Maintaining Outperform — Blackwell in Full Ramp, $4.5B H20 Charge Absorbed Without Cracking the Trajectory; Sovereign AI Emerges as a New Growth Engine

Rating Action

Maintaining Outperform. Walking the arc:

• Initiating coverage at Hold (Q2 FY25, August 2024): Initiated at Hold on a powerful franchise running into a stretched setup. Concerns: (1) Hopper-to-Blackwell transition risk (yield, customer adoption, ASP transition), (2) China policy overhang on H20 / future SKUs, (3) the "everyone’s already long" positioning problem, (4) a multiple in the high-30s on FY26 estimates that left limited cushion for any execution wobble.
• Upgrading to Outperform (Q3 FY25): 13,000 Blackwell samples shipped with hyperscaler system stand-ups underway and Q4 explicitly exceeding prior estimates; GM trough quantified at 71-72.5% with mid-70s recovery specified for 2H calendar 2025; Jensen's three-scaling-laws framework and the 100,000-Blackwell next-gen-model floor materially expanded the multi-year compute TAM. The cleanest thesis-favoring print of the cycle resolved the central uncertainties anchoring the Hold.
• Maintaining Outperform (Q4 FY25): The framework-reset Q4 print extended the Q3 upgrade thesis. Blackwell delivered $11B in its first full production quarter (the fastest product ramp in company history), DC +93% YoY, FY26 framework guide of Q1 $43B reset positively, and the three-scaling-laws reframe became the structural answer to the post-DeepSeek bear case.
• Today (Q1 FY26): Maintaining Outperform. The constructive case continues to compound.
  • (a) Blackwell margins are recovering on schedule. Ex-charge non-GAAP GM 71.3% in Q1 (vs. ~71% prior); Q2 guided 72% mid; mid-70s targeted late this year. Colette Kress: "We are continuing to work towards achieving gross margins in the mid-seventies range late this year." This is the GM trajectory we needed to see.
  • (b) China policy is now resolved — in the bear direction, but resolved. The H20 export ban is final; NVDA has taken the writedown; $8B of Q2 H20 revenue is gone; the $50B China TAM is closed. The franchise still grows sequentially in Q2 (+$45B). The overhang is now quantified rather than open-ended.
  • The unexpected upside catalyst: sovereign AI was not on our prior-quarter checklist and has emerged as a multi-year, multi-country pipeline that materially offsets the China loss in scale.

Net: every operating thesis point we underwrote at the Q3 FY25 upgrade has confirmed or improved. The post-print level near $135 prices the China hit at face value and underweights the sovereign-AI offset and the GM recovery trajectory.

Results vs. Consensus

Quality of Beat

• Revenue mix: Beat is broad-based once you adjust for the H20 cliff. The $44B headline includes $4.6B of H20 recognized prior to April 9; ex that, core (non-China) Data Center grew at an even more accelerated rate. Gaming — $3.8B vs. ~$3.0B Street — was the largest single source of upside relative to expectations, driven by the RTX 50-series ramp and improved supply availability. Networking re-accelerating to $5B with NVLink shipments alone exceeding $1B is the read-through that scale-up is now a discrete revenue category, not a feature.
• Gross margins: The ex-charge 71.3% beats the prior outlook ("slightly above our outlook at the beginning of the quarter," per Kress) and sets up the 72% Q2 guide and mid-70s late-year target. This is the operational story under the GAAP-line noise. Bear case argued Blackwell would compress gross margins through 2H FY26; print and guide refute it.
• The $4.5B charge: Cleanly explained — H20 inventory and purchase obligations tied to orders received before the April 9 export-control date. Notably, "the $4.5 billion charge was less than what we initially anticipated as we were able to reuse certain materials." Reuse is a partial silver lining; the materials going into the broader Blackwell/networking flow.
• Q2 guide: $45B ± 2% is a guide that absorbs $8B of foregone H20 revenue and still grows sequentially. That is the most informative number on the call. If you back out the China headwind, implied core sequential growth is roughly $9B Q/Q — an extraordinary number at this scale. Modest sequential growth across "all platforms" per Kress.
• Operating expense: +6−7% sequential, with full-year FY26 OpEx growth guided in the mid-30% range. Investing aggressively into infrastructure and AI solutions; consistent with a company seeing demand acceleration, not deceleration. Capital return: a record $14.3B returned to shareholders in Q1 (buybacks + dividend).

