NVIDIA Q1 FY2027 Earnings Deep Dive: The Technical Signals and Strategic Gambits Behind $81.6B

Published: May 21, 2026 · Author: Phoenix Lee (locsic.com) Based on NVIDIA official earnings (quarter ended 2026.04.26) and earnings call Core thesis: Agentic AI demand has gone "parabolic," but NVIDIA faces a triple squeeze — customers turning into competitors, China revenue going to zero, and LPU being downgraded

---

One-Line Takeaway

NVIDIA delivered an $81.6B quarter in Q1 FY2027, beating Wall Street consensus by ~$2.5B — but the real story isn't the number itself, it's the structural shifts behind it: networking revenue is morphing from an "accessory" into a "profit center" within data center revenue; Hyperscale and ACIE revenues are nearly equal for the first time; the Vera CPU is opening up a $200B new market; and Jensen Huang's language around the LPU has been downgraded from "the seventh chip" at GTC to "niche product" on the earnings call.

The buildout of AI factories is the largest infrastructure expansion in human history — this isn't rhetoric, it's fact. The real question is how long NVIDIA can capture the full红利 of this expansion.

---

Core Data at a Glance

Shareholder Returns:

• Buybacks + dividends this quarter: $20B
• New buyback authorization: $80B
• Quarterly dividend: $0.01 → $0.25 (25x increase)
• Private company / infrastructure fund investments: $18.6B (this quarter)

Important context: The Q2 guidance of $91B excludes China data center revenue, and the company expects ~$8B in losses due to H20 export restrictions. Q1 GAAP figures include an ~$5.5B H20 inventory write-down.

---

Historical Revenue Trend: Where We Are on the Growth Curve

Trend Analysis:

1. YoY growth re-accelerated in Q1 FY27. After declining from 78% in Q4 FY25 to 69% in Q1 FY26, YoY growth started rising again in Q2 FY26, reaching 85% in Q1 FY27. This tells us AI infrastructure demand, after a brief "growth normalization," has entered a second acceleration — driven by a shift from "training" to "training + inference + Agentic AI."

2. Gross margin stabilizing around 75%. After falling from 78.4% in Q1 FY25 to 73.0% in Q4 FY26, it rebounded to 75.0% in Q1 FY27. This V-shaped recovery is driven by higher ASPs on Grace Blackwell systems and product mix improvement from the rising share of networking revenue.

3. Explosive networking revenue growth. Networking revenue appeared "stagnant" — $11.0B in Q4 FY25 to $10.9B in Q4 FY26 — then surged to $14.8B in Q1 FY27 (+35% QoQ, +199% YoY). This jump reflects the massive NVLink Switch demand driven by Grace Blackwell NVL72 shipments.

---

1. Behind the $81.6B: What the Numbers Tell Us

1.1 Quality Analysis of the Guidance Beat

NVIDIA has beaten Wall Street consensus for seven consecutive quarters. This time the beat was ~3% ($81.6B vs $78.8-79.2B). On the surface, it's a "routine beat" — but for a company approaching $100B in quarterly revenue, a 3% miss means ~$2.5B in extra revenue, more than many mid-cap semiconductor companies generate in an entire year.

But what's more worth analyzing is the structure of the beat.

Compute revenue $60.4B (YoY +77%) — growth is slower than data center overall (+92%), indicating compute revenue growth is "normalizing." Grace Blackwell's capacity ramp has reached a steady state, no longer enjoying the "scarcity premium" from supply-constrained demand.

But "normalizing" doesn't mean "slowing down." $60.4B in quarterly compute implies a ~$242B annualized run rate — still a dizzying number. Considering Vera Rubin's ASP may be even higher than Grace Blackwell (because it includes the Vera CPU + more NVLink components), compute revenue could re-accelerate in Q3-Q4 FY27.

Networking revenue $14.8B (YoY +199%, QoQ +35%) — this is the most underrated number in the entire earnings report. $14.8B in quarterly networking implies a $59B+ annualized run rate. If spun out as an independent business, NVIDIA's networking division would already be larger than most standalone semiconductor companies. More importantly, the 199% YoY growth signals that NVLink and Spectrum-X are transitioning from "free bundling" to "independently priced" profit centers.

Edge Computing $6.4B (YoY +29%) — relatively modest growth, but given Edge's already substantial base, 29% is still healthy. This segment includes autonomous driving (DRIVE Thor), robotics (Isaac), and edge AI inference. Worth noting: Edge isn't directly impacted by China export restrictions (main customers are Western auto and industrial companies), so it may serve as a "cushion" against the China cutoff in coming quarters.

EPS $1.87 (YoY +140%) — EPS growth far exceeds revenue growth (+85%), showing operating leverage continues to work. NVIDIA's OpEx growth is far below revenue growth — a hallmark of a "platform company": once the infrastructure (CUDA, software stack, customer relationships) is built, the marginal profit on every incremental dollar is extremely high.

1.2 Breaking Down Data Center $75.2B: Compute vs. Networking

Networking's share has risen from ~13% a year ago to nearly 20%. This trend will only accelerate — because Grace Blackwell NVL72 and the upcoming Vera Rubin NVL72 are deeply dependent on NVLink interconnects, and the number of NVLink Switches per rack keeps growing.

A number worth pondering: NVIDIA's networking business, annualized from Q1 ($59B), is already close to or may exceed 8x Arista Networks' revenue (market cap ~$120B, annual revenue ~$7B). NVLink is becoming NVIDIA's most underestimated moat — it's not just a technical advantage, but an independent profit center.

Another overlooked angle: Networking revenue growth isn't just "more NVLink Switches." It also includes contributions from Spectrum-X Ethernet switches and ConnectX NICs. As AI Cloud providers (CoreWeave, Lambda, etc.) build out new data centers at scale, they need to build networking infrastructure from scratch — creating massive incremental demand for Spectrum-X.

1.3 New Reporting Structure: Hyperscale $38B vs ACIE $37B

CFO Colette Kress disclosed the data center revenue split by customer type for the first time on the earnings call:

• Hyperscale: ~$38B, ~50.5% of data center revenue
• ACIE (AI Clouds / Consumer Internet / Enterprise): ~$37B, ~49.5% of data center revenue

This breakdown is significant and worth analyzing point by point.

First, customer concentration is lower than expected. Hyperscale accounts for "only" about half, meaning NVIDIA's customer base is more diversified than the market assumes. There's been persistent concern about over-reliance on a handful of hyperscale customers (Microsoft, Google, Meta, Amazon), but the ACIE segment — including AI Cloud companies like CoreWeave, Lambda, NEBIUS, plus consumer internet and enterprise customers — contributes nearly as much revenue. This reduces "single customer loss" risk.

Second, the rise of AI Cloud is stunning. CoreWeave as one of NVIDIA's top two customers (confirmed by Jensen at GTC 2026) represents a new type of "middleman" customer: they don't develop their own ASICs but rely on NVIDIA's full-stack solution to provide AI compute services. These customers have much higher stickiness than Hyperscale — because they lack the scale and talent to pursue custom silicon. CoreWeave's business model is essentially "NVIDIA compute reseller," and its interests are fully aligned with NVIDIA's.

Third, the Consumer Internet share within ACIE (Meta, ByteDance, etc.). While these companies are also developing custom chips (Meta's MTIA v3), their in-house progress lags behind Google, and their near-term dependence on NVIDIA remains strong. Meta's FY27 CapEx guidance is ~$60-65B, mostly for AI infrastructure — a significant share of that flows to NVIDIA.

Fourth, the near-balance between Hyperscale and ACIE provides NVIDIA a "pricing power buffer." When Microsoft or Google threaten to replace GPUs with in-house TPU/ASIC, NVIDIA can redirect capacity to AI Cloud and enterprise customers. This buffer is especially valuable in a semiconductor upcycle — because capacity is always tight, and NVIDIA has no trouble selling.

But there's a risk to note: AI Cloud companies in the ACIE segment (like CoreWeave) are themselves financially fragile. If AI inference monetization underwhelms, these companies could face debt pressure — directly impacting NVIDIA's revenue quality. NVIDIA invested $18.6B this quarter in private companies and infrastructure funds, partly to "lock in" these AI Cloud customers, but this also increases NVIDIA's credit risk exposure.

1.4 GAAP vs Non-GAAP: The Real Impact of the $5.5B Write-down

Before diving into gross margin analysis, it is worth clarifying the GAAP vs Non-GAAP differences in Q1:

GAAP net income of $42.96B ($1.76/share) vs prior year $18.8B ($0.76/share), YoY growth of +129%. Even after the $5.5B H20 write-down, NVIDIA's GAAP profit growth is historic.

The $5.5B write-down reflects H20 chips already manufactured or on order before the export license requirement was imposed. These chips can no longer be sold to Chinese customers.

The $8B Q2 loss projection is not a write-down but rather foregone revenue. NVIDIA could have sold these chips to China, but now cannot. The Q2 guidance of $91B already excludes this $8B, meaning the true global demand figure is closer to $99B.

This is a critical insight: NVIDIA's Q2 guidance of $91B actually implies non-China global demand of approximately $99B, which is even stronger than the headline number suggests.

1.5 Sustainability of 75% Gross Margin (with Historical Trend)

NVIDIA's Non-GAAP gross margin has stayed in the 73-76% range for multiple consecutive quarters. Q1 came in at 75.0%, and Q2 guidance is also 75.0% ±50bps.

Three factors supporting the 75% gross margin:

Factor 1: Continuous product mix shift toward the high end. Grace Blackwell NVL72 systems carry far higher ASPs than the previous-generation Hopper, and NVLink networking components typically enjoy higher margins than pure GPU compute. NVIDIA is shifting from "selling chips" to "selling systems" — system-level solutions command much stronger pricing power than bare chips, because switching costs are far higher. A customer already running Grace Blackwell NVL72 can't easily switch to AMD MI400 or Google TPU — because the entire software stack (CUDA + Dynamo + NemoClaw) and operations framework would need to be rewritten.

Factor 2: Software lock-in effect. The CUDA ecosystem, Dynamo inference engine, and NemoClaw Agent platform — these software stacks make it extremely difficult for customers to do apples-to-apples substitution, thereby supporting NVIDIA's pricing power. When a customer has trained models on CUDA, used Dynamo for inference optimization, and deployed Agents with NemoClaw — switching to AMD isn't just a hardware swap, it's a full software stack rebuild. This "ecosystem lock-in" allows NVIDIA to sustain a hardware premium.

Factor 3: Supply-constrained "selective selling." When demand far exceeds supply, NVIDIA can prioritize high-margin customers and product mix. Vera Rubin's capacity ramp is expected to continue throughout FY27, meaning for at least the next 4-5 quarters, NVIDIA remains in a "seller's market." In a seller's market, gross margins are naturally protected.

Three risks to gross margin:

Risk 1: Long-term substitution effect from customer-designed ASICs. Once Google's TPU JV begins operations (500MW capacity expected online in early 2027), NVIDIA's share in Google data centers may decline. Google might demand lower prices to maintain the NVIDIA relationship — or simply reduce GPU procurement. Similarly, Amazon's Trainium 3 has already achieved cost/performance parity with NVIDIA GPUs on certain inference workloads.

