NVIDIA Q2 Earning Results-Deep Analysis

Revenue $96.2bn, +106% YoY. First-ever annual guide: ~70% growth in FY2028. And a gross margin that management now says bottoms at 71–72%.

---

Sec. 01 — Key Takeaways

Ten things that mattered this quarter

01 · The quarter beat, but the quarter was not the story. Revenue of $96,221m came in ~4.5% above the ~$92.1bn consensus; non-GAAP EPS of $2.22 beat the ~$2.09–2.10 consensus by ~6%. Data Center revenue of $89.0bn grew 117% year over year. All of this was inside the range of outcomes the market had already modelled.

02 · NVIDIA guided a full year for the first time ever. Kress: “We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook.” Street consensus sat near 44%. On our arithmetic that reframes FY2028 from roughly $570bn to roughly $690bn. Huang’s own gloss — “It is the case that we’ve never forecasted, never guided to a year in advance” — tells you management understood exactly what it was doing.

03 · The 70% is a floor, not a forecast. Kress said customer forecasts “point to our growth doubling next year.” Huang, asked directly what unconstrained demand looks like, answered: “The unconstrained would be a lot higher.” NVIDIA is guiding to what it can build, not what it can sell.

04 · Gross margin is going down, and management said so before being asked. The disclosed path is 75.0% actual in Q2 → 74.0% ±50bp guided for Q3 → 71–72% in Q4 → 72–73% for FY2028. That is a ~350bp trough and a partial, not full, recovery. Management volunteered the trajectory in prepared remarks rather than waiting for the Q&A — a deliberate attempt to control a difficult narrative.

05 · The cause is memory, and it is getting worse. Kress: “We are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year.” Her framing is notably self-implicating — memory scarcity is “being driven in large part by the AI build-out itself.”

06 · The fix is price increases NVIDIA has already put through. The FY2028 recovery to 72–73% is attributed to executed price increases landing in fiscal Q1. This is the mechanism the entire margin story now rests on, and it is the most attackable assumption in the guide.

07 · Vera Rubin is ramping faster than anything NVIDIA has shipped. Production began in August across CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. It is expected to be ~20% of Q3 Data Center revenue. The economic claim that matters: revenue opportunity per gigawatt of deployed capacity rises from ~$18bn (Hopper) to $25bn (Grace Blackwell) to $40bn (Vera Rubin).

08 · The customer base has genuinely broadened. Hyperscalers contributed ~$49bn (+13% QoQ). The bucket NVIDIA calls ACIE — AI-native clouds, enterprise and sovereign — contributed ~$40bn, up 25% sequentially and 138% year over year, and is expected to reach roughly half of Data Center revenue. This is the most important structural change in the model and the strongest single rebuttal to concentration bears.

09 · Networking had a record quarter and is now a real business. Networking revenue grew 18% sequentially to a record; Spectrum-X Ethernet grew 2.6x year over year. NVIDIA declined, again, to disclose a networking revenue figure in dollars — a disclosure gap we think is now conspicuous.

10 · China contributed almost nothing and is guided to zero for the third consecutive quarter. Kress: “In Q2, we shipped less than 1% of our total data center revenue in Hopper 200 products to customers based in China… there is no China data center compute revenue in our forward outlook.” This stance held even a week after Beijing approved limited H200 shipments on 19 August.

---

Sec. 02 — Headline Financials

The reported quarter

Source: NVIDIA Q2 FY2027 press release, 26 August 2026. Note that from Q1 FY2027 NVIDIA no longer excludes stock-based compensation from its non-GAAP measures; prior periods have been restated on the same basis.

Segments

Six months, and the run-rate

Capital returns and the balance sheet

• $26.0bn returned in the quarter on NVIDIA’s own measure. The cash flow statement shows $19,732m of buybacks and $6,047m of dividends actually paid, or $25.8bn, against $21.3bn of free cash flow — roughly 121% of free cash flow, funded in part by debt.
• $99.0bn remains under the repurchase authorisation.
• Next dividend of $0.25 per share, payable 1 October 2026 to holders of record on 10 September 2026.
• Long-term debt rose to $32,366m from $7,469m at the January year-end, reflecting the ~$25bn investment-grade bond offering NVIDIA launched on 15 June — its first since 2021 and the largest debt raise in its history.
• Net cash (cash, marketable debt and marketable equity securities less total debt) stands at roughly $66bn, up from approximately $54bn at the January year-end — the bond proceeds and the appreciation in the equity portfolio together more than covered the buyback.

