Dolphin Research's transcript of NVDA FY27 Q2 earnings call:

Detailed review:NVDA: Another 70% next year; the compute central bank is running red-hot

I.$NVIDIA(NVDA.US) Key takeaways

1. Shareholder returns: record $26bn in a single quarter, above the stated floor

a. Repurchased $20bn. Paid $6bn in quarterly dividends ($0.25/share).

b. Initially committed to return ≥50% of FCF. YTD actual is 60%.

c. After reserving for strategic uses, excess FCF will be returned, with a larger allocation. Management will keep increasing returns when capacity allows.

2. Outlook: Q3 revenue guide $108bn, and first full-year guide — FY28 growth approx. 70%

a. Q3 revenue guided to $108bn, ±2%. QoQ growth is primarily driven by ACIE, while hyperscalers are set to re-accelerate in Q4 and FY28 as Vera Rubin supply scales; Vera Rubin is expected to contribute about 20% of data center revenue in Q3.

b. The ~70% for FY28 is on a supply-constrained basis. Customers point to demand doubling next year, and management expects supply to remain a bottleneck at least through end-FY28.

c. CPU revenue is expected to more than double in FY28. Management believes this will make the company one of the major server CPU suppliers globally.

3. GPM: memory inflation beats expectations; trough at 71%–72% in Q4, back to 72%–73% in FY28 via pricing

a. Q2 GAAP and non-GAAP GPM were both 75%, roughly flat QoQ given similar mix. Q3 GPM guided to 74%, ±50bps.

b. Management called memory pricing 'extremely elevated', already above prior expectations and likely higher next year. They chose to lay out the GPM path upfront to remove suspense.

c. Price increases already enacted will take effect in FY28 Q1, lifting GPM back to 72%–73%. Memory tightness is seen as a spillover from AI buildouts, unlike ordinary components that raise cost without matching benefits.

4. Other financials

a. Q3 GAAP/non-GAAP OpEx approx. $9.2bn/$9.0bn. Full-year OpEx growth lifted to just above 50%, driven by broader product lines and increased use of AI tools.

b. Q2 non-GAAP effective tax rate was 16% (higher YoY mainly on revenue growth). FY27 GAAP and non-GAAP tax rates remain guided at 16%–18%.

c. Inventory rose to $32bn, stocked for Vera Rubin ramp. DSO increased to 60 days, reflecting extended terms on some large, investment-grade, cross-quarter purchases.

d. China: Hopper-200 shipments to China, under U.S. government approval, were under 1% of data center revenue in Q2 and diluted overall GPM. Forward guidance includes no China data center compute revenue.

II. Earnings call details

2.1 Management remarks

1. Data center: both sub-segments accelerating, non-hyperscale to about half

a. Data center revenue rose 18% QoQ to $89bn. Hyperscalers delivered $49bn, up 13% QoQ, driven by continued Blackwell ramp.

b. ACIE (including neocloud, industrial and enterprise) reached $40bn, up 25% QoQ and 138% YoY. Growth came from neocloud capacity additions and non-hyperscale customers augmenting self-build with external capacity.

c. Management expects non-hyperscale (sovereign, regional neocloud, enterprise, edge and air-gapped data centers) to account for roughly half of data center. This mix is becoming structurally balanced.

d. Industry view: cloud backlogs exceed $2tn. Top-5 hyperscalers' capex is estimated near $800bn in 2026 and $1.3tn in 2027.

e. Neoclouds using NVIDIA DSX reference designs go live faster with lower token cost. Installed base is expected to reach 8GW by year-end and about 3GW by end-2025.

2. Vera Rubin and Vera CPU: mass production and shipments start this month

a. Vera Rubin has orders from all major hyperscalers, AI clouds and system OEMs. Management expects the fastest product ramp in history; vs. Grace Blackwell Ultra, throughput per MW is 30x and token cost is 35x lower.

b. Vera Rubin spans Vera CPU, Rubin GPU, NVLink, InfiniBand or Ethernet, plus the Groq LPU announced this week. Revenue per GW opportunity rises from $25bn on Blackwell to $40bn.

c. Vera CPU is at full production: on SPEC, agentic tasks complete 1.8x faster, and per-watt bandwidth is 5x other data center CPUs. Shipments have gone to OCI and SpaceXAI, with AWS supply starting this quarter.

d. Total server CPU demand observed is about $20bn. Grace CPU revenue exceeded $5bn over the past 12 months.

3. Groq and networking

a. Groq 3 LPX is the first rack-scale LPU system and is in full production. On the Artificial Analysis benchmark its tokens/sec are nearly 4x the next-best setup; volume shipments to early customers will start later this quarter, with Nebius first.

b. Networking set another quarterly record, up 18% QoQ. Spectrum-X Ethernet grew 2.6x YoY.