Segment Performance

Data Center

The headline number ($39B / +73% YoY) reads "in line" against an extraordinary Street bar but hides the operational story. Blackwell is now ~70% of Data Center compute revenue, with Hopper transition nearly complete. The ramp pace — "fastest in our company’s history" per Kress — is corroborated by the deployment cadence: major hyperscalers each putting up roughly 1,000 NVL72 racks (roughly 72,000 Blackwell GPUs) per week. Microsoft has deployed tens of thousands of Blackwell GPUs and is on track to hundreds of thousands of GB200s, with OpenAI as a key customer.

The next-cycle visibility is what re-rates the franchise: GB300 sampling at major CSPs already began earlier this month, with production shipments later this quarter. GB300 leverages the same architecture, footprint, and electrical/mechanical specs as GB200 — a "drop-in design" that lets CSPs transition on existing manufacturing tooling. B300 GPUs carry 50% more HBM and deliver another 50% increase in dense FP4 inference compute vs. B200.

"Our Blackwell ramp, the fastest in our company’s history, drove a 73% year-on-year increase in data center revenue. Blackwell contributed nearly 70% of data center compute revenue in the quarter, with a transition from Hopper nearly complete... On average, major hyperscalers are each deploying nearly 1,000 NVL72 racks, or 72,000 Blackwell GPUs per week, and are on track to further ramp output this quarter." — Colette Kress, CFO

Networking is the under-appreciated story: $5B in Q1 with sequential growth resumed (+64% Q/Q). Three pillars (NVLink scale-up, InfiniBand, Spectrum X scale-out Ethernet), now arguably four with BlueField. NVLink alone exceeded $1B in Q1 shipments, with NVLink Fusion announced at Computex enabling semi-custom CPUs/accelerators (MediaTek, Marvell, Alchip, Astera Labs, Fujitsu, Qualcomm). Spectrum X annualizing >$8B; added Google Cloud and Meta as new CSP customers.

"Spectrum X has been really quite frankly a home run in this last quarter." — Jensen Huang, CEO

Assessment: The Data Center number was modestly light vs. an extreme Street bar, but the underlying operational signals are uniformly bullish: Blackwell ramp on schedule, GB300 already sampling, hyperscaler deployment cadence accelerating, networking emerging as a $20B+ run-rate business in its own right. The China revenue gap is the only meaningful pull on the segment; ex-China, growth re-accelerated.

Gaming and AI PC

$3.8B record (+42% YoY, +48% Q/Q). Blackwell gaming ramp is also "our fastest ever." RTX 5060 / 5060 Ti launched at $299, with laptops starting at $1,099. Improved supply availability vs. last quarter. Notably, the Nintendo Switch 2 leverages NVIDIA’s neural rendering and AI tech with a custom RTX GPU + DLSS — a multi-year platform tailwind given Switch’s 150M+ cumulative installed base.

Assessment: Gaming was a clean upside surprise. The segment has flipped from "supply-constrained but growing" to "supply-improved and accelerating," and the Switch 2 design win compounds for years. Not a thesis-mover by itself, but it’s the kind of broad-based segment strength that signals the franchise is firing on all cylinders.

Pro Visualization, Automotive

Pro Viz $509M (+19% YoY), flat sequentially with tariff-related uncertainty temporarily impacting systems demand. Sequential growth resumes in Q2. DGX Spark (1 PFLOPS desktop) and DGX Station (20 PFLOPS, GB300-powered) were the headline product disclosures; Spark in calendar Q3, Station later this year. Omniverse continues to compound with TSMC, Foxconn, Pegatron, Quanta usage.