---

Sec. 03 — What the cash flow statement says that the income statement does not

This is the section most sell-side previews will skip, and it is where the quarter is genuinely interesting.

Free cash flow conversion fell to 36%

Operating cash flow of $24,077m was lower than the $50,344m generated in the April quarter despite revenue growing 18% sequentially. The gap is working capital, and it is large:

• Accounts receivable rose $22,346m in a single quarter — more than the entire sequential revenue increase of $14,606m. Receivables now stand at $63,059m against $38,466m at the January year-end.
• Inventories rose $5,784m to $31,575m.
• Prepaid expenses and other assets consumed $5,497m.

Days sales outstanding rose to approximately 60 days, from roughly 51 days at the January year-end. Inventory days sit near 119, against approximately 114 on our estimate of the January-quarter cost base.

How to read it

There are two honest interpretations and we think both are partly true.

The benign reading. Vera Rubin shipments began in August — the final weeks of a quarter that ended 26 July, with the ramp weighted heavily to the back end. Rack-scale systems carry long invoicing and acceptance cycles compared with board-level product. A back-end-loaded quarter of very large, very expensive systems mechanically inflates receivables. Inventory build ahead of a ramp NVIDIA calls the fastest in its history is exactly what you would want to see.

The reading that deserves monitoring. NVIDIA is now a significant provider of credit support, equity capital and lease guarantees to the customers buying its product. The balance sheet shows marketable equity securities of $42,783m and non-marketable securities of $51,157m — $93.9bn combined, up from $35.1bn six months ago. When a supplier’s receivables, its equity stakes in its customers, and its off-balance-sheet guarantees all expand together, receivable quality stops being a purely mechanical question.

We are not calling this a red flag this quarter. We are saying it is the metric to watch in Q3, because a second consecutive quarter of DSO expansion alongside decelerating cash conversion would change the character of the story.

Earnings quality: GAAP EPS is flattered by mark-to-market gains

GAAP EPS of $2.46 exceeded non-GAAP EPS of $2.22 — an unusual inversion. The driver is $7,771m of net gains from equity securities, which represented 10.9% of pre-tax income in the quarter and 16.8% across the first half ($23,707m of $141,410m).

These are real gains on real investments. They are also non-operating, non-recurring, non-cash and reversible. Any model that runs GAAP EPS forward without stripping them will overstate earnings power. NVIDIA’s own non-GAAP presentation correctly excludes them — which is why, this quarter, the non-GAAP number is the conservative one.

---

Sec. 04 — Guidance: the number NVIDIA had never given before

Q3 FY2027

At the midpoint that is +12.2% sequentially and +89% year over year against the $57.0bn NVIDIA reported in Q3 FY2026 — and roughly $4bn above the ~$104bn the street was carrying.

The FY2028 guide

The disclosure that reset the stock:

“We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook.” — Colette Kress

Our arithmetic. First-half revenue of $177.8bn plus the $108.0bn Q3 guide gives $285.8bn through nine months. A fourth quarter in the $115–125bn range — consistent with the sequential cadence implied by the ramp — puts FY2027 at roughly $400–411bn. Applying ~70% gives an FY2028 revenue range of approximately $681–698bn.

For scale: that would mean NVIDIA adding roughly $285bn of annual revenue in a single year, which is more than the company’s entire FY2026 revenue of $215.9bn.

Why guide a year at all?

Three reasons, in our view, and only the first is the one management gave.

1. Supply visibility genuinely improved. Kress cited cloud industry backlog “now greater than $2 trillion” and top-five hyperscaler capex reaching “nearly $800 billion in 2026 and $1.3 trillion in 2027.” With multi-year supply agreements across HBM, foundry and optics now signed, NVIDIA can see its own capacity further out than it once could.
2. It reframes the margin disclosure. A 350bp margin decline announced alone is a bad headline. Announced alongside a 70% growth guide, it becomes a rounding error in the gross profit dollar bridge. The sequencing was not accidental.
3. It defends against the demand-durability bear case pre-emptively. The single most persistent short thesis on NVIDIA is that AI capex is a 2024–2027 bubble. An annual guide is the most credible instrument management has to push the debate out a year.