4. Expanded AWS partnership

a. From this quarter through FY29 Q2, AWS will deploy 2mn GPUs, paired with Vera CPUs. Some will be integrated with Rubin, and others will be standalone.

b. AWS will offer Nemotron open-source models on Bedrock and SageMaker. Amazon will also use Omniverse, Cosmos, Isaac and Jetson to power its warehouse robotics stack.

5. Sovereign AI and neocloud: new revenue-sharing model

a. Sovereign AI (primarily via regional neoclouds) rose 35% QoQ and more than doubled YoY. Many countries or regions prefer to allocate land and power to regional cloud partners rather than foreign hyperscalers, while NVIDIA remains cloud-neutral.

b. New model: NVIDIA makes take-or-pay commitments on part of facility capacity, providing minimum revenue guarantees to enable project underwriting. In return, it shares rent above the floor — effectively two revenue streams: hardware sales and rent sharing.

c. Management stressed independent capital still underwrites based on project merit, and they do not make loans. This model could yield multi-bn, usage-linked recurring revenue over time.

6. Frontier AI labs: nearly $50bn invested, plus a $500bn financing platform with six partners

a. Bottlenecks for frontier labs are in compute, not tech or demand. Growth outpaces what their balance sheets and credit can support, lacking long-term infra contracts and investment-grade financing.

b. Nearly $50bn has been invested, a small portion of the expected FCF over the period. NVIDIA also formed a third-party financing platform of over $500bn with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.

c. Partnered with SoftBank Energy to secure land, power and facilities at the PORTS-Pike campus, exclusively hosting NVIDIA compute. The first phase is about 4.25GW for OpenAI, with each system generation tied to ~1.5mn GPUs and multi-cycle upgrades over 20 years.

d. OpenAI has committed to large-scale deployments through 2030. Existing and planned commitments equate to roughly 12GW of compute; another frontier lab has an optional credit enhancement up to nearly 2GW.

e. On concerns about circular financing, management argues the compute platform is reusable and portable, limiting risk. Next year, labs requiring NVIDIA balance sheet support are expected to account for about one-quarter of the business, and output will ultimately be absorbed by investment-grade buyers or those with IG backing.

7. Enterprise and AI-native ecosystem

a. Over the past 12 months of on-prem revenue, autos reached $8bn. Financial services, manufacturing and healthcare combined were $7bn.

b. Customers include Hudson River Trading, Jane Street, Samsung Electronics (cuLitho improves computational lithography performance by up to 20x) and Bristol Myers Squibb.

c. Global AI VC funding in H1 2026 exceeded $400bn (about 70% spent on compute), above $265bn in all of 2025. Companies with annualized revenue over $1bn are approaching 20, up from 13 in Q4 last year.

2.2 Q&A

Q: What gives confidence to issue a full-year guide for the first time? Where is the gap between 70% and 100% demand growth, and can it narrow over time?

A: Demand growth is far above 70%, but supply only supports 70%. The exception was made to align expectations. Demand has two sources: first, agent workloads require 15x–100x the compute of human-driven tasks, depending on task type, with larger and smarter models involving multi-round tool calls for inference and planning.

The other half is not visible externally — sovereign AI, regional AI, neoclouds, AI startups and enterprises account for about half the business and are growing 100% YoY. These customers do not buy custom chips or assemble themselves, and they need full AI factory platforms; management believes this segment could be larger than cloud over time.

Delivery cadence has changed too. One cannot stand up infra by just buying tech anymore, as land, power and facilities must be secured 2–3 years in advance; therefore, NVIDIA is extending upstream into memory and power, and downstream into land, power and facilities, increasing visibility across the stack.

The full-year guide was given to ensure customers, shareholders and the supply chain invest against the same information set. Management wants all parties aligned on a single outlook.

Q: How do agentic workloads evolve, and what happens to inference share? How does TAM lift each generation across the stack, and how to view Groq 3 LPX?

A: The AI lifecycle has four stages, and a single NVLink 72 stack spans them, accelerating growth. The four stages are data prep (synthetic, real and human labeling), pre-training, post-training and agentic inference.

NVLink 72 rack-scale architecture is in its third generation. NVIDIA rebuilt supply chain, systems and software to create a general, switchable system that runs the same assets from data creation through agentic inference.

Another reference point: capex per GW for a full data center has risen from about $30bn five years ago to $60bn today. Because one system can run diffusion, state space and attention-based models of all sizes, customer investments last longer.

Groq 3 holds the token interaction speed record with very low latency. In recent months, the NVLink architecture was integrated as the core engine; its throughput is far lower with higher per-token cost, fitting high-ASP, high-interaction services as an add-on, while most data centers will stick with Vera Rubin NVLink 72.

