A complete strategic analysis of a historic keynote. GTC Taiwan / Computex 2026. Token economics, competitive landscape, geopolitics, microarchitecture, historical parallels.
The keynote opens not with slides, but with a poetic film. The camera glides through an AI factory against the narration: «This is how intelligence is made. A new kind of factory, generator of tokens, the building blocks of AI. Tokens have opened a new frontier, turning data into knowledge, reason, action...» Tokens are shown as a new fundamental unit of reality, like atoms from which intelligence is built. The final frame: «And here in Taipei is where it all begins.» Not a single product, not a single slide, pure ideology. This is not a presentation. This is a film.
Five layers of meaning, from surface to fundamental
Jensen Huang took the stage in Taipei, not just with products, but with a manifesto for a new computing era. If two years ago he presented the concept, now he said: «Agentic AI has arrived.» Not a hypothesis. An established fact.
Five key layers of the keynote:
AI factory unit economics, TCO, payback, and why tokens/watt is the new religion
| Year | Commits (M) | YoY Growth | Context |
|---|---|---|---|
| 2023 | 300 | — | Pre-agentic AI. Manual coding. |
| 2024 | 400 | +33% | Early AI assistants (Copilot, Codex) |
| 2025 | 500 | +25% | Agentic coding — AI as co-pilot |
| Q1 2026 | ~3× | +200% | Explosion. Agents generating code |
$3T in salaries × 3× multiplier = $9T output
30 million developers with a combined salary of $3 trillion produce output equivalent to $9 trillion. A 3× multiplier. AI does not replace engineers, it makes each one 3 times more productive. «AI will reduce jobs» — «complete nonsense» (verbatim).
| Parameter | Grace Blackwell (2024–25) | Vera Rubin (2026–) |
|---|---|---|
| Factory Scale | 1 GW | 1 GW |
| CAPEX per GW | $30–50B | $50–80B → $100B |
| Time to First Token | Months | Weeks (via DSX SIM) |
| Power Overprovisioning | Up to 40% | ~0% (DSX Max LPS) |
| Cooling | Liquid / Air | Hot liquid 45°C |
| Rack Assembly | 2 hours (cables) | 5 minutes (midplane PCB) |
Electricity is a fixed resource. 1 GW = 1 GW. You cannot add more. Therefore tokens per watt = revenue. The choice of architecture is not a question of chip price. It's a question of revenue per watt. Jensen, without naming names: «Choosing the wrong architecture just because the chips are cheaper doesn't make sense.» If a chip is 30% cheaper but delivers half the tokens per watt, at the scale of a $100B factory, that's a catastrophe.
Jensen introduces the concept of a four-dimensional AI factory efficiency curve:
This is not a slogan. It's an economic law in a world where every token is worth money. If your CAPEX generates tokens/watt above market, you're printing money. If below, you're burning it.
AMD MI400, hyperscaler custom chips, Cerebras, Groq, Intel, the battle map for AI compute
Jensen rarely names competitors by name. But every word in this keynote is a direct strike. Let's break it down by target.
| Target | What the Competitor Does | NVIDIA's Response | Jensen's Weapon |
|---|---|---|---|
| AMD MI400 | Cheap GPUs with open ROCm | «Choosing the wrong architecture just because chips are cheaper doesn't make sense» | Tokens/watt as the absolute metric. CUDA lock-in. 1000+ CUDA X libraries. |
| Google TPU / Amazon Trainium / Meta MTIA | Custom chips for internal use. Vertical integration. | «We built the entire infrastructure. We didn't just come up with a PowerPoint slide.» | DSX — turnkey AI factory. Custom chips don't solve the factory problem (networking, power, cooling, orchestration). |
| Cerebras (CS-3) | Wafer-scale chip for inference | Emphasis on disaggregated computing — one giant chip cannot be flexible | Vera Rubin disaggregated: GPU rack + CPU rack + storage rack + network rack. Scales by component. |
| Groq | LPU for ultra-low latency inference | Absorption. NVIDIA integrates Groq LPU into Vera Rubin as LPX tray | If you can't beat them — embed them into your platform. LPX provides latency, NVL72 provides throughput. A dual strategy. |
| Intel Gaudi / x86 | Gaudi for AI + x86 dominance in CPUs | Vera CPU — a new CPU market for agents. Not competition — creation of a new category | World-class IPC, bandwidth at reticle limits. Not «better than x86» — «for a different task.» |
Rhetoric, reality, and the stakes, why «Taiwan is where it all begins»
Jensen opens the keynote not with products, but with his parents. «I brought my parents home.» He brings his mom and dad on stage to applause. This is not sentimentality. This is a message.
