Earnings Highlights: NVIDIA (NVDA) | Q2 FY2027 — Quarterly Revenue Tops $96.2B, FY2028 Guided +70%, and the Price of a GW Jumps from $25B to $40B

2 days ago

9 min read

NVIDIA delivered $96.221 billion in revenue this quarter, up 106% YoY, its fourth straight quarter of accelerating growth. But what's worth remembering isn't that number — it's three things. First, Vera Rubin lifts the revenue NVIDIA captures per 1 GW of data center from $25 billion with Grace Blackwell to $40 billion, and almost all of the increment is CPUs and networking. Second, non-hyperscale customers (ACIE) already account for 45% of data center revenue, up 138% YoY — this is the real order pool for Taiwan's optical communications vendors. Third, the company for the first time gave guidance a full year out — FY2028 revenue up about 70% YoY, explicitly described as supply-constrained, with demand actually close to doubling. The cost is gross margin: memory price hikes push gross margin from 75% down to 71–72% in Q4, recovering to 72–73% only after price increases take effect in Q1.

1. Three Core Takeaways

Takeaway 1: this quarter's growth engine isn't hyperscalers — it's NeoClouds and sovereign AI. Of the $89.021 billion in data center revenue, Hyperscale was about $49 billion, up 13% QoQ; ACIE (NeoCloud, enterprise, industrial, sovereign) was about $40 billion, up 25% QoQ and 138% YoY. Management said outright that non-hyperscale will eventually be half of data center. For the optical supply chain, that means order sources are spreading from four or five mega-customers to a whole row of regional clouds that must buy “turnkey” facilities.

Takeaway 2: what NVIDIA sells is getting thicker — price per GW is the main thread of this call. Hopper-generation was about $18 billion per GW, Grace Blackwell $25 billion, Vera Rubin $40 billion. The extra $15 billion isn't higher GPU prices; it's the Vera CPU, NVLink, InfiniBand or Ethernet, plus the Groq LPU announced just this week, all packed into the same rack. NVIDIA is shifting from “selling accelerators” to “selling entire AI factories”; we saw the full technical version of this in 2026 OCP APAC Summit | NVIDIA | Gilad Shainer | Scaling the AI Factory, and this quarter is its financial version.

Takeaway 3: the bottleneck has moved — from advanced packaging to memory. For two years everyone watched CoWoS; this quarter the CFO spent a whole passage on how “memory price increases exceeded our earlier expectations, and will be higher still next year.” Gross margin guidance was reset as a result. It was the most candid passage of the quarter, and the one most worth tracking.

2. Revenue and Financials

On a GAAP basis (figures from the SEC 8-K earnings press release):

Metric	Q2 FY2027	YoY	QoQ
Revenue	$96.221B	+106%	+18%
Gross margin	75.0%	+2.6pp	+0.1pp
Operating income	$63.734B (operating margin 66.2%)	+124%	+19%
Net income	$59.688B	+126%	+2%
Diluted EPS	$2.46	+128%	+3%

Non-GAAP: net income $53.954 billion (+118%), EPS $2.22 (+120% YoY, +19% QoQ). GAAP net income exceeds non-GAAP because valuation gains on equity investments are included in GAAP; for the underlying operations, use the non-GAAP $2.22. The same reason explains why GAAP net income rose only 2% QoQ (a high base from valuation gains booked last quarter).

Segment breakdown (three mutually exclusive segments summing to total revenue):

- Data Center – Hyperscale: about $49 billion, up 13% QoQ, with Blackwell continuing to ramp

- Data Center – ACIE: about $40 billion, up 25% QoQ, up 138% YoY

- Edge Computing: $7.2 billion, up 27% YoY and 13% QoQ

Data center totaled $89.021 billion, up 117% YoY and 18% QoQ. Networking also hit a record, up 18% QoQ, with Spectrum-X Ethernet up 2.6x YoY — NVIDIA now calls itself “the world's largest and fastest-growing networking company.”

Two cash-side numbers are worth noting: inventory rose to $32 billion (building for Vera Rubin), and days sales outstanding rose to 60 days (extended payment terms for investment-grade large customers, with shipments spread over several quarters). The company returned $26 billion to shareholders this quarter ($20 billion in buybacks, $6 billion in dividends, $0.25 per share), a 60% payout ratio year to date.

