Nvidia Deep Dive: Blackwell Peak Pricing, Rubin on Deck, and the ASIC Siege — What Changes Before the May 20 Earnings

OurAlpha Scorecard

Composite 6.75/10 — worth a deep dive, but not a no-brainer-buy window.

I. Pre-Earnings 5/20: $78B Guide, Raised Estimates, and the Potential Setup

NVIDIA's Q1 FY27 earnings will be reported on May 20, 2026. Management's Q1 FY27 outlook, issued during the Q4 FY26 call, was $78.0B ±2%^{[NVIDIA Newsroom]}. But the Street has since revised estimates up to roughly $79B—meaning the real bar this quarter isn't simply "did they hit official guidance," but rather whether Q2 guidance comes in above the already-elevated consensus.

Why this print matters more than usual:

1. FY26 full-year revenue of $215.9B (+65% YoY) is now in the books^[CNBC], with Q4 Data Center revenue of $62.3B (+75% YoY / +22% QoQ)—raising the comp base further. Even if Q1 FY27 comes in at $78B, sequential growth is only mid-single digits; YoY growth can still exceed 60%, but that reflects base effects compounding with the Blackwell ramp—not a trend that extrapolates cleanly into FY28
2. Networking is the underappreciated story: Q4 Networking revenue hit $11B (+3.5x YoY)—Ethernet, NVLink, and Spectrum-X are collectively becoming a standalone profit center. The quality of this growth is higher than GPU silicon sales alone
3. Q4 GAAP gross margin 75.0%, non-GAAP 75.2%: Management had guided to a 73–75% range; Q4 came in at the top end—the market is watching whether the B300 ramp can push gross margins above 75%
4. GB300 / Blackwell Ultra entering material revenue contribution is a key focus on 5/20 (see Section II)—whether management breaks out GB300 revenue for the first time will be a critical signal for a valuation re-rating
5. Customer concentration disclosure: The FY26 10-K has been updated with full-year figures (see Section VI)—whether management emphasizes diversification narratives such as "40% of revenue from non-hyperscaler customers" on the call will affect the valuation premium (management has repeatedly highlighted enterprise, sovereign AI, industrial, and neocloud demand alongside hyperscalers)

OurAlpha View: If Q1 revenue beats to $80B+, Q2 guidance steps up again, and GB300 revenue contribution is clearly signaled, it means ASIC substitution hasn't meaningfully cut in yet → the narrative stays intact. If Q1 lands at $78–79B and Q2 guidance is cautious, a valuation transition will kick in immediately—the market will rotate its anchor from FY27 EPS to FY28 EPS, and volatility will expand in the interim.

II. Blackwell Enters Peak Pricing Territory: Demand Isn't Slowing — the Market Is Pricing in Rubin Early

Blackwell is the fastest-ramping product cycle in NVIDIA's history^{[NextPlatform]}. But note: rather than saying Blackwell's fundamentals have already peaked, it's more accurate to say the market has begun treating Blackwell as a cycle already being harvested, and has pivoted its valuation anchor to Rubin ahead of time. On the demand side, NVIDIA management continues to emphasize strong Blackwell/Rubin demand and an ongoing GB300 ramp.

Current product lineup:

A note on B300 single-chip compute figures: the official DGX B300 documentation discloses system-level numbers (8 chips combined, 144 PFLOPS FP4 inference); some third-party sources back-calculate a per-chip figure of ~15–18 PFLOPS dense FP4 — variances across dense/sparse, NVFP4/FP4, and system-efficiency assumptions are significant, so per-chip numbers should not be cited as fact.

