August 26, 2026 11 min read

NVIDIA (NVDA) — Q2 FY2027 Analyst Q&A Breakdown

Source: transcripts/NVDA/Q2_2027.txt — NVIDIA Q2 FY2027 earnings call, August 26, 2026.

Q1. Joseph Moore, Morgan Stanley

The Question: (a) Give color on the confidence behind guiding a full fiscal year out. (b) What gap separates the ~70% fiscal 2028 plan from the roughly 100% demand growth he references — what is the key constraint, and can it close over time? The 100% demand figure is analyst-provided (derived from the company’s doubling forecast); management confirmed demand sits above the plan without adopting the figure.
Management’s Response: Jensen’s most strategic answer of the session. New data: agents need 15–100x the compute of a human user depending on the task; the non-hyperscaler block (sovereign, AI natives, NeoClouds, start-ups, enterprises) is about half the business and growing about 100% a year; procurement now requires securing land, power, and shell two to three years out, so NVIDIA works upstream (memory suppliers, power generators, land/power/shell companies) and downstream. He noted NVIDIA has never guided a year ahead before and is doing it so customers, shareholders, and suppliers work from the same information. The guide itself was restated as supply-capped — demand “much greater than 70%” against supply that “allows us to confidently deliver 70%” — with no new mechanics for closing the gap.
The “Analyst’s Take”: Direct and information-dense on the demand structure; soft on the gap. The new ACIE growth data and procurement-lead-time framing go well beyond the script, but the confidence to guide a year out is justified as a communications choice — “everybody has the same set of information” — rather than a forecasting one, and no single binding constraint is named. He closed on momentum: “we’ve got a huge year coming up next year.”

Q2. CJ Muse, Cantor Fitzgerald

The Question: On inference market share: how do agentic workloads evolve and how does NVIDIA’s share trend over time, factoring in the growing TAM per full-stack generation, expected ACIE growth, and Groq 3 LPX?
Management’s Response: Jensen split the AI life cycle into four phases (data preparation, pre-training, post-training, agentic inference) and argued that one fungible rack-scale system — NVLink 72, now in its third generation — serves all four, which is why customers can invest at scale. New Q&A detail: NVIDIA’s revenue opportunity per gigawatt reaches $40B once CPUs, multiple networking types, and Groq are included; Groq is a high-interactivity bolt-on for latency-sensitive, high-ASP services while the vast majority of data centers stay on Vera Rubin and NVLink 72. His per-gigawatt framing then slipped into a different series — data center investment rising from “$30 billion… to… $60 billion today” — which does not reconcile with the prepared-remarks ladder of NVIDIA revenue opportunity from Hopper to Vera Rubin (see Contradictions).
The “Analyst’s Take”: Direct but structural rather than quantitative — no share number, no inference-mix figure, no timeline. The genuinely new content is the Groq niche and the reaffirmation that the core platform carries most inference; the $60B aside gestures at rising project size but muddies the economics story and is left unreconciled.

Q3. Stacy Rasgon, Bernstein Research

The Question: (a) Walk through the contributors to fiscal 2028 growth — Vera, CPUs, Groq, pricing among them — given the guide implies a roughly $200B step-up versus the prior “$1 trillion over three years” outlook framing (he then says “$200 million more,” a slip); (b) what would the number be if not supply-constrained? The $1T baseline and $200B/$200M figures are analyst-provided; management did not confirm or dispute them.
Management’s Response: Unconstrained “would be a lot higher,” and the company grew 100% YoY this year — the only new number in the answer. Jensen declined a product-level bridge, reframed around two halves — hyperscalers (whose backlog scale he repeated from prepared remarks) and the invisible ACIE half that cannot use custom chips and needs a full factory platform — repeated the per-gigawatt ladder (see Q2), and asserted customer economics are “very profitable.” The pricing question went unaddressed.
The “Analyst’s Take”: Evasive on the bridge that was actually requested: narrative architecture replaced the arithmetic, and the pricing angle the prepared remarks introduced was ignored. The one new datapoint reframes the debate as purely supply-side, but the analyst’s math — including his own $200B/$200M stumble — was left unexamined.