---

Sec. 05 — The margin problem

The disclosed path

What management said

“We are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year.”

“Memory scarcity today is being driven in large part by the AI build-out itself… we are working closely with [memory suppliers] to further increase the capacity.”

Our read

This is a structurally lower plateau, not a V. The recovery to 72–73% does not restore the 75% level. Management is signalling that a portion of the memory cost increase is permanent within the current architecture generation, and that NVIDIA absorbs some of it.

The recovery mechanism is price, and price is the weak link. The FY2028 improvement is explicitly attributed to price increases NVIDIA has already executed, landing in fiscal Q1. This is a bet that customers will absorb higher prices without reducing units. In an environment where hyperscalers are simultaneously funding internal silicon programmes specifically to reduce NVIDIA dependence, and where enterprise AI ROI remains contested, that bet is not free.

The sharpest external pushback we found came from Bill Birmingham of Rex Financial, who observed that the credit-default swap market repriced NVIDIA’s five-year risk from 40bp to 82bp when reports of a large OpenAI capacity guarantee surfaced, and put the objection squarely: “It’s dangerous to raise prices when ROI for AI at the customer level is still unknown.”

The mitigations are real but slow. NVIDIA announced a multiyear memory technology partnership with SK hynix on 7 June, expanded into a broader SK Group agreement on 24 July covering long-term AI memory supply and next-generation HBM co-development. Samsung, SK hynix and Micron were all reported certified for Vera Rubin HBM4 supply in June. On 26 August NVIDIA disclosed NVHBM — moving its own memory controller into the HBM base die for up to 30% greater bandwidth, 15% lower HBM power and up to 25% more usable compute-die area, with Amazon’s Annapurna Labs as first collaborator from Trainium4. None of these change the FY2027 cost curve. All of them are aimed at FY2029 and beyond.

What was not disclosed. No dollar or basis-point quantification of the memory headwind. No contracted volumes, pricing mechanics or HBM supply-agreement terms. No breakdown of how much of the 350bp trough is memory versus Rubin ramp inefficiency versus mix. This is a meaningful transparency gap on the quarter’s most important variable.

---

Sec. 06 — The Q&A, exchange by exchange

Eight analysts asked questions. Below is what each asked, what management actually answered, and — where relevant — what they declined to answer. Toshiya Hari, VP of Investor Relations, hosted.

Theme A — Demand durability and the shape of the guide

1 · Joseph Moore, Morgan Stanley — asked how NVIDIA gets confidence sufficient to guide a full year, and what constrains the number versus what demand implies.

Huang anchored the answer in agentic workload intensity: “The amount of compute necessary for an agent versus a human using it is probably 15 to 100 times.” He then pivoted to differentiation — NVIDIA sells “an entire AI factory platform, a full stack system,” not a chip. Substance: the guide is set to achievable capacity; demand materially exceeds it.

2 · Stacy Rasgon, Bernstein — the sharpest question on the call. He asked what actually drives the 70%, what unconstrained demand would be, and probed pricing as a contributor: “I assume some of this is pricing…”

Answer: “The unconstrained would be a lot higher.” Hyperscalers are approximately half the business; ACIE is growing near 100% annually; revenue per gigawatt scales from $25bn to $40bn on architecture plus ecosystem breadth. Note what happened here: Rasgon opened the door to a pricing-versus-volume decomposition of the 70% guide, and management did not walk through it. Given that price increases are the stated mechanism for the FY2028 margin recovery, how much of 70% is price and how much is units is a first-order modelling question that remains unanswered.

3 · Ben Reitzes, Melius — asked about recursive self-improvement and AGI-driven compute demand.

Huang: agents proliferate without human gating — “400,000 agents, 4 million agents” — and “when the world goes to agentic, fully agentic systems, you are going to have agents running all the time.” Substance: the demand thesis has shifted from training scale to always-on inference population.

