--- title: "NVDA: Delivery Over Specs — Who Is Jensen Huang's Cost Calculus For?" type: "Topics" locale: "en" url: "https://longbridge.com/en/dolphin/post/44480806.md" description: "Below is Dolphin Research's Trans of the NVIDIA GTC Europe (Oct 21, 2026; Berlin) CEO keynote. This session was a keynote with no Q&A.Core takeaways as organized by Dolphin Research: Jensen Huang's remarks focused on the 'European narrative', 'AI factories', 'agents and software coexisting', and 'physical AI'.1) European narrative: given the Europe-based venue, he emphasized Europe should shift from 'using AI' to 'building AI', following a five-layer stack from energy → compute → AI factories → models → applications..." datetime: "2026-10-22T06:07:01.000Z" locales: - [en](https://longbridge.com/en/dolphin/post/44480806.md) - [zh-CN](https://longbridge.com/zh-CN/dolphin/post/44480806.md) - [zh-HK](https://longbridge.com/zh-HK/dolphin/post/44480806.md) author: "[Dolphin Research](https://longbridge.com/en/dolphin.md)" generator: "portal-rs" --- # NVDA: Delivery Over Specs — Who Is Jensen Huang's Cost Calculus For? **Below is Dolphin Research's**$NVIDIA(NVDA.US) **GTC Europe (Oct 21, 2026, Berlin) CEO keynote Trans. There was no Q&A.** **Key takeaways compiled by Dolphin Research:** **Jensen Huang's Berlin keynote centered on the 'Europe narrative', 'AI factories', 'agents alongside software', and 'physical AI'.** **1) Europe narrative:** Given the local setting, he urged Europe to move from using AI to building AI, with five layers from energy → compute → AI factories → models → apps, and local companies at each layer.NVDA invested nearly $10 bn in Europe over the past year, and last week backed 14 startups across France, the U.K., Sweden and Germany, with 12k+ partners in the region. **2) AI factories: Rubin is in full mass production. The only new product announced was DSX POD,** a pre-integrated bundle of 7 Vera Rubin liquid-cooled racks plus 17 km of cabling.Assembly moves from on‑site to the supply chain for integration, testing and crating, cutting on‑site deployment from 4–5 months to roughly 1 month. **The focus is not peak perf but economics and delivery.** NVDA frames a 1 GW AI factory at ~$50 bn capex and ~$40 bn annual revenue, implying a 3‑month acceleration can unlock ~$10 bn of revenue.Shorter time to revenue is the key ROI driver for customers. **3) Agents and software can coexist: Contrary to the view that AI will end software and tools, management argues the opposite.** More people will use tools than ever, and every tool company will ship agents that excel at their own tools.Agents get work done through tools, so tool vendors will evolve into agent companies and enterprises will retain ownership of their data and tools. **4) Physical AI: New steps in Robotaxi; humanoids still early.** **i) Robotaxi: Hyperion as a unified reference (14 cameras, 9 radars, 3 lidars, dual GPUs)** and new programs including VW Commercial Vehicles, Jaguar Land Rover's new models, Volvo fleets and Mercedes' next-gen series connecting into Robotaxi networks.Huang sizes the TAM as ~1 tn miles per year globally at $1–2 per mile, and NVDA is also building AV software. **ii) Humanoids: Reference robot Agile ONE targets a Q1 2027 launch (with preloaded compute and software),** and humanoids could become the highest‑volume robot form factor once standards converge.Standardization is the precondition for scale. **iii) Telecom is also physical AI**: Cell sites will evolve into a distributed AI network (~100k sites globally, ~100 GW).NVDA and Nokia completed the first commercial AI phone call on T‑Mobile's network last week. **Overall, Rubin is now fully in mass production. Versus raw perf, this GTC Europe emphasized 'economics', with DSX POD the lone new product to shorten time‑to‑deploy, opening $10–20 bn of potential revenue for customers,** while dedicating more airtime to opportunities in Europe, agents, and physical AI. **NVDA has held up better in the recent sector pullback, largely because Rubin is in a full‑ramp product cycle and the valuation is relatively undemanding.** At a $5.35 tn market cap, it trades at ~13x PE on FY2028 post‑tax core OP (assumes 2‑yr revenue CAGR +89%, GPM 73.5%, tax rate 16.5%). **We see the sub‑15x PE cap constrained by share‑loss risk to rivals and the broader AI correction.