Risk 2: The gross margin impact of rising networking share is unclear. Although networking revenue is growing fast, NVLink Switch and Spectrum-X switches have different margin characteristics than GPUs. NVLink Switches don't need expensive HBM, but they require large volumes of high-speed SerDes and advanced packaging — which also carry costs. If networking margins are lower than GPU margins, then the rising networking share could actually dilute overall gross margin.

Risk 3: Financial impact of geopolitical risk. The $5.5B inventory write-down and $8B revenue loss from H20 export restrictions are already reflected in Q1's GAAP figures (GAAP net income $42.96B, GAAP gross margin ~73.5% vs Non-GAAP 75.0%). If export restrictions extend to more countries (e.g., certain Middle Eastern nations), similar write-downs could recur.

My judgment: 75% gross margin is sustainable in the medium term (FY27-28), but faces downward pressure in the long term (FY29+). The key variables are:

1. Pricing trajectory of Grace Blackwell Ultra and Vera Rubin. If NVIDIA continues selling as "system-level solutions" (GPU + NVLink + Vera CPU + software) rather than "bare chips," margins can hold or even improve. System-level solutions command far higher pricing power than bare silicon.

2. Competitive landscape in inference. Inference has lower GPU stickiness than training — because inference doesn't require CUDA (any compatible inference framework works). If customer-designed ASICs break through in the inference market, NVIDIA's inference pricing power will erode, pulling down gross margins.

A number worth watching: NVIDIA's Q2 gross margin guidance is 75.0% ±50bps — despite losing the China market (which typically carried lower margins), gross margin didn't increase; it stayed flat. This suggests NVIDIA may face some pricing pressure in non-China markets, or the dilution effect from higher networking share has already begun to show.

---

2. Vera Rubin: A Stronger Starting Point Than Grace Blackwell

2.1 Analyzing Jensen's Exact Words

Jensen Huang's comments about Vera Rubin on the earnings call are worth parsing word by word:

"Vera Rubin is off to a tremendous start." "It will be even more successful than Grace Blackwell." "It will be constrained throughout its entire life."

These three sentences send three different signals:

"Tremendous start" — confirms Vera Rubin has entered full production (Jensen personally announced "full production" at CES 2026). Microsoft and CoreWeave will be first-wave deployment customers. Some partners are already running next-gen AI models on Vera Rubin systems. This is the first time Jensen has used "tremendous" to describe a new product line's launch — his description of Grace Blackwell was "off to a great start."

"More successful than Grace Blackwell" — an exceptionally bold prediction. Grace Blackwell is one of NVIDIA's most successful products ever; Q1 FY27 data center revenue was almost entirely driven by Blackwell. Jensen predicting Vera Rubin will surpass this implies his read on the demand curve is that AI factory buildout has barely begun.

"Constrained throughout its entire life" — this is the most important signal. Jensen is saying that from production launch to retirement, demand will exceed supply for Vera Rubin's entire lifecycle. This isn't marketing spin; it's based on his real assessment of customer orders and capacity planning. If you put this alongside NVIDIA's simultaneous announcement of $80B in new buyback authorization — management has extremely high confidence in generating excess cash flow in coming quarters.

An additional signal: Jensen also said on the call —

"The buildout of AI factories — the largest infrastructure expansion in human history — is accelerating at extraordinary speed."

Connecting all of Jensen's statements, the narrative is: AI factory buildout is "the largest in human history," demand growth is "extraordinary," and Vera Rubin will have lifetime demand exceeding supply. This is an extremely bullish judgment — but given NVIDIA has beaten consensus for seven consecutive quarters, Jensen's optimism has at least been validated by the data so far.

2.2 Delivery Challenges of 1.3M Components, 72 GPUs + 36 CPUs

Vera Rubin NVL72 is an extraordinarily complex system:

• 1.3 million discrete components — more than double Grace Blackwell NVL72's ~600K components.
• 72 Rubin GPUs + 36 Vera CPUs — NVIDIA's first rack combining self-designed GPU and self-designed CPU.
• 33.6 billion transistors per Rubin GPU — single-chip compute density reaches new heights.
• HBM4 memory — each Rubin GPU equipped with 288GB HBM4, 22 TB/s bandwidth. A single NVL72 system carries over 20TB of total HBM.
• NVLink 6 interconnect — 3.6 TB/s GPU-to-GPU bandwidth, a quantum leap over NVLink 5.
• ICMS storage — each NVL72 needs 1,152 TB of NAND storage, which could pressure global NAND supply.

The core delivery challenge lies in "system-level complexity":

Grace Blackwell's capacity ramp took over a quarter to stabilize, during which there were "liquid cooling design adjustments" and "NVLink interconnect debugging" issues. Vera Rubin's complexity far exceeds Grace Blackwell's, and delivery challenges will only be greater.

Specifically, Vera Rubin faces these key delivery challenges:

Challenge 1: Liquid cooling. 72 Rubin GPUs + 36 Vera CPUs + 36 NVLink Switches may draw over 120kW per rack. Grace Blackwell NVL72 draws ~100-120kW; Vera Rubin could be higher. This requires data centers with direct-to-chip liquid cooling capability — and the global share of data centers with this capability remains low.

Challenge 2: HBM4 supply. Rubin GPU uses HBM4 (not Grace Blackwell's HBM3e). HBM4 is the next generation of high bandwidth memory, and both SK Hynix and Samsung are still ramping production. If HBM4 supply is tight, Rubin GPU shipments will be constrained — this may be a key reason Jensen says "constrained throughout its entire life."

Challenge 3: System integration of 1.3M components. 1.3 million discrete components means supply chain management, quality control, and field deployment complexity grows exponentially. A delay or defect in any single component can delay the entire system's delivery.

Challenge 4: ICMS storage's impact on NAND supply. Each NVL72 needs 1,152 TB of NAND storage (~1.1 PB). If NVIDIA ships thousands of NVL72 systems in H2 FY27, NAND demand would reach exabyte scale — potentially creating perceptible pressure on global NAND supply and pushing up prices.

But NVIDIA learned important lessons from the Grace Blackwell ramp:

1. Liquid cooling systems — Vera Rubin uses liquid cooling throughout, requiring deep integration with data center cooling infrastructure. NVIDIA is working closely with data center operators to plan liquid cooling infrastructure in advance.
2. Vertical supply chain integration — NVIDIA is signing strategic agreements with optical interconnect vendors like Coherent, Corning, and Lumentum to secure optical module supply. HBM4 supply agreements with SK Hynix and Samsung are also being locked in early.
3. System integration partners — Moving from "selling chips" to "selling systems" requires stronger system integration capabilities. NVIDIA is expanding its professional services team to help customers deploy and maintain NVL72 systems.

My judgment: Vera Rubin's initial deliveries (H2 2026) will almost certainly see delays and capacity bottlenecks — this is virtually a sure thing. But Jensen saying "constrained throughout its entire life" means even with delays, customers' willingness to wait remains strong. In the AI infrastructure arms race, no customer wants to wait for the next-next generation.

Investor takeaway: Vera Rubin delivery delays aren't necessarily a negative signal — because "delays" means "demand exceeds supply," which itself supports high pricing and high gross margins. The real worry would be "quality issues" — if NVL72 experiences large-scale failures at customer sites, that would be material risk.

2.3 Comparison with Grace Blackwell Capacity Ramp

NVIDIA is clearly compressing product cycles — from Grace Blackwell's 9 months to Vera Rubin's ~6 months. This acceleration is both a competitive advantage (maintaining technical leadership) and a risk source (shorter validation cycles may lead to more early-stage issues).

A noteworthy detail: Vera Rubin's first-wave customers shifted from Google and Meta to Microsoft and CoreWeave. This may reflect Google redirecting more resources to in-house TPU (Blackstone JV), while Meta's MTIA v3 is also maturing. NVIDIA's core customer base is shifting from "primarily Hyperscale" to "Hyperscale + AI Cloud in tandem."

2.4 Practical Implications of 10x Performance/Watt

NVIDIA claims Vera Rubin delivers roughly a 10x improvement in inference cost vs. Grace Blackwell. Let's unpack that number:

• Compute performance: Rubin GPU FLOPS are approximately 2.5-3x Blackwell's.
• Memory bandwidth: HBM4 bandwidth improvement of ~2x (22 TB/s vs 12 TB/s).
• System-level optimization: NVLink 6 (3.6 TB/s vs 1.8 TB/s) + Vera CPU + Dynamo 1.0 contribute the remaining efficiency gains.
• "10x inference cost" is more of a system-level metric, incorporating inference engine (Dynamo) optimization, network bandwidth improvements, and CPU-GPU co-compute efficiency.

Practical implication: If NVIDIA delivers on its 10x inference cost reduction promise, per-token compute cost will go from "cheap" to "nearly negligible." This will directly accelerate AI commercialization — from "experimental deployments" to "production at scale."

The commercial significance of 10x, quantified:

Assuming Grace Blackwell's inference cost is $0.15 per million tokens (roughly the market price in late 2025), Vera Rubin's would be $0.015 per million tokens. At this price level:

• An AI service processing 1 billion tokens per day sees inference cost drop from $150,000/day to $15,000/day
• Annualized inference cost drops from $54.75M to $5.475M
• This means inference cost is no longer something to "optimize" — it becomes a negligible operating expense

If this 10x cost reduction is real, it will fundamentally transform AI business models. Many AI applications today are "unprofitable" primarily because inference costs are too high. If inference costs drop by an order of magnitude, AI application economics shift from "burning money for growth" to "healthy software margins."

---

3. Vera CPU: The $200B New Battlefield NVIDIA Just Opened

3.1 Why CPUs Matter Again in the Agentic AI Era

Jensen Huang spent considerable time on the earnings call discussing the Vera CPU, and this was no accident. His core argument:

"The 'thinking' part of AI uses the GPU, but Agent execution runs primarily on the CPU. Agents use CPUs to complete their assigned tasks, and in the future there will be CPUs specifically designed for Agents — Agent-driven PCs."

The technical foundation for this argument:

1. Agentic AI workload distribution. An AI Agent's lifecycle includes: perception (receiving input) → reasoning (calling LLM, GPU-intensive) → planning (logical reasoning, CPU-intensive) → execution (calling tools, APIs, file operations, CPU-intensive) → feedback (returning to perception).

Over 80% of an Agent's runtime actually executes on the CPU.

Why? Because most of an Agent's work isn't "generating tokens" (GPU's strength) but "calling tools" and "orchestrating tasks" (CPU's strength). A typical Agent workflow:

• Receive user request (network I/O, CPU)
• Call LLM to generate initial response (GPU, lasting hundreds of milliseconds to seconds)
• Call search engine for information (network I/O, CPU)
• Call LLM to generate final response based on search results (GPU)
• Write results to file or send email (file I/O and network I/O, CPU)
• Log audit trail (database I/O, CPU)

In this workflow, the GPU is invoked only twice, while the CPU is involved throughout the entire lifecycle. As Agent complexity increases (multi-step reasoning, multi-tool invocation, multi-Agent collaboration), CPU workload share will only grow.

2. Token processing becomes a core CPU task. Vera CPU is "the world's first CPU designed specifically for Agentic AI" — its design core is processing tokens as fast as possible. In an Agent's inference-execution loop, the CPU needs to continuously move, format, and route tokens between GPU and CPU. Traditional x86 CPUs are inefficient at this because they were designed for general-purpose computing (databases, web servers, virtualization), not high-speed token processing.

