Jensen Huang at Morgan Stanley: NVIDIA Nears $100B Quarter, Denies Rubin Ultra Delay, and Reveals the ASIC Reversal

In early July 2026, NVIDIA CEO Jensen Huang, alongside CFO Colette Kress, appeared at a Morgan Stanley Non-Deal Roadshow (NDR) in California. Before an audience of institutional investors, Huang delivered a message of striking confidence: NVIDIA’s quarterly revenue is approaching $100 billion — and growth is still accelerating.

The timing was critical. Over the preceding month, NVIDIA had faced a barrage of market skepticism on three fronts: rumors that the next-generation Rubin Ultra chip would be delayed to 2028, fears that custom ASICs were systematically eating into GPU market share, and mounting questions about whether the AI capex cycle had peaked.

Huang addressed all three head-on with data, technical specifics, and an uncharacteristically combative tone. Morgan Stanley emerged from the meeting maintaining its Overweight rating and $288 price target — representing approximately 42% upside — while reiterating NVIDIA as its top semiconductor pick.

Here is what was said, what it means, and why it matters for enterprise AI strategy.

1. The $100 Billion Quarter: Inside NVIDIA’s Accelerating Growth Curve

NVIDIA is approaching a milestone that even Huang described as “hard to fathom”: a $100 billion quarter.

The most recent fiscal data (Q1 FY2027, reported May 2026) paints the picture:

Q2 guidance stands at approximately $91 billion — notably excluding any Data Center sales to China, suggesting potential upside. The $100 billion quarterly threshold is no longer a question of “if” but “when.”

Morgan Stanley analyst Joseph Moore provided longer-range projections:

Calendar 2027 EPS is estimated at $13.08. At 22× earnings, the $288 price target requires no valuation multiple expansion — pure earnings-driven upside. For enterprises evaluating GPU investments, 🔗 understanding NVIDIA’s pricing power and margin structure is essential to forecasting your own AI infrastructure costs.

2. Rubin Ultra On Track: The Technical Truth Behind the Delay Rumors

In late June, SemiAnalysis published a report that shook NVIDIA’s stock: the original 4-die Rubin Ultra packaging was reportedly canceled due to TSMC CoWoS-L substrate warping and signal integrity failures, forcing a downgrade to a 2-die design, with shipments potentially slipping to 2028.

At the Morgan Stanley roadshow, Huang was unequivocal: “Rubin Ultra will ship next year, on schedule.”

He explained that the Kyber rack system is being replaced with an improved architecture designed to support larger compute domains — but framed this as routine system-level optimization, not a product delay. Three facts support this timeline:

1. 800V high-voltage power delivery is progressing on schedule — a critical enabler for rack-scale Rubin Ultra deployments
2. Inter-rack optical interconnects are replacing copper, enabling million-GPU clusters
3. TSMC CoWoS capacity is expanding ~80% year-over-year, providing the physical foundation for mass production

🔗 NVIDIA’s GTC 2026 Rubin architecture reveal and 🔗 Blackwell vs. MI300X performance benchmarks provide the technical context for understanding Rubin Ultra’s generational leap.

3. The ASIC Reversal: When Custom Chip Customers Come Back to GPU

The most explosive disclosure of the roadshow concerned ASIC competition. Huang revealed that a frontier AI lab previously running its flagship models almost entirely on custom ASICs — widely believed to be Anthropic — has shifted nearly 50% of its compute back to NVIDIA GPUs.

This single data point upends the prevailing narrative that custom silicon is on an inexorable march to replace general-purpose GPUs. The ASIC camp’s most prominent customer is buying GPUs again — and in large quantities.

Huang’s explanation: “The customer’s evaluation criteria isn’t chip unit price — it’s total cost per token.” At deployment scale, NVIDIA’s ecosystem maturity (CUDA, cuDNN, TensorRT), supply chain reliability (TSMC CoWoS capacity), and developer efficiency (millions of CUDA engineers) translate into a genuine total-cost advantage.

