NVDA — Q3 FY26 Earnings Review (muji)

Quarter: Q3 FY26 (Oct-25), reported 2025-11-19 Written: 2026-02-22 (pre-Q4 FY26 earnings, Feb 25) Long NVDA: 0% (watchlist — not in current wsm portfolio) Atlas baseline: ~/.agents/skills/atlas/analyses/NVDA/NVDA_earnings-review_Q3_FY26.md

Links

• NVIDIA Newsroom — Q3 FY26 PR
• Motley Fool Transcript — Q3 FY26 (2025-11-19)
• CFO Commentary PDF

The Numbers

| | Q4 FY23 | Q1 FY24 | Q2 FY24 | Q3 FY24 | Q4 FY24 | Q1 FY25 | Q2 FY25 | Q3 FY25 | Q4 FY25 | Q1 FY26 | Q2 FY26 | Q3 FY26 | | | Jan-23 | Apr-23 | Jul-23 | Oct-23 | Jan-24 | Apr-24 | Jul-24 | Oct-24 | Jan-25 | Apr-25 | Jul-25 | Oct-25 | |---|---|---|---|---|---|---|---|---|---|---|---|---| | Revenue ($m) | 6,051 | 7,192 | 13,507 | 18,120 | 22,103 | 26,044 | 30,040 | 35,082 | 39,331 | 44,062 | 46,743 | 57,006 | | YoY % | -20.8% | -13.2% | 101.5% | 205.5% | 265.3% | 262.1% | 122.4% | 93.6% | 77.9% | 69.2% | 55.6% | **62.5% ^^** | | QoQ % | 2.0% | 18.9% | 87.8% | 34.2% | 22.0% | 17.8% | 15.3% | 16.8% | 12.1% | 12.0% | 6.1% | **+22.0% !!** | | GM [GAAP] | 63.3% | 64.6% | 70.1% | 74.0% | 76.0% | 78.4% | 75.1% | 74.6% | 73.0% | 60.5%* | 72.4% | **73.4%** | | GM [Non-GAAP] | — | — | — | — | — | — | — | 75.0% | — | 70.6%* | 75.7% | **73.6%** | | Op Margin [GAAP] | 20.8% | 29.8% | 50.3% | 57.5% | 61.6% | 64.9% | 62.1% | 62.3% | 61.1% | 49.1%* | 60.8% | **63.2%** | | Net Margin [GAAP] | 23.4% | 28.4% | 45.8% | 51.0% | 55.6% | 57.1% | 55.3% | 55.0% | 56.2% | 42.6%* | 56.5% | **56.0%** | | FCF ($m) | 1,736 | 2,643 | 6,048 | 7,042 | 11,246 | 14,936 | 13,483 | 16,787 | 15,519 | 26,135 | 13,450 | 22,089 | | DC Revenue ($m) | 3,616 | 4,284 | 10,323 | 14,514 | 18,404 | 22,563 | 26,272 | 30,771 | 35,580 | 39,112 | 41,096 | 51,215 |

*Q1 FY26 margins depressed by $4.5B H20 inventory charge. Normalized [Non-GAAP] GM was ~70.6%.

Revenue trajectory: -20.8% → 101.5% → 265.3% → 122.4% → 93.6% → 77.9% → 69.2% → 55.6% → 62.5% ^^

Q4 FY26 guide: $65.0B ± 2% | GM guide: 75.0% [Non-GAAP] | Implied YoY: +65.3% ^^

Beat-and-Raise Summary

The two genuine surprises: (1) $10.3B sequential revenue add — largest in NVIDIA's history. (2) Q4 GM guide of 75% [Non-GAAP], signaling Blackwell cost structure maturing faster than feared.

Platform Assessment: NVIDIA as "Scale in Platform"

Here's the thing — NVIDIA is one of the few semiconductor companies I'd call a true platform company by the "Scale in Platform" framework. Let me break this down:

What is the NVIDIA platform?