Vera CPU's 88 Olympus cores deliver 1.5x IPC (instructions per clock) improvement, with single-thread performance claimed to be the fastest on the market. This means Vera CPU can be 50-100% faster than Intel Xeon or AMD EPYC at token routing and orchestration tasks.

3. Agents need their own "PCs." Jensen predicts that in the future every Agent will need a dedicated computing device — not necessarily a traditional PC, but an independent compute unit with CPU + memory + network. This is an entirely new demand category.

The context for this prediction: most AI Agents today run on shared server clusters, isolated through containers or VMs. But if Agents need to run long-term ("always-on assistant") and need access to large amounts of local data and tools, shared clusters become less efficient — because of resource contention and isolation overhead. Dedicated CPU racks can provide independent compute environments for each Agent (or Agent team), improving execution efficiency and security.

3.2 $20B Target This Year vs. Intel/AMD's Existing Market

Jensen claimed on the earnings call:

"Vera opens an entirely new TAM of $200 billion for Nvidia — a market that didn't exist before." "Approximately $20 billion worth of Vera CPUs have already been sold so far this year."

What $200B TAM means: This number is roughly equal to the total global server CPU market. Jensen isn't saying Vera CPU will replace all x86 servers — he's saying Agentic AI creates an entirely new CPU demand category whose size could be comparable to the entire traditional server CPU market.

$20B already sold: This number is fairly credible because Vera CPU can be sold independently (Vera CPU Rack = 256 liquid-cooled CPUs in a pure-CPU cabinet) or bundled with Rubin GPU (Vera Rubin NVL72 = 72 Rubin GPUs + 36 Vera CPUs). Given NVIDIA's Q1 data center compute revenue of $60.4B, $20B from Vera CPU is reasonable.

What does $20B mean in context?

For comparison:

• Intel's 2025 Data Center and AI Group (DCAI) full-year revenue: ~$16B
• AMD's 2025 Data Center segment full-year revenue: ~$13B
• NVIDIA's Vera CPU reached $20B in its first half-year on the market

If this number is real, it means NVIDIA's CPU revenue in a single quarter has already surpassed Intel and AMD's full-year data center CPU revenue. This is a historic inflection point — a GPU company's CPU revenue exceeds that of CPU companies.

But a caveat: The "$20B in Vera CPUs already sold" may include the value allocation of Vera CPU in bundled sales. If a Vera Rubin NVL72 has a total price of $X, NVIDIA may allocate $Y of that to Vera CPU. This internal pricing approach makes the definition of "CPU revenue" somewhat flexible.

3.3 Strategic Significance of the CPU-Only Rack

The Vera CPU Rack is the most underrated new product in NVIDIA's lineup:

• 256 liquid-cooled Vera CPUs in a single cabinet
• 6x CPU throughput improvement (vs. traditional x86 cabinets)
• 2x Agentic AI workload performance

The strategic significance of the CPU-only rack:

1. Lowering the customer entry barrier. Not all AI workloads need GPUs. For Agent execution layers, data processing layers, and orchestration layers, pure CPU cabinets suffice. This lets NVIDIA enter markets that "don't need GPUs but need AI CPUs" — traditional enterprise IT departments, mid-size AI startups, and customers who only need Agent execution capability (not model training).

2. Expanding NVIDIA's share within data centers. A typical AI data center might already use NVIDIA at the compute layer (GPU racks) and networking layer (NVLink/Spectrum-X). If the CPU layer also uses Vera, NVIDIA penetration across the entire data center would exceed 90%. This means a customer's "de-NVIDIA-ification" cost would escalate from "replacing GPUs" to "replacing the entire data center compute layer."

3. Direct competition with AWS Graviton and Google Axion. AWS Graviton and Google Axion (ARM-based server CPUs) have been eating into traditional x86 share. NVIDIA entering this market with Vera CPU goes beyond "defending GPU territory" — it's saying "the entire data center compute layer should be NVIDIA's."

4. Foundation for "inference racks." In the Agentic AI era, a typical data center may have three types of racks: GPU racks (training and inference), CPU racks (Agent execution and orchestration), and networking racks (NVLink/Spectrum-X). NVIDIA already dominates GPU racks and networking racks. If CPU racks also use Vera, NVIDIA will control all compute layers across the entire AI data center.

3.4 Actual Impact on Intel and AMD

From a timeline perspective:

• Short-term (FY27-28): Limited impact. Vera CPU's primary customers are NVIDIA's existing AI infrastructure customers, who weren't core Intel/AMD targets anyway. Intel and AMD's traditional x86 markets (databases, ERP, web servers) won't immediately shrink due to Vera CPU's appearance.
• Medium-term (FY29-30): Impact starts to show. If Agentic AI truly scales as Jensen predicts, enterprise customers may start replacing traditional x86 deployments with Vera CPU. Especially customers already using NVIDIA GPUs — if their AI racks already contain Vera CPUs (as part of Rubin), extending to non-AI workloads on Vera is a natural progression.
• Long-term (FY31+): Depends on software ecosystem. If OpenShell and NemoClaw successfully establish the standard runtime environment for Agentic AI, Vera CPU will become the optimal hardware carrier — just as CUDA made the GPU the standard carrier for AI training. But this "long-term" prediction assumes massive commercialization of Agentic AI — an uncertain assumption.

AWS/Meta's CPU choices will also shape the competitive landscape. AWS is aggressively pushing Graviton (in-house ARM CPU), and Meta is developing its own ARM server CPU. If these Hyperscalers choose in-house CPUs over Vera, NVIDIA's $200B TAM will shrink.

Risk note: The $200B TAM assumption rests on the premise of large-scale Agentic AI commercialization. If Agent deployment falls short of expectations, Vera CPU's TAM will contract significantly. This is a "if you build it, they will come" bet — NVIDIA is investing heavily to build hardware and software infrastructure, betting customers will show up.

---

4. LPU: From "The Seventh Chip" to "Niche Product"

4.1 Analyzing Jensen's Exact Words

This is the passage from the entire earnings call most worth re-reading:

"LPU is designed for low latency and high token rate, but its throughput is low. The use case for LPU is not broad. It will be a niche product for some time."

Put this alongside Jensen's high-profile LPU promotion at GTC just two months earlier:

• GTC 2026 (March): LPU debuted as the "seventh chip" of the Vera Rubin platform. Groq 3 LPU, fabricated by Samsung 4nm, equipped with 512MB on-chip SRAM, slated for Q3 2026 shipment. Jensen positioned it as a "co-processor" for Vera Rubin inference acceleration, claiming it would "accelerate every layer of every token."
• Q1 Earnings Call (May): Jensen voluntarily downgraded LPU to "niche product," explicitly acknowledging its "throughput is low" and "use case is not broad."

The contrast between these two moments is the most subtle signal NVIDIA released in these earnings. Jensen rarely makes negative comments about his own newly launched products in public — especially not on an earnings call facing investors. His choice to acknowledge LPU's limitations at this particular moment may have two reasons:

1. Managing expectations. Jensen may have realized that if the market's LPU expectations ran too high ("revolutionary product for inference"), actual contributions falling short would hurt the stock. Lowering expectations proactively is a defensive strategy.

2. Customers have already voted with their wallets. In the two months after GTC, NVIDIA likely received pushback from key customers about LPU's actual value. If the most important customers (Microsoft, Google, Meta) aren't interested in LPU, Jensen needed to adjust the narrative before investors started asking.

4.2 Contrast with the High-Profile GTC 2026 Launch

At GTC 2026, NVIDIA integrated Groq's LPU technology into the Vera Rubin platform as a dedicated "decode-phase co-processor." The external read at the time was that NVIDIA — through its ~$20B acquisition of Groq (completed December 2025) — had acquired core low-latency inference technology, a major bet on the inference market.

NVIDIA also showcased the Groq LPX Rack at GTC — a dedicated LPU cabinet featuring LP30 chips (512MB SRAM, 150 TB/s bandwidth), planned for Samsung 4nm fabrication, Q3 2026 shipment. This was NVIDIA's first time positioning a non-GPU architecture chip as part of a core product line — interpreted externally as a major commitment to the "inference-specific chip" market.

But Jensen's language on the earnings call indicates NVIDIA's internal positioning of LPU has shifted. From "core component" downgraded to "niche product," the reasons may include:

1. Customer feedback underwhelmed. In the two months after GTC, NVIDIA likely received pushback from major customers about LPU's actual value. For large customers already deploying Grace Blackwell NVL72, an additional LPU accelerator may not be necessary — because Blackwell's inference performance is already strong enough.

2. Vera Rubin's inference performance is strong enough on its own. If Vera Rubin + Dynamo 1.0 can already achieve inference latency close to LPU (for most scenarios), the additional LPU accelerator becomes redundant. Especially since Dynamo 1.0's "up to 7x" inference optimization may have already solved most inference performance issues.

3. Cost-effectiveness analysis. LPU relies on large amounts of on-chip SRAM (512MB/chip), high manufacturing costs, and low yields. Within the overall Vera Rubin system, LPU may not be the most cost-effective component. NVIDIA may have done the math: spending the silicon area and budget for LPU on additional Rubin GPUs or NVLink Switches could deliver more customer value.

4.3 Technical Reasons Why LPU Isn't Suited for General-Purpose Scenarios

The LPU (Language Processing Unit) architecture inherently limits its applicability:

1. The Natural Constraint of SRAM Cost

LPU's core advantage is the ultra-high bandwidth provided by on-chip SRAM (150 TB/s). But 512MB of SRAM on 4nm process occupies significant die area — at Samsung 4nm density, 512MB SRAM takes up roughly 200-250mm². This means LPU's manufacturing cost is disproportionate to the compute capacity it actually provides.

Compare the Rubin GPU: while much larger (~2000mm²+), it delivers 33.6 billion transistors of general-purpose compute + 288GB HBM4 of massive memory capacity. From a price/performance perspective, GPU delivers far more compute per dollar than LPU.

2. MoE Models Challenge SRAM Capacity

Modern large language models increasingly adopt Mixture of Experts (MoE) architectures — including DeepSeek V4 (1 trillion parameters, ~37B active parameters), GPT-5 series, etc. MoE models distribute parameters across multiple Experts, with each token activating only a subset. This means the LPU needs to cache all Expert parameters in SRAM — but MoE models' total parameter counts can reach trillions, far exceeding LPU's 512MB SRAM capacity.

For Dense models (like early GPT-4), LPU could load the entire model or key layers into SRAM for ultra-low-latency inference. But for MoE models, LPU's SRAM capacity is insufficient, requiring frequent data loads from HBM — at which point its bandwidth advantage is negated.

A simple calculation: Assume an MoE model has 128 Experts, each ~8B parameters (2 bytes/parameter = 16GB). Total parameters across 128 Experts: 128 × 16GB = 2TB. LPU's SRAM is only 512MB — it can't even store 3% of a single Expert. In this scenario, LPU's SRAM advantage completely disappears.

3. Long Context Bandwidth Bottleneck

LPU's design assumes model parameters can reside in on-chip SRAM. But in real deployments, as context length grows from 128K to 1M or beyond, KV Cache capacity requirements increase dramatically. A 1M-context KV Cache might require 10-50GB of storage — far exceeding LPU's 512MB on-chip storage.

