AI stock valuations may have peaked. Trivariate Research thinks these shares can keep working

Trivariate Research expects AI-stock valuation multiples to decline, potentially permanently, as higher interest rates and AI-policy risks weigh on the sector, but argues revenue growth and ongoing earnings upgrades can offset the compression. The Fed raised its policy rate to 3.75%-4.00% after strong jobs and higher August inflation, increasing financing pressure for debt-heavy AI infrastructure companies. Trivariate favors Iren, Nebius, CoreWeave, AMD, Marvell and Oracle; AMD rose 9% on the day and is up 184% in 2026, while Marvell is up about 200% year to date.

Analysis

The key dispersion is between asset-light AI beneficiaries and balance-sheet-intensive compute providers. AMD and MRVL can convert incremental accelerator/networking demand into gross-margin and earnings revisions without funding multiyear power and facility commitments; CRWV, IREN, and NBIS must prove that contracted demand earns above their rising cost of capital. A higher discount rate therefore creates a two-stage outcome: near-term earnings-revision support for semis, followed by multiple compression once backlog growth no longer exceeds capex and financing growth.

ORCL is the more defensible infrastructure exposure because its installed enterprise base can monetize AI through database, cloud and application attach rates rather than solely selling scarce GPU capacity. The less obvious risk is that power interconnection delays and local permitting constraints shift bargaining power toward utilities, transmission equipment and existing powered data-center sites, raising the value of already-operational capacity but reducing returns on announced greenfield builds. Watch electrical equipment proxies ETN and VRT for confirmation; their order trends should remain strong if the buildout is real rather than simply compute-reservation inflation.

Over the next 1-3 months, upward revenue and EPS revisions can sustain leadership despite valuation pressure, especially for AMD and MRVL. The 6-18 month risk is that AI workload economics disappoint: if enterprise inference revenue fails to scale, customers will optimize utilization and defer new capacity, leaving leveraged GPU-cloud operators with fixed lease, depreciation and power costs. This thesis is falsified by sustained backlog conversion, improving utilization and declining financing spreads at CRWV/IREN/NBIS; conversely, any capex increase without corresponding contracted-revenue growth should be treated as a negative signal.

Consensus is too broad in treating all AI exposure as one duration trade. Falling multiples need not imply equal downside: semis with earnings revisions can compound through compression, while companies whose equity story depends on perpetual external capital can suffer nonlinear downside if rates or regulation delay capacity by even one or two quarters.

Key Decisions for Investors

• Initiate a 3-6 month pair: long AMD and MRVL, short a basket of CRWV and IREN in equal dollar beta-adjusted weights. The trade isolates higher-quality earnings conversion from financing/permitting risk; reassess if CRWV/IREN report rising utilization and contracted backlog sufficient to cover next-12-month capex.
• Accumulate ORCL on weakness ahead of the next earnings and cloud-bookings update, with a 6-12 month horizon. Risk/reward depends on cloud growth accelerating without a disproportional capex step-up; exit if remaining performance obligations and OCI growth decelerate while capex guidance rises.
• Use VRT or ETN as a small 6-12 month second-order long rather than adding to leveraged compute lessors. These suppliers benefit from power-density upgrades regardless of which cloud provider wins, but reduce exposure if data-center order growth or utility interconnection approvals weaken materially.
• Do not chase CRWV, IREN, or NBIS after momentum rallies until financing terms, debt maturities, power costs and customer concentration are verified. Set an alert for widening high-yield spreads or delayed capacity commissioning: either would pressure equity value well before revenue estimates are formally cut.