It's NOT just the GPU. The moat is the full stack:

1. CUDA — The real building block. 4M+ developers, 5000+ applications, 20 years of optimization. This is the "install base" that makes NVIDIA sticky in the same way Salesforce's data network or Cloudflare's global anycast makes them sticky. Every major AI model — GPT, Gemini, Claude, Llama — trains on CUDA. You can't just swap it out.
2. NVLink / NVSwitch — The interconnect that makes multi-GPU coherent compute possible. Scale-up networking. GB200 NVL72 at 72 GPUs with NVSwitch coherency is something ASICs cannot replicate for training.
3. Spectrum-X / InfiniBand — Scale-out networking. This is the "picks and shovels" play — NVIDIA is selling the interconnect fabric that ties GPU clusters together. $8.2B in networking at +162% YoY. This segment is becoming a second growth engine !! and most people are still treating it as a rounding error.
4. DGX Cloud / NVIDIA AI Enterprise — The software layer. Recurring revenue against the hardware install base. Early, but this is where SaaS-like economics eventually emerge.
5. CUDA X Libraries — cuDNN, cuBLAS, NCCL, Triton interop. These are the "building blocks" that make every ML framework faster on NVIDIA hardware. PyTorch and JAX both default-optimize for CUDA kernels first.

NUANCE: Jensen's "three platform shifts" framing (CPU→GPU, generative AI, agentic AI) is marketing, but it's correct marketing. These are real demand waves. Agentic AI in particular is genuinely net-new inference demand — it's not cannibalizing prior workloads, it's additive. Cursor, Harvey, Waymo, Tesla FSD — all running inference at scales that didn't exist 18 months ago.

Cloud-native? Self-serve? Building block?

Not a typical SaaS, but evaluating on equivalent dimensions:

• Architecture: Cloud-agnostic (runs on AWS, Azure, GCP, on-prem). GPU clusters are becoming the new cloud primitive.
• GTM: SI-dependent for enterprise deployment, BUT the demand pull is developer-first. Developers choose CUDA. Procurement follows. This is the same pattern as Cloudflare — bottom-up adoption forces top-down spend.
• Building block vs process tool: NVIDA hardware is the deepest possible building block — you cannot run AI without it right now. Your product is embedded in their models, their training infra, their inference clusters. Switching costs are 10 years of CUDA optimization.

My stance on platform quality: This is a Tier 1 platform company by any reasonable definition. CUDA is the most defensible developer ecosystem in infrastructure. The question isn't if NVIDIA is a platform — it's whether the platform moat survives ASIC competition.

Product Cadence

This is what I track most closely. NVIDIA has been running an annual architecture cadence — the most aggressive in semiconductor history:

GB300 at 67% of Blackwell mix in Q3 is a big deal. It means customers are skipping B200 and going straight to the higher-margin, higher-compute GB300. That's why GM recovered to 73.6% [Non-GAAP] — the product mix is richer than bears expected.

Rubin on track for H2 CY2026. Silicon is back from TSMC. This is the most important leading indicator for FY27 revenue. The 5x inference improvement over Blackwell means customers who bought Blackwell for inference will have a compelling upgrade cycle. Training upgrades will follow.

NUANCE: Annual architecture cadence is both a strength and a risk. Strength: creates recurring upgrade demand. Risk: every new architecture has a transition quarter where GMs compress (we saw this with Blackwell in Q1 FY26 — the H20 charge + ramp costs took GM to 60.5% [GAAP]). The Rubin ramp will cause another dip. Watch FY27 Q1-Q2 GMs carefully.

Leading Indicators — The Data That Matters

NVIDIA doesn't give NRR or customer cohort data like SaaS. But there are equivalent signals:

DING DING DING on the $26B in multi-year cloud agreements, up from $12.6B. That's not a purchase order — that's a hyperscaler sitting down and saying "we're committed to this infrastructure for 2+ years." Amazon, Google, Microsoft, Meta — all writing long-term supply agreements. This is the closest equivalent to an NRR signal NVIDIA has.

Key CC Quotes

Jensen Huang (CEO):

"The demand for Blackwell is extraordinary. Our supply is sold out well into next year. And our customers are not just buying GPUs, they're building next-generation AI factories."

"Three simultaneous platform shifts are happening: CPU-to-GPU accelerated computing, generative AI, and now agentic AI. Each one is a multi-hundred-billion-dollar infrastructure opportunity."

"Rubin is on track. Silicon is back. The team is doing an extraordinary job."

Colette Kress (CFO):

"Q4 gross margins guided to 73% GAAP / 75% Non-GAAP. For FY27, we expect to be in the mid-seventies. But input costs are rising."

The "input costs rising" caveat from Kress is the only thing in this transcript that gave me pause. TSMC pricing power, HBM4 costs (SK Hynix and Samsung both raising prices), CoWoS packaging constraints — these are real and will pressure FY27 GMs. The "mid-seventies" guidance is honest. Don't expect a return to 78% peak anytime soon.

NUANCE Section

NUANCE 1: ASIC is a 2027+ risk, not a 2026 risk — but the roster is expanding.