Signals that Blackwell has entered "peak pricing territory":

1. B200/GB200 backlog ~3.6M units, sold out through mid-2026 — third-party supply chain estimate^{[FinancialContent]}, not official NVIDIA disclosure — the largest backlog in the company's history, but also meaning there is limited room for near-term upside surprises
2. B300 is ahead of the original schedule: third-party sources indicate B300 began shipping in January 2026 (NVIDIA has not separately disclosed revenue recognition timing in earnings) — an accelerated product cadence could pressure B200 ASPs in the near term
3. Rubin takes over in 2H 2026, leaving Blackwell an effective "exclusive window" of just 6–9 months
4. Scattered GB200 delay signals emerged in late 2025^{[IO Fund]}; liquid-cooling supply chain complexity is higher than prior cycles — GB300 carries the same dependency

Strategic significance of Blackwell Ultra (B300):

B300 marks the first time NVDA has made a major hardware upgrade mid-cycle within a primary generation — previously, H100 → H200 was a smooth step-up; B200 → B300 jumps HBM capacity from 192GB to 288GB and lifts FP4 throughput (system-level) by roughly 50% versus GB200. This can be read as a direct response to the ASIC threat: by tightening the cadence of hardware generations, NVIDIA is signaling to customers that "in-house silicon will never keep pace with NVDA."

The core tension: Blackwell shipments are still accelerating, yet the market is already pricing in the next generation, Rubin. This valuation handoff creates a period of "multiple compression limbo" — the peak itself is not the concern, but volatility will amplify during the transition.

III. Rubin: CES 2026 Launch, 2H 2026 Mass Production — Valuation Story Is Shifting

Rubin (the Vera Rubin platform) was officially unveiled at CES 2026^{[ServeTheHome]}. The officially confirmed direction is a full-stack platform upgrade — HBM4 + NVLink 6 + Vera CPU + DPU/NIC/Switch — with memory bandwidth nearly 3x that of Blackwell^{[NVIDIA Blog]}.

Details disclosed by select tech media (not all explicitly listed on NVIDIA's official pages):

⚠️ Due to varying definitions of FP4/NVFP4 and system-level vs. chip-level measurements across sources, the figures above should be read as publicly available technical data and should not be treated as equivalent to earnings-reported metrics.

Why Rubin Matters for the Industry:

1. HBM4 is Rubin's defining feature — the generational leap is not raw compute but memory bandwidth. 3x bandwidth = a step-change in large-model inference throughput
2. Rubin has moved from roadmap to production/availability — the market will now focus on initial hyperscaler deployment cadence, particularly whether Microsoft, Meta, Oracle, AWS, and other major customers commit to bringing Rubin systems online in 2H 2026
3. The full platform includes Vera CPU + ConnectX-9 + BlueField-4 + NVLink 6 + Spectrum-X — NVLink 6's inter-GPU bandwidth is a system-level advantage that ASICs cannot replicate within two years

HBM4 Supply Allocation:

• SK Hynix ~2/3 / ~70% of NVDA's HBM4 allocation (per supply chain research)^[TrendForce], using conventional MR-MUF process
• Samsung >30% — began shipping HBM4 in February^[Digitimes], via hybrid bonding process
• Micron primarily supplies mid-tier inference chips (Rubin CPX), not the flagship Rubin
• NVDA has required suppliers to deliver 16-layer HBM4 before Q4 2026 — in preparation for Rubin Ultra

Valuation Implications:

• The market is already toggling between FY27 and FY28 P/E. Forward P/E of 26–29x is based on FY27 EPS estimates — if Rubin lifts data center revenue another leg in FY28 (some sell-side models project $300B+), current valuations still have upside
• Conversely, if Rubin production is delayed, HBM4 yields disappoint, or 16-layer HBM4 faces supply difficulties, the valuation re-rating cycle stalls

IV. China Reentry Window: H200 Approved but Undelivered — Breaking Down the Market Share

In April 2025, the U.S. imposed license requirements on the H20.^{[NVIDIA Newsroom]} Direct impact:

• Q1 FY26: $4.5B inventory write-down + $2.5B in lost revenue
• Q2 FY26 guidance: H20-related revenue loss of ~$8.0B

The picture shifted in 2026: China moved from "fully banned" to a gray zone of "approved but undelivered."