This means LPU still depends on external memory (HBM or DDR) when handling long contexts. Its bandwidth advantage is negated — because the bottleneck shifts from "on-chip bandwidth" to "external memory bandwidth."

4. Inherent Lack of Versatility

GPUs became the standard carrier for AI compute because they excel simultaneously at training and inference, Dense and MoE models, short and long contexts. LPU excels at only one specific scenario: short-context, Dense model, low-latency inference. This scenario has value (real-time conversation, code completion), but it's not enough to sustain an independent product line.

Summary: LPU's technical architecture was cutting-edge in 2023-2024 (when mainstream models were Dense, short-context), but in 2026's MoE + long-context era, its applicable scope has significantly narrowed. NVIDIA's $20B acquisition of Groq may have missed LPU technology's "optimal window."

4.4 Reflections on the Groq Acquisition ROI

NVIDIA completed the ~$20B "acquihire" (talent + technology license acquisition) of Groq in December 2025. The logic at the time: Groq's LPU technology could fill NVIDIA's gap in low-latency inference.

The language shift on the Q1 earnings call implies this acquisition's ROI may be below expectations.

Possible reasons:

1. The $20B price was too high. Groq's pre-acquisition valuation was ~$3-5B. NVIDIA paid a 4-6x premium, primarily to acquire Jonathan Ross (Groq founder, former core member of Google's TPU team) and his team. But with LPU downgraded to niche product, the core technology value of Ross and his team is correspondingly diminished.

2. Technology integration went less smoothly than expected. Groq's software stack (based on its proprietary compiler and scheduler) needs more time to integrate with NVIDIA's CUDA/Dynamo ecosystem. Groq's compiler was optimized for Groq's dataflow architecture, while NVIDIA's GPUs use SIMT architecture — the programming model differences between the two architectures are significant.

3. Market positioning downgraded. From "inference accelerator" to "niche product" means LPU's revenue contribution will fall far short of initial expectations. If LPU is only a "scenario-specific co-processor" rather than "the main product for inference," the $20B investment payback period extends significantly.

But there are positive aspects:

1. Talent value. Jonathan Ross and the Groq team remain inside NVIDIA. Their chip design experience (especially low-latency inference optimization) could be applied to future GPUs or other architectures.
2. Technology reserve. LPU's SRAM-intensive design may find new applications in future products — such as Agent token routing and cache optimization.
3. Defensive acquisition. Even with LPU downgraded, acquiring Groq prevented it from falling into competitor hands (like AMD or Intel). This was a "lose if you don't buy" bet — though the bet may have been oversized.

Implications for Samsung Foundry: Groq 3 LPU was to be fabricated by Samsung 4nm — NVIDIA's first time assigning a core product to a non-TSMC foundry. With LPU downgraded to niche product, Samsung Foundry's strategic importance correspondingly declines. NVIDIA's core products (Rubin GPU, Vera CPU, NVLink Switch) remain with TSMC.

---

5. China Cutoff: Q2 with $0 Data Center Revenue

5.1 Actual Impact of Export License Requirements

In April 2026, the Trump administration notified NVIDIA that chips exported to China and certain other countries would require export license applications. This effectively banned NVIDIA from selling H20 and other data center chips to China.

Financial impact timeline:

• Q4 FY2026: $5.5B H20 inventory write-down and purchase commitments (partially reflected in this quarter's GAAP figures)
• Q1 FY2027: GAAP net income impacted by $5.5B write-down ($42.96B GAAP vs higher Non-GAAP)
• Q2 FY2027: Expected additional ~$8B revenue loss
• Jensen's exact words: "In China, we have now dropped to zero"

Historical context — Three years of NVIDIA's China market evolution:

Key number: NVIDIA once held ~95% AI chip share in China (pre-2022), with annualized revenue of ~$10-15B. This market went from "core revenue source" to "$0" in three years.

A notable detail: The Q2 guidance of $91B is a figure that "excludes China data center revenue." This means if the China market were to recover (though extremely unlikely), actual revenue could be higher. But more importantly — $91B excluding China means the strength of demand in the rest of the world is being underestimated.

5.2 Jensen's Attitude Toward the China Market

Jensen's language on the earnings call and in recent public appearances is worth analyzing.

On the call:

"The Chinese government has to decide."

The subtext: NVIDIA, as a U.S. company, has no choice in this matter. Jensen chose to pass the ball to the Chinese government — implying this is a political issue, not a commercial one.

But Jensen was more direct in an interview with the Special Competitive Studies Project (early May 2026):

"In China, we have now dropped to zero."

Jensen has been one of the most public critics in Washington of chip restrictions on China. His core argument: restrictions will only accelerate China's development of its own AI chip industry, ultimately harming long-term U.S. interests. He has reiterated this view on multiple occasions — including warning in a 2025 interview that if NVIDIA is barred from selling chips to China, "China will have to make these chips themselves — and they will."

Jensen's "China anxiety" has commercial logic:

1. Direct revenue loss. $8-15B/year in China revenue is no small sum — even for NVIDIA.
2. Long-term ecosystem risk. Once Chinese developers migrate from CUDA to Huawei CANN/Cambricon BANG and other domestic frameworks, they won't return even if sanctions are lifted. This is an "irreversible" ecosystem shift.
3. Global South competition. If China's AI chip ecosystem matures, it could export to budget-sensitive markets in the Middle East, Southeast Asia, and Latin America — markets currently NVIDIA's growth frontier.

5.3 Accelerating Huawei Ascend, Cambricon, and Other Domestic Alternatives

The biggest beneficiary of NVIDIA's China exit is Huawei Ascend.

Huawei Ascend's status (mid-2026):

• Capacity: Plans to produce ~600K Ascend 910C chips in 2026, double 2025's output. Total product line target: 1.6M units.
• Product roadmap: Ascend 950PR launched Q1 2026, followed by 960 and 970 planned, with performance doubling each generation.
• Ecosystem: DeepSeek V4 has completed large-scale migration to Ascend 950PR, with code directly rewritten for Huawei's CANN architecture. Additionally, 8 domestic AI chip makers — Cambricon (SiYuan 690), Hygon, Moore Threads, Biren, MetaX, Kunlunxin, T-Head — have all achieved DeepSeek V4 compatibility.
• Cluster solution: CM384 cluster (384 Ascend 910C fully interconnected) can deliver 300 PFLOPS BF16 compute on specific tasks (1.7x NVIDIA NVL72), 49.2 TB total HBM (3.6x NVL72), but at nearly 4x the power consumption.
• Market forecast: Huawei expects to capture ~50% of China's AI chip share by 2026.

Key insight: Huawei's strategy is "brute force works wonders."

A single Ascend 910C still lags behind NVIDIA Blackwell — but at the system level, by packing in more chips, consuming more power, and using more complex interconnects, Huawei can approach or even surpass NVIDIA's performance on specific tasks. This strategy is viable because China has abundant and cheap power resources (coal and hydro).

CM384 vs NVL72 Technical Comparison:

Huawei's system surpasses NVIDIA in absolute performance (compute, memory capacity, bandwidth), but at the cost of 4x power consumption and 16x space. This is acceptable in scenarios with abundant power and space (like data centers in western China), but impractical where power or space is constrained (like urban data centers).

DeepSeek V4's migration is a landmark event.

DeepSeek V4 is China's most advanced large language model launched in 2026 (1 trillion parameters, MoE architecture). It runs entirely on Huawei Ascend 950PR, completely abandoning NVIDIA's CUDA ecosystem. This means:

1. Algorithm-hardware co-optimization is mature. The DeepSeek team spent months collaborating with Huawei, rewriting low-level code specifically for the CANN architecture. This isn't simple "porting" — it's deep optimization.
2. "No chips available" pressure catalyzed innovation. DeepSeek shocked Silicon Valley in early 2025 with its MLA (Multi-head Latent Attention) mechanism and extreme algorithmic compression — driving training costs down to a fraction of OpenAI's. Now the same innovative spirit has been applied to hardware adaptation.
3. A positive feedback loop has started. When China's best AI team (DeepSeek) successfully trains and deploys the most advanced model on domestic chips, other teams have a clear reference path. This lowers both the psychological and technical barriers to "migrating to domestic chips."

Long-term impact on NVIDIA:

Jensen Huang put it best himself:

"The U.S. has forced China to develop a complete hardware and software stack independent from the American one. Once this ecosystem is established and builds user stickiness, even if the U.S. completely lifts sanctions in the future, Chinese tech giants will never again easily hand their core infrastructure back to a foreign company."

This is the fundamental reality NVIDIA faces in China — even if sanctions are lifted, Chinese customers won't come back. Three years of "domestic substitution" has created irreversible ecosystem momentum.

But this doesn't mean NVIDIA will lose its position globally. China's AI chips are still primarily focused on the domestic market. Expansion to the Middle East, Southeast Asia, and Latin America — the "Global South" — will take time. NVIDIA's advantages in these markets — CUDA ecosystem, full-stack solutions, customer relationships — remain impregnable.

An overlooked risk: If China's AI chip ecosystem matures enough to export to the Global South, NVIDIA will face an entirely new competitor — one pricing at 1/3 to 1/2 of NVIDIA, with lower performance and weaker software ecosystem, but "good enough" for budget-sensitive markets. This is Jensen's greatest "China anxiety" — not the China market itself, but the proliferation of Chinese technology to third countries.

5.4 U.S. Think Tank and NVIDIA Management Perspectives

CSIS (Center for Strategic and International Studies) assessment:

"America's blockade hasn't locked down China — instead, it has served as the harshest form of 'natural selection.' It has forced China's AI army onto an entirely different technology tree branch of 'high-efficiency, low-consumption, software-hardware co-optimization,' and at this moment that branch is flourishing."

SemiAnalysis estimates:

Without sanctions, Huawei Ascend's annual capacity could exceed 5 million units, but is currently limited to ~800K-1M. Chinese AI labs still have an objective gap in long-context training stability and absolute compute density for frontier models — but the gap is closing fast.

NVIDIA management's response: Jensen didn't spend much time discussing the China market on the earnings call — probably because he considers it a "priced-in risk." The market already knows China revenue is going to zero, and Q2 guidance of $91B also excludes China. Jensen preferred to spend his time on growth opportunities from Vera Rubin and Vera CPU — things he can control.

---

5.4 NVIDIA's China Revenue: Historical Evolution

A look at the complete timeline of NVIDIA's China data center revenue from "core market" to "$0":

Estimated cumulative losses over three years:

• 2022-2024: Revenue losses from export restrictions (price gap between derated and full versions): ~$30-50B
• 2025-2026: Revenue losses from H20 ban and full cutoff: ~$100-150B
• Total: ~$130-200B in potential revenue losses

In NVIDIA's historical context, this means: if the China market had remained open, NVIDIA's FY27 full-year revenue could have reached $450-500B (vs ~$385-400B). China market losses reduced NVIDIA's potential growth rate by approximately 15-20%.

5.5 Analysis of the Trump Administration's "25% Tax" Proposal

Before the full cutoff, the Trump administration had proposed an alternative: allow NVIDIA and AMD to sell more advanced H200 chips to China, provided that 25% of the revenue from each chip sold was turned over to the U.S. Treasury.