** But at 80%+ growth, ~13x PE suggests limited downside, helping explain the relative resilience.Reports that OpenAI will run GPT‑6.1 Sol's Ultrafast mode on NVDA GPUs continue to highlight the product edge. **Do note NVDA's scale is now $5 tn+; a major leg up requires sizable incremental capital.** With the AI supply chain under pressure (e.g., Anthropic ARR), shares may oscillate in the 10–15x PE band, roughly mapping to a $180–260 range. Below are the main items from the GTC Europe keynote: **I. Key info recap** 1\. Europe investment: Nearly $10 bn deployed over the past year across startups and partnerships; 14 new startup investments announced last week across France, the U.K., Sweden and Germany, with 12k+ European partners.Execution remains focused on local ecosystems. 2\. AI factory economics a. Cost: **Approx. $50 bn total capex for a 1 GW AI factory,** covering land, power, buildings, and compute, networking, storage, power and cooling.This frames the unit economics for large‑scale AI infra. b. Delay cost: Each GW is expected to generate tens of billions in annual revenue; at **\~$40 bn per year, every 3‑month delay implies ~$10 bn revenue foregone.**Time‑to‑revenue is the binding constraint. c. Company view: **DSX POD is expected to cut the interval from equipment arrival to token output from 4–5 months to ~1 month,** reducing idle GPUs and idle capex and improving profitability for all parties.Delivery certainty becomes a competitive edge. ***Note: DSX POD is a prefabricated, containerized Vera Rubin data‑hall reference design.** NVDA provides the design, partners build and assemble in factory. The container houses a complete Vera Rubin AI factory, fully integrated and validated before shipment, **then connects on‑site to pre‑provisioned power, cooling and networking.** Compute, fabric, cabling, rack‑level power and liquid cooling are all integrated upstream.* 3\. Ramp and demand a. Ramp: Vera Rubin is in full mass production; Spectrum‑X CPO switches are the world's first fully mass‑produced co‑packaged optics product.Supply chain readiness has broadened. b. Demand: Agents need to think, reason, iterate and invoke tools, driving sharp token consumption and GPU demand; this was qualitative only with no quantified data.Workloads are shifting toward agent pipelines. c. Delivery: NVL72 rack‑level assembly has moved into the supply chain; global ODMs/OEMs are building NVL72, enabling hundreds of data centers to be built in parallel worldwide. **II. Keynote in detail** 1\. Platform stack and perf cadence a. Framework: Treat AI as manufacturing, with five layers from energy, compute, AI factories, models to apps; NVDA sits at the compute layer, with all layers expanding.This underpins the AI economy view. b. Full‑stack hardware: Vera Rubin GPUs, NVLink switching, Vera CPUs, Spectrum‑X Ethernet (management calls NVDA the world's largest networking company), Groq 3 decoders for token generation, BlueField‑4 for storage accel, and RTX compute.A breadth play across data center to client. c. Software: AI factory system software, 350 CUDA‑X libs beyond cuDNN, plus domain models; CUDA‑X is the go‑to‑market to open new categories.Software remains a core moat. d. Perf: Six architectures over a decade lifted AI perf by 1,000,000x vs. ~100x for Moore's Law; for LLMs and agents, Blackwell is ~30x Hopper, and Vera Rubin targets another ~30x over Blackwell within two years.Step‑function gains continue. e. Scope: Blackwell pioneered rack‑scale computing; Vera Rubin extends from model‑centric to agent‑centric compute: GPUs for LLMs, Vera CPUs for the harness (agent orchestration), plus systems for data processing and CUDA‑X tools.End‑to‑end agent pipelines are in scope. 2\. Vera Rubin hardware a. Compute tray: No cables, **100% liquid‑cooled (vs. 80% on Grace Blackwell),** cutting single‑node assembly from ~2 hours to under 5 minutes, improving turns, reliability and perf.Manufacturing quality is higher. b. NVLink: Switch trays stitch 72 GPUs into one logical system; each GPU delivers 20 petaFLOPS.Tight coupling reduces latency. c. Vera CPU: Eight instructions per cycle, LPDDR memory for bandwidth and efficiency, with what management calls the fastest on‑die interconnect; focused on data processing and agent sessions.CPU complements GPU for agents. d. CPO switches: 512×200 Gb/s ports; lasers coupled via optics directly to the die, with silicon photonics attached using the COUPE process co‑developed with TSMC, bypassing PCB and pluggables for higher data rates and signal integrity. 