Reuters / CNBC reporting^{[CNBC/Reuters]} — key facts:

• The U.S. has approved roughly 10 Chinese companies to purchase H200s, including Alibaba, Tencent, ByteDance, JD.com, Lenovo, Foxconn, and others
• Per-company purchase cap: 75,000 units
• As of the May 14 report, no actual deliveries had occurred
• The Chinese government has advised buyers to hold off, citing concerns that the H200 may contain "hidden backdoors"

Accordingly, Q1 FY27 should not assume any meaningful China revenue recovery; the more actionable signal will be whether management reinstates China revenue in its Q2/Q3 guidance.

NVDA's China share must be interpreted by the right lens:

• Installed base / cumulative historical share: Some third-party estimates put this at ~55–60% (including legacy H100/A100 units, etc.)
• Current new sales of high-end AI accelerators under restrictions: Jensen Huang stated publicly in April that NVDA's share in China has "effectively fallen to zero"^{[Tom's Hardware/Reuters]}, down from a historical peak of 90%+

These two metrics cannot be conflated. "55% installed base" reflects cumulative shipments over the past five years; "near-zero new sales" reflects the current reality of a restricted market.

Huawei Ascend 910C and domestic substitution:

That said, real-world large-scale training efficiency remains constrained by software ecosystem maturity, cluster interconnect, yield rates, and supply chain.

Huawei 2026 production targets:

• Plans to produce 600,000 910C units in 2026, double the 2025 output^[Bloomberg]
• Total Ascend-series die output could reach 1.6 million units
• Chinese model companies such as DeepSeek have publicly emphasized compatibility with and deployment on domestic AI chips, reinforcing Ascend's position in China's inference and domestic-substitution use cases

OurAlpha take: The H200 is a relatively dated product (pre-Blackwell), and even if Chinese customers ultimately receive it, they're getting a downgraded SKU. Once the Ascend 910C/910D closes the performance gap to 80%+ with stable supply, pricing pressure will compress NVDA's China margins well below its global average. The bear case flip: Huawei's ramp has its own ceiling — SMIC 7nm yields remain constrained, and the CUDA vs. MindSpore developer ecosystem gap is roughly 10:1. Chinese customers running large training workloads will likely still reach for NVDA. Ascend's primary battlefield will be inference.

V. The ASIC Siege: What Is Custom Silicon Actually Changing?

This is the biggest long-term variable in the NVDA story for 2026. The ASIC siege has moved from narrative to execution, but certainty varies widely across projects. The table below breaks each player into "official/verifiable" vs. "market expectation/supply-chain sourced":

The real threat from custom ASICs is not "replacing NVDA" — it's diverting incremental orders:

1. Training: Rubin remains nearly irreplaceable for frontier model training — precision, ecosystem, and CUDA lock-in all create meaningful moats
2. Inference: this is where ASICs fight their main battle — predictable workloads, single-model repeated execution, low ecosystem dependency
3. Inference already accounts for ~2/3 of total AI compute (per some sell-side estimates) — the natural pool for ASIC encroachment

Broadcom is one of the clearest beneficiaries of the ASIC substitution narrative: Google's long-term TPU contract (through 2031), the Meta 2nm AI accelerator partnership, and Anthropic's TPU capacity expansion with Google/Broadcom all point to the same conclusion — hyperscaler custom silicon is not a closed loop within individual companies, but is coalescing into a Broadcom-led custom silicon supply chain. AVGO has emerged as the single biggest winner of the "AI alternatives" trade.

