wrapper: an open-weight partner plus the vendor's own governance/security layer is how a platform sells control while keeping the workload.
4. Follow the capital commitment — a multibillion-dollar purchase of a partner's regional capacity indicates the vendor believes the demand is durable and jurisdictional.

• Named regulated-sector wins in Europe; whether disconnected deployments carry different (usually better) pricing; and any policy change that restricts which model families can be deployed locally.

6. Strip investment gains out of "adjusted" earnings — including the ones management kept in

1. Reconcile GAAP to non-GAAP line by line and identify every equity-investment mark, then check which were excluded and which were not.
2. Selective exclusion is the signal: if one stake's gain is removed and a larger one is retained, adjusted EPS is carrying non-operating income by choice.
3. Re-run the beat on a fully stripped basis before comparing to consensus.
4. Apply the same test to any company holding stakes in private AI labs, where marks are large, lumpy and driven by other people's funding rounds.

• Consistency of the exclusion policy quarter to quarter; the size of private-lab marks relative to the EPS beat; and whether a down-round would be excluded as readily as an up-round.