Q4. Vivek Arya, Bank of America Securities

The Question: (a) Is the ~$500B aggregate of ecosystem commitments in the CFO commentary the ballpark for total investments, with more ahead? (b) What is the cash portion in fiscal 2028? (c) How does NVIDIA square heavy investment in frontier labs (OpenAI, Anthropic) that are designing competing custom chips — with OpenAI’s “Jalapeno” claimed to beat Blackwell? All figures and the chip claim are analyst-provided.
Management’s Response: Jensen took (c): XPUs are inference-specific chips for one cloud or service; NVIDIA is a full AI factory platform that runs everywhere, so “I have 100% confidence” the labs remain customers — and the investments are once-in-a-generation opportunities (his only regret is not investing more, sooner; two of the labs will likely go public soon). Colette added on (a)/(b): the supply commitments are essential to raising Vera Rubin this year and next, with the largest portions falling in the first three years, and they underpin revenue confidence. Neither the $500B tally nor the fiscal 2028 cash portion was confirmed or broken out.
The “Analyst’s Take”: Confident and personal rather than analytical: the custom-silicon threat is met with platform assertion, and the analyst’s specific benchmark claim is never mentioned. Colette’s supplement is the closest to new substance, but it answers the money question with role and phasing rather than amount — leaving both the commitment tally and cash timing open.

Q5. Timothy Arcuri, UBS

The Question: Does the rise of open-source models help or hurt NVIDIA, given that much end demand comes from frontier labs and some investors read open models as pressure on those companies?
Management’s Response: The world needs both types of models and both are skyrocketing, with open models nearly all running on NVIDIA hardware thanks to the CUDA footprint (PCs, edge devices like DGX Spark, robots, workstations, on-prem). New detail: frontier-lab economics (profitable tokens, compute-limited growth) apply equally to open models; open models are what let enterprises and countries build proprietary, domain-specific AI — cybersecurity being his marquee example — and NVIDIA claims to be the only platform running every frontier model, closed or open.
The “Analyst’s Take”: Direct and thesis-proof by design — NVIDIA wins on either model path, and Jensen says so plainly. What the answer does not offer is sizing: no open-model revenue mix, no read on pricing or margin mix implications, and no risk scenario. Relaxed and deliberately agnostic.

Q6. Ben Reitzes, Melius Research

The Question: (a) With recursive self-improvement progressing and OpenAI talking about AGI by year-end, does industry demand inflect further? (b) How is NVIDIA treating RSI/AGI as a demand catalyst? The AGI timeline is analyst-provided (attributed to OpenAI).
Management’s Response: “It’s going to inflect further.” New: Jensen claims most AI usage crossed into agentic over the last month; NVIDIA plans from 40,000 employees toward a future of “400,000 agents, 4 million agents” running continuously, with personal agents on DGX-class machines running 24/7. He defines the moment as coarse-grained recursive self-improvement — agents updating their own skill files after every run — such that “for many tasks, we could say that we’ve already achieved AGI,” and calls the milestones “kind of senseless”; what matters is useful work, profitable tokens, and more compute producing more of both.
The “Analyst’s Take”: Full embrace of the premise with an operational reframe: AGI defined down to today’s economics strengthens the demand narrative but is unfalsifiable for tracking purposes. No uplift is quantified, and the deliberately provocative framing (“kind of senseless”) substitutes conviction for evidence.

Q7. Jim Schneider, Goldman Sachs

The Question: Rank the most acute constraints on the plan versus demand — data center/power/shell, DRAM, wafer foundry — which are biggest? The demand figure he references is analyst-provided (the doubling forecast).
Management’s Response: Jensen declined to rank, defusing a pun (“There’s something funny I could say, but I’m going to just not”) and digressing about investors’ dinner plans moving stocks. Substance: the entire supply chain is challenged and running flat out; capacity arrives continuously rather than all at once; yield improvement is a lever; and NVIDIA has given suppliers exactly the information it has given investors (“the exact same song sheet”). On quantification he said supply covers the guided plan while demand runs “much higher” — but the phrasing garbles it, claiming both more supply than the plan and about the plan in the same breath (see Contradictions).
The “Analyst’s Take”: An outright non-answer to a direct ranking request, softened with humor. The process assurances are familiar refrains rather than new data, and the lone falsifiable statement about supply-to-plan match arrived garbled and was not corrected. Evasive.

Q8. Aaron Rakers, Wells Fargo

The Question: (a) Beyond Vera Rubin (including Vera Rubin Ultra), should per-gigawatt opportunity scale 40 → 60 → 80? (b) Is capacity scaling linear, or is there an unlock that lets NVIDIA supply more through fiscal 2027 and 2028?
Management’s Response: Directionally yes on (a), with a new floor to the series: general-purpose computing was roughly $3–5B per gigawatt, versus the current ladder (see Q2). The aspiration was stated as “the perfect answer is actually infinity per gigawatt” — put as much compute on the land as physics allows. On returns, he cited ROIC of under a year on $50B data-center projects, crediting productivity, durability, and fungibility. On (b), no supply unlock was identified beyond the implication that each generation raises compute density per site.
The “Analyst’s Take”: Direct on direction, thin on mechanism: the capacity question was effectively answered with the monetization question — “higher” per gigawatt is a demand-side argument, not a supply-side unlock. The ROIC datapoint is the session’s strongest new evidence for cycle durability, though sourced anecdotally (“I heard the other day”).