Automotive $567M (+72% YoY) on self-driving ramp and NEV demand. Mercedes Benz full-stack solution in production starting with the new CLA. GM partnership for next-gen vehicles, factories, robots. Isaac GROOT N1 (humanoid foundation model) and Cosmos (world foundation models) released.

Assessment: Both segments are small relative to Data Center but compounding. Automotive’s +72% YoY is notable as a multi-year emergence story; Pro Viz is the early read on the on-prem enterprise AI opportunity (DGX Spark/Station + RTX Pro server). Neither segment is a near-term swing factor; both are call options.

Key Topics & Management Commentary

Overall Management Tone: Confident and direct. Colette Kress walked the H20 charge with unusual specificity (numbers, timing, materials reuse, foregone revenue). Jensen Huang devoted a full prepared-remarks block to AI policy — export controls, AI diffusion rule rescission, onshoring — in a way that read as deliberately strategic, not reactive. Q&A was unusually focused on China and the cadence of growth through the year; management answered directly without hedging.

The H20 Charge and the China Cliff

The cleanest disclosure on the call. Per Kress: $4.6B H20 recognized in Q1 prior to April 9 export-control change; $4.5B writedown on inventory and purchase obligations; $2.5B that could not be shipped. Q2 outlook reflects $8B of foregone H20 revenue (the orders that would have shipped this quarter). The new export-control rules "did not provide a grace period" for sell-through. The $4.5B charge "was less than what we initially anticipated as we were able to reuse certain materials."

"In Q1, we recognized $4.6 billion in H20 revenue, which occurred prior to April 9, but also recognized a $4.5 billion charge as we wrote down inventory and purchase obligations tied to orders we had received prior to April 9. We were unable to ship $2.5 billion in H20 revenue in the first quarter due to the new export controls." — Colette Kress

Jensen’s framing was sharper:

"The H20 export ban ended our Hopper data center business in China. We cannot reduce Hopper further to comply. As a result, we are taking a multibillion-dollar write-off on inventory that cannot be sold or repurposed. We are exploring limited ways to compete, but Hopper is no longer an option." — Jensen Huang

On the longer-term China-AI TAM: management sized it at "close to about $50 billion in the future," now effectively closed to NVIDIA. On a future replacement SKU: "We don’t have anything at the moment. But we’re considering it... the limits are quite stringent at the moment. And we have nothing to announce today."

Assessment: The China cliff is clean, quantified, and absorbed. From a modeling standpoint this is the best possible outcome short of a policy reversal — the bear case ("China overhang continues to drag indefinitely") is now dead. Either NVIDIA gets a future-compliant SKU (upside), or the $50B TAM stays closed (today’s baseline). The franchise still grows sequentially in Q2 with the $8B already gone.

Inference Demand — Reasoning Models as the Step-Function

Reasoning AI was the structural-thesis topic of the call. Kress quoted Microsoft processing >100 trillion tokens in Q1, a 5x YoY jump. NVL72 delivers up to 30x higher inference throughput vs. an 8-GPU H200 submission on the Llama 3.1 benchmark; software optimizations alone have lifted Blackwell performance another 1.5x in the last month.

"Reasoning is compute-intensive and requires hundreds to thousands more tokens per task than previous one-shot inference. Reasoning models are driving a step function surge in inference demand. AI scaling laws remain firmly intact not only for training but now inference too, requiring massive scale compute." — Jensen Huang

On Joe Moore’s opening question (how big is inference, and do you need NVL72 racks for reasoning inference):

"Compared to Hopper, Grace Blackwell is some forty times higher speed and throughput. And so this is going to be a huge benefit in driving down the cost while improving the quality of response with excellent quality of service at the same time. So that’s the fundamental reason. That was a core driving reason for Grace Blackwell NVL72." — Jensen Huang

Assessment: The "AI scaling laws are dead" narrative that surfaced after DeepSeek’s release in January is now decisively refuted by the data. Scaling has shifted from training-only to training-plus-inference, and inference volumes are growing exponentially. The fact that reasoning models consume 100−1,000x the tokens per task is structurally bullish for compute demand, and NVL72’s 30x throughput advantage is the right architecture for the workload.