Analysis of this proposal:

1. Impact on NVIDIA: If an H200 is sold to a Chinese customer at $30,000/chip, the 25% "tax" means $7,500/chip goes to the U.S. government. NVIDIA's actual revenue would be $22,500/chip — 25% less than without the tax, but better than not being able to sell at all.

2. Impact on Chinese customers: If the H200's actual cost after the 25% tax is $37,500/chip (assuming NVIDIA passes the tax to customers), Chinese customers would pay 25% more for the same chip. Given that Chinese customers were willing to pay 50%+ premiums for H100s in 2023, a 25% surcharge may not deter purchases.

3. Why didn't this proposal get implemented? Possible reasons:

   • U.S. domestic political pressure ("shouldn't sell any advanced chips to China")
   • Chinese government's reaction (might reject this "protection money" arrangement)
   • Concerns about technology leakage (even with a tax, H200 could still be reverse-engineered)

Jensen's attitude toward this proposal: While he hasn't stated publicly, Jensen likely would have preferred to accept this arrangement — because a 25% tax is better than a 100% loss. But ultimately, the full cutoff became reality.

5.6 Long-Term Impact of the Cutoff on China's AI Industry

Short-term impact (2024-2026): Chaos and adjustment. Chinese companies forced to migrate from the NVIDIA ecosystem to domestic alternatives, efficiency drops 30-50%. DeepSeek V4's delayed launch may be partly attributable to this migration.

Medium-term impact (2026-2028): Domestic ecosystem matures. Huawei Ascend, Cambricon, and other vendors accelerate product iteration; software ecosystems (CANN, BANG, etc.) gradually improve. DeepSeek V4 successfully running on Ascend is a milestone.

Long-term impact (2028+): China could develop a fully independent AI chip ecosystem — from hardware design to manufacturing (SMIC, etc.) to software frameworks to model training. Once this ecosystem matures, it will no longer depend on any U.S. technology — even if sanctions are lifted, there's no reason to go back to NVIDIA.

Impact on the global AI landscape: The world could split into two parallel AI ecosystems —

1. U.S./Western system: Centered on NVIDIA GPU + CUDA + Dynamo, covering North America, Europe, Japan, South Korea, etc.
2. China system: Centered on Huawei Ascend + CANN, covering China, and potentially expanding to the Global South

This split has profound implications for NVIDIA — it means the global AI chip market's "ceiling" has been permanently lowered. If the China market (~25%) + Global South (~10-15%) ultimately shifts to the Chinese ecosystem, NVIDIA's addressable market shrinks by ~35-40%.

6. Competitive Landscape: Customers Turning into Competitors

6.1 NVIDIA's 10-Q Formally Acknowledges Customer ASIC Risk for the First Time

In its Q1 FY2027 10-Q filing, NVIDIA added two new risk disclosures:

1. Customer-designed ASIC competition: "Some of our largest customers are developing or may develop their own AI accelerators, which could reduce their demand for NVIDIA products."
2. Customers competing for wafer capacity: "If our customers are also our competitors, they may compete for the same wafer capacity, affecting our supply."

These two risk disclosures have never appeared in previous 10-K and 10-Q filings — their inclusion marks NVIDIA management formally acknowledging the structural trend of "customers becoming competitors."

The first is triggered by Google (TPU + Blackstone JV) and Amazon (Trainium 3). Google and Amazon are both major NVIDIA customers while simultaneously developing in-house AI chips. When these customers' custom silicon starts handling more inference workloads, NVIDIA's revenue growth faces structural pressure.

The second trigger is more subtle. NVIDIA's customers (Google, Amazon, Apple, etc.) are also major TSMC customers — competing for the same advanced-node capacity (3nm, 2nm). If these customers allocate more capacity to in-house chips, NVIDIA's TSMC capacity share could be compressed. This is a "zero-sum game" — with advanced-node capacity constrained, every wafer used for a customer's in-house chip is potentially one less wafer for NVIDIA.

6.2 Google + Blackstone TPU JV: $5B Equity Investment

On May 18, 2026 — two days before NVIDIA's earnings release — Blackstone announced a joint venture with Google, investing $5B in equity to build a 500MW TPU cloud data center, scheduled to go live in 2027.

This JV's strategic implications far exceed the $5B itself:

1. TPU goes to third-party sales for the first time. Google's TPU was previously only used internally within Google Cloud. Through the Blackstone JV, TPU will be sold externally as "compute-as-a-service" — meaning Google is officially entering NVIDIA's core market. Google Cloud CEO Thomas Kurian stated:

"This joint venture with Blackstone helps meet growing demand for TPUs, which are optimized specifically for efficiency and performance in the AI era."

2. Blackstone's role. Blackstone is the world's largest alternative asset manager (over $1.3T in AUM) and one of the world's largest data center providers. It brings capital and real estate resources — TPU data centers need significant land, power, and cooling infrastructure. Benjamin Treynor Sloss was appointed JV CEO, signaling this is a serious commercial effort.

3. Timeline: 500MW capacity coming online in 2027. A 500MW data center can accommodate roughly 50,000-100,000 TPUs — equivalent to a mid-size AI training cluster. If operations succeed, Blackstone plans to "scale significantly over time."

Actual impact on NVIDIA:

Google has been one of NVIDIA's largest customers. If Google starts shifting AI workloads from GPUs to in-house TPUs, NVIDIA's data center revenue growth may slow. But several buffer factors:

1. Google's GPU usage is still growing. Even with the TPU JV live, Google's GPU demand remains strong in the near term — particularly for training the next generation of Gemini models. TPU is primarily for inference; training still runs mostly on GPUs.
2. TPU's software ecosystem still trails CUDA. Google's JAX/PyTorch/XLA compiler stack works well internally but has limited appeal to external customers — because most AI developers are already familiar with CUDA.
3. 500MW capacity is limited in scale. Compared to the multi-gigawatt total global AI data center footprint, a 500MW TPU cloud is a drop in the ocean. But if the JV succeeds and scales up, long-term impact could be greater.

6.3 Cerebras' $95B IPO Debut: What It Signals

On May 14, 2026, Cerebras Systems completed the largest IPO of 2026:

• Offering price: $185/share, raising $5.55B
• First-day close: $331.07/share, +68%
• First-day market cap: $95B
• Contract backlog: $24.6B (primarily from OpenAI)
• Projected 2028 revenue: ~$5.5B
• Forward P/S: ~17x

The $95B market cap sends three signals:

Signal 1: The market believes in AI chip diversification. Investors no longer see NVIDIA as the only AI chip winner. Cerebras' wafer-scale chip (Wafer-Scale Engine) represents a fundamentally different technical approach from GPUs — and has top-tier customer validation from OpenAI.

Signal 2: OpenAI's NVIDIA dependence is decreasing. OpenAI is Cerebras' largest customer (primary source of the $24.6B backlog). If OpenAI simultaneously uses NVIDIA GPUs and Cerebras WSE, its bargaining power versus NVIDIA strengthens. More importantly, OpenAI could shift more training workloads to Cerebras in the future — directly impacting NVIDIA's training revenue.

Signal 3: The XPU market TAM is being repriced. Cerebras projects ~$5.5B in 2028 revenue, corresponding to a $95B market cap implying ~17x forward P/S. This valuation only makes sense if the market believes the AI chip market is large enough to support multiple players. Futurum Group's analysis noted that Cerebras' $5.5B in 2028 revenue represents only ~3.4% of the projected XPU total market — meaning the market thinks "3.4% share is worth $95B."

But the stock dropped the next day — showing the market is also skeptical. Cerebras' business model depends on a few large customers (OpenAI accounts for most of the backlog), creating extreme customer concentration risk. If OpenAI cuts orders or switches to other suppliers, Cerebras' revenue would shrink dramatically.

6.4 How the "Frenemies" Landscape Impacts NVIDIA's Long-Term Pricing Power and Margins

NVIDIA's competitive landscape is shifting from "pure competition" to "frenemies":

Long-term impact at three levels:

1. Pricing power. When Hyperscale customers have alternatives (in-house ASICs or Cerebras), NVIDIA's pricing power will erode. Inference will be impacted first, because inference has lower GPU stickiness than training — inference doesn't require CUDA; any compatible inference framework (like vLLM, TensorRT-LLM) can run on different hardware.

2. Gross margin. If NVIDIA needs to cut prices to maintain customer share, the 75% gross margin will face pressure. But in the near term (FY27-28), with Vera Rubin supply constrained, NVIDIA still holds pricing power. Real gross margin pressure may not materialize until FY29+.

3. Complexity of customer relationships. NVIDIA needs to "make customers dependent" (through CUDA ecosystem and full-stack solutions) while not "making customers hate them" (through excessive price increases). This balance will become increasingly difficult to maintain. Jensen always emphasizes "collaboration" and "empowerment" in public, but privately, NVIDIA's pricing strategy is viewed by some major customers as "monopoly premium" — a potential PR and political risk.

---

6.5 Amazon, Microsoft, and Meta's In-House Chip Progress

Beyond Google, other Hyperscale customers' in-house chip programs are advancing rapidly:

Amazon (AWS):

• Trainium 2 is already deployed at scale on AWS, claiming cost/performance parity with NVIDIA GPUs on certain inference workloads
• AWS signed a large Trainium instance contract with Meta in early 2026 — an important signal that Trainium is not just for internal AWS workloads but is starting to attract major external customers
• Trainium 3 (expected 2027) will further narrow the performance gap with NVIDIA GPUs
• Risk to NVIDIA: Medium-High. AWS is a major NVIDIA customer but is aggressively promoting Trainium as an NVIDIA GPU alternative. Meta choosing Trainium instances is an important market validation

Microsoft:

• Maia 100 in-house chip progress is relatively behind — reportedly, Microsoft has encountered challenges in chip design
• Microsoft remains one of NVIDIA's largest single customers; Azure's AI infrastructure is almost entirely NVIDIA GPU-based
• Risk to NVIDIA: Low-Medium. Microsoft's in-house chip progress lags Google and Amazon; near-term impact on NVIDIA is limited. But Microsoft has ample capital and talent reserves, and will inevitably accelerate in-house development over time

Meta:

• MTIA v3 (Meta Training and Inference Accelerator) primarily used for inference workloads
• Meta's 2026 CapEx guidance is ~$60-65B, mostly for AI infrastructure — the majority still flows to NVIDIA
• But Meta is also using AWS Trainium instances (under the contract signed with Meta), indicating a "multi-vendor" strategy
• Risk to NVIDIA: Medium. Meta's in-house chips primarily impact the inference market; training remains NVIDIA-dominated

Apple:

• Apple is also developing in-house AI chips (for server-side inference for Apple Intelligence), but currently at small scale
• Apple's CapEx guidance increased significantly from ~$10B in FY2026 to ~$15-20B in FY2027, mostly for AI infrastructure
• Risk to NVIDIA: Low. Apple's AI infrastructure scale is relatively small, and Apple prefers in-house chips over third-party solutions

Summary: The four major Hyperscalers' in-house chip progress varies — Google is most aggressive (TPU + Blackstone JV), Amazon next (Trainium 2/3 + Meta contract), Meta is catching up (MTIA v3 + multi-vendor strategy), Microsoft is furthest behind (Maia 100 delayed). But all major customers are moving in the same direction — reducing single-source dependence on NVIDIA. This is an irreversible structural trend.

