Current · technology · operating model
AI Margin Compression
The defining motion of 2025–2027 — where the cost of producing intelligence outruns its price.
Momentum
↑ Accelerating
+0.42 velocity
Belief
61 / 100
consensus building
Maturity
Consensus
where on the adoption curve
Numen reads this Current
Markets are pricing a story the data is only now starting to confirm. GPU capex commitments by the hyperscalers crossed three hundred billion in the trailing twelve months. Software companies — the buyers of that capacity — are showing the first ARR-to-cost spread compression in a decade. The Current is no longer emerging; it is what the smart capital is already repositioning around.\n\nThe believer side is dense. Equity strategists at Goldman, JPMorgan, and Morgan Stanley have all published margin-compression theses in 2026. NVIDIA's gross margin guidance softened in the Q1 print. The CFO survey from Brainyard shows 58% of software CFOs flagging AI infrastructure costs as a top-three margin concern.\n\nThe skeptic side is narrower than it used to be. The principal counter-argument — that inference costs collapse faster than capex amortization — is structurally true but operationally lagged. Three years of GPU depreciation will compress margins before the inference-curve catches up. The Current accelerates through that gap.
Believers
Hyperscaler capex (trailing 4Q sum)
$314B
SaaS 100 ARR / cost-of-revenue spread (YoY change)
-310 bps
Brainyard CFO Survey — AI as top-3 margin concern
58% (n=412)
NVDA gross-margin guidance (forward Q)
74.8% → 73.5%
Skeptics
H100 spot rental price (weekly)
$2.85/hr
Leading actions
01
Stewards are repricing software multiples on a 12–24 month margin-compression overhang. Public-market PMs are rotating from picks-and-shovels (NVDA, AMD) toward application-layer SaaS that has already absorbed the compression hit.
02
CFOs in software companies are accelerating BYO-LLM architectures to convert AI costs from a vendor margin drag to a customer pass-through line. Anthropic and Azure OpenAI customer-keyed deployments crossed 40% of enterprise AI seats in Q1.
03
Private equity buyouts of compute-heavy AI vendors slowed sharply in 2026. The LBO model assumes margin expansion, not compression — buyout funds are waiting for cost-curve clarity.
Methodology
Composite of: hyperscaler capex announcements (50% weight, quarterly), software ARR / cost-of-revenue ratio across SaaS 100 (30% weight, monthly), GPU spot pricing (10% weight, weekly), CFO survey margin-concern frequency (10% weight, monthly). Belief = market-pricing composite (60%) + survey signal (40%). Maturity classified by velocity slope + believer-skeptic spread; transitioned to 'consensus' Q2 2026 when major sell-side desks aligned. Refresh: weekly methodology, daily readings. Sources cited inline on components.