| Component | Taiwan Partner | Why It's Critical |
|---|---|---|
| 3nm Vera Rubin chips | TSMC | The only company in the world capable of this scale and process node |
| CoWoS-R / CoWoS-L packaging | TSMC | 6 trillion transistors on a board — impossible without this technology |
| MGX rack assembly | Foxconn, Quanta | 1.3M components in a rack. Assembly speed: 5 minutes |
| CPO (Co-Packaged Optics) | TSMC COUPE | 200 Gbit Ethernet switch with indium phosphide lasers |
| PCB midplane | Taiwan ODM | Critical innovation: assembly from 2 hours to 5 minutes |
100% of NVIDIA's advanced chip manufacturing is in Taiwan. China considers Taiwan its territory. A military conflict in the Taiwan Strait is an existential threat to the entire AI industry. That's precisely why (1) the CHIPS Act in the US, (2) TSMC is building fabs in Arizona, Japan, Germany, (3) Jensen so demonstratively emphasizes the partnership with Taiwan, it's both gratitude and a political signal.
Jensen brings his parents on stage to applause: «I brought my parents home.» And immediately, scale: 70 watch parties across Taiwan, 70 parallel conferences watching the keynote. «So many of you here today. We are broadcasting right now to 70 other watch parties across Taiwan.» The personal and the global in one frame. Parents, a symbol of roots. 70 broadcasts, a symbol of scale.
Jensen never mentions China directly. But he says many times: «Taiwan is where it all begins.» «Taiwan was with us at the beginning.» «Together, we reinvented computing.» This is a declaration of loyalty. NVIDIA cannot afford to lose the Taiwanese supply chain, and it cannot afford to ignore the Chinese market ($17+ billion in 2024 before sanctions). Balancing between these two realities is the company's most complex geopolitical challenge.
Microarchitectural analysis, why SSM+MoE and Olympus Core are not marketing
| Characteristic | Vera (Olympus Core) | Best x86 | Delta |
|---|---|---|---|
| Decode width | 10-wide | 6-wide (Zen 5) / 8-wide (Lion Cove) | 1.7–2.5× |
| IPC (instructions/clock) | 10 (target) | ~5–6 (SPECint) | ~2× |
| Branch predictor | Neural · 2 taken branches/cycle | Traditional TAGE/Perceptron · 1/cycle | 2× throughput |
| Prefetcher | Graph engine · Anticipates data paths | Stride/Stream-based | — |
| Core interconnect | Monolithic Mesh · 3.6 TB/s | Chiplet (Infinity Fabric / EMIB) | 50% faster core-to-core |
| Memory | LPDDR5X · 1.2 TB/s · On-die ECC | DDR5-6400 · ~200 GB/s | 6× bandwidth |
| Memory latency | 40% lower peak latency vs x86 | — | 1.7× |
| PCIe | Gen 6 (first CPU) | Gen 5 | 2× bandwidth |
The key decision of Vera's architects: abandon chiplets. All 88 cores on a single monolithic die with a mesh interconnect. The reason: in agentic workloads, CPU cores work together on a single task rather than serving isolated tenants. Chiplet architectures (AMD, Intel) are optimized for multi-tenancy, while Vera is optimized for cooperative low-latency processing. Hence the 50% advantage in core-to-core latency.