3\. AI factory delivery: DSX and DSX POD a. DSX: A full‑stack AI factory reference co‑designed with PTC, Dassault Systèmes, Schneider Electric, Jacobs, Siemens and Procore; performance, power, water and scale are validated first in Omniverse digital twins.This de‑risks construction. b. Power: End‑to‑end control loops dynamically allocate power by load, cutting energy use by 30% for the same output.Efficiency is system‑level. c. Cooling and water: Vera Rubin uses closed‑loop liquid cooling with coolant temps up to 45°C; with dry coolers in suitable climates, heat is rejected to air without water evaporation, saving 54 mn liters/day at 2 GW, equal to 180k households' daily use.Water is a first‑order constraint. d. **DSX POD (new): Seven Vera Rubin NVL72 liquid‑cooled racks plus 17 km of Spectrum‑X cabling, integrated, wired and tested in‑factory and shipped in transportable containers.**Integration and validation are upstreamed. e. Deployment: Dry coolers instead of evaporative cooling and 800V DC to the rack; once site utilities are ready, on‑site work drops from months to days. Following NVL72, wiring, bring‑up and burn‑in are now also moved into the supply chain.Time‑to‑token is the KPI. 4\. Moats and customer mix a. General‑purpose arch: Positioned as the only general‑purpose accelerated computing stack, not just for Transformers but for fluids, particles, quantum chemistry, imaging and data processing; the architecture has remained consistent for 25 years with expanding installs and domains.Breadth underwrites durability. b. Full‑stack delivery: Many aspire to design chips, but most enterprises just want AI infra to run their biz; designing chips, building supercomputers and writing all software is beyond their scope. CUDA plus the full AI factory opens markets otherwise unreachable.TAM expands beyond hyperscalers. c. Customer map: Hyperscale remains large; also includes **AI‑native companies, 'new clouds' (regions ring‑fencing local energy for local firms such as Mistral, Nebius, Firebird in Central Asia and Cassava in Africa), and enterprise OEMs/ODMs.**Diversity reduces concentration risk. d. AI labs: A year ago the focus was OpenAI; now OpenAI, Anthropic, Meta's Muse, Grok and Gemini are all in the leading pack.Frontier research is competitive. e. End markets and longevity: Autos, manufacturing, healthcare and finance are all becoming AI companies; the same infra spans data processing, training, post‑training, inference and agent workflows, with software updated over time for flexible, durable use.Asset life is extended. 5\. Agents: Five elements and safety a. Five elements: Agents comprise the model, the harness, a safety sandbox, knowledge and tools; the open‑source agent OpenClaw earlier this year helped catalyze the wave.Architecture is modular. b. OpenShell: An Apache 2.0 open‑source secure runtime enforcing policies at syscalls, file access and networking; each policy update is formally verified against privilege ceilings, with secrets kept outside the sandbox and all outbound requests reviewed by an external supervisor.Defense‑in‑depth by default. c. Sentry: Optional hardware isolation on BlueField‑4 monitors agents and can sever model connections within milliseconds upon anomalies, logging tamper‑evident records.Hardware‑rooted safety. d. Adoption: Frontier labs, OS vendors, cybersecurity and platform companies have announced OpenShell support, and NVDA co‑launched Claude OpenShell agents with Anthropic; permissions must be authorized, executed and monitored dynamically by task and user.Granular least‑privilege is required. 6\. Agents: Models and knowledge a. Models: Closed‑source, open‑source and custom models will coexist; NVDA has open‑sourced Nemotron (language), Cosmos (world model), GR00T (humanoids), Alpamayo (AD), Ising (quantum‑classical), and Earth‑2 (physics), with training data and repro scripts.A portfolio approach is pursued. b. Positioning: NVDA concedes its open models, especially language, may not be SOTA, but can reach frontier levels via post‑training on private data, as seen with AI21 and Alice, Baseten and CodeRabbit, Trajectory and Harvey, Scale and Skydance, Palantir and Tampa General Hospital.Customization closes the gap. c. Knowledge: The Agent Knowledge Foundation spans Nemotron Parse (unstructured doc parsing), Embed (vector retrieval) and Rerank (result re‑ranking), with NVDA citing leaderboard‑top results.Retrieval quality drives outcomes. d. Structured data: Kumo builds predictions atop structured DBs, e.g., combining supplier delay notices to forecast stock‑outs and lead times, then using vector retrieval to fetch engineering docs for substitute parts.Tabular + RAG for planning. 