Market share forecasts split into two camps:

• Bear case on share loss: NVDA data center share drifts from 86% to ~75% by 2026, but with the overall market still growing 30%+, absolute revenue keeps expanding
• Bull case on disruption: Inference market share could fall from 90%+ to 20–30% by 2028 — a minority view from more aggressive analysts (sourced via Introl citing New Street Research et al.); not NVIDIA's own guidance, not current consensus — better treated as a stress-test scenario than a base case

OurAlpha's take: The conservative scenario is more probable. The CUDA ecosystem, software stack, and customer switching costs dwarf any price-tag advantage — but ASICs have moved from "theoretical threat" to "real numbers." The premium in NVDA's valuation is being slowly compressed. The market is repricing NVDA from "monopoly premium" to "high-share incumbent" — and those two framings carry an implied EV/Sales gap of at least 3–5x.

六、客户集中度：必须区分单季与全年口径

这是 NVDA 财报里最危险也最容易被误读的一行数字——口径不分清就会犯错。

FY26 Q3 single-quarter 10-Q basis^{[Motley Fool]}:

FY26 full-year 10-K basis: only 2 direct customers exceeded 10%, at 22% and 14% respectively.

In other words, the concentration looks extreme on a single-quarter basis (4 customers, 61%), but the full-year picture is far less dramatic — an important distinction that prevents reading Q3 as the new normal.

On customer identity: NVIDIA does not disclose names. The market commonly speculates that the top two customers may be linked to Microsoft / Meta / Oracle / Amazon / xAI / OpenAI, based on capex scale, AI cluster purchases, and supply-chain signals — but NVIDIA has not confirmed this. It would be inaccurate to state outright that "the top two = MSFT + Meta."

Risk transmission:

• If any top customer publicly announces "50% of 2027 inference workloads moving to ASICs" → NVDA faces an immediate derating
• Meta's collaboration with Broadcom on a 2nm AI accelerator is an on-the-record signal
• AMZN Trainium 3 + Anthropic closed loop (see OurAlpha AMZN deep dive) = Anthropic's future training compute will gradually diversify (including Google TPU + AWS Trainium)
• Once Microsoft's internally developed ASICs for OpenAI reach production, it could theoretically reduce dependence on NVDA

The counter-argument: high customer concentration is not necessarily a negative — deep hyperscaler relationships mean long-term contracts and multi-year commitments. Management has also repeatedly emphasized demand from enterprise, sovereign AI, industrial, and neocloud customers beyond the cloud providers, as part of an effort to proactively diversify the customer base.

VII. Hyperscaler Capex: $600B Total, $450B AI-Related — Mind the Definitions

2026 Big Five hyperscaler aggregate capex is widely projected to exceed $600B (~$602B, +36% YoY)^[MUFG]. However, roughly 75%, or ~$450B, is directly attributable to AI infrastructure — the $600B figure represents total capex across all uses and should not be treated as a GPU/AI chip procurement pool in its entirety.

2026E capex / AI infrastructure spend rough estimates (not company-audited segment figures):

⚠️ The figures above are sell-side model / media estimates, not GAAP segment capex guidance. Amazon's $200B reflects management commentary but is not purely AI-directed; Meta, Microsoft, and others are closer to sell-side aggregates.

The real question is how much of that capex flows to NVDA:

• 2024–2025: 80%+ went to NVDA (GPUs were the only viable option)
• From 2026 onward: that share is declining — hyperscalers are directing a portion of capex toward in-house ASICs, as well as networking, storage, power, and cooling
• Scenario estimate: using $450B in AI-related capex as the base and assuming NVIDIA captures 45–55% of it, the implied addressable revenue pool is roughly $200B–$250B — this is a rough scenario and should not be equated directly with NVIDIA's reported revenue, as it encompasses systems, networking, memory, services, channel dynamics, delivery lead times, and revenue recognition differences

Bank of America's $320 price target rests on a projected 2027 AI data center TAM of $1.7T^{[24/7 Wall St]} — as long as that total market materializes, NVDA's absolute revenue could still double even as its market share declines.

