Here Are the Companies in Nvidia's $99 Billion Artificial Intelligence (AI) Portfolio. These 3 Are the Best of the Bunch

Nvidia reported a roughly $99 billion investment portfolio as of the July 26 fiscal-Q2 close, split among $47.7 billion of public equities, $47.9 billion of non-marketable securities, and $3.3 billion of equity-method investments. Its largest AI-related private commitments include $30 billion to OpenAI, up to $10 billion to Anthropic, and $5 billion to Safe Superintelligence, while public and convertible holdings span Intel, CoreWeave, Coherent, Lumentum, Marvell, Nokia, Nebius, and SpaceX. The article identifies CoreWeave, Coherent, and Marvell as the most investable public proxies for AI compute leasing, optical interconnects, and scale-up networking demand.

Analysis

The investee map is more important as a revenue-quality signal than as a stock-picking endorsement. Capital deployed into compute buyers can pull forward accelerator orders and protect utilization, but it also increases circular-financing scrutiny: if receivables, purchase commitments, or equity-method marks rise alongside reported data-center revenue, NVDA's earnings multiple faces a 1-3 month de-rating risk even without a demand collapse. The key diligence item is the extent to which investee-funded GPU purchases are backed by end-customer contracts rather than equity-linked vendor financing.

COHR and MRVL offer cleaner second-order exposure than GPU-rental operators because optical interconnect content rises with cluster size, bandwidth, and rack density regardless of whether workloads sit at hyperscalers or neoclouds. Over 6-18 months, scale-out architectures should shift a growing portion of AI infrastructure BOM from compute toward networking, DSPs, lasers, and transceivers; this is a margin-expansion opportunity if supply remains tight. LITE is a plausible sympathy beneficiary, but its exposure is less differentiated and more vulnerable to optical-component price competition.

CRWV remains the highest-beta expression of AI capex, but its economics are fundamentally a leveraged spread between contracted compute revenue, GPU depreciation, power, and financing costs. Consensus may underappreciate that additional GPU supply or a customer concentration event can compress that spread before headline AI spending slows; equity upside requires both sustained utilization and benign refinancing conditions. NBIS is a smaller, less liquid alternative neocloud proxy and should be treated as event-driven rather than a core infrastructure holding.

Contrarianly, strategic investments may be defensive rather than proof of unconstrained demand: NVDA has incentive to finance ecosystems that broaden its addressable market and counter custom-silicon substitution. MRVL benefits from that ecosystem buildout but also gains if customers diversify away from merchant GPUs through custom accelerators, making it a more balanced exposure than a pure NVDA satellite.

Key Decisions for Investors

• Initiate a 6-12 month long COHR / short CRWV pair, sized beta-neutral. COHR captures rising interconnect content with materially lower utilization and refinancing sensitivity; CRWV is vulnerable if GPU rental pricing or funding spreads weaken. Reassess if CRWV discloses contracted backlog conversion and operating cash flow sufficient to fund capex without incremental equity or expensive debt.
• Accumulate MRVL on pullbacks ahead of the next two hyperscaler capex updates; target a 9-12 month holding period. Thesis is custom-silicon plus optical/networking content growth, but exit if management reduces AI revenue trajectory, or if gross-margin guidance implies mix dilution rather than operating leverage.
• Maintain NVDA exposure but hedge 1-3 month revenue-quality risk with a small NVDA put spread around the next earnings date rather than adding outright. The hedge is warranted if receivables growth materially exceeds data-center revenue growth, customer prepayments decline, or disclosures indicate increased related-party concentration.
• Avoid adding CRWV purely on strategic-investor signaling. Upgrade only after verifying utilization, average contract duration, customer concentration, debt maturities, and fixed-price power exposure; absent that data, the appropriate action is a watch alert, not a directional long.
• Use LITE only as a tactical catch-up basket component with COHR and MRVL, not a standalone core position. A broad optical upcycle could lift all three over 3-6 months, but commoditization risk makes LITE the first candidate to short against COHR if pricing deteriorates.