7\. Agents: Tools and software companies a. Tools: CUDA‑X accelerates the sims, solvers and data pipelines agents call by 10–100x vs. CPU‑only; tools are the key to agents doing real work.Compute meets domain logic. b. Cases: Lila Sciences uses BioNeMo to design B‑cell lymphoma therapies; Siemens' Erlangen e‑factory on‑boards new products with Cosmos; Dassault uses Nemotron and SIMULIA to morph the VW Golf door to Tiguan.Verticals demonstrate ROI. c. Software companies: **Addressing the claim that 'AI will end software and tool companies', management argues the opposite: more users than ever, and each tool company will ship agents proficient in its tools. Cadence, CrowdStrike, Dassault Systèmes and IQVIA are all building agents.** 8\. Physical AI: Telecom and healthcare a. **Telecom: AI will meaningfully raise spectral efficiency,** and telcos' ~100k sites and **\~100 GW of distributed base stations and data centers are set to evolve into distributed AI infra.**Networks become AI fabrics. b. Progress: After ~8 years of R&D, NVDA and Nokia placed the first commercial AI phone call on T‑Mobile's network last week, layering AI atop CUDA‑handled voice.Telecom AI is moving to production. c. Healthcare: Adaptive radiotherapy needs real‑time tumor/body motion sensing, precision segmentation and guidance; NVDA frames it as a form of robotic surgery.Latency is mission‑critical. 9\. Physical AI: Autonomous driving and humanoids a. Robotaxi: Automakers are expected to connect vehicles to ride‑hail platforms (Uber, Bolt, Grab, Lyft, etc.); ~1 tn miles are driven annually, with a large portion likely shifting to Robotaxi at $1–2 per mile, a huge but fragmented opportunity.Platform economics will vary by region. b. **Hyperion: The standard arch includes 14 cameras, 9 radars, 3 lidars and multiple ultrasonics, with dual NVIDIA GPUs,** now adopted across fleets; NVDA is also an AV software provider.Standardization lowers integration cost. c. New models: VW vans, new Jaguar Land Rover models, Volvo fleets and Mercedes‑Benz next‑gen vehicles will be Hyperion‑compatible and connect to Robotaxi networks.Ecosystem breadth is expanding. **d. Humanoids**: Likely the next very large industry and the highest‑volume robot form; standard form factors enable scale and a tech flywheel, letting SMEs teach robots by demonstration without in‑house robotics teams. Europe has deep mechatronics and precision manufacturing know‑how.Industrial base is an advantage. e. Reference robot: Physical AI needs three computers for training, digital‑twin skill learning and the robot body; **the reference robot Agile ONE with Agile Robots is slated for a Q1 2027 launch with preinstalled compute and software for out‑of‑box use.**Time‑to‑work is simplified. 10\. PCs and workstations a. RTX Spark: Co‑developed with Microsoft for four years around the N1X chip, bringing desktops, notebooks and workstations now in market; ~1 petaFLOPS and 128GB memory, akin to the 2016 DGX‑1 that took two people to carry into OpenAI.Client AI enters the mainstream. b. DGX Station: Runs Windows with ~20 petaFLOPS and 748GB memory, supporting CATIA, Autodesk, Adobe, Cadence and Synopsys, expected by year‑end.A desk‑side AI workstation tier. c. Demo: On RTX Spark, a Hermes harness with Nemotron and OpenAI's GPT‑6 Astra called FreeCAD, Blender and Isaac Sim to model a small Pollen Robotics bot, trained in the cloud and deployed to the physical robot.Agent‑to‑robot workflow was showcased. \ **Risk disclosure and statement:**[**Dolphin Research Disclaimer and General Disclosures**](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### Related Stocks - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [DSX.US](https://longbridge.com/en/quote/DSX.US.md) - [VLVLY.US](https://longbridge.com/en/quote/VLVLY.US.md) - [00175.HK](https://longbridge.com/en/quote/00175.HK.md) - [80175.HK](https://longbridge.com/en/quote/80175.HK.md) - [GELYY.US](https://longbridge.com/en/quote/GELYY.US.md) - [DSX-B.US](https://longbridge.com/en/quote/DSX-B.US.md) - [HGMD.SG](https://longbridge.com/en/quote/HGMD.SG.md) - [GELHY.US](https://longbridge.com/en/quote/GELHY.US.md) - [VERA.US](https://longbridge.com/en/quote/VERA.US.md) --- > **Disclaimer: This article is for reference only and does not constitute any investment advice.**