Neotron 3 Ultra is the world's first model based on a hybrid State Space Model + Mixture of Experts architecture. This is not two separate tricks, it's synergy:
| Component | Role | Advantage |
|---|---|---|
| SSM (State Space Model) | Processing long sequences | Linear complexity in context length (O(n)) instead of quadratic (O(n²)) for attention |
| MoE (Mixture of Experts) | Activating a subset of parameters per token | More total parameters at the same inference FLOPs |
| Hybrid | SSM for long context + MoE for capacity | 5× faster + 30% cheaper at frontier quality |
Jensen says: «We made it fast so that you could think fast. When you think fast, you could think longer at the same cost.» This is the key design principle of agentic models: the agent must think iteratively, observe → reason → plan → act. Every iteration consumes tokens. The faster the model, the more iterations for the same budget, the smarter the agent.
Jensen gives the most detailed description of the agentic loop in NVIDIA's history. This is not just «an LLM answering a question», it's a full cognitive cycle:
Midway through the CUDA X presentation, an unexpected pause. The screen fills with a heat visualization: orange and red waves to a hypnotic beat. Jensen: «A round of applause for math. Math is beautiful.» This is not a special effect, it's a reminder that beneath all AI lies mathematics. GPUs heat up, and that heat is the physical embodiment of computation. The most unexpected and memorable moment of the keynote.
1000+ libraries, 20 years of development, each with an «instruction manual» for AI
Jensen calls the CUDA X libraries «Nvidia's treasure», and it's no exaggeration. 20 years of investment in domain-specific acceleration are now becoming available to agents. Each library receives «skills», an instruction manual that AI reads and understands. «The AI reads it and goes: aha, that's how you use it.»
| Library | Domain | What It Does | Why It Matters for Agents |
|---|---|---|---|
| cuLitho | Computational lithography | Computes chip masks. Replaces 40,000 CPU servers with 350 GPUs | TSMC uses it for 3nm. Without cuLitho — Vera Rubin impossible. Agents will be able to design chips. |
| CuOpt | Decision optimization | Solves NP-hard problems: routing, logistics, scheduling | A supply chain agent will find the optimal route in seconds instead of hours. |
| Parabricks | Genomics | Genome analysis 80× faster than CPU | Medical agents: real-time patient DNA analysis. |
| Aerial | AI RAN (Radio Access Network) | AI for 5G/6G base stations | Agents on base stations: dynamic spectrum optimization, energy savings. |
| Warp | Differentiable physics | Physical simulations with gradients | Robot agents: learning the physics of the world through simulation. |
| AIQ | Deep research | Search across structured and unstructured documents | Research agents: finding insights in corporate data. |
| QDSS | Direct sparse solvers | Solves sparse systems of equations | Engineering agents: structural simulation, CFD, electromagnetism. |
CUDA X is an integral part of NVIDIA's moat. Competitors can build a fast chip. But they cannot replicate 20 years of domain-specific libraries optimized for every industry. Now these libraries get an «agent interface», and become tools that AI agents will use more effectively than humans. This transforms the moat from a defensive ditch into an offensive weapon.
IBM PC 1981, Windows 95, iPhone 2007, and why Jensen is right in his bet
| Era | Interface | Model | Who Won |
|---|---|---|---|
| 1970s | Terminal / Command Line | Time-sharing | IBM, DEC |
| 1980s–90s | GUI / Windows | Personal Computer | Microsoft, Intel |
| 2000s–10s | Touch / App Store | Smartphone | Apple, Google |
| 2010s–20s | Cloud / Browser | SaaS / Mobile Apps | AWS, SaaS Giants |
| 2026– | Intent / Agent | Agentic Computing | NVIDIA (+ Microsoft?) |
The pattern: each successive transition changes not the device, it changes the mode of interaction. Command line → GUI → Touch → Intent. In the world of agents, you don't «launch an app», you explain what you want to do, and the agent executes.