Q: The 70% implies roughly $200bn above the prior 'three years, $1tn' view. Which products drive the delta, how much is pricing, and what is the unconstrained level?

A: No unconstrained figure was given, only that it is 'much higher'. Pricing contribution was not addressed. Growth is 100% YoY this year, and the unconstrained volume is significantly higher; the only path is adding capacity.

The split remains in two halves: hyperscalers see revenue and profit rise alongside NVIDIA compute going live, and the compute itself is highly profitable, driving a race to deploy. The other half is ACIE — enterprises, neoclouds and sovereign AI — who need turnkey factory platforms rather than bespoke or per-chip buys.

Management reiterated that revenue per GW rises each generation alongside order-of-magnitude productivity gains. That is why customers rush to the next generation.

Q: Summing the CFO's commitments suggests about $500bn. Is that the full scope of ecosystem investment over the next few years, and what cash outlay corresponds to FY28?

A: Management did not confirm that sum or provide FY28 cash numbers. The CFO reframed the topic as 'supply commitments'. The largest portion of commitments are supply, supporting Vera Rubin ramps this year and next; most of the commitments are front-loaded in the first three years.

Because supply and capacity have been aligned in advance, management has confidence in growth and revenue. This alignment underpins the full-year guide.

Q: OpenAI and Anthropic are building in-house chips (OpenAI recently claimed Jalapeno outperforms Blackwell). How do you balance investing in the ecosystem versus competing products inside it?

A: No comment on Jalapeno's claims, except that it is a single-cloud inference chip. These XPUs are custom inference chips for one cloud and one service, while NVIDIA sells an AI factory platform that spans the full lifecycle and runs across any cloud.

The implication is AI services will be global, and those data centers may not be built by the labs themselves, so they will still run on NVIDIA. Management has 100% confidence in the economics from data processing, training and post-training to agentic inference, and expects those labs to remain customers and partners.

On investing, management called this a once-in-a-generation opportunity. The only regret is not investing more and earlier, and two portfolio companies may list soon, with others to follow.

Q: Do open-source models help or hurt NVIDIA, given that demand is currently led by closed-source frontier labs?

A: Both closed and open source are booming, and nearly all open-source models run on NVIDIA, driving sales on both fronts. NVIDIA has the broadest deployment footprint and the most general architecture, from PCs and DGX Spark at the edge to robots, workstations and on-prem data centers.

Frontier labs are seeing surging sales and strong margins with profitable tokens, constrained only by compute. Open source matters because every large enterprise, country and startup needs domain-specific AI, and frontier-level open models make that viable; cybersecurity is a prime use case for large-scale distributed, continuously autonomous defenses.

Management believes NVIDIA is likely the only platform that can run all frontier models. Most models are built on NVIDIA in the first place.

Q: If RSI and AGI materialize, what happens to industry demand, and what does that mean for NVIDIA?

A: Demand will inflect higher; and the primary mode of AI shifted last month from human prompting to agentic. Every company will run many agents in the background; using NVIDIA's ~40k employees as a proxy, that could be 400k or even 4mn agents.

Personal AI agents at the edge already run 24/7 on DGX Spark and DGX Station. There is always work to do.

On RSI, management believes it is already recursive to some extent, albeit at coarse granularity. Agents reflect after each run and update skill files (markdown documentation), improving on the next execution.

On AGI, management thinks many tasks already meet an AGI bar, and such milestone debates are less meaningful. What matters is that AI delivers productive work, produces profitable tokens, and more compute yields more profitable tokens.

Q: Can you rank the tightest constraints between 70% and 100% — power and facilities, DRAM, wafer foundry capacity?

A: No ranking given, only that the entire supply chain is tight. Current supply roughly supports 70%. All links are running at full tilt; new capacity comes online daily rather than at one point in time, alongside ongoing yield improvements.

Supply is a bit above 70% but around that level, while demand is far higher. The whole chain must scale together to avoid disappointing customers, and the demand figures shared publicly match what has been told to suppliers.

Q: After $40bn per GW, does it go to $60bn or $80bn? Is capacity expansion linear, or is there a unlock point?

A: No answer on whether capacity scaling is linear, but revenue per GW will rise beyond Vera Rubin. The objective is to pack as much compute as possible into one site and 1GW; the theoretical limit is infinite revenue per GW, and the direction is moving toward that boundary.

The trajectory for revenue per GW has been: $3bn–$5bn in general compute during the Moore's Law era, then $18bn for Hopper, $25bn for Grace Blackwell, and $40bn for Vera Rubin. Management expects further upside.

Why customers can absorb it: as productivity, durability and generality improve, customers will invest in assets that generate revenue and profits. Payback periods are reportedly under one year, even for $50bn-scale data centers.

<End of text>

Risk disclosure and statement:Dolphin Research disclaimer and general disclosure