However, the ASIC threat is not imaginary. The data tells a nuanced story:

🔗 ASIC vs. GPU: the technology battle is fundamentally about the bifurcation of the AI compute market into training and inference. NVIDIA’s training dominance (90%+ share) remains nearly unassailable, but inference is where Google TPUs, Amazon Trainium, Microsoft Maia, and Broadcom custom ASICs are gaining ground.

🔗 Google TPU vs. NVIDIA GPU: a complete comparison illustrates why purpose-built silicon can win on cost in specific inference workloads — though Huang’s point stands that when you factor in the full software stack and engineering overhead, GPU economics often still prevail.

4. Three Growth Engines: Mapping the Global Demand for Compute

Huang segmented NVIDIA’s revenue into three distinct growth engines, a strategic framing that itself carries analytical weight:

The third engine — sovereign AI — deserves particular attention. Huang emphasized that governments are treating AI infrastructure as a national security asset, with an investment logic distinct from commercial cloud: “They’re not optimizing for cost. They’re optimizing for autonomy.”

An underappreciated highlight: NVIDIA’s CPU business. Huang projected ~$20 billion in CPU revenue this fiscal year, nearly half from standalone Vera CPU racks — NVIDIA’s first serious push into the general-purpose server CPU market, directly challenging Intel Xeon and AMD EPYC.

🔗 GPU, NPU, TPU, LPU: the complete 2026 AI processor landscape explains why NVIDIA is expanding from GPUs into CPUs and DPUs — not diversification, but platform construction.

5. Risks and Challenges: Why the Market Stays Anxious

Despite Huang’s confident presentation, Morgan Stanley’s report enumerated four key risks:

1. Faster-than-expected hardware supply release — if TSMC CoWoS expansion outpaces demand or competitors (Samsung 2nm GAA) achieve yield breakthroughs, GPU oversupply could erode NVIDIA’s pricing power
2. Dramatic reduction in AI R&D costs — a DeepSeek-style architectural innovation that slashes training compute requirements would reduce per-unit demand
3. Competitor breakthrough products — AMD, Intel, or emerging ASIC vendors could deliver solutions significantly better than GPUs in specific use cases
4. Cloud customers becoming competitors — Google TPU is the most mature ASIC ecosystem; Meta’s Iris chip enters mass production in September; the risk of major customers turning into rivals persists

🔗 Meta Compute: Zuckerberg’s $145B AI bet and the global semiconductor sell-off demonstrated how “compute oversupply” fears can vaporize hundreds of billions in market value in a single day — even when those fears prove unfounded.

Yet each of these risks, paradoxically, underscores the depth of NVIDIA’s moat. Every threat scenario presumes competitors first reaching technological, ecosystem, and scale parity with NVIDIA — which remains the hardest part of the equation.

Conclusion: From GPU Supplier to Computing Platform

What Huang presented at Morgan Stanley was, at its surface, a rebuttal of three bear cases. But the deeper message was NVIDIA’s evolution from a GPU supplier into an integrated computing platform spanning GPUs, CPUs, DPUs, and networking.

The strategic implication: when customers buy rack-scale systems — Vera CPU to Rubin Ultra GPU to Spectrum-X networking — the substitutability of any single component plummets. NVIDIA isn’t selling chips; it’s selling an integrated computing architecture that resists disaggregation.

For enterprises planning AI infrastructure, the roadshow’s core takeaways are:

1. Compute supply is moving from scarcity to abundance, but frontier compute remains concentrated — procurement strategy must distinguish “commodity compute” from “frontier compute”
2. ASICs and GPUs are not a zero-sum game — both will coexist; choose based on workload (training vs. inference)
3. NVIDIA’s pricing power derives from the ecosystem, not the silicon — CUDA’s network effects resist near-term replication
4. Sovereign AI and industrial AI represent the next wave — these segments have more fragmented customer bases and more diverse demand, implying a more stable revenue foundation for NVIDIA

Huang’s closing line captured the moment best: “I don’t know anyone in the industry who thinks they have too much compute.”

This article was compiled by the INFINITIX team. Sources include Morgan Stanley NVIDIA Research Report, DataConomy Roadshow Coverage, EdgeN ASIC Competition Analysis, SemiAnalysis Rubin Ultra Technical Analysis, and multiple financial media reports.