Synthesis

Clusters

Concern	Analysts (Q#s)	Uniform management answer	Signal
Can supply expand beyond the plan, and what unlocks it	Joe Moore (Q1), Stacy Rasgon (Q3), Jim Schneider (Q7), Aaron Rakers (Q8)	Demand exceeds supply (“a lot higher”); entire supply chain running flat out; capacity comes online continuously; no quantified gap or unlock	Four separate probes produced zero mechanics or numbers beyond “higher” — investors are testing whether the plan is a floor, and management keeps the ceiling open-ended
Inference-share defense against custom silicon	CJ Muse (Q2), Vivek Arya (Q4)	Platform fungibility versus single-cloud point solutions; every model runs on NVIDIA; labs remain long-term customers	Competitive anxiety persists even in a record quarter; NVIDIA’s counters stay qualitative — no share metric, no benchmark rebuttal
Cost and capital intensity of the model	Stacy Rasgon (Q3, pricing), Vivek Arya (Q4, cash commitments)	Prepared remarks pre-empted margin concerns; price increases are in motion; supply commitments are strategic and front-loaded	Questions about monetization quality (pricing, cash timing) draw narrative answers; margin/cash skepticism is not resolved in Q&A
Durability of the demand base beyond hyperscalers	Timothy Arcuri (Q5), Ben Reitzes (Q6)	Both model types and both human/agent demand paths expand the pie; NVIDIA wins on either	Management is comfortable with unfalsifiable demand narratives and declines to size them

Deflections

Topic (Q#)	What was asked	Deflection technique
Constraint ranking (Q7)	Rank DRAM vs power/shell vs foundry	Declined outright; substituted “entire supply chain is challenged”
Unconstrained revenue (Q3)	Quantify growth if supply were free	“A lot higher”; no figure
Fiscal 2028 revenue bridge and pricing (Q3)	Break down contributors, including pricing	Reframed to hyperscaler-versus-ACIE halves; pricing not addressed
Cash portion of commitments (Q4)	Fiscal 2028 cash outlay	Shifted to the role and three-year phasing of supply commitments
Capacity scaling mechanics (Q8)	Linear versus step-change unlock	Restated per-gigawatt ambition; no mechanism given

Contradictions

Contradiction	Prepared-remarks claim (Q#/source)	Q&A walk-back
Per-gigawatt economics (Q2, Q8)	Colette Kress: NVIDIA revenue opportunity of $18B per gigawatt with Hopper, $25B with Blackwell, $40B with Vera Rubin	Jensen: “each gigawatt of data center increased from, say, $30 billion about 5 years ago to now $60 billion today” — a different series (seemingly total project investment) used interchangeably with NVIDIA’s own revenue per gigawatt; never reconciled, and Q8 returns to the prepared ladder without addressing it
Supply quantification (Q7)	Prepared remarks: the ~70% guide is explicitly supply-constrained, with supply supporting it	“we have supply for 70%. We have more supply than 70%, but about 70%” — self-contradictory on its face; not clarified
Fiscal 2028 uplift math (Q3)	Analyst-recalled prior framing: ~$1T over three years; the new plan implies a ~$200B step-up	Analyst’s own next clause drifts to “$200 million more” (analyst-provided slip); management neither corrected the baseline nor engaged the math

Unanswered

Joe Moore (Q1): no timeline or mechanism for closing the supply-demand gap.
CJ Muse (Q2): no inference-share trajectory or metric, despite the direct share question.
Stacy Rasgon (Q3): no unconstrained revenue figure, no product-level bridge, no pricing contribution.
Vivek Arya (Q4): no confirmation of the ~$500B ecosystem commitment tally and no fiscal 2028 cash breakout.
Jim Schneider (Q7): no constraint ranking or severity order across DRAM, power/shell, and foundry.
Aaron Rakers (Q8): no answer on whether capacity scaling is linear or unlocks via some mechanism.

Count check: 8 entries (Q1–Q8) against 8 question units in the transcript — Joseph Moore (Morgan Stanley); CJ Muse (Cantor Fitzgerald); Stacy Rasgon (Bernstein Research); Vivek Arya (Bank of America Securities); Timothy Arcuri (UBS); Ben Reitzes (Melius Research); Jim Schneider (Goldman Sachs); Aaron Rakers (Wells Fargo). No follow-up turns, clarifications, or inline asks.