6.6 Broadcom's Role in AI Chips

When discussing customer-designed ASICs, Broadcom's role cannot be ignored — it's the design partner for most Hyperscale in-house chips:

• Google TPU: Broadcom participated in TPU design support
• Meta MTIA: Broadcom provides custom chip design services
• Other customers: Broadcom is working with multiple Hyperscalers on AI inference chip development

Broadcom's model is "designing custom chips for customers" — it doesn't compete directly with NVIDIA, but helps NVIDIA's customers develop in-house chips faster. This makes Broadcom an "amplifier" in NVIDIA's competitive landscape — accelerating customer migration from NVIDIA GPUs to in-house ASICs.

Impact on NVIDIA: Broadcom's existence lowers the barrier to customer-designed chips — even customers without deep chip design expertise can quickly launch in-house chips through Broadcom partnerships. This means the in-house ASIC threat may arrive faster than NVIDIA management expects.

6.7 Quantitative Analysis of Customer Concentration Risk

Based on NVIDIA's Q1 FY27 data, of the $75.2B in data center revenue:

• Hyperscale (top 4: Microsoft, Google, Meta, Amazon) contributed ~$38B
• ACIE contributed ~$37B

Assuming the top 4 customers each contributed ~$9-10B (average), single-customer revenue concentration is ~12-13%. This is a relatively healthy level — below the typical "single customer dependency" risk threshold of 20%.

But if we consider compute revenue ($60.4B) rather than total data center revenue, concentration may be higher. Networking revenue ($14.8B) is typically procured by data center infrastructure teams and isn't heavily influenced by individual customers. Excluding networking, the top 4 customers' concentration in compute revenue could reach 50-60% — roughly 12-15% per customer.

The real risk isn't "one customer leaves" — it's "multiple customers simultaneously start replacing with in-house chips." If Google migrates 30% of GPU workloads to TPU, Amazon migrates 20% to Trainium, and Meta migrates 15% to MTIA — NVIDIA could lose 25-30% of Hyperscale revenue. This is a medium-term (2028-2030) risk.

NVIDIA's mitigation strategy: Diversify Hyperscale risk through AI Cloud customers (CoreWeave, Lambda, etc.). ACIE already contributes ~50% of data center revenue — if this ratio continues rising, NVIDIA's Hyperscale dependence will further decrease.

7. Networking: The Underestimated Growth Engine

7.1 $14.8B Networking Revenue, 199% YoY, 35% QoQ

Networking revenue was one of the standout numbers in Q1:

A $59B annualized run rate means: If NVIDIA's networking business were an independent company, it would be one of the world's largest semiconductor companies — nearly twice AMD's total company revenue (~$28B in 2025).

But networking revenue growth isn't just from NVLink. Q1's $14.8B in networking revenue includes contributions from:

1. NVLink Switch (NVLink 5): Grace Blackwell NVL72 requires large numbers of NVLink Switches. As NVL72 shipments increase, NVLink Switch revenue grows.
2. Spectrum-X Ethernet switches: Used in AI Cloud provider and enterprise AI data center Ethernet networks.
3. ConnectX NICs: For connecting GPU servers to the network.
4. InfiniBand (Quantum series): For traditional HPC and high-end AI training clusters.

7.2 NVLink's Transition from "Free Interconnect" to "Profit Center"

In the Hopper (H100) era, NVLink was treated as a GPU accessory — customers "got" NVLink interconnect capability "for free" when buying GPUs. But with Grace Blackwell and Vera Rubin, this has fundamentally changed:

1. NVLink Switch became an independent product. Grace Blackwell NVL72 requires 36 NVLink Switches, each a complex chip (containing NVLink switching silicon and numerous high-speed SerDes). These Switches are no longer "freebies" but independently priced components.

2. Pricing power comes from technical monopoly. NVIDIA has virtually no competitors in NVLink interconnect technology. PCIe bandwidth is far from sufficient (PCIe 6.0 bidirectional bandwidth ~256 GB/s, vs NVLink 6 at 3.6 TB/s — a 14x gap). Other interconnect solutions (like AMD's Infinity Fabric) also fall well short of NVLink bandwidth. Without alternatives, customers must pay for NVLink.

3. Networking margins may exceed compute. While networking components have lower absolute selling prices than GPUs, NVLink Switch and Spectrum-X switch gross margins may be higher — because their manufacturing costs are relatively low (no expensive HBM), while pricing power is extremely strong.

A key calculation: Assuming a Grace Blackwell NVL72 total price of ~$3-4M, NVLink Switches account for ~20-30% ($600K-1.2M). If NVIDIA ships 10,000 NVL72s, NVLink Switch revenue alone would be $6-12B — explaining why networking revenue surged to $14.8B in Q1.

7.3 Spectrum-X Ethernet vs. InfiniBand: Route Selection

NVIDIA has two parallel networking paths for data centers:

InfiniBand (Quantum series):

• Advantage: Ultra-low latency, ultra-high bandwidth, naturally suited for HPC and large-scale training
• Customers: Traditional HPC users and top AI labs
• Positioning: High-end market
• Representative customer: Stargate project (OpenAI + SoftBank)

Ethernet (Spectrum-X series):

• Advantage: Compatible with existing Ethernet infrastructure, lower deployment barrier, lower cost
• Customers: Enterprise AI, AI Cloud providers, large-scale inference clusters
• Positioning: Mass market
• Performance: Spectrum-X claims 1.6x network performance improvement over standard Ethernet
• Representative customer: CoreWeave and other new-build AI Cloud providers

NVIDIA's dual-track strategy means: Regardless of which networking technology customers choose, NVIDIA covers both. This further consolidates NVIDIA's monopoly position in data center networking.

A trend worth watching: Spectrum-X may grow faster than InfiniBand. The reason is AI Cloud providers (CoreWeave, Lambda, NEBIUS) prefer Ethernet when building new data centers (because more operations talent is available, costs are lower, and compatibility is better). As these AI Cloud companies' revenue share increases (ACIE at 49.5%), Spectrum-X's share will also rise.

7.4 Optical Interconnect Investments (Coherent/Corning/Lumentum Strategic Agreements)

In Q1 FY2027, NVIDIA signed strategic agreements with multiple optical interconnect companies:

• Coherent: Providing optical transceivers and optical modules for NVLink and Spectrum-X long-distance interconnects.
• Corning: Providing fiber optic cables for intra-data-center and inter-data-center connectivity.
• Lumentum: Providing lasers and photonic chips for next-generation optical interconnects.

Why optical interconnects matter so much:

When NVLink bandwidth reaches 3.6 TB/s, copper interconnects are physically approaching their limits (signal attenuation, power consumption, wiring complexity). Copper's effective transmission distance is ~1-2 meters — meaning NVLink can only be used "within the rack." To extend to "cross-rack" and "cross-data-center," optical interconnects are necessary.

Strategic significance of optical interconnects:

1. Unlocking "multi-rack" AI training. Today's largest AI models (trillion-parameter scale) require clusters of thousands to tens of thousands of GPUs. If NVLink is limited to within-rack (72 GPUs), cross-rack communication must use other interconnect methods — creating a bottleneck. Optical interconnects can break NVLink beyond rack boundaries, enabling high-speed cross-rack communication.

2. Foundation for "distributed AI factories." Jensen has repeatedly mentioned the "AI factory" concept — but AI factories don't need to be co-located. If optical interconnects are fast enough, multiple geographically distributed AI factories could work together as one virtual supercomputer.

3. New revenue stream. The market for optical interconnect equipment and services could reach tens of billions of dollars annually. If NVIDIA controls the optical interconnect supply chain through strategic investments and agreements, it will further expand its share of data center infrastructure.

---

7.5 Networking Competitive Landscape: NVIDIA's Advantages and Challenges

While NVIDIA dominates AI data center networking, the competitive landscape is also shifting:

NVIDIA's three major advantages:

1. NVLink technology monopoly. No competitor can provide NVLink-level bandwidth (3.6 TB/s). AMD's Infinity Fabric bandwidth is ~500 GB/s (only ~14% of NVLink's), and Intel's UPI bandwidth is even lower. In the GPU-to-GPU high-speed interconnect niche, NVIDIA has no real competitor.

2. Full-stack integration effect. NVLink is deeply integrated with NVIDIA GPUs, Vera CPUs, and Dynamo inference engine. Customers using the NVIDIA full-stack solution can maximize network performance — while customers using third-party networking need to handle GPU-network adaptation and optimization themselves.

3. Mellanox's patents and talent. Through the 2020 acquisition of Mellanox for $6.9B, NVIDIA acquired extensive InfiniBand and high-speed networking patents and engineering talent. These assets now form the technical foundation of NVIDIA's networking business.

Three competitive challenges:

1. Ultra Ethernet Consortium (UEC). A group of networking companies (Arista, Broadcom, Cisco, etc.) is pushing "Ultra Ethernet" — an Ethernet standard optimized for AI workloads. If UEC successfully establishes an open high-performance Ethernet standard, it could become an alternative to Spectrum-X — because customers may prefer open, multi-vendor standards over NVIDIA's proprietary solution.

2. In-house networking chips. Both Google and Amazon are developing in-house networking chips (for TPU and Trainium cluster interconnects). If these chips mature and become available in the third-party market, they could reduce dependence on NVIDIA networking components.

3. Optical interconnect newcomers. Several startups (like Ayar Labs) are developing entirely new optical interconnect approaches — using silicon photonics for chip-to-chip high-speed communication. If these technologies mature, they could challenge NVIDIA's copper-based interconnect solutions in the future.

My judgment: In the near term (FY27-28), NVIDIA's networking position is unassailable. But in the medium term (FY29+), as UEC matures and in-house networking chips progress, NVIDIA's networking share may face pressure — particularly in the Ethernet space (Spectrum-X), because Ethernet is a more open and competitive market.

7.6 Mellanox Acquisition ROI in Retrospect

Looking back at NVIDIA's 2020 acquisition of Mellanox for $6.9B:

• At acquisition (2020), Mellanox annual revenue was ~$1.4B
• Q1 FY27 networking revenue $14.8B (single quarter) = 10.6x Mellanox's annual revenue
• Networking business annualized run rate $59B = 8.6x the acquisition price

This acquisition's ROI has exceeded everyone's expectations. What Jensen saw in 2020 was that AI computing's bottleneck wasn't single-GPU performance, but communication speed between GPUs. Mellanox's InfiniBand and Ethernet technology solved that bottleneck — and the value of solving a bottleneck grows exponentially as AI cluster scale expands.

This story parallels CUDA — Jensen started investing in CUDA in 2006, when GPUs were primarily used for gaming. Many questioned why NVIDIA would invest heavily in general-purpose computing "unrelated to gaming." But 15 years later, CUDA became the standard platform for AI computing — and NVIDIA's cumulative ROI on CUDA is in the thousands of times.

Jensen's strategic pattern: Invest in "bottleneck-solving technology" before the market recognizes the need. Mellanox (networking) was the 2020 example. Groq (LPU) may have been the 2025 attempt — but based on the Q1 earnings call language, this attempt's results may be below expectations.