Sovereign AI — The Hidden Growth Engine

This was the most surprising disclosure of the call. NVIDIA-powered AI factories: ~100 in flight this quarter (2x YoY); average GPU count per factory also doubled. ~800 in planning. Saudi Arabia 500 MW project (announced on the Trump Middle East tour). UAE 5 GW AI campus. Taiwan’s first AI factory partnership with Foxconn + Taiwan government. Sweden national AI infrastructure launched the prior week. Pipeline includes Japan, Korea, India, Canada, France, UK, Germany, Italy, Spain.

"We have a line of sight to projects requiring tens of gigawatts of NVIDIA AI infrastructure in the not-too-distant future." — Colette Kress

And from Jensen on the AI diffusion rule rescission:

"Sovereign AI is a new growth engine for NVIDIA." — Jensen Huang

Assessment: A "tens of GW" pipeline is a meaningfully larger TAM than is on Street numbers today. Even if a tenth of the 800 planned AI factories materialize at modest scale, the implied multi-year revenue contribution is substantial. The AI diffusion rule rescission removed the regulatory friction; the Trump administration’s Middle East tour locked in headline deals; and Jensen’s European tour next week is positioned to stack more. This is the upside vector that materially offsets the China loss in scale — and it was largely off the analyst checklist a quarter ago.

Gross Margin Trajectory — Mid-70s Late This Year

Kress was unusually direct on the GM walk. Ex-charge Q1 71.3%; Q2 72% (mid) on better Blackwell profitability; mid-70s targeted late this fiscal year. The mechanism: Blackwell yield improvement (continued), software-driven performance gains (Hopper improved 4x over its lifecycle), NVLink mix accretion, and continued networking ramp.

"We expect better Blackwell profitability to drive modest sequential improvement in gross margins. We are continuing to work towards achieving gross margins in the mid-seventies range late this year." — Colette Kress

Assessment: This is the GM recovery trajectory we wanted to see at Q4. Mid-70s late this fiscal year (~late calendar 2025 through early calendar 2026) implies Q3/Q4 GM of 73−75%, which gets us back to pre-Blackwell levels with the ramp behind us. The ASP pricing power thesis on Blackwell is intact.

Onshoring and the U.S. Manufacturing Build-Out

Jensen’s prepared remarks devoted a notable block to U.S. onshoring: TSMC building six fabs and two advanced packaging plants in Arizona for NVIDIA chips, with volume production by year-end. Spil + Amkor in Arizona for packaging/assembly/test. Foxconn partnering on a million-square-foot factory in Houston; Wistron building similar in Fort Worth. "Substantial long-term purchase commitments" backing the investments. Goal: chip-to-supercomputer built in America within a year.

Assessment: The manufacturing onshoring narrative serves three purposes: (1) tariff-resilience (real near-term), (2) policy alignment with the Trump administration (politically valuable in the export-controls negotiation), (3) supply-chain redundancy for what is increasingly critical infrastructure. Each GB200 NVL72 rack contains 1.2M components and weighs ~2 tons — the engineering complexity is its own moat.

Guidance & Outlook

The Q2 guide is the most important number on the call. $45B at the mid — up sequentially — while absorbing $8B of foregone H20 revenue. Implied core (ex-China) sequential growth is roughly $9B Q/Q. This is the de-risking event for the FY26 trajectory.

Implied 2H ramp: The mid-70s GM target for late FY26 + continued Blackwell + GB300 production + sovereign AI deals + networking acceleration suggest Q3 likely $48–52B and Q4 $52–58B in our framework. FY26 lands $190–205B in our base case — vs. prior Street consensus near $185B. EPS power on the GM expansion + continued OpEx leverage points to $4.50–5.00 FY26 vs. prior ~$4.20.

Guidance style: NVDA has historically guided modestly conservatively and beaten by 2–5%. The Q2 $45B mid is in the same spirit despite absorbing $8B in H20. Track record suggests Q2 prints somewhere in the $46–47B range.