3. Technology and Business Highlights: How $40 Billion per GW Adds Up

Vera Rubin began volume shipments this month and will be about 20% of data center revenue in Q3. The official figures are 30x throughput per MW and 35x lower cost per token versus Grace Blackwell Ultra. All major hyperscalers, AI clouds and system OEMs have placed orders, and the company says it will be the fastest product ramp in its history.

The real focus is the “revenue opportunity per GW” line:

- Hopper generation: about $18 billion/GW (Hopper + InfiniBand)

- Grace Blackwell: about $25 billion/GW

- Vera Rubin: about $40 billion/GW (Vera CPU + Rubin GPU + NVLink + InfiniBand or Ethernet + Groq LPU)

Jensen Huang added a bigger frame in Q&A: the total investment in a 1 GW data center has grown from about $30 billion five years ago to about $60 billion today, and NVIDIA takes $40 billion of it. That is the financial definition of the “AI factory platform.”

The three new revenue sources each mean something different for the supply chain:

(1) Vera CPU becomes a standalone business. Grace CPU revenue already exceeded $5 billion over the past 12 months; Vera is in volume production, shipping first to OCI and SpaceX AI, with AWS joining this quarter. The company sized overall server CPU demand at about $20 billion and said FY2028 CPU revenue will “more than double”. Agentic workloads are pushing CPU:GPU ratios back up — we broke this down in 2026 OCP APAC Summit | From 1:4 to 1:1: AMD's OCP 2026 Talk Wasn't About Open, It Was About the CPU's Comeback in the Agentic Era, and now NVIDIA has confirmed it with guidance.

(2) Networking goes from one kind to five. This quarter Jensen Huang inventoried NVIDIA's networking into five categories: NVLink (scale-up), InfiniBand and Ethernet (scale-out), scale-in security networking, and cross-campus scale-across. That is exactly the map in After Copper Runs Out for AI: Seven Paths for Scale-Up Optical Interconnect, except now every path is written into the same BOM. For optical module and optical engine makers, the answer to “how many kinds of optics does a rack need” is going from 1 to 3 or more.

(3) Groq 3 LPX enters volume production. This is NVIDIA's first rack-scale LPU system, built for ultra-low latency and very high token interactivity; official figures put tokens/sec at nearly 4x the next-best solution. Shipments began this quarter, with Nebius as the first customer. Jensen Huang was blunt: throughput is lower and cost per token is higher, but it can command high-ASP services; most of the world's data centers will still be Vera Rubin NVL72.

As for China: this quarter only shipped Hopper 200 amounting to less than 1% of data center revenue, and that batch was dilutive to gross margin. The guidance includes no China data center revenue at all.

4. Management Outlook: A “Supply-Constrained” Annual Guide

Next quarter (Q3 FY2027):

Revenue $108 billion ± 2% (implying about 89% YoY and about 12% QoQ growth)
GAAP / non-GAAP gross margin 74.0% ± 50bps
GAAP / non-GAAP operating expenses about $9.2 billion / $9.0 billion
Growth comes mainly from ACIE; Hyperscale re-accelerates only in Q4 and FY2028 as Vera Rubin supply expands

FY2028 full-year revenue up about 70% YoY. This is NVIDIA's first full-year guide, and the wording should be read on three levels:

“Our customers' forecasts point to our growth doubling next year” → this is the demand side
“We expect growth of about 70% because we are supply-constrained” → this is the supply-side commitment
“We expect supply to remain a constraint at least through the end of FY2028” → this is the timeline

Jensen Huang explained why they dared give a one-year guide this time: buying AI compute today isn't just buying chips — land, power, shells, cooling and construction all have to be locked two to three years ahead, NVIDIA has been forced deep into both upstream and downstream, and “everyone needs to see the same information to commit resources at the same time.” Translated into supply-chain language: your 2027 capacity plan — customers want it to be the same plan as NVIDIA's.