Theme B — Competition and the ecosystem

4 · C.J. Muse, Cantor Fitzgerald — inference and agentic workload share, including the Groq LPX product and ACIE expansion.

Management framed the AI lifecycle as data preparation → pre-training → post-training → agentic inference, served by “one fungible system” that moves across phases. Vera Rubin’s claimed 30x throughput per megawatt and 35x lower token cost versus Grace Blackwell Ultra were deployed as the competitive answer.

5 · Vivek Arya, BofA — asked whether NVIDIA’s $500bn-plus ecosystem investments conflict with frontier labs building their own silicon.

Huang: “We’re building something very different” — a full AI factory platform available across every cloud globally, versus competitors building “inference-specific chips for one cloud.”

This was the closest any analyst came to a circular-financing question, and Huang answered it as a competitive-positioning question rather than a financial-risk question. We score this a partial dodge. The financial question — what is NVIDIA’s aggregate at-risk exposure across equity stakes, warrants, lease guarantees and credit support, and how is it reserved — was neither asked in this form nor answered.

6 · Timothy Arcuri, UBS — whether open-source model growth threatens the frontier-lab customer base.

Huang: “Nearly all open models run on NVIDIA.” Both closed and open model usage is “skyrocketing.” Open models expand the addressable market into enterprises and startups rather than cannibalising it. This is a strong answer and, in our view, correct.

Theme C — Supply

7 · James Schneider, Goldman Sachs — asked management to rank the binding constraints: data centre shell, power, DRAM, or foundry wafers.

Answer: “Our entire supply chain is challenged.” “Everybody is really running flat out.” NVIDIA has “supply for 70%” while demand is materially higher.

This was a non-answer to a well-constructed question. Which constraint binds first determines whether the 70% guide has upside (if it is a shell-and-power problem, it eases as campuses come online) or downside (if it is HBM4 wafer allocation, it does not). Management declined to rank. Investors should treat the composition of the supply constraint as unresolved.

8 · Aaron Rakers, Wells Fargo — whether revenue per gigawatt keeps scaling beyond $40bn.

Huang: “It is very simple” — maximise compute density per site. “The perfect answer is actually infinity per gigawatt.” Trajectory given as $3–5bn historically → $40bn now → higher in future generations.

What nobody asked

Three absences are notable on a call of this consequence:

• No clean gross margin bridge question. With margin guided down 350bp, not one analyst asked management to decompose the decline into memory, mix and ramp costs.
• No question on depreciation schedules or customer useful-life assumptions — a live debate across the hyperscaler complex and directly relevant to whether the 2027 capex figures NVIDIA cites are sustainable.
• China went unasked entirely. It was addressed only in prepared remarks, a week after Beijing approved limited H200 shipments. Nobody probed the upside case.

Management closed by noting Huang will appear at the Goldman Sachs Communacopia & Technology Conference in San Francisco on 10 September — the next scheduled catalyst.

---

Sec. 07 — Six months of news flow: how NVIDIA got here

The quarter did not appear from nowhere. Reading NVIDIA’s newsroom from February through August 2026 shows a company executing four parallel campaigns: locking supply, converting nations into customers, industrialising the financing layer, and pushing compute out of the data centre and into the network.

The timeline that mattered

---

Sec. 08 — Hurdles

We think six risks matter, ranked by how underpriced we believe they are.

1 · Memory cost, and whether price increases stick (most underpriced)

Already covered in Sec. 05. The specific vulnerability is that NVIDIA’s margin recovery requires customers to absorb price increases at exactly the moment those customers are being asked to absorb the largest capex step-up in their histories. If even one large buyer pushes back publicly, the FY2028 margin guide comes under immediate pressure.

2 · Composition of the supply constraint

Management explicitly declined to rank shell, power, DRAM and foundry. The 70% guide is a capacity number. Without knowing which input binds, investors cannot judge whether that number carries upside or downside. Power and grid interconnection — NVIDIA itself cites >200GW in US interconnection queues — is the constraint least responsive to capital.

