Nvidia is happy to take your money whether you're using open or closed models, says Jensen Huang

Nvidia CEO Jensen Huang endorsed both closed frontier models and open models, emphasizing that specialized AI can create substantial value while Nvidia hardware powers the underlying infrastructure. Fireworks.ai received 25,000 Nvidia GPUs to support expansion in Europe; Huang also highlighted agent harnesses as a recent breakthrough and forecast safe agent deployment and robotics hardware-in-the-loop as future advances. At the event, Genspark said its slide-generation model delivers results comparable to Anthropic's Opus 5 at $0.30 per million tokens versus $5 per million.

Analysis

The key investment question is whether lower inference cost creates enough new usage to offset fewer GPU-hours per task. If adoption expands faster than compute efficiency improves, NVIDIA (NVDA) can benefit even as open and specialized models take share from proprietary APIs; if not, cheaper tokens may pressure accelerator utilization and cloud-provider pricing. This makes inference volume, GPU-hours per workload, and realized pricing more important than model openness alone.

Near term (days to weeks), the remarks are more supportive of NVDA’s demand narrative than evidence of incremental orders: customer deployment claims and strategic endorsements should not be treated as revenue guidance. Over 1–3 months, watch hyperscaler and AI-platform commentary on inference growth, utilization, and capex. A sustained shift toward lower-cost self-hosted models could weigh on Anthropic (ANT)’s pricing power, but specialized performance, reliability, and tool integration may limit substitution. MiniMax Group (100) gets a possible product-validation signal from downstream use of its model, not proof of material revenue or durable share gains.

Over 6–18 months, agent deployment could expand inference demand; robotics could add a distinct physical-AI workload, but remains a longer-dated optionality rather than a near-term earnings catalyst. Contrarian point: efficiency is not automatically bearish for GPUs—lower unit costs can unlock workloads—but the market may over-credit that elasticity before it appears in utilization or orders. Falsifiers for the NVDA demand thesis include weaker inference growth, falling GPU utilization, or capex guidance reductions despite rising AI usage.

Key Decisions for Investors

• NVDA: Treat this as supportive narrative, not a standalone catalyst. Consider adding only on weakness if subsequent earnings calls or disclosures confirm rising inference workloads and sustained accelerator utilization; reassess if capex is cut or utilization weakens.
• ANT: Avoid trading on the open-model discussion alone. Track evidence of customer migration, API price reductions, or weaker retention; absent those signals, specialized-model substitution remains a risk scenario rather than an established earnings hit.
• 100: Put on a watchlist, not a size-up. Verify whether downstream adoption of MiniMax M3 translates into disclosed customer growth, recurring usage, or revenue before treating the association as a commercial catalyst.
• Monitor the next 1–3 months for inference-token growth, GPU-hours per task, realized pricing, and capex guidance. The trade hinges on whether usage expansion outruns efficiency-driven compute savings.