Nvidia Quarterly Revenue Nears $100B: Why GPU Scarcity Still Drives AI API Prices Up

Near $100 Billion in a Single Quarter

At the Morgan Stanley Technology, Media & Telecom conference, Nvidia CEO Jensen Huang disclosed that the company's quarterly revenue is approaching $100 billion. To put this in perspective: Nvidia's entire annual revenue in fiscal year 2023 was $27 billion. The company is now generating nearly four times that amount every quarter, driven almost entirely by AI infrastructure demand.

Morgan Stanley maintains an overweight rating with a price target of $288. The analyst consensus is that demand continues to outstrip supply across all GPU product lines — which is the single most important signal for developers budgeting AI API costs.

Why GPU Revenue Directly Affects Your API Bill

Every AI API call you make — whether to Claude Fable 5 at $10/$50, GPT-5.6 Sol at $5/$30, or even budget options like Tencent Hy3 at $0.14/$0.58 — runs on GPU infrastructure. When GPU supply is constrained and prices remain high, API providers cannot reduce their per-token costs below the hardware floor.

Nvidia's near-$100B quarter confirms that demand has not softened. Every major AI lab — OpenAI, Anthropic, Google, Meta, xAI — is competing for the same limited GPU supply. This competition keeps hardware costs elevated, which propagates directly into the token prices developers pay.

Rubin Ultra on Track: Next-Gen Costs Still 12+ Months Away

Huang confirmed that Rubin Ultra is not delayed and remains on schedule for next year. Rubin Ultra represents Nvidia's next major architecture leap, promising significant performance-per-watt improvements that could eventually reduce the cost floor for AI inference.

However, "on track for next year" means these efficiency gains will not reach production AI services until late 2027 at the earliest. The typical timeline from GPU launch to widespread API cost reduction spans 12-18 months as providers purchase hardware, build out clusters, optimize software stacks, and amortize existing Hopper/Blackwell deployments. Do not budget for hardware-driven price drops in 2026.

A Major ASIC Customer Shifts to GPUs

Perhaps the most significant signal from the roadshow: Huang mentioned that a major customer previously committed to custom ASICs is shifting workloads back to Nvidia GPUs. Industry analysts speculate this is Anthropic, which had been exploring custom silicon partnerships to reduce inference costs for Claude models.

If an AI lab the size of Anthropic determined that custom ASICs cannot match Nvidia GPU efficiency for their workloads, it signals that the GPU tax on AI inference will persist longer than optimists hoped. Custom silicon was supposed to be the path to dramatically cheaper inference — if that path is narrowing, the pricing floor for models like Claude Opus 4.8 ($5/$25) and Claude Fable 5 ($10/$50) may hold longer than expected.

Nvidia CPU Business and Server Market Expansion

Huang disclosed that Nvidia's CPU business is tracking toward approximately $20 billion this fiscal year, with the Vera CPU entering the general server market. This diversification matters because it means Nvidia is capturing revenue across the entire AI inference stack — not just GPUs but also the CPUs handling orchestration, preprocessing, and serving.

For AI API pricing, vertically integrated Nvidia hardware (GPU + CPU + networking) can actually improve price-performance ratios by eliminating bottlenecks between components. However, it also increases Nvidia's market power, reducing competitive pressure that might otherwise drive hardware costs down faster.

What This Means for 2026-2027 AI Coding Budgets

Based on the signals from this roadshow, here is how to think about AI coding costs over the next 12-18 months:

• No hardware-driven price drops in 2026 — GPU supply remains constrained, Rubin Ultra is still a year from deployment, and demand continues to outpace supply
• Price reductions will come from software efficiency — MoE architectures (like Tencent Hy3), quantization, speculative decoding, and prompt optimization are the near-term cost levers
• Premium models will stay premium — Claude Fable 5 ($10/$50) and GPT-5.6 Sol ($5/$30) are unlikely to see significant price cuts while GPU costs remain elevated
• Budget optimization means model selection — the biggest cost lever available today is choosing the right model per task, using cheaper models for simple coding and reserving expensive ones for complex reasoning
• ASIC alternatives are not materializing — if major AI labs are returning to Nvidia GPUs from custom silicon experiments, the competitive pressure on GPU pricing weakens

The Bottom Line for Developers

Nvidia's financial performance is a proxy for AI infrastructure scarcity. When Nvidia revenues grow this fast, it means every GPU being produced is immediately consumed — leaving no surplus capacity to drive prices down. Your AI coding budget should assume current token prices as the floor through at least mid-2027, with optimization coming from smarter model routing and architecture improvements rather than raw hardware cost reductions.

Plan accordingly: invest in tooling that selects optimal models per task (Grok 4.5 at $2/$6 for simple edits, Opus 4.8 at $5/$25 for moderate complexity, Fable 5 at $10/$50 only for the hardest problems), implement caching and context management to reduce redundant token consumption, and budget for current prices persisting through at least the next fiscal year.