8. Software Moat: Dynamo 1.0 + NemoClaw + OpenShell

8.1 Dynamo 1.0's "Up to 7x" Acceleration and Its Value for Customer Stickiness

NVIDIA launched Dynamo 1.0 at GTC 2026 — an open-source AI inference operating system:

• Core functionality: Manages GPU cluster inference workloads, optimizing model distribution and scheduling across multiple GPUs. Includes dynamic batching, KV Cache management, prefix caching optimization, etc.
• Performance improvement: Officially claims "up to 7x" inference acceleration (vs. unoptimized baseline).
• Open-source strategy: Dynamo 1.0 is open source, similar to CUDA's early strategy — lower adoption barriers through open source, then build a moat through ecosystem lock-in.
• Deep optimization for NVIDIA hardware: While Dynamo is open source, it's most deeply optimized for NVIDIA GPUs (leveraging Tensor Cores, NVLink, GPU Direct, and other proprietary features).

What "up to 7x" means:

This number isn't a universal speedup — it's a peak improvement in specific scenarios (like multi-turn inference, KV Cache management, and GPU communication optimization). Average improvement in real deployments might be 2-3x. But even so, this is a significant efficiency gain.

Value for customer stickiness:

If a customer uses Dynamo 1.0 to optimize its NVIDIA GPU cluster's inference performance, the cost of migrating to other hardware (like AMD MI400 or Google TPU) isn't just "replacing hardware" — it also requires "rewriting the inference scheduling layer." Dynamo manages more than inference scheduling — it also manages KV Cache distribution and recycling, GPU communication optimization, and inference request routing. These functions are implemented completely differently on different hardware — migration costs are extremely high.

Comparison with competitors: AMD's ROCm inference stack and Google's TPU inference stack have no equivalent Dynamo product. AMD relies on open-source vLLM and TensorRT-LLM (the latter being NVIDIA's), and Google relies on its internal XLA compiler. Dynamo's emergence gives NVIDIA another "CUDA-like" moat in inference.

8.2 NemoClaw (OpenClaw Agent Platform) Strategic Intent

NemoClaw is NVIDIA's enterprise-grade Agentic AI platform launched at GTC 2026:

• Based on OpenClaw: OpenClaw is the open-source Agent framework that exploded in early 2026 (one of the fastest-growing repositories in GitHub history). NVIDIA built an enterprise-grade version on top of it.
• Core components: OpenShell (secure runtime) + Nemotron (open-source models) + Privacy Router (privacy controls)
• Partnership with CrewAI: CrewAI provides multi-Agent orchestration capabilities; NemoClaw provides the secure execution environment.
• One-command deployment: NemoClaw installation requires just one command — significantly lowering the enterprise adoption barrier.

NemoClaw's strategic intent:

NVIDIA is transitioning from "a company that sells hardware" to "a company that sells AI infrastructure." NemoClaw's positioning is — if enterprises want to deploy production-grade AI Agents, NVIDIA provides a full-stack solution from underlying hardware (Vera Rubin + Vera CPU) to middleware (Dynamo + NemoClaw) to upper-layer models (Nemotron).

The brilliance of this positioning:

1. Locking in enterprise customers. Once enterprises build Agent workflows on NemoClaw, switching to another platform is extremely costly — because Agent security policies, access controls, and audit logs are all embedded in OpenShell.
2. Driving hardware sales. NemoClaw is most deeply optimized for Vera CPU and Vera Rubin — while theoretically runnable on other hardware, performance is far inferior to the NVIDIA full-stack experience.
3. Establishing new standards. If NemoClaw/OpenShell becomes "the standard runtime for enterprise AI Agents," NVIDIA will control the most important software layer of the Agentic AI era — just as Microsoft controlled the PC operating system.

8.3 OpenShell's Secure Runtime Positioning

OpenShell is part of the NVIDIA Agent Toolkit — a secure execution environment for running autonomous AI Agents:

• Sandboxed execution: When Agents run in OpenShell, their behavior is constrained within security boundaries. Agents cannot access unauthorized file systems, network resources, or system calls.
• Access control: Enterprises can finely control which data and systems Agents can access — similar to a set of "capability tokens."
• Audit trail: All Agent operations have complete audit logs — a requirement in regulated industries like finance and healthcare.

OpenShell addresses the biggest pain point for Agentic AI adoption — security. The core reason enterprises hesitate to deploy AI Agents isn't immature technology, but fear of Agents going rogue — accidentally deleting data, sending inappropriate emails, or executing unauthorized transactions. OpenShell reduces this risk by providing a "security sandbox."

Why this matters:

Agentic AI commercialization faces a "chicken and egg" problem — enterprises need security tools to deploy Agents, but security tools need Agents running in production to be validated. OpenShell attempts to break this cycle — providing a "security-first" runtime so enterprises can confidently deploy Agents.

If OpenShell succeeds, it could become the "container" of the Agentic AI era — just as Docker containers became the standard runtime for the microservices era. This isn't just a technical tool but an ecosystem gateway — controlling the runtime means controlling the entire ecosystem.

8.4 Software Lock-In Is Harder to Replace Than Hardware

The software moat NVIDIA is building can be represented as a layer diagram:

Application Layer: NemoClaw (Agent Platform)
  ↓
Runtime: OpenShell (Secure Sandbox) + Dynamo (Inference Engine)
  ↓
Framework Layer: CUDA (Compute Framework) + NVLink (Communication Framework)
  ↓
Hardware Layer: Vera Rubin (GPU) + Vera CPU + NVLink Switch

Every layer increases customer stickiness:

Increasing replacement difficulty: Replacing NVIDIA hardware (with AMD MI400 or Google TPU) is only step one. Replacing CUDA is step two (very difficult, because millions of developers and billions of lines of CUDA code exist worldwide). Replacing Dynamo + NemoClaw + OpenShell is step three (currently nearly impossible, because no equivalent alternatives exist).

This is what Jensen means by "full-stack" moat — not leading at one layer, but leading at all layers, so that replacement at any single layer becomes incomplete.

An analogy: replacing NVIDIA's GPU is like switching from iPhone to Android — you can swap hardware, but if all your apps (CUDA), photos (KV Cache), and passwords (OpenShell) live on iOS, the switching cost is far more than just buying a new phone.

---

8.5 Quantifying the Software Moat's Value

How much commercial value is NVIDIA's software ecosystem worth? While not directly disclosed, it can be estimated:

Method 1: Gross margin premium approach. Without the CUDA ecosystem and full-stack software lock-in, NVIDIA's pricing power would decline significantly. Assuming no software moat would reduce gross margin from 75% to 65% (10 percentage points), the software moat's annualized value would be:

FY27 estimated revenue $390B × 10% = ~$39B/year

Method 2: Replacement cost approach. If a Hyperscale customer were to migrate from NVIDIA's full stack to an AMD + Arista + open-source inference engine alternative, the required investment would be:

• Hardware replacement cost: $B-scale
• Software rewrite cost (CUDA → ROCm): Thousands of person-years
• Dynamo inference engine replacement: Need to develop in-house or use open-source (performance may drop 2-3x)
• NemoClaw/OpenShell replacement: No equivalent products currently exist
• Personnel training and organizational restructuring: $100M+

Comprehensive estimate: NVIDIA's software moat creates ~$30-50B in annual value — reflected in higher gross margins (vs. hardware companies without software lock-in), higher customer retention, and higher incremental revenue.

8.6 Software Ecosystem Comparison with Competitors

NVIDIA's software ecosystem leads all competitors in both depth and breadth. Particularly Dynamo (inference engine) and NemoClaw (Agent platform) — these two products currently have no equivalent alternatives, representing NVIDIA's unique advantage in inference and Agentic AI.

9. Q2 Guidance of $91B: What $91B Excluding China Really Means

9.1 Full-Year Projection: $81.6B → $91B → Subsequent Quarter Growth

If NVIDIA's FY27 follows this path:

If Q3/Q4 maintain QoQ growth, FY27 full-year could reach $380-400B.

Compared to FY26 full-year revenue of ~$196B, this implies ~94-104% YoY growth.

But QoQ growth is decelerating. From +20% in Q1 to +12% in Q2 guidance. If this trend continues, Q3 might be +8-10% and Q4 +5-8%. Growth is "normalizing" — but note this "normalization" is on a $100B+ base.

9.2 If H2 Continues QoQ Growth

NVIDIA has achieved QoQ growth for multiple consecutive quarters, and Q2 guidance of $91B continues this trend. If H2 continues QoQ growth (even decelerating to 5-8%), FY27 full-year would exceed $400B.

What does $400B mean?

• NVIDIA's FY27 revenue could approach 5x Intel's peak annual revenue ($79B in 2021).
• FY27 data center revenue (~$350B) would exceed the global smartphone market's annual revenue.
• NVIDIA's net income (assuming 55% Non-GAAP net margin) could reach ~$210B — one of the highest annual profits for a single company in history, second only to Saudi Aramco's peak-oil-price profits.
• NVIDIA's FY27 free cash flow could reach ~$180B — explaining management's confidence in announcing $80B in new buyback authorization and a 25x dividend increase.

9.3 What $91B Excluding China Means — Implied Global Demand Strength

This is the most underrated number in the Q1 earnings.

NVIDIA's Q2 guidance of $91B is based on the assumption of zero China data center revenue. If the China market were still open (assuming Q1 China contribution of ~5-10%), Q2 "true demand" could be $95-100B.

This means non-China global AI infrastructure demand strength is underestimated by ~5-10%.

When the market discusses "NVIDIA beat consensus by $2.5B," few mention this beat comes despite losing a market that once represented 20-25% of revenue. Adding China back, NVIDIA's actual demand strength far exceeds the headline number.

Another angle: If China is permanently lost, NVIDIA's future growth will depend entirely on non-China regions. But this didn't stop NVIDIA from issuing $91B guidance — suggesting Jensen has extremely high confidence in global demand outside China.

What $91B implies:

1. Global AI investment isn't slowing. Even without China, $91B guidance means AI investment elsewhere is accelerating — North American, European, Middle Eastern sovereign AI investments, and Southeast Asian emerging AI infrastructure are all contributing incremental demand.
2. Inference demand is starting to contribute meaningful revenue. Q2 is the first quarter where Agentic AI demand translates into hardware revenue. As more enterprises deploy AI Agents, inference server demand will continue growing.
3. Vera Rubin initial shipments. While Vera Rubin's mass shipments come in Q3-Q4, Q2 may include some initial deliveries — contributing incremental revenue.

9.4 Shareholder Return Signals

NVIDIA's Q1 shareholder return announcements sent strong signals:

$80B new buyback authorization:

• This isn't a "quarterly" buyback — it's a "long-term" authorization executable over multiple quarters.
• Management clearly believes the stock is undervalued — otherwise they wouldn't commit $80B to buybacks.
• Jensen told CNBC the stock's performance is "one of the mysteries of the universe," implying he thinks the market isn't fully reflecting NVIDIA's fundamentals.

Quarterly dividend raised from $0.01 to $0.25 (25x):

• This is a "qualitative shift" — from a token dividend to meaningful cash return.
• At current share price, annualized yield is ~0.25-0.30% — not high, but the direction is clear.
• More importantly, a 25x increase signals management's extremely high confidence in future cash flow.