3 · The financing architecture

NVIDIA is now simultaneously a supplier, an equity investor, a warrant holder, a lease guarantor and a credit-support provider to its own customer base. The visible footprint: $93.9bn of equity and non-marketable securities on balance sheet, up from $35.1bn in six months; roughly $50bn invested in frontier labs; a reported $105bn payment-obligation cap on the OpenAI/SB Energy structure; and a $500bn third-party capital platform still at MOU stage.

Kress addressed it head-on — “we know some will call this circular financing. We see it differently” — arguing that frontier labs are constrained by compute access rather than end demand, that the platform is “fungible and durable and can be redeployed” if a partner fails, and that neocloud partners’ capital is independently underwritten. These are reasonable arguments. They are not the same as disclosure. NVIDIA has not published an aggregate at-risk exposure figure, a reserve methodology, or a sensitivity analysis. Until it does, this risk is unquantifiable from outside — and unquantifiable risks tend to get repriced abruptly rather than gradually.

4 · Customer concentration, improving but real

ACIE at 138% year-over-year growth is the genuine mitigant, and it is working: hyperscalers are now roughly 55% of the Data Center business, against close to 60% a year ago. But the ACIE bucket contains neoclouds and frontier labs that are themselves financed by NVIDIA or by capital NVIDIA helped arrange. Diversification of counterparty name is not the same as diversification of underlying credit.

5 · China: capped upside, non-zero downside

Three consecutive quarters of guiding China Data Center compute to zero means any approval is pure upside to the model. But the 19 August approvals were narrow — roughly 10,000 H200s each to ByteDance and Tencent, most chips required to stay outside the mainland, per-order NDRC approval, and H200 already two generations behind the leading edge. NVIDIA reportedly holds ~500,000 H200s earmarked for Chinese customers. The realistic outcome is a modest, slow, politically reversible revenue stream. The downside risk — retaliatory action, or a domestic-preference mandate — is smaller in revenue terms precisely because guidance already assumes zero.

6 · Competition, mostly on a lag

AMD and Intel are targeting inference; hyperscalers continue funding internal silicon; Chinese vendors are building proprietary parts. None of this is visible in Q2 numbers, and Huang’s fungibility argument is genuinely strong for training and mixed workloads. The exposure is narrower than the bear case suggests but real at the margin: high-volume, latency-sensitive, single-model inference is the workload class where a purpose-built accelerator on one cloud can win on total cost. NVIDIA’s own answer to this is Groq 3 LPX — which is an implicit acknowledgement that the threat is real enough to need a dedicated product.

---

Sec. 09 — What is in this earnings report for telecom

This is the section most general-market coverage will not write, and there is more in the quarter for operators and network vendors than the absence of the word “telecom” from the press release suggests.

9.1 · There is no telecom line item — and that is the first finding

NVIDIA does not disclose telecom revenue. It does not disclose AI-RAN revenue. It does not disclose networking revenue in dollars, only growth rates. Telecom-related revenue sits inside two disclosed buckets:

• ACIE ($40.3bn, +138% YoY) — where sovereign AI factories built by operators land. SK Telecom’s 2GW DSX facility, NAVER’s GAK Sejong build and Indosat’s Indonesian sovereign infrastructure all report here.
• Networking (record quarter, +18% QoQ, Spectrum-X +2.6x YoY) — data centre fabric, not RAN.

Practical implication for anyone modelling telecom exposure to NVIDIA: the only telecom-attributable revenue currently in NVIDIA’s P&L at scale is operators buying AI factories, not operators buying AI-RAN. Those are different businesses with different buyers, different budgets and different timelines.

9.2 · The operator is being repositioned from customer to infrastructure provider

The single clearest telecom signal in the six-month news flow is Huang’s June formulation at the SK Telecom announcement: “Telecom networks are becoming national AI infrastructure.”

The pattern repeats across geographies:

Why this matters strategically. An operator that builds a 2GW AI factory is not buying network equipment — it is buying into a compute business with a different cost of capital, a different customer set and a different margin structure than connectivity. For incumbent RAN vendors, this is simultaneously an opportunity (operators with capital deploying at scale) and a threat (operator capex reallocated from radio to compute). NextGComm readers tracking Nokia and Ericsson should note that the largest new capex line at several Asian operators is now GPUs, not radios.

