Nvidia's $78B Guidance Hides an $8B China Hole — Beijing Summit Is the Variable

Nvidia's Q1 FY2027 guidance of $78 billion is the highest revenue guidance any semiconductor company has ever issued. It also assumes zero recovery from China's H20 chip market, which cost Nvidia an $8 billion charge in a single quarter. The Beijing summit on May 14-15 — where the US and China are negotiating semiconductor export control terms — is the single variable that could make Nvidia's next fiscal year significantly better or significantly worse than that $78 billion baseline.

Jensen Huang called the H20 ban "deeply painful." That is an understatement applied to a situation where a single export control decision by the US government cost a company more revenue in one quarter than most Fortune 500 companies generate in an entire year.

The Q1 FY2026 Actuals

Nvidia's Q1 FY2026 results (reported in May 2025, covering the February-April 2025 quarter):

• Revenue: $44.1 billion, up 69% year over year
• Data Center revenue: $39.1 billion, up 73% year over year — the segment that includes AI GPU sales
• Gross margin: 75.1% — extraordinarily high for a semiconductor company
• Net income: $18.8 billion

The H20 China ban hit in April 2025 — within the Q1 FY2026 reporting period. The charge: $4.5 billion in inventory write-downs for H20 chips that could no longer be legally shipped to China, plus $2.5 billion in revenue that had been anticipated from Chinese customers who had purchase orders for H20s. Total impact: approximately $7 billion in Q1, with further ongoing lost revenue in subsequent quarters.

The $78B Q1 FY2027 Guidance

Nvidia guided Q1 FY2027 (the February-April 2027 quarter, reporting in May 2027) to approximately $78 billion ± 2%. That guidance was issued in the context of:

• No H20 recovery assumed
• Blackwell GPU architecture ramp continuing through 2026-2027
• GB200 NVL72 rack-scale systems shipping to hyperscalers at scale
• Inference demand growth from rapidly expanding AI deployment

The $78 billion figure represents approximately 77% revenue growth from Q1 FY2026's $44.1 billion. That is a continuation of Nvidia's growth trajectory even without China.

The implicit question: if $78 billion is the no-China recovery scenario, what is the upside if the Beijing summit produces H200 export license approvals for Chinese commercial customers?

The China Revenue Opportunity

Before the H20 ban, China accounted for approximately 17% of Nvidia's data center revenue — roughly $7-8 billion per year. The H20 ban eliminated that entirely and imposed the inventory charge on top.

The H20 ban specifically targeted the H20, which Nvidia had designed to comply with the previous export control thresholds (the H20 was specifically detuned from H100 capability to fall below the export control limits). After the H20 ban, there is no compliant Nvidia data center GPU that can be sold to Chinese customers under current rules.

The Beijing summit discussion is about H200 export licenses — whether Chinese commercial AI customers could apply for and receive licenses to purchase H200s under a case-by-case review regime. This is a higher-capability chip than the H20, which makes the US government's willingness to approve such licenses politically sensitive. But Chinese customers have demonstrated they can build frontier models without H200s — DeepSeek V4 and the other May model releases are evidence of that.

The potential revenue recovery if the summit produces H200 case-by-case licensing:

Even at 50% of the previous China run rate (conservative, given the two-year absence), that represents $3.5-4 billion per year in additional Nvidia revenue. Against a $78 billion baseline, that is a 4-5% upside. For a company trading at high multiples, 4-5% additional revenue is not trivial.

Why Blackwell Is More Important Than China for 2026

The China discussion is important but secondary to the Blackwell architecture ramp for Nvidia's 2026 numbers.

The Blackwell GPU architecture (GB200, B200) began shipping in volume in late 2025 and is ramping through 2026. The GB200 NVL72 — a rack-scale system combining 72 Blackwell GPUs with NVLink interconnect — is the product hyperscalers are buying for their 2026 training clusters. Amazon, Microsoft, Google, and Meta are collectively spending $695-725 billion on AI infrastructure in 2026; a substantial fraction of that is Blackwell-based Nvidia systems.

Blackwell improves on Hopper (H100) in two dimensions that hyperscalers care most about: training throughput per watt (approximately 4x improvement) and inference throughput for large models (approximately 30x improvement versus H100 for transformer models using FP4 precision). These improvements translate directly to the hyperscaler economics that are driving the AI capex surge.

The $78 billion guidance is a Blackwell story, not a China recovery story. China recovery would be incremental upside to a trajectory that is already driven by hyperscaler demand for Blackwell capacity.

The Developer and Infrastructure Impact

For organisations planning AI infrastructure purchases:

GPU pricing: Nvidia maintains substantial pricing power as long as Blackwell demand exceeds supply. H100 spot prices on the cloud GPU market have remained elevated despite Blackwell availability — the older-generation chips continue to be used for inference workloads where Blackwell's training advantages are less relevant. Expect continued tight supply through H2 2026.

The Beijing summit as a catalyst: If the summit produces H200 licensing news, watch for a Nvidia share price reaction. The market has priced in zero China recovery — any positive news is upside relative to current expectations. The opposite is also true: if the summit fails and antitrust investigations resume, China-related uncertainty could create short-term pressure on the stock.

Alternatives getting more interesting: Cerebras's IPO and its OpenAI inference deal, AMD's MI300X ramp, and Intel's Gaudi 3 are all making progress in the inference market specifically. Nvidia dominates training but the inference market is more contested. For organisations building inference-heavy applications, the competitive landscape is broadening.

Rare earth supply chains: China's restrictions on gallium, germanium, and antimony exports affect compound semiconductor inputs. A partial relaxation from the summit reduces supply chain risk for power electronics and RF components used in data center power distribution and server components — a second-order infrastructure impact.

Key Takeaways

• Q1 FY2026 actuals: $44.1B revenue, +69% YoY; Data Center $39.1B, +73%; gross margin 75.1%; H20 ban cost $4.5B inventory charge + $2.5B lost revenue = ~$7B total impact
• Q1 FY2027 guidance: $78B ± 2% — highest semiconductor revenue guidance ever; assumes zero China H20/H200 recovery; driven by Blackwell architecture ramp to hyperscalers
• China upside scenario: H200 case-by-case licensing at 50% of previous run rate = $3.5-4B/year additional revenue; 4-5% upside to $78B baseline; Beijing summit is the trigger event
• Blackwell is the real story: GB200 NVL72 shipping to hyperscalers; 4x training throughput per watt vs. H100; 30x inference throughput for transformers; $695-725B hyperscaler capex in 2026 is largely Blackwell orders
• Competition broadening: Cerebras, AMD MI300X, Intel Gaudi 3 gaining traction in inference market; Nvidia dominates training but inference alternatives are more viable
• Beijing summit watch: Zero China recovery priced in; positive news = upside catalyst; antitrust investigation resumption = short-term uncertainty

For the Beijing summit semiconductor negotiations context, read US-China Trade Truce May 12: Chip Export Controls at the Beijing Summit. For the Cerebras IPO competing in the inference market, read Cerebras IPO 2026: $3.5B Raise, WSE-3 vs NVIDIA H100.