NVIDIA $96B Earnings Validate Supercycle as OpenAI Chip Fractures Inference Market

Nvidia's $96.2 billion quarter confirms AI capex durability through 2028, but OpenAI's Jalapeño ASIC and Anthropic's $45 billion Nscale lease expose a rapid bifurcation toward custom inference silicon and vertically integrated private clouds that will erode Nvidia's long-term pricing power.

Nvidia's $96.2 billion second-quarter revenue, up 106% year-over-year, shatters valuation skepticism and confirms the AI infrastructure supercycle has structural durability extending through fiscal 2028. The company's execution velocity is unprecedented: Nvidia expects to sell $20 billion of Vera Rubin systems in Q3 alone, capturing 20% of data center revenue, while Dell has already deployed the first Vera Rubin NVL72 rack to CoreWeave. This hardware ramp proves supply chain bottlenecks are resolving faster than anticipated, allowing hyperscalers and neocloud operators like Nebius, which surged 34% post-earnings, to accelerate cluster builds without waiting for next-gen availability. Nvidia is effectively monetizing the entire compute stack, turning its $105 billion Ohio liability cap into a moat that smaller players cannot replicate.

However, Nvidia's dominance is facing a precision strike on the inference side, where OpenAI's debut of the 700W Jalapeño ASIC fundamentally alters the efficiency calculus. Co-developed with Broadcom, Jalapeño delivers 1.9 times higher throughput per kilowatt and 3.6 times lower latency than Nvidia's 1,400W GB300, proving that custom silicon can outperform general-purpose accelerators on cost-per-token and speed metrics. This efficiency leap directly addresses the thermal and grid constraints plaguing deployments like OpenAI's leased 10-gigawatt SoftBank campus, signaling a strategic migration away from proprietary GPUs for inference workloads. As frontier labs prioritize power density and unit economics, OpenAI's move threatens to fragment the accelerator ecosystem and redirect future hyperscaler capex toward bespoke designs that bypass Nvidia's architecture entirely.

The capital markets are responding to this consolidation by rewarding private cloud operators who secure guaranteed compute access through massive, long-dated contracts. Anthropic's $45 billion, six-year agreement with Nscale for the Monarch Compute Campus anchors revenue visibility for the neocloud provider while insulating the AI lab from spot-market volatility ahead of its anticipated IPO. Similarly, IREN locked $6.5 billion in GPU financing and $4 billion in AI cloud ARR despite booking a $684 million loss on its mining wind-down, highlighting the industry's pivot to asset-backed debt structures to fund Blackwell Ultra deployments. CoreWeave's 19% stock jump following its earnings beat underscores investor confidence in these models, where neoclouds are vertically integrating procurement to guarantee cluster access independent of traditional hyperscaler allocations.

Upstream, Nvidia is leveraging its scale to restructure the memory supply chain, squeezing margins from traditional manufacturers while locking customers into its packaging ecosystem. The company's NVHBM strategy exploits the fact that base dies cost three to four times more than core dies, capturing the majority of HBM value and pushing SK Hynix, Micron, and Samsung toward commodity pricing dynamics. While SK Hynix broke ground on a $4 billion advanced memory hub in Indiana to address throughput bottlenecks, Nvidia's vertical integration tactic risks compressing supplier returns and altering competitive dynamics. With power density becoming the hard constraint—evidenced by Nvidia's Ohio risk allocation and OpenAI's gigawatt-scale leases—the winners will be those who control both the silicon efficiency and the power delivery infrastructure required to run dense racks at scale.

Looking ahead, the critical inflection points are the production yields of OpenAI's Jalapeño at volume and the margin impact of Nvidia's NVHBM strategy on memory vendors. Investors must monitor whether Vera Rubin NVL72 deployments meet performance targets at CoreWeave and if Anthropic's infrastructure spend accelerates further ahead of its IPO. The divergence between Nvidia's training monopoly and the rising tide of custom inference silicon will define the next cycle; any delay in Jalapeño scaling or a failure to secure power permits for multi-gigawatt campuses could trigger a repricing of the neocloud valuations currently riding high on CoreWeave and Nebius momentum.