9.3 · AI-RAN: technically proven, commercially not yet in the numbers

The technical progress across the six months is substantial and worth cataloguing:

• T-Mobile US ran an over-the-air field trial of concurrent AI and RAN processing using Nokia’s CUDA-accelerated software on AirScale massive MIMO at 3.7GHz — carrying video streaming, generative AI applications and AI video captioning alongside live 5G traffic.
• SoftBank demonstrated an industry-first 16-layer massive MIMO on fully software-defined 5G via AITRAS, plus repurposing spare AI-RAN compute for third-party AI workloads.
• Indosat completed Southeast Asia’s first AI-powered 5G call using Nokia vRAN on NVIDIA AI-RAN.
• SynaXG ran fully software-defined AI-RAN on a single GH200 server — the world’s first AI-RAN on FR2 mmWave, at 36Gbps throughput and sub-10ms latency across 20 component carriers.
• DeepSig demonstrated an AI-native air interface delivering up to ~2x higher throughput.
• 26 of the 33 AI-RAN Alliance demos at MWC 2026 were built on NVIDIA AI Aerial. NVIDIA open-sourced the Aerial CUDA-accelerated RAN libraries on GitHub.
• The AI-RAN Alliance now has 130+ members, and a 6G coalition of BT, Deutsche Telekom, SK Telecom, SoftBank and T-Mobile signed up to open, AI-native platforms.

9.4 · The edge inference architecture is now specified

The 16 March T-Mobile announcement is, in our view, the most operationally concrete telecom item of the half-year because it names the hardware tier by tier:

• RTX PRO 6000 Blackwell Server Edition in mobile switching offices — the aggregation tier
• ARC-Pro (RTX PRO 4500 Blackwell) at power-constrained cell sites — the far edge
• Running over T-Mobile’s 5G Standalone and 5G-Advanced network
• Workloads: Metropolis VSS v3 for video search and summarisation, with partners Fogsphere, LinkerVision, Levatas, Vaidio and Siemens Energy; City of San Jose among first assessors

This is a repeatable reference architecture, and it answers the question operators have been asking since 2019 about what actually runs at the edge. The answer is not consumer AR. It is municipal and industrial computer vision — utility line inspection, facility management, industrial safety, smart-city operations. Those are budgeted, unglamorous, and buyable today.

The commercial caveat is equally clear: no deployment timeline and no commercial terms were disclosed. These remain pilots.

9.5 · The “AI grid” is NVIDIA’s telecom TAM, and the number is large

From the 17 March AI Grids announcement, NVIDIA’s own framing of distributed network infrastructure:

• ~100,000 distributed network data centres worldwide
• offering more than 100GW of new AI capacity over time
• Spectrum: 1,000+ edge data centres, hundreds of megawatts, reaching 500 million devices at under 10ms
• Akamai: 4,400+ edge locations globally
• Participating operators: AT&T (with Cisco, IoT-focused), T-Mobile, Comcast, Spectrum, Indosat

Apply NVIDIA’s own revenue-per-gigawatt arithmetic to this and the number is arresting. At the Vera Rubin figure of $40bn per gigawatt, 100GW of distributed capacity implies a theoretical $4 trillion of equipment opportunity. That figure should not be taken literally — edge sites will not be populated with NVL72 racks, densities and configurations differ by orders of magnitude, and the 100GW is explicitly “over time.” But it explains precisely why NVIDIA is investing this much executive attention in an industry that contributes no disclosed revenue today.

9.6 · The transport and optics layer is where telecom money is being spent now

While AI-RAN sits in trials, NVIDIA has been spending heavily on the optical and transport supply chain — and this is where telecom-adjacent vendors are seeing real order flow:

• Spectrum-6 (21 July): 102.4 Tb/s, double the prior generation, supporting both pluggable and co-packaged optics, sustaining up to 95% network efficiency across deployments exceeding 100,000 GPUs. First adopters include CoreWeave, Microsoft, Nebius, OCI, SpaceXAI and Tesla.
• $2bn equity into Coherent and $2bn into Lumentum (2 March), each with multibillion-dollar purchase commitments and future capacity access rights, funding US-based laser and optical component fabs.
• Corning (6 May): 10x increase in US optical connectivity manufacturing capacity, >50% increase in US fibre production, three new plants in North Carolina and Texas.
• Marvell $2bn (31 March) via NVLink Fusion, with Octeon baseband processors integrating with NVIDIA GPUs on a unified software platform — explicitly advancing Aerial AI-RAN.

