Nvidia's Jensen Huang denies Rubin delay, sees ASIC rival shift 50% to GPUs

Nvidia Corp. is not only maintaining its AI chip dominance — it is winning back ground from custom-chip rivals, according to the company's top executives at a Morgan Stanley investor roadshow this week.

Chief Executive Officer Jensen Huang, Chief Financial Officer Colette Kress and investor relations head Toshiya Hari met with institutional investors in California on Tuesday, directly addressing three of the biggest questions hanging over the stock: product delays, custom-chip competition and whether growth can continue at a nearly $1 trillion quarterly revenue run rate.

"Rubin Ultra is on track to ship next year," Huang said, according to Morgan Stanley analyst Joseph Moore's note, denying market chatter that the next-generation chip had slipped to 2028. The company is redesigning the Kyber rack system for the Rubin platform — swapping in what Huang described as a better architecture that supports larger compute domains — but the 800-volt power delivery and optical interconnects between racks remain on schedule.

The most striking disclosure came from the customer side. One frontier AI lab that had been developing its flagship model primarily on custom ASICs — widely believed to be Anthropic, whose backer Amazon has pushed its Trainium chip — has shifted nearly 50% of its compute to Nvidia GPUs, Moore wrote. The shift directly challenges the bear case that hyperscaler in-house chips would erode Nvidia's market share. "The comparison is not chip price but total cost per token," Moore said, citing industry research showing Nvidia's solutions still deliver lower per-token costs across many workloads.

Three growth engines, not one

Nvidia's revenue is becoming less concentrated. Morgan Stanley estimates AI labs account for about 20% of total demand, while traditional hyperscalers — Microsoft, Meta, Amazon and Google — contribute roughly half. The third bucket, which includes sovereign AI projects, enterprise deployments and new AI cloud providers, is growing faster than both, constrained more by power and data center construction timelines than by demand.

Sovereign AI projects, where governments build domestic model infrastructure for data security and industrial policy reasons, are particularly insulated from the ASIC competition narrative, Moore noted. These buyers tend to prefer fully integrated systems over custom silicon.

The company's CPU business is also expanding beyond its traditional role. Nvidia reiterated a roughly $20 billion CPU revenue target for the current fiscal year, with nearly half coming from standalone Vera CPU racks — not just management processors inside GPU servers. Vera is designed for single-threaded workloads with larger die area, fewer cores and memory optimizations for AI inference.

The valuation case for a new investor base

Moore maintained his overweight rating on Nvidia with a $288 price target, implying about 42% upside from the July 9 close of $202.78, which valued the company at roughly $4.97 trillion. He expects revenue to grow 82% in fiscal 2026 and 52.4% in fiscal 2027.

Nvidia is also broadening its investor outreach beyond growth funds, many of which are near single-stock concentration limits. Moore said the company could direct more than 50% of free cash flow to buybacks and dividends, giving it the cash-flow profile of a value stock while maintaining high growth.

The risks remain real. If supply catches up with demand faster than expected, data center revenue growth could decelerate sharply. A significant drop in AI development costs, more competitive products from rivals or faster deployment of custom silicon by hyperscalers could all pressure Nvidia's position. For now, the company's main challenge is not whether AI demand exists — it is converting that demand into shippable systems within the constraints of memory supply, network bandwidth, power availability and data center construction.

This article is for informational purposes only and does not constitute investment advice.