$18.6B in private company / infrastructure fund investments:

• NVIDIA isn't just returning capital to shareholders — it's actively investing in the ecosystem, particularly AI Cloud providers and infrastructure projects.
• These investments are both a "customer lock-in" strategy and a direct bet on future AI infrastructure demand.

---

8.5 Quantifying the Software Moat's Value

(See Section 8.5 above — this section was merged)

8.6 Software Ecosystem Comparison with Competitors

(See Section 8.6 above — this section was merged)

10. Investment Judgment

10.1 Bull Case

Core assumptions:

• Vera Rubin deliveries proceed smoothly, becoming the primary growth driver in H2 FY27
• Agentic AI achieves large-scale commercialization, driving exponential inference demand growth
• Networking revenue continues to exceed expectations, becoming an independent profit center
• Vera CPU opens up $200B TAM, contributing significant revenue starting FY28
• Customer in-house ASIC substitution proceeds slower than the market expects
• NVIDIA software stack (Dynamo + NemoClaw + OpenShell) successfully establishes ecosystem lock-in
• Global AI investment continues to accelerate, FY27 full-year exceeds $400B

FY27 full-year revenue: $400-420B FY27 Non-GAAP EPS: $8.0-8.5 Target market cap: $5.5-6.5 trillion (based on 30-35x forward P/E)

10.2 Bear Case

Core assumptions:

• Vera Rubin delivery delays or initial quality issues impact H2 FY27 revenue
• Agentic AI commercialization falls short of expectations, inference demand growth slows
• Google + Blackstone TPU JV erodes NVIDIA share after 2027 launch
• Gross margins decline to 72-73% due to increased competition and product mix changes
• China market permanently lost, and domestic alternatives begin exporting to Global South markets
• AI investment enters a "digestion period," Hyperscale CapEx growth slows

FY27 full-year revenue: $360-380B FY27 Non-GAAP EPS: $6.5-7.0 Target market cap: $3.5-4.0 trillion (based on 25-28x forward P/E)

10.3 Base Case

Core assumptions:

• Vera Rubin deliveries see minor delays but proceed overall smoothly
• AI infrastructure demand remains strong but growth gradually normalizes
• Customer in-house ASICs make limited progress in inference market; training remains NVIDIA-dominated
• Gross margin holds at 74-75%
• China market doesn't recover, but doesn't worsen further
• Networking revenue continues high growth, but growth rate gradually slows

FY27 full-year revenue: $385-395B FY27 Non-GAAP EPS: $7.5-8.0 Target market cap: $4.5-5.0 trillion (based on 28-32x forward P/E)

10.4 Key Risk Matrix

10.5 Stock Price Reaction Pattern: Will the "Beat but Drop" Curse Continue?

Historical pattern: NVIDIA has beaten consensus for six consecutive quarters, but the stock fell after earnings in four of those six. Polymarket priced beat probability at 97.2% before earnings — but when a beat is a "certainty," the beat itself isn't a catalyst.

Q1 FY2027 situation:

• After-hours reaction: +~6% (next-day pre-market data)
• Next trading day: continued strength; market reacted positively to Q2 guidance of $91B
• Market reaction to the beat was positive — likely because Q2 guidance of $91B significantly exceeded the $86B consensus

Why the "beat but drop" curse exists:

1. Options market impact. NVIDIA options volume is enormous, and post-earnings Gamma effects frequently cause reverse price moves. Large numbers of call option sellers need to hedge after earnings — creating selling pressure.
2. Already priced in. When beats become "certain events" (7 consecutive quarters), beats themselves aren't catalysts. The market has already priced in the beat — what really drives the stock is Q2 guidance and Jensen's earnings call commentary.
3. Management guidance conservatism. NVIDIA's guidance typically comes in 3-4% below actual results; the market knows this.

My judgment: This beat's quality was high ($91B guidance vs $86B consensus, exceeding by $5B), and the 25x dividend increase was a positive surprise that may break the "beat but drop" curse. But in the short term (first week post-earnings), stock movement is more influenced by macro factors (Trump tariff policy, Fed rate expectations, Treasury yields) than by NVIDIA's fundamentals.

Long-term stock drivers: NVIDIA's long-term stock price depends on two questions —

1. Persistence of AI demand. If AI investment starts slowing in FY28-29 (similar to the 2000 dot-com bust), NVIDIA's valuation would face a major correction. But if AI investment continues growing (even at a decelerating rate), NVIDIA's revenue and profit will continue setting records.

2. Evolution of the competitive landscape. If NVIDIA's full-stack moat (hardware + software + networking) continues strengthening, and customer in-house ASIC substitution proceeds slower than expected, NVIDIA can sustain high margins and high growth. But if Google's TPU JV succeeds, Cerebras' technology path is validated, or China's alternatives are exported globally — NVIDIA's growth prospects will be reassessed.

---

Summary

Three Most Important Conclusions

1. AI infrastructure demand is real, sustained, and accelerating — but NVIDIA's share of the红利 is being eroded.

$81.6B in revenue and $91B in Q2 guidance prove AI factory buildout isn't a bubble. But Google's TPU JV ($5B, 500MW), Cerebras' $95B market cap IPO, and NVIDIA's first-ever 10-Q acknowledgment of customer ASIC risk — all point to one trend: NVIDIA's customers are becoming NVIDIA's competitors, and they have money, talent, and motivation. The question isn't "whether" but "how fast."

2. NVIDIA is transitioning from "GPU company" to "AI infrastructure company" — Vera CPU, networking, and software are the three pillars of this transformation.

GPU remains NVIDIA's core, but Vera CPU opens a $200B new market ($20B already sold in the first months of this year), networking revenue is annualizing at $59B with growth far exceeding compute, and the software stack (Dynamo + NemoClaw + OpenShell) is building a moat even harder to replace than CUDA. If this transformation succeeds, NVIDIA's long-term growth runway extends far beyond the GPU market's boundaries. NVIDIA will no longer be a "chip company" but an "AI-era operating system company" — providing complete infrastructure from hardware to software.

3. The permanent loss of the China market is NVIDIA's biggest structural risk — not just lost revenue, but the creation of an independent competitive ecosystem.

Jensen is right: U.S. sanctions helped China "shatter its comprador-style technology fantasies," catalyzing a complete alternative ecosystem from hardware (Ascend 910C/950PR) to software (CANN/CUDA compatibility layers) to models (DeepSeek V4). Once this ecosystem matures and exports to the Global South, NVIDIA won't face a missing market — it'll face a new competitor. Huawei's CM384 cluster already surpasses NVIDIA NVL72 on specific tasks — albeit at 4x power consumption and 16x space, but in budget-sensitive markets, this may be acceptable.

Next Observation Points

---

Appendix A: Key Data Sources

1. NVIDIA Q1 FY2027 Official Press Release (2026.05.20)
2. NVIDIA Q1 FY2027 Earnings Call Transcript (2026.05.20)
3. NVIDIA Q1 FY2027 10-Q Filing (pending SEC submission)
4. NVIDIA FY2026 10-K Annual Report (2026.03)
5. NVIDIA GTC 2026 Keynote (2026.03.17)
6. Jensen Huang CES 2026 Presentation (2026.01.06)
7. Blackstone + Google TPU JV Announcement (2026.05.18)
8. Cerebras IPO Pricing and First-Day Trading Data (2026.05.14)
9. CNBC/Yahoo Finance: NVIDIA Q1 FY2027 Earnings Live Coverage
10. SemiAnalysis: CM384 Cluster vs. NVL72 Comparison Analysis
11. Coherent/Corning/Lumentum Strategic Agreement Announcements
12. Tom's Hardware: Vera Rubin, Vera CPU, and Groq 3 LPU Technical Analysis
13. Futurum Group: Cerebras IPO Analysis Report
14. The Information: DeepSeek V4 Migration to Huawei Ascend Coverage
15. Jon Peddie Research: NVIDIA-Groq Technology Integration Analysis

---

This article is based on publicly available information and does not constitute investment advice. All data as of May 21, 2026.

---

Appendix B: Hyperscale CapEx and NVIDIA Revenue Correlation Analysis

NVIDIA's data center revenue is highly correlated with Hyperscale customers' CapEx. Below are major customers' FY2026-27 CapEx guidance:

Key inference: Assuming ~40-50% of Hyperscale CapEx goes to AI compute hardware (GPU/TPU/ASIC + networking), FY2027 Top 5 Hyperscale AI hardware spending would be ~$140-190B.

NVIDIA's Q1 FY27 Hyperscale data center revenue was ~$38B, annualized ~$152B — broadly consistent with the above estimate. This means NVIDIA's share of Hyperscale AI hardware spending is roughly 80-100% — factoring in non-NVIDIA hardware (in-house TPU/Trainium, networking equipment, power and cooling), NVIDIA's actual share may be 70-80%.

This share could decline in FY28-29 — but the pace depends on in-house chip maturity. If Google's TPU JV succeeds and scales, NVIDIA's Google share could drop from ~80% to ~50-60%. If Amazon's Trainium 3 wins more inference customers, NVIDIA's AWS share could also decline.

Sensitivity analysis for NVIDIA revenue:

In the "medium share decline" scenario, NVIDIA's Hyperscale revenue is still growing — just at a slower rate. The real risk lies in the "large share decline" scenario, which would require multiple in-house chips to simultaneously mature and achieve broad customer adoption.

---

Appendix C: Vera Rubin Technical Specifications Summary

Core improvements summary:

• Compute: GPU transistor count +62%, FLOPS ~ +2.5-3x
• Memory: HBM capacity +50%, bandwidth +83%
• Interconnect: NVLink bandwidth +100%
• Storage: ICMS NAND capacity +100%
• CPU: From Grace (ARM v9, non-in-house cores) to Vera (ARM v9 Olympus, in-house cores)
• System complexity: Discrete components +117%

---

Appendix D: Vera CPU Rack Technical Specifications

Vera CPU Rack's design goal is clear: not to replace traditional x86 servers for general computing, but to provide dedicated CPU clusters for Agentic AI's execution layer.

---

Appendix E: Quarterly Financial Data Summary (FY25-FY27)

Trend analysis:

1. Revenue growth re-accelerated to +85% YoY in Q1 FY27. This breaks the deceleration trend from Q1 FY25 (122% → 94% → 78% → 69% → 70% → 94% → 73% → 85%). The re-acceleration indicates AI demand has entered a second growth wave.

2. Q1 FY27 QoQ growth of +20% is the highest in the past 5 quarters. This shows Grace Blackwell's full shipment meaningfully boosted revenue.

3. EPS growth (+140% YoY) continues to exceed revenue growth (+85% YoY). Operating leverage continues to work — every incremental dollar of revenue carries higher marginal profitability.

---

Appendix F: NVIDIA FY27 Full-Year Scenario Analysis

Bull scenario assumptions: Vera Rubin delivery smooth, Agentic AI demand exceeds expectations, networking revenue sustains high growth. Base scenario assumptions: Vera Rubin delivery with minor delays but overall smooth, AI demand remains strong but growth normalizes. Bear scenario assumptions: Vera Rubin delivery delays, AI investment growth slows, customer in-house ASICs begin substitution. Recession scenario assumptions: Macroeconomic recession, AI investment cuts sharply, H2 shows QoQ decline.

---

Appendix G: Key Terminology Glossary

---

This article is based on publicly available information and does not constitute investment advice. All data as of May 21, 2026.