«The agent is the last breakthrough in computer science.» Why the last? Because it's the final form of abstraction: Machine code → Assembler → C → OS → Applications → Cloud → Agents. The agent is an autonomous digital entity that understands intent and executes it. After this level of abstraction, only AGI remains.
Full breakdown: models, harnesses, tools/skills, runtimes
| Component | What NVIDIA Offers | Alternatives | Strategy |
|---|---|---|---|
| Model (LLM) | Neotron 3 Ultra (open) Future: Neotron 4, 5, 6 |
Llama, Mistral, Qwen, Claude, GPT | Open model as a base for fine-tuning. Companies create proprietary super-agents. |
| Harness | OpenShell (open source) Support: Claude Code, Codex, Hermes |
LangChain, AutoGen, CrewAI | Enterprise-grade security + isolation + identity. Red Hat, Canonical, Microsoft already onboard. |
| Tools + Skills | CUDA X Libraries (1000+) With «instructions» (skills) for agents |
MCP servers, REST APIs, SDKs | 20 years of libraries — now available to agents. «The AI reads the manual and goes: aha, that's how you use it.» |
| Runtime | OpenShell Runtime Cloud / On-prem / On-device |
Docker, K8s, WASM | Unified runtime for GPU (NVL72), CPU (Vera), PC (RTX Spark), robots (Jetson Thor). |
The most concrete and powerful example from the keynote:
«Nvidia has thousands of chip designers. We are going to hire hundreds of thousands of Cadence super agents.»— Jensen Huang▸ HERMES ON RTX SPARK: LIVE DEMONSTRATION
In the RTX Spark demo, an agent designs a house: OpenShell sandbox + Hermes harness + Claude Sonnet (cloud) + local tools (Rhino, Blender, Flux 2). The agent opens Rhino on its own, models the site, generates the floor plan, places doors and windows — and «detects its own mistakes and fixes them.» This is the first public showing of Hermes as a production harness on NVIDIA hardware. A key signal: «open claw Hermes» (Jensen's verbatim quote) — one of three named harnesses alongside Claude Code and Codex.
CrowdStrike, Palantir, SAP, ServiceNow + AI clouds from startups to Google
Jensen directly addresses the «AI will kill software companies» narrative: «Exactly the opposite. There are going to be so many agents. The world is no longer limited by the number of people.»
| Partner | Domain | What They're Doing with Agents |
|---|---|---|
| CrowdStrike | Cybersecurity | Agents for real-time threat detection and response |
| Palantir | Analytics / Defense | Agents for intelligence analysis and operational planning |
| SAP | ERP / Enterprise | Agents for automating business processes at the world's largest companies |
| ServiceNow | ITSM / Workflow | Agents for automating IT support and service processes |
| Cadence | EDA / Chip Design | Super-agents for chip verification (40× acceleration) |
| Adobe | Creative Tools | Photoshop and Premiere with MCP server — agents control the tools |
Jensen proudly showcases AI clouds built on NVIDIA: «Every one of these are small companies just recently, and now CoreWeave is worth $50-70 billion.»
| AI Cloud | Region | Key Clients | Significance |
|---|---|---|---|
| CoreWeave | USA | Cursor, Black Mountain Labs, World Labs, Revolut, Shopify | From mining farm to $50-70B AI cloud |
| Nbius | USA/Israel | Frontier AI companies | Ultra-fast growth |
| NScale | Europe | British Telecom, Google, Thinking Machines | Google rents compute from NScale on NVIDIA — a massive signal |
| Neighbor Cloud | Korea | Bank of Korea, Hyundai | National champions on NVIDIA |
| Yotta | India | — | Indian AI sovereignty |
| Together AI | Singapore / Australia | AI Singapore | SEA regional hub |
| IndoAT | Indonesia | — | Indonesian AI sovereignty |
| GMI Cloud | Taiwan | — | Home AI provider |
NVIDIA doesn't compete with AWS/GCP/Azure directly. Instead, it creates a new layer of AI clouds, niche, regional, specialized. Each such cloud is an order for thousands of GPUs. Each client of such a cloud is a workload on CUDA. «AI is going to run everywhere. Every company will be powered by it. Every region will build it.»