Gross margin was the quarter's only bad news, and the company chose to lay it all out at once:

Period: Q3 FY2027; gross margin guide: 74.0% ± 50bps
Period: Q4 FY2027; gross margin guide: 71%–72% (trough)
Period: FY2028; gross margin guide: 72%–73% (after price increases take effect from Q1)

The CFO's wording is notable: memory price increases “exceeded our earlier expectations, and will be higher still next year,” while stressing that “the memory tightness is itself caused by the AI build-out — it's not a component that only adds cost without corresponding benefit.” That is the tone of “acknowledging the cost but refusing to call it structural erosion”; the tracking point is whether the Q1 FY2028 price increases can be fully passed through.

5. Supply Chain and Customer Clues: Funding Structure Is This Quarter's Biggest Variable

AWS expanded the partnership with an additional 2 million GPUs, shipping from this quarter through FY2029 Q2, and is also adopting the Vera CPU (partly integrated with Rubin, partly standalone). This is the most concrete single-customer signal of the quarter.

NeoCloud installed capacity is set to grow from about 3 GW at the end of 2025 to 8 GW by the end of this year. This is the single number the optical supply chain should remember most — 5 GW of new build-out in one year, and unlike hyperscalers these customers don't develop their own interconnect; they buy the complete package based on NVIDIA's DSX reference design. Sovereign AI was also strong, up 35% QoQ and more than 3x YoY, with a roster ranging from Firebird in Armenia, Cassava in Africa, GMI Cloud in Taiwan, Yotta and Neysa in India, Firmus in Australia, to YTL AI Cloud in Malaysia.

But what really needs a sober look this quarter is that NVIDIA's financial involvement is becoming institutionalized. Taken together:

Nearly $50 billion invested in frontier AI labs
A financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR able to raise more than $500 billion in third-party capital
Through SB Energy, the Portsmouth campus is locked in, with an initial 4.25 GW dedicated to NVIDIA compute for OpenAI; OpenAI's existing and planned commitments total about 12 GW
Optional credit enhancement of nearly 2 GW for another frontier lab
A revenue-sharing structure for NeoClouds: NVIDIA provides take-or-pay backstops for part of the capacity so banks will lend, and shares rental income above the backstop

The CFO herself acknowledged “some will call this circular financing,” and offered a rebuttal: compute is a fungible, redeployable asset with limited risk. But she also gave a number that must go on the watch list — “AI labs that need our balance sheet support will contribute roughly a quarter of our business next year.” Read alongside $32 billion in inventory and DSO stretched to 60 days, the key to this quarter's earnings quality isn't the income statement — it's the balance sheet. If this “seller arranging financing for the buyer” model shows any crack in 2027, the first to see valuations cut won't be NVIDIA but the entire optical supply chain — we saw the exact same transmission path in Valuations Falling, Shipments Surging: The Week NVL576 Slipped, Taiwanese Vendors Posted Record Revenue.

Another set of macro numbers: the cloud industry's backlog exceeds $2 trillion; the top five hyperscalers' 2026 capex is close to $800 billion, and $1.3 trillion in 2027; global AI venture funding exceeded $400 billion in the first half (versus $265 billion for all of 2025), with about 70% spent on compute.

6. Conclusion

NVIDIA had no surprises this quarter; the surprise was how far ahead it was willing to talk: next year's revenue guide, the gross margin trough at year end, and the funding source for a quarter of its business. Saying it all means it needs the entire supply chain to invest on the same set of assumptions.

For Taiwan's optical communications vendors, this call boils down to one sentence: NVIDIA is making “how much it can sell per GW” its core KPI, and the increment falls almost entirely on CPUs and networking — exactly the turf of optical engines, optical modules, FAUs, lasers and connectors. The ACIE and NeoCloud build-out curve from 3 GW to 8 GW is closer to your shipment schedule than any hyperscaler's capex figure.

Three metrics to track next quarter:

1. Whether Q4 gross margin holds at 71–72%, and whether the Q1 FY2028 price increases can be fully passed through — the dividing line for whether memory inflation becomes structural erosion

2. Vera Rubin's share of data center revenue (Q3 guide about 20%); the ramp speed directly sets the pull-in pace for 1.6T optical modules and NVL rack components

3. The trend in receivable days and inventory, and the actual progress of the quarter of the business that “needs NVIDIA balance sheet support”

Between the supply-constrained 70% and the demand-side 100%, those 30 percentage points are what the whole supply chain will be fighting over next year.

This article is for technology and industry trend analysis only and does not constitute investment advice.