The read-through for telecom supply chains is direct. Optical component capacity that AI data centres are consuming is the same capacity that serves telecom transport. Lead times and pricing for coherent optics, lasers and fibre are being set by AI demand, not by operator demand. Operators planning 2027–2028 transport upgrades should assume tighter supply and firmer pricing than historical cycles would suggest.

9.7 · The memory warning is a telecom BOM warning

This is the read-through we think telecom planners are most likely to miss.

Kress said memory prices have exceeded expectations and are “headed even higher into next year.” NVIDIA — the largest and most sophisticated memory buyer in the industry, with multiyear agreements at SK hynix, Samsung and Micron — cannot fully insulate its own margin.

Radio units, baseband units, transport equipment, routers, optical line systems and customer premises equipment all contain DRAM. Telecom vendors buy that memory in far smaller volumes and with far less leverage than NVIDIA does.

The implication: 2027 BOM cost inflation for network equipment is likely to be materially worse than the current guidance from RAN and transport vendors implies. Operators negotiating 2027–2028 frame agreements should expect vendor price increases or margin warnings. Investors modelling Nokia, Ericsson, Ciena, Juniper or Casa should stress-test memory cost assumptions against NVIDIA’s disclosure rather than against historical DRAM cycles. NVIDIA has just told the market that the AI build-out is itself the cause of the scarcity — which means it does not resolve while the build-out continues.

9.8 · Power is the constraint operators are best positioned to arbitrage

NVIDIA’s DSX materials cite >$300bn in equipment backlogs and >200GW in US interconnection queues, and claim power-flexible AI factories could unlock up to 100GW across the US power system. NVIDIA has invested $1.5bn in SB Energy and guaranteed an 8 IT-GW Ohio campus to secure land, power and shell.

Telecom operators hold something structurally scarce here: tens of thousands of powered, cooled, fibre-connected, permitted sites already inside the grid interconnection perimeter. Central offices, mobile switching offices, headends and aggregation hubs were built for equipment that no longer needs the space. The strategic question for every operator is whether that estate is a stranded asset to be sold or an AI-adjacent asset to be monetised — and the SK Telecom and NAVER announcements suggest that at least in Asia, the answer has already been decided.

---

Sec. 10 — NextGComm inference: what we are watching

Our read of where this leaves the next-generation communications sector.

1 · The AI-RAN investment case survives this print intact, but its clock is 2028, not 2027. Nothing in NVIDIA’s numbers validates or invalidates AI-RAN, because AI-RAN is not in the numbers. The Nokia commercial-release timeline of 2027 is the date that matters, not NVIDIA’s fiscal 2028 guide. We expect continued technical milestones and coalition growth through 2027 with negligible revenue attached, and we would treat any vendor forecast implying material AI-RAN revenue before calendar 2028 as aggressive.

2 · The operator-as-AI-infrastructure-provider thesis is now the highest-conviction telecom trade in the complex. SK Telecom committing 2GW, NAVER attracting up to $9bn from Brookfield, and Japan building national infrastructure with METI are not pilots — they are capital allocation decisions of a size that reshapes an operator’s balance sheet. We expect at least one European or Middle Eastern operator to announce a comparable gigawatt-scale commitment before MWC 2027. The financing template — NVIDIA equity plus an infrastructure fund plus the operator — is now demonstrated and repeatable, and the 10 August $500bn platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR exists precisely to scale it.

3 · Memory is the sector-wide 2027 risk that is not yet in consensus telecom models. This is our strongest non-consensus call from the quarter. NVIDIA’s disclosure is the most credible forward signal on DRAM pricing available to the industry, and it points to worsening conditions into 2027. We would expect network equipment gross margin guidance across the vendor cohort to come under pressure in H1 2027, and we think the risk is currently underpriced because telecom analysts are not reading semiconductor calls.

