The revised outlook rests on a specific argument Huang made at GTC: "The inference inflection has arrived." That phrase carries weight once you understand what it means for compute demand.

Training AI models is periodic and bounded. You train a model, ship it, then train the next one. Inference is different. Every query a user sends, every task an AI agent executes in real time the chips run continuously. The shift from training-dominated spending toward inference-dominated spending implies a structurally larger and more persistent demand curve than training cycles alone.

That argument has numbers behind it. AI inference token generation surged tenfold in a single year, per Nvidia's Q1 fiscal 2026 earnings release from ten months ago. The forecast revision isn't projecting a future that hasn't arrived it's catching up to usage patterns already visible in Nvidia's own infrastructure data.

Emarketer analyst Jacob Bourne read the revised forecast as a signal of durable rather than cyclical demand, noting it reflects Nvidia "sustaining its leadership in the AI chip market while the overall AI industry expands beyond early experimentation into large-scale deployment," per Reuters. Wedbush analyst Dan Ives called it "a much needed confidence boost for tech investors navigating a very tricky tape," adding that the outlook signals the AI buildout is "accelerating, not decelerating, despite the market noise," per Yahoo Finance.

The upgrade from $500 billion to $1 trillion is best understood as a claim about what Nvidia is now selling, not just how much. Two developments account for most of the change.

First, the inference architecture itself. Huang described a two-layer stack pairing Nvidia's own silicon with technology from Groq, a chip startup from which Nvidia licensed technology for $17 billion in December. Nvidia's Vera Rubin chips handle the "prefill" stage, converting a user's request into the token language AI systems process. Groq's chips handle the "decode" stage that generates the actual output, per Reuters. Capturing both ends of a real-time inference workload means Nvidia is billing for the full transaction, not just the compute core. That's a different revenue model than selling training GPUs.

Second, the product mix has physically changed. Huang unveiled a standalone Vera CPU at GTC, calling it "already for sure going to be a multi-billion-dollar business." Bob O'Donnell of Technalysis Research described the shift directly: Huang used to arrive at conferences with a new GPU chip; now he shows up with five racks of equipment, via Bloomberg. A rack sale isn't a chip sale. The contract size, the customer relationship, the margin profile all different. Bank of America analysts noted the scale of the forecast implies demand across the full AI supply chain, spanning servers, storage, networking, and optics, with potential beneficiaries including Dell, Super Micro Computer, Hewlett Packard Enterprise, Sandisk, Corning, and Amphenol, per Yahoo Finance.

Whether Nvidia captures a proportional share of that expanded pool is the question to stress-test. A company selling infrastructure commands a different multiple than one selling components. The open issue is whether the margin profile holds as product mix shifts and on that, management hasn't given shareholders a clean answer.

Huang's projections carry weight because the underlying financial performance gives them a foundation. January-quarter revenue came in at $68.13 billion, up 94% year over year, beating estimates of $66.21 billion. Adjusted EPS of $1.62 cleared the $1.53 consensus. Nvidia guided fiscal Q1 revenue to $78 billion against a Wall Street expectation of $72.6 billion, per Reuters. Ken Mahoney of Mahoney Asset Management described it as "the usual for Nvidia" while noting much of the upside had already been priced in.

That tension a company consistently beating and raising into a stock that barely moves tells you something about where expectations sit.

Supply concerns, previously a real investor worry, received a direct answer last month: Nvidia confirmed it has secured sufficient chip inventory and TSMC manufacturing capacity to meet demand beyond the next several quarters, per Reuters. One credible obstacle is gone.

Margins are the legitimate open question. Non-GAAP gross margin ran at 61% in Q1 fiscal 2026, depressed by a $4.5 billion inventory charge tied to China export restrictions. Strip that out and the underlying margin was 71.3%, per Nvidia's Q1 press release. As Nvidia shifts more of its mix toward full systems, CPUs, and inference hardware, whether that 71% baseline holds is genuinely unknown. Selling infrastructure at scale can compress margins even as it expands revenue.

On capital returns: Nvidia distributed $37 billion to shareholders through buybacks and dividends in the first nine months of fiscal 2026, with $62.2 billion still authorized for repurchase as of late October 2025, per Nvidia's SEC filing. CFO Colette Kress has been consistent that reinvestment takes priority over near-term cash return, per Reuters. Reasonable given current growth rates. Shareholders seeking more aggressive distributions will keep waiting.

Three scenarios worth holding simultaneously:

• Bullish case: Inference and systems revenue expands the addressable pool faster than consensus assumed, and Nvidia captures a disproportionate share with margins intact.
• Neutral case: Much of the optimism is already priced in. Shares closed up just 1.2% after initially jumping on the GTC news, per Bloomberg.
• Risk case: Margin pressure from product mix shift and China policy instability erode earnings quality even if revenue targets are met.

China is not a side note. It caused a quantifiable hit to Nvidia's recent results, reversed course rapidly, and remains the least predictable variable in the near-term picture.

When the U.S. government imposed new H20 export licensing requirements in April 2025, Nvidia absorbed a $4.5 billion inventory and purchase-obligation charge, forfeited $2.5 billion in H20 shipments it couldn't send, and projected an $8 billion drag on the following quarter's revenue, per Nvidia's Q1 earnings release. Real numbers, not hypothetical risks.

What followed was unusual. By early March 2026, Nvidia had paused China-targeted H200 production and shifted TSMC capacity to Vera Rubin, signaling no expectation of near-term China sales, per the Financial Times. Within roughly two weeks, Huang reported at GTC that Nvidia had received U.S. licenses for many Chinese customers, collected substantial purchase orders, and restarted H200 manufacturing, per the Financial Times.

The licensing structure itself is without precedent. Nvidia and AMD agreed to remit 15% of chip revenues from Chinese sales to the U.S. government as a condition of the export licenses, a structure that export control experts said had no prior equivalent, per the Financial Times. Nvidia's response "we follow rules the U.S. government sets for our participation in worldwide markets" neither confirmed nor denied the terms.

Nvidia's February guidance included zero expected China data center revenue, so any Chinese sales that materialize represent upside not captured in current models. The less comfortable read: a policy environment capable of reversing twice in three weeks can reverse again.

• Data Center revenue growth and product mix will be the primary signal. The segment hit $51.2 billion in fiscal Q3 2026, up 66% year over year, per Nvidia's investor release. Genuine scaling from inference and systems should show up in both the growth rate and any management commentary on mix.
• Gross margin will tell shareholders whether the shift to full systems is value-accretive or margin-dilutive. Watch for whether 70%-plus non-GAAP margins prove sustainable as CPUs and racks become a larger portion of the revenue base.
• Any disclosed China revenue clarifies whether Huang's GTC restart announcement represents real bookings or optimistic positioning. February guidance assumed zero; the first quarter showing meaningful China data center revenue changes the model materially.
• Product roadmap continuity matters to the long-term thesis. Huang outlined a path extending to the Feynman architecture in 2028, following Vera Rubin Ultra, per Reuters. The investment case doesn't depend on a single product cycle. Near-term credibility, though, gets established or challenged in the next two earnings reports not at a developer conference.