Cosmos 3 (simulation), Alpamo 2 (autonomy), Isaac Groot (robots)
| Product | Domain | Model | Uniqueness |
|---|---|---|---|
| Cosmos 3 | Physical AI Foundation | Mixture of Transformers: AR + Diffusion | 4 modes: VLM, World Model, Simulator, Policy. «Compute is data» — generates synthetic data. |
| Alpamo 2 | Self-Driving | Reasoning AV model | The first «reasoning» car. «It talks to itself all the time — that's called thinking.» |
| Isaac Groot | Humanoid Robots | Teleop → Sim → Train → Deploy | Reference platform. 31 DOF. Full stack from simulation to deployment. «Open. Use ours, or swap in your own.» |
Jensen emphasizes: the same agentic pattern works for everything. Cloud agent, PC agent, robot, autonomous vehicle, all use Model + Harness + Tools/Skills + Runtime. Only the harnesses and runtimes change depending on the environment. This is architectural elegance that makes NVIDIA's bet symmetric: a win in one vertical strengthens all the others.
First-person vs third-person, teleoperation, simulation, bootstrap, how NVIDIA solves the data problem for robots
Jensen highlights a fundamental distinction: data for language models vs data for physical AI. All text on the internet is written by humans for humans, that's natural data for LLMs. But video on the internet is shot from a third-person perspective (someone filming someone else), while a robot needs data from a first-person perspective (the robot sees through its own eyes).
For language models, data is a finite resource (the internet is finite). For physical AI, data is a generated resource. The more compute, the more synthetic data through simulation. This makes NVIDIA the only company that controls both compute and data generation. «Data plus compute gives you AI. Now that we have AI, compute is data.»
Spectrum X, CX9, CPO, Quantum X, how interconnect became a strategic weapon
«Today NVIDIA is the largest networking company in the world. I'm so proud of the networking team.»— Jensen Huang
This statement is easy to miss, but it is colossal. NVIDIA bought Mellanox for $6.9 billion in 2019. Now the networking division is the world's largest. Why this matters:
| Product | Role | Innovation |
|---|---|---|
| Spectrum X | Ethernet for AI | The world's first 200 Gbit CPO (co-packaged optics) Ethernet switch. Indium phosphide lasers. |
| CX9 (ConnectX-9) | SuperNIC | Next-generation network adapter. Embedded in Vera Bluefield STX. |
| Quantum X | InfiniBand | For the most demanding clusters. NVLink + InfiniBand = full stack. |
| NVLink 72 | GPU-to-GPU | 72 GPUs in a single coherent memory domain. Foundation of Vera Rubin. |
| MVLink C2C | Chip-to-chip | Connects Vera CPU to GPU directly, bypassing PCIe. |
In a disaggregated agentic system, the network is the computer. Data must move between GPUs (NVLink), CPU (MVLink C2C), storage (CX9), and the outside world (Spectrum X) with minimal latency. Control over the entire chain, from chip to optical transceiver, gives NVIDIA the right to dictate the architecture of AI factories. The TSMC COUPE process for CPO is yet another example of vertical integration at the manufacturing level.
Bluefield DPU, at rest / in motion / in use, why the model is more valuable than the factory
Jensen calls the AI model «so precious», and this explains the unprecedented level of security in Vera Rubin:
At rest, all data on disks is encrypted.
In motion, all data between components is encrypted (NVLink, CX9, Spectrum X).
In use — confidential computing. Even during GPU processing, data is protected from the factory operator.
The role of Bluefield 4 DPU: a specialized security processor that isolates tenants. In a multi-tenant AI factory, clients must not see each other's data. Bluefield provides hardware-level isolation. Moreover: Vera Bluefield STX simultaneously manages both storage and security, accelerating agent memory access (KV-cache, retrieval) without compromising security.