$600B is both a floor (the pie is still growing) and a ceiling (growth will eventually converge):

• Physical constraints: power availability, transformer capacity, data center construction timelines, HBM4 / TSMC N3 supply — these are all hard caps
• Depreciation cycle: hyperscalers are broadly extending AI server useful lives from 5 to 6 years — a near-term EPS tailwind, but also a signal that customers don't want hardware to turn over too quickly (a bearish read-through for Rubin Ultra / Rubin Next demand)

Whether you frame NVDA as a cyclical, a growth stock, or an infrastructure bellwether, the three valuation frameworks yield wildly different price targets.

VIII. Valuation & Price-In: $271 Consensus, $320 BofA, $195 Bear Case

Current share price approximately $236 (as of 2026-05-14 close / 5-15 UTC)^{[Stock Analysis]}. Sell-side distribution^[MarketBeat]:

85% Buy ratings, Strong Buy consensus. Yet the dispersion (high-vs-low ratio of 1.64x) is already well above the typical Mag7 name—the market's long-term narrative on NVDA is beginning to fracture.

What's already priced into consensus:

• Blackwell cycle delivers cleanly through mid-2026
• Blackwell Ultra (B300) picks up the baton with a smooth Q1–Q2 FY27 ramp
• Rubin enters volume production in 2H 2026 without material delay
• China revenue recovers quarter by quarter (market assumes H200 deliveries materialize in Q2–Q3)
• ASIC displacement is "gradual"—share loss stays below 15 percentage points over five years
• Hyperscaler capex continues climbing through 2027

Tail risks not yet priced in (each capable of derating the multiple 10%+):

• Rubin HBM4 supply or yield issues
• Hyperscalers shift ASIC share in inference to 50%+ in a single move
• China re-ramp slower than expected (even if H200 gets cleared, Beijing may continue to discourage purchases; Huawei Ascend has already partially filled the gap)
• AI capex cycle "quality adjustment"—the market starts asking about the real ROIC on $600B of spend
• A key customer (e.g., Meta) publicly announces in-house chips handling 70%+ of inference
• Geopolitical reversal—next round of China export restrictions tightens further

OurAlpha Take: A forward P/E of 26–29x for a hardware company whose growth is stepping down but whose revenue base keeps expanding is fair-to-rich. Not cheap, but not absurd—provided Rubin executes cleanly and ASIC displacement stays gradual rather than accelerating.

NVDA is no longer a no-brainer buy, but it's far from a sell. It's entering a phase of multiple compression + narrative refinement, during which the stock will oscillate ±10–20% around every earnings print and hyperscaler capital allocation update.

IX. Hopper → Blackwell → Rubin: Three-Generation Cycle Math

To frame NVDA's medium-term valuation, look at the cycle math across three product generations:

Key observations:

1. Each cycle is getting longer——Hopper ran 24 months; Blackwell is now tracking to 30. Blackwell Ultra (B300) is essentially NVDA harvesting a "mid-cycle refresh" as a second revenue wave within the same generation.
2. Revenue scales up by an order of magnitude each generation——Hopper peaked at ~$110B in datacenter revenue in FY25; Blackwell is tracking $215B+ across FY26–FY27 ($215.9B for full-year FY26); the sell-side bull case for Rubin in FY28–FY29 is $300B+.
3. Generations stack, they don't replace——many customers are running H100 + B200 + B300 simultaneously. That coexistence is a structural moat for NVDA's cash flows.

Risk factors:

• If the Rubin cycle "only holds for 24 months" — because custom ASICs accelerate share gains in 2028 — the FY29+ narrative breaks down.
• If Rubin Ultra (HBM4 16-layer, late 2027) hits volume production snags, NVDA loses its cadence advantage in the next leg of the products-vs-ASICs race.

That's why the Blackwell Ultra gross margin signal in the 5/20 earnings matters more than any single quarter's revenue — it tells you whether NVDA can manufacture a mid-generation upgrade cycle within each architecture.