The ASIC competitive threat is real. I've tracked it since Google TPU v3. Here's my updated read:

• Training ASICs: not competitive. Training requires massive all-reduce communication, dynamic sparsity, and model architecture flexibility that ASICs can't handle efficiently. Google's TPUs do 75%+ of Gemini inference but Gemini was trained on NVIDIA hardware. This distinction matters enormously.
• Inference ASICs: increasingly competitive for narrow, fixed-topology models. Google TPU v6, AWS Trainium 2, Meta MTIA. But these are for their own models, at their own scale. Third-party inference (OpenAI running on Azure, Anthropic on GCP) still routes through NVIDIA.
• The new risk: OpenAI + Broadcom ASIC partnership (announced Feb 2026). OpenAI is NVIDIA's single largest inference customer. If OpenAI migrates meaningful inference volume to custom silicon, that's a demand headwind. Not existential, but non-trivial. Timeline: 2026 delivery, volume ramp 2027.

My take: 2027-2028 ASIC headwinds are likely real but bounded. Jensen's response ("ASICs optimize for yesterday's model architecture, NVIDIA optimizes for the next one") is clever and partially true. Model architecture is still evolving rapidly. But at scale, hyperscalers have both the volume and talent to make custom ASICs work. This risk doesn't go away.

NUANCE 2: The networking business is being underappreciated.

$8.2B at +162% YoY in a single quarter. That's a $30B+ annual run rate business growing faster than the GPU business was at its peak. NVLink and Spectrum-X are not accessories — they're the glue that makes GPU clusters function. As cluster sizes scale from 1,024 GPUs to 16,384 GPUs (NVL-72 rack configurations), the interconnect complexity — and NVIDIA's monetization — scales with it. Every GB300 sold pulls through NVLink and Spectrum-X revenue. This is the best "building block" dynamic in the whole stack.

NUANCE 3: Sovereign AI is a meaningful demand floor, underappreciated by the market.

5M GPU announcements in Q3 from sovereigns (KSA, UAE, France, Japan, India, etc.). Saudi Arabia signed for 400-600K GPUs over 3 years. Sovereign governments building national AI compute are not subject to hyperscaler CapEx cyclicality. They're budget-line items in national security plans. This demand doesn't disappear if Google has a down quarter. It's a structural demand floor that partially offsets hyperscaler concentration risk.

Risks Scorecard

My Stance

NVIDIA is the most dominant platform company in technology infrastructure today. Full stop.

At $57B in quarterly revenue, +62.5% YoY, +22% QoQ, with 73.4% GAAP gross margins and $22B of quarterly FCF — this is the most profitable hardware platform ever built at this scale. The revenue trajectory just reaccelerated (55.6% → 62.5% ^^) and Q4 is guided to 65.3% further acceleration. That is extraordinarily rare at $57B quarterly revenue.

The technologist in me is most impressed by two things: (1) the CUDA ecosystem moat — 20 years of developer lock-in that no ASIC can replicate for general-purpose training workloads — and (2) the networking business emerging as a genuine second growth engine. $8.2B at +162% YoY means NVIDIA is no longer selling GPUs, they're selling entire AI factory infrastructure.

IMHO, the forward P/E of ~25x for 60%+ growth and 56% net margins is objectively reasonable. The market is sitting on this sideways because it needs confirmation that Q4 FY26 (reporting Feb 25, 2026) sustains the acceleration. If NVDA guides Q1 FY27 to $75B+, the stock re-rates.

For Q4 FY26 (Feb 25, 2026) — watch these specifically:

1. Q1 FY27 guidance. Street at ~75B.Anything≥76B = buy signal.
2. Rubin timeline confirmation. Any pull-forward is !! Any delay is a problem.
3. ASIC commentary. Jensen's response to the OpenAI/Broadcom ASIC news will set the narrative.
4. Networking — can it continue above $8B? If it approaches $10B, this segment alone is remarkable.

My stance: NVDA is Tier 1. Not currently in the wsm portfolio but deserves a watchlist position with a clear entry thesis — either a pullback on the Rubin transition quarter (FY27 Q1-Q2 GM dip) or confirmation of sustained 65%+ growth post-Q4.

The ASIC risk keeps me from a max-conviction add today. But make no mistake — this is the best platform company in infrastructure computing. The supply commitment data alone ($50.3B, +63% QoQ) tells you demand is not slowing.

Long NVDA: 0% (watchlist)

-muji

Written 2026-02-22, pre-Q4 FY26 earnings (Feb 25, 2026). Based on Q3 FY26 data (Oct-25, reported Nov 2025) + Atlas baseline analysis.