Nvidia's Nemotron 3 Ultra becomes the smartest open US model, but China still leads

Nvidia's Nemotron 3 Ultra ranks as the highest-scoring open US AI model on Artificial Analysis at 48 points, ahead of Gemma 4 31B (39), Nemotron 3 Super (36), and gpt-oss-120b (33). The model has about 550 billion total parameters with roughly 55 billion active and reportedly exceeds 300 tokens per second on DeepInfra. Nvidia says it will be released on June 4 via Hugging Face, OpenRouter, and other platforms, though it still trails leading Chinese open models such as Kimi K2.6 (54) and the top closed model Opus 4.8 (61).

Analysis

This is less a model launch story than a signal that NVDA is extending its moat from silicon into software-mediated demand capture. A high-performing open model with strong inference throughput widens the addressable market for Nvidia’s CUDA/serving stack and makes its platform harder to displace, because the company now benefits whether enterprises choose closed, open, or hybrid deployment paths. The second-order effect is that “open” model adoption can actually reinforce GPU utilization rather than commoditize it, especially if the benchmark gap compresses time-to-deployment for regulated and cost-sensitive customers.

The near-term equity relevance is not the benchmark score itself but what it implies for inference workload migration over the next 1-2 quarters. Faster tokens/sec means lower latency and better economics, which expands feasible use cases in agentic workflows, customer support, and code assist; that tends to pull forward enterprise spend across both NVDA hardware and adjacent networking/rack-level content. The competitive pressure is more likely to show up in model providers and inference hosts with weaker unit economics than in Nvidia’s core revenue stream.

The contrarian risk is that investors may overread this as a pure NVDA win while underestimating the margin mix issue: if open models become “good enough,” bargaining power can shift from model owners to cloud operators and inference aggregators. The real medium-term vulnerability is not demand destruction but pricing normalization in the software layer, which could cap upside for names selling premium model access even as compute demand stays healthy. Also watch for China’s continued benchmark lead as a reminder that leadership is fragmented; that reduces the odds of a clean U.S.-centric narrative premium and may keep competition intense into 2026.

Catalyst-wise, the key window is the next 30-90 days as enterprise pilots and API providers integrate the release. If early adoption validates materially better throughput per dollar, NVDA can see a multiple tailwind from renewed “picks-and-shovels plus platform” positioning; if usage is shallow, the stock may revert to hardware-cycle sensitivity. The setup favors owning NVDA into launch/availability, but being selective on adjacent AI software names with exposed pricing power.

Key Decisions for Investors

• Long NVDA into the June 4 release window; target a 1-3 month hold as the market digests whether open-model adoption lifts inference demand. Risk/reward: positive skew if enterprise deployment commentary confirms broader GPU utilization; trim if the event is treated as purely incremental.
• Pair trade: long NVDA / short an AI software name with model monetization risk over the next 2-4 months. The thesis is that compute demand is sticky while software margin capture is more vulnerable to open-model commoditization.
• Buy NVDA call spreads 6-10 weeks out to express upside from launch-driven sentiment without paying full delta. Use a defined-risk structure because benchmark-driven moves can fade quickly if adoption metrics disappoint.
• If already long high-multiple AI application names, reduce exposure on strength over the next 1-2 weeks; this release increases the odds that value accrues to infrastructure rather than application-layer pricing power.
• Monitor hyperscaler capex commentary over the next quarter; if they highlight stronger inference buildouts tied to open model deployment, add to NVDA and AI networking exposure. If not, treat the launch as a sentiment event rather than a fundamental re-rate.