Analyst Q&A Highlights

Inference Sizing and Reasoning Compute

• Joe Moore, Morgan Stanley: Asked how much of the surging reasoning-inference demand NVIDIA can serve and whether NVL72 rack-scale solutions are required for reasoning inference going forward. Jensen confirmed NVL72 is the ideal architecture for reasoning workloads, citing the 100−1,000x token amplification of reasoning vs. one-shot inference and Blackwell’s 40x throughput advantage over Hopper.
  Assessment: The framing question of the call. The answer establishes that reasoning AI is structurally an NVL72-class workload, not something that can be served by older or competing architectures. This is the "inference moat" data point.

China Modeling and the Trillion-Dollar Build-Out

• Vivek Arya, Bank of America: Two-part — clarification on Q2/run-rate China impact for Colette, and a question for Jensen on where the trillion-dollar AI build-out stands and whether digestion periods should be expected. Kress confirmed Q1 unable-to-ship $2.5B + recognized $4.6B = "should have been $7B" in Q1; Q2 H20 orders had been planned at $8B, all gone. Future quarters will reflect "a bigger issue regarding the amount of the market that we will not be able to serve" — the ~$50B China-AI TAM. Jensen reframed: AI is essential infrastructure for every country and every industry; "we’re really at the very beginning."
  Assessment: The $7B "should have been" Q1 H20 number is useful for reconciliation. Jensen’s answer was deliberately macro — he is positioning the build-out as multi-decade infrastructure, not a capex cycle. Whether the Street accepts that framing is a different question.

GPU Cluster Visibility and Lead Times

• C.J. Muse, Cantor Fitzgerald: Asked about announced large-scale GPU cluster orders (Saudi, UAE, Oracle, xAI), unannounced orders of similar magnitude, and Blackwell lead-time visibility. Jensen confirmed orders are running ahead of GTC framing; supply chain ramping; "umpteen AI factories being planned"; ~800 AI factories in some stage. He used the Internet/electricity-as-infrastructure analogy to frame the duration.
  Assessment: The "800 AI factories" number and "umpteen" unannounced sovereign deals suggest the order book is materially deeper than the named-customer disclosure. This is the visibility data point that supports the no-digestion case.

Q2 Bridge and Conviction in Sequential Growth

• Ben Reitzes, Melius: Pressed Kress on the math: the $8B H20 hole implies the rest of the business is doing $2–3B better than Street to hit the $45B mid — what is driving the upside? And asked Jensen on multi-quarter conviction given AI diffusion rescission + sovereign demand. Kress confirmed the mechanic and pointed to Blackwell strength. Jensen named "four positive surprises" since GTC: (1) reasoning AI step-function inference demand, (2) AI diffusion rule rescission, (3) enterprise AI traction (RTX Pro server, agents working), (4) industrial AI / Omniverse / robotics emerging.
  Assessment: The four-pillars framing is the strongest forward-looking signal of the call. Jensen rarely volunteers conviction language — "four positive surprises since GTC" is meaningful.

China SKU Replacement

• Timothy Arcuri, UBS: Asked directly whether NVIDIA has been approved to ship a modified version into China and whether the ~$7–8B/qtr China run-rate could be recovered. Jensen was unambiguous: Hopper is "the end of the road" under the new limits; "we don’t have anything at the moment. But we’re considering it... the limits are quite stringent... we have nothing to announce today."
  Assessment: The clearest possible "no" on near-term China-SKU recovery. This sets the floor: assume zero China-AI revenue, model upside if/when a compliant SKU emerges. We treat this as the baseline.