Photoshop and Premiere rearchitected for RTX Spark, 2× faster, with a built-in MCP server
One of the most concrete partner announcements: Adobe rearchitected Photoshop and Premiere specifically for RTX Spark:
The key moment: Adobe embedded an MCP server (Model Context Protocol) directly into applications. This means an AI agent on your RTX Spark can directly control Photoshop: open a file, apply a filter, export, without clicks. «It can now interact with agents on your laptop.»
This is the beginning of a trend: every application will have an MCP interface. NVIDIA promotes a skills-based approach (instructions for agents), while Adobe uses MCP (an interaction protocol). Together they form a new standard: applications not for humans, but for agents. Applications controlled by voice and intent, not clicks.
| Risk | Probability | Impact | What NVIDIA Is Doing |
|---|---|---|---|
| Taiwan Conflict | Medium | Catastrophic | Diversification: TSMC in Arizona, Japan, Germany. But it will take years. |
| Customer Concentration | High | High | AI Cloud partners (CoreWeave, Nbius, Yotta). But hyperscalers are building their own chips. |
| Energy Crisis | Medium | Medium | DSX Flex (grid cooperation). 45°C cooling. But 100 GW is the energy consumption of a small country. |
| 3× Multiplier | Low | High | GitHub data is compelling, but early 2026 is a small sample. If the trend slows — the argument weakens. |
| Open Source Competition | High | Medium | Neotron and Cosmos are open. But if Llama/Mistral pull ahead — the «open but ours» strategy is at risk. |
The closing number where the NVIDIA CEO raps about tokens, Vera, and Taiwan
The keynote ends with a unique moment: Jensen walks to center stage to a beat and raps. This is not a joke, it's a poetic summary of the entire keynote, compressed into 90 seconds. The rap lyrics are, in essence, the executive summary of the keynote:
«Taipei agents used to be misunderstood. Only movie stars had them in Hollywood. Now we all got teams making dreams come true. Building companies from living rooms, but they need so much compute. We hear you. That's why we created Vera. The cheapest tokens coming through.»— Jensen Huang, closing rap
Key lines and their meaning:
| Rap Line | What's Behind It |
|---|---|
| «The cheapest tokens coming through» | Vera Rubin NVL72 — tokens/watt leadership |
| «50% faster, that's outrageous. Not for Vera.» | Vera CPU — 1.8× sandbox performance vs x86 |
| «OpenShell keeps the sand tight» | Secure runtime for enterprise agents |
| «Biggest PC moment in 40 years» | RTX Spark — the reinvention of the PC |
| «Cosmos built worlds that robots need» | Cosmos 3 — simulation for physical AI |
| «Every optimized for you, so you can have your cake and eat it too» | NVIDIA Agent Toolkit: models + harness + tools + runtime |
The rap is not entertainment. It's a mnemonic anchor. The viewer will remember «Taipei agents», «cheapest tokens», «PC moment in 40 years», exactly the messages Jensen wants to cement. This is top-tier corporate communication: a CEO who can explain his strategy in verse. And the finale — «Oh, he don't sing yet», self-deprecation that makes him human.
Jensen Huang delivered a keynote that will be studied in textbooks. But this is not just a product presentation. This is a transition manifesto, from the era of AI models to the era of AI agents.
Three fundamental theses:
«The computer industry has been completely changed in the last 6 months. Everything changed. Everything changed because agents were realized and it converged with the latest frontier models and it made possible the AI to now do useful work.»— Jensen Huang, GTC Taiwan 2026
Jensen is not selling the future. He is saying: «Here's GitHub data, agents are already here. Here's Vera Rubin in full production. Here's the PC we reinvented with Microsoft. Here's a model you can take and fine-tune. Here's an AI factory we'll help you build.» This is not a promise, it's a statement of an established fact. Useful AI has arrived.
ARLI membership: the notes from the workshop
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.