4 · Optical and transport is where AI capex touches telecom today. Coherent, Lumentum, Corning, Marvell and the Spectrum-6 co-packaged optics transition are the live commercial vectors, not the RAN. Component supply tightness and pricing power sit with the optical vendors, and telecom buyers are now the marginal — not the primary — customer in that market. Plan procurement accordingly.

5 · The concentration risk telecom should actually worry about is not NVIDIA’s — it is the industry’s. A single vendor now sets the roadmap for AI compute, the reference architecture for edge inference, the standards agenda for 6G through the AI-RAN Alliance and OCUDU, and the financing structure through which operators fund AI infrastructure. The open-source releases — Aerial CUDA libraries on GitHub, Nemotron 3 Large Telco Model at 30B parameters, Cosmos 3, Alpamayo 2 — are genuine and useful, and they are also the most effective possible mechanism for making an ecosystem dependent on one company’s silicon. Operators should press hard on multi-vendor GPU abstraction in AI-RAN procurement now, while they still have negotiating leverage.

6 · Watch three specific data points in the November print. Whether DSO retreats from ~60 days (working capital normalisation, or not); whether Q4 gross margin lands at or below the 71–72% guide (price increases holding, or not); and whether ACIE crosses 50% of Data Center revenue (diversification thesis confirmed, or stalling).

---

Sec. 11 — Key dates

---

Methodology and sources

How this analysis was produced. All financial figures are taken directly from NVIDIA’s Q2 fiscal 2027 press release and the accompanying condensed consolidated financial statements filed with the SEC on 26 August 2026. Growth rates, margin calculations, days-sales-outstanding, inventory days, free-cash-flow conversion ratios and the FY2028 revenue range were computed independently by NextGComm Insights from those primary statements; where our derived figures appear, they are labelled as such. Earnings call quotations were verified against two independent transcript sources before use. News-flow items were sourced from NVIDIA’s own newsroom and investor relations releases between 1 February and 27 August 2026, and cross-checked against partner announcements and primary regulatory sources where available. Figures we could not verify to a primary source — including the reported value of the Groq transaction and NVIDIA’s aggregate at-risk financing exposure — are explicitly flagged as unverified in the text.

Primary sources

• NVIDIA, Q2 FY2027 financial results press release, 26 August 2026
• NVIDIA, Q2 FY2027 press release as filed with the SEC
• NVIDIA Q2 FY2027 earnings call, 26 August 2026 — transcripts via Investing.com and AlphaStreet
• NVIDIA, Q4 and FY2026 results, 25 February 2026; Q1 FY2027 results, 20 May 2026

NVIDIA announcements referenced

• 6G coalition with global telecom leaders · Software-defined AI-RAN · T-Mobile physical AI on AI-RAN-ready infrastructure · Telecom AI grids · Telecom AI agents, DTW Ignite · Autonomous networks and telco reasoning models
• Vera Rubin platform, GTC 2026 · Vera Rubin full production · Spectrum-6 · Groq 3 LPX full production · NVLink Fusion with NVHBM
• Coherent partnership · Lumentum partnership · Corning US manufacturing · Marvell and NVLink Fusion
• SK Telecom AI infrastructure · SK Group and SK hynix expansion · NAVER, NVIDIA and Brookfield · Japan national AI infrastructure · Europe’s 35 AI supercomputers
• $500bn financing platforms · SB Energy PORTS-Pike guarantee · AWS 2 million GPUs · IREN partnership · Nebius partnership · Meta partnership

Third-party and regulatory sources

• US Bureau of Industry and Security, revised license review policy for semiconductors exported to China, 13 January 2026
• Business Standard, on Beijing’s limited H200 approvals, 19 August 2026
• Fortune, on the revised OpenAI data centre commitment, 18 August 2026 · Fortune, on the FY2028 guide
• Groq, non-exclusive inference technology licensing agreement with NVIDIA, December 2025
• Mobile World Live, Nokia AI-RAN momentum at MWC 2026 · Light Reading, on a possible 6G radio unit chip
• Axios · Benzinga on Kress and FY2028 · Reuters via Investing.com

---