Networking Strength

• Aaron Rakers, Wells Fargo (Jake on for): Asked about networking strength, Ethernet adoption at CSPs, attach rate trends. Jensen walked the four-pillar networking platform: NVLink scale-up; InfiniBand; Spectrum X (Ethernet enhanced with low-latency, congestion control, adaptive routing); BlueField (control plane + storage + security). Cluster utilization improving "from as low as 50% to as high as 85%, 90%" with Spectrum X — on a $10B cluster, that is $4B of equivalent value. Two new significant CSPs added to Spectrum X this quarter (Google Cloud, Meta).
  Assessment: Networking is increasingly a discrete revenue lever, not a tag-along to GPU sales. The cluster-utilization economics ("$4B of equivalent value on a $10B cluster") explain why hyperscalers are paying for the full-stack solution rather than cobbling Ethernet themselves.

What They’re NOT Saying

1. No specific FY26 revenue guide. NVIDIA does not give annual guides; the $45B Q2 mid is the closest forward signal. Implied $190B+ FY26 trajectory is investor-derived, not management-confirmed.
2. No quantification of sovereign AI revenue contribution. ~100 AI factories in flight, ~800 in planning, "tens of gigawatts" line of sight — but no $-amount on what this contributes to FY26/FY27. This is the largest unquantified upside vector in the model.
3. No GB300 unit/revenue color. "Sampling at major CSPs" + "production shipments later this quarter" but no breakout of GB300 vs. GB200 mix in Q2. Material question for Q3/Q4 modeling.
4. No commentary on hyperscaler capex assumptions or visibility windows. Jensen avoided the "what does FY27 hyperscaler capex look like" question; reasonable, but the Street will keep asking. Hyperscaler capex inflection is a gating risk.
5. No explicit commentary on competing custom silicon (TPU, MTIA, MAIA, Trainium). The implicit answer is "Blackwell wins on programmability + ecosystem + NVL72 architecture," but management chose not to engage directly. Bull case argues this is unnecessary; bear case argues custom silicon is a long-tail share-loss vector.
6. No update on Hopper installed-base decommissioning timeline. Hopper transition "nearly complete" on the new-shipment side, but the existing installed base of H100/H200 will continue to consume software/services revenue for years — not quantified.
7. No commentary on Stargate or other named hyperscale projects beyond the prepared-remarks list. Notable given the headline $500B Stargate disclosure earlier in the year.
8. No comments on a future-compliant China SKU timeline. Jensen explicitly said "nothing to announce." The optionality exists but is not in plan.

Market Reaction

• Pre-print setup: NVDA closed near $135 ahead of the print, off the recent highs in the $150 area but well off the early-April low near $86 (during the China-export-controls news flow and the broader tariff-driven AI-stock drawdown). Implied move on options: ~7%. The setup was tense — bears pointing to the H20 charge as the catalyst for a multi-quarter re-rating; bulls expecting the print to clear the China overhang.
• After-hours move: Stock initially down ~1–2% on the headline GAAP miss (the $4.5B charge dragged GAAP GM to 60.5%) before reversing to up ~4% as the call digested the (a) ex-charge GM upside, (b) Q2 guide absorbing $8B in H20 revenue, (c) sovereign AI scale, and (d) Jensen’s four-pillars framing. Stock heading into May 29 cash session at roughly $140–142.

This is a print-clears-the-overhang reaction. The market repriced for "China cliff is taken; Blackwell ramp is on; sovereign AI is a real new vector" — the three things bears and bulls had been arguing about for two quarters. The post-print levels still leave room: at ~$140 stock, NVDA trades at roughly 28x our updated FY27 EPS power, vs. multi-year average closer to 35x. The China hit is priced; the sovereign-AI upside isn’t.

Street Perspective

Debate: Is the Blackwell Ramp Past Peak Growth?

Bull view: Blackwell is ~70% of DC compute with Hopper transition still completing — meaning the substitution effect that flatters YoY growth has another 1–2 quarters to run. GB300 sampling now at major CSPs with production shipments this quarter; the next-architecture cycle is overlapping with the current ramp, smoothing the transition. The ~800 AI factories and tens-of-GW pipeline argue this is a multi-year build-out, not a peak-cycle.