十、接下来 3 个季度，市场真正会看什么？

5 个可执行指标：

1. 5/20 Earnings: Whether B300 / GB300 Ramp Enters Meaningful Revenue Recognition——Whether management breaks out GB300 revenue contribution for the first time; whether Q4 GAAP gross margin of 75% expands further post-B300 ramp
2. Q2 FY27 China Revenue Run-Rate——Whether management is willing to put a number on it. The key question: whether H200 moves from "approved" to "delivered"
3. The Specific Quarter Rubin Revenue Begins to Be Recognized——Consensus expects Q3–Q4 FY27; any pull-forward or pushout will trigger a reprice
4. Hyperscaler-Reported ASIC vs. GPU Deployment Mix (Meta / AWS / Google are growing more transparent in their own filings)——If any one of them discloses inference ASIC share above 40%, NVDA faces an immediate derate
5. TSMC N3 / N3P Capacity Allocation——How much capacity Rubin secures is zero-sum against ASIC customers (who are also competing on 3nm)

十一、风险清单

• Valuation re-rating risk: As the market rolls from FY27 to FY28 EPS, Rubin expectations are already priced in — any negative surprise will be amplified
• Accelerating ASIC displacement: If any hyperscaler publicly states "60% of 2027 inference runs on ASICs," the "gradual-shift" narrative breaks immediately
• Rubin production delays or HBM4 16-layer yields: HBM4 supply constraints, TSMC N3 yield issues, or any bottleneck will push out revenue recognition
• China geopolitical volatility: The grey zone of H200 approvals-pending-delivery could evolve into sustained Chinese rejection post-approval; the next round of export controls may be even stricter; meanwhile Huawei Ascend ramping in 2026 could further erode market share
• AI capex cycle quality reassessment: If a major hyperscaler publicly questions the pace of AI ROIC realization, the entire sector faces a valuation reset
• Customer concentration shift: Either of the top two customers publicly diversifying their compute supply would immediately pressure NVDA's valuation
• Key-man risk: Jensen Huang's personal influence over NVDA's culture, customer relationships, and technical roadmap far exceeds the typical CEO — any uncertainty around that amplifies valuation volatility

XII. What Are You Really Buying When You Buy NVDA?

Buying NVDA is no longer the "only gateway to AI." At its core, buying NVDA today is a bet on three things:

1. The total AI compute market keeps expanding ($600B → $1T+), so even with share erosion, NVDA's absolute revenue can keep doubling
2. The CUDA moat + ecosystem lock-in holds for frontier training for at least another 3–5 years, with hyperscaler ASICs primarily diverting inference workloads
3. The Rubin → Rubin Ultra → next-gen roadmap executes without major missteps, resetting the AI infrastructure standard with each generation

If you believe all three → $236 is still a reasonable entry (or even cheap); if you're wavering on the second (ecosystem moat) → start trimming exposure; if you're wavering on the first (market expansion) → revisit the entire AI theme.

The 5/20 earnings report isn't a single event — it's a window where multiple signals converge: Blackwell Ultra ramp cadence, China re-entry trajectory, customer concentration shifts, Rubin production timing, and capex guidance from the top two customers. Each signal will be amplified by the market.

The highway determines the outcome.

Additional notes on data sources

• Share price & valuation: $236 / Forward P/E 26–29x as of 2026-05-14 after U.S. market close / 05:15 UTC; subject to real-time market fluctuation
• "Top two customers = MSFT/Meta" is a market inference based on capex scale; NVIDIA does not disclose customer names
• "Jensen's 40% non-cloud customer comment": management has repeatedly highlighted enterprise, sovereign AI, industrial, and neocloud demand outside the hyperscaler base in response to market concerns about customer concentration
• All "OurAlpha perspective / opinion / scenario estimate" sections represent editorial judgment and do not constitute investment advice

This content is for informational purposes only and does not constitute investment advice, trading advice, or any guarantee of returns.