Bear view: Hyperscaler capex on AI infrastructure has tripled in two years; mean-reversion arguments apply. Custom silicon (TPU, MTIA, Trainium) is gradually accumulating share at the largest customers. The DeepSeek-cost narrative, while temporarily refuted by reasoning-inference demand, will resurface if reasoning models prove less compute-intensive than current pace suggests. Comps get progressively harder through FY26.

Our take: The bull view has the better evidence in the Q1 print. The 1.5x software-driven Blackwell performance lift in a single month is a reminder that the moat is software + architecture + ecosystem, not just raw silicon. Custom silicon will take share at the margin but is not a structural threat through FY26–FY27. Mean-reversion on hyperscaler capex is the real risk — we acknowledge it but don’t see it landing in this fiscal year.

Debate: Is Sovereign AI a Real Number or Press Releases?

Bull view: ~100 AI factories in flight (already shipping), ~800 in planning, named deals in Saudi, UAE, Taiwan, Sweden + multi-country European pipeline, line-of-sight to "tens of gigawatts." Jensen is on a multi-week European tour locking in additional country deals. The AI diffusion rule rescission removed the regulatory friction. Even modest realization rates produce material multi-year revenue.

Bear view: Sovereign AI deals have long lead times, complex financing, and political-cycle risk. Country-scale projects have been announced before (smart cities, etc.) and underdelivered. Without quantified guidance, the upside is speculative.

Our take: Sovereign AI is real but the timing is more back-end-loaded than the headline "tens of GW" suggests. Realistic FY26 contribution ~$3–6B; FY27 contribution potentially $10B+; long-tail meaningful. The mistake is to ignore it; the mistake is also to model it as smooth ramps. We treat it as upside optionality, not base case.

Debate: How Much Should the China-AI TAM Loss Cost the Multiple?

Bull view: $50B closed TAM is a ceiling on FY26–FY27 lost revenue, much of which would have come at structurally lower margin (H20 was below the corporate average). Q2 absorbs $8B in foregone H20 and the franchise still grows sequentially — the operating leverage on the rest of the business is so positive that the China cliff is, in earnings terms, a one-time hit rather than a structural drag. Sovereign AI offsets the absolute scale.

Bear view: $50B TAM lost is real money. Worse, China is not just lost as a customer market but is increasingly a competitor — Huawei’s Ascend ramp benefits from the regulatory shelter. Long-term geopolitical-stack-fragmentation risk is real and undermines the "one global AI infrastructure" thesis.

Our take: The bear view is correct in absolute terms but priced wrong. Stock at ~$140 already discounts the China cliff. The competitive-shelter-for-Huawei concern is real but plays out over 3–5 years, not the rating window. We size the impact at ~10–12% off our FY26 revenue estimate vs. a no-export-control counterfactual; that is meaningful but absorbed.

Model Implications

Valuation impact: Our prior fair-value framework was $115–140, set at the Q4 FY25 print. We raise to $145–180:
• Low end $145 = ~30x forward FY27 EPS of ~$4.85 (mid)
• High end $180 = ~37x forward FY27 EPS of ~$4.85 (in line with multi-year average)
The new range starts modestly above the post-print level near $140 and gives meaningful room without requiring multiple expansion. The China hit is priced in; the sovereign-AI upside and continued GM expansion are not. Conviction is high on the operating story; multiple-compression risk is real if the broader AI capex cycle digests.

Thesis Scorecard Post-Earnings

Overall: Thesis materially de-risked. The two open questions from Q4 FY25 (Blackwell GM trajectory; China-policy resolution) have both landed: GM recovering on schedule; China resolved in the bear direction but quantified. Sovereign AI is the upside vector that wasn’t on the prior-quarter checklist. The franchise is firing on more cylinders than at any point in the last 12 months.

Action: Upgrading to Outperform from Hold; raising fair value to $145–180 from $115–140. Post-print level near $140 is at the lower bound of the new range, leaving room without requiring multiple expansion. We do not chase but would build positions on any pullback. Holders should reassess sizing if the stock breaks above ~$175 (top half of fair value) or if hyperscaler capex shows clear digestion in upcoming reporting.