Jensen Huang — Milken Global Conference 2026 Investment Thesis

Source: Leading in the Age of AI: A Conversation with NVIDIA CEO Jensen Huang | Global Conference 2026, Milken Institute (Becky Quick, CNBC), May 4, 2026.

The Framework: From Generative AI to Agentic AI — Five-Layer Cake, Thousand-X Step-Up

Huang reframes the last ~24 months as a capability ladder: generative AI (token creation) unlocked reasoning, which enabled agentic AI (understand → plan → use tools → act). Alongside the five-layer cake (energy → chips → systems → models → apps), he stresses a second axis: compute per useful task explodes once agents replace “press play on prerendered media” with interactive, intention-driven workloads.

Layer	Huang’s emphasis (this session)	Investable signal
Intent / agents	Coding agents (e.g. Claude Code cited) prove “AI became useful” in recent months—not just demo-worthy	Workflow software, EDA, and corporate automation budgets re-rate
Compute depth	Agentic stacks need ~1000× more compute than plain generative use	Accelerator, memory, photonics, liquid-cooled rack attach
Physical systems	Vera Rubin–class racks = $4–5M, ~3 tons, ~1.5M parts, silicon photonics, 3D packaging	Subsystem oligopolies—not only the GPU badge
Capital formation	$0.5T committed orders to pull supply chains onshore	US re-industrialization + bankable offtake stories
Safety / policy	Export tiering (e.g. H200 framing) + open-weight cyber “swarms” vs monolithic “super weapons”	Regulated dual-use SKUs; OSS model + security tooling demand

Investment Thesis #1: Agentic AI Is the Thousand-X Multiplier on Installed Accelerator Demand

The argument: Huang describes a discontinuity: once models chain reasoning and invoke tools, unit work consumes far more sequential/interaction compute than single-shot generation—analogized to needing 1000× more “cars” or 2000× more “airplanes” in two years versus a static industry mental model.

"And so when you use the browser, you have to generate words, control something else. those two ideas. The moment that happened, the entire industry raced off… and then now agentic AI, which is the ability for AI to understand, reason, plan, use tools to do something useful… the amount of computation necessary compared to generative AI is like a thousand times more."

Street models still extrapolate linear “tokens out” growth; Huang implies effective FLOPs per user session stair-steps with agent loops + tool hops.

Trigger: Hyperscaler & neocloud disclosures of interactive vs batch inference mix; rising attach of reasoning/thinking SKUs priced above base chat.

Names: Nvidia (NVDA); memory vendors (Micron MU, SK Hynix HXSCF) tied to agent context growth.

Investment Thesis #2: “Good Wine” GPUs — Depreciation Narratives Underestimate Useful Life & Resale Power

The argument: Scarce throughput + rising software capability increases economic value of old accelerators—contrary to classic 3-year straight-line GPU obsolescence.

"…even GPUs we sold four or five years ago now are rising in price faster than, you know, good wine."

Bears model hardware as fast-depreciating commodities; Huang claims inverse price aging when intelligence automation TAM expands faster than wafer supply.

Trigger: Secondary/renewal pricing for Hopper / prior-gen clusters; lengthening lease renewals on ~3–4 year-old fleets (echoing other SemiAnalysis-era datapoints Huang aligns with here).

Names: CoreWeave (CRWV) and other GPU lessors with long-dated contracts; Nvidia supply-controlled remarketing economics.

Investment Thesis #3: Systems, Not Dies — Vera Rubin Economics Crown photonics, Memory, Cooling Vendors

The argument: Huang stresses seven chip types per next-gen “computer,” silicon photonics, advanced memory, 3D packaging, liquid cooling, and multi-million-dollar racks scaled to football-field data centers—bottleneck whack-a-mole across suppliers.

"When we say Vera Rubin, it's, you know, twice the width of this stage. Each one of the racks is about four or five million dollars. Three tons, one and a half million parts inside one of these racks… These systems have silicon photonics inside. It's got the most advanced memories… three dimensional packaging, liquid cooling…"

Investors laser-focus wafer counts; Huang says integration + electromechanical content is half the moat.

Trigger: Rack-level ASP disclosure trends; supplier prepayment lines (Huang notes rising cash return on invested capital effects from prepaids).

Names: TSMC (TSM) (advanced packaging attach); Lumentum (LITE) / Coherent (COHR) (optical networking); Vertiv (VRT) (thermal & power chain “land power and shell” adjacent).

Investment Thesis #4: Half-Trillion Order Book as Industrial Policy — Onshoring Chip, System & “AI Factory” CapEx

The argument: Huang recounts committing ~$500B of orders to suppliers to convince ecosystem players to build in the United States, complementing CHIPS-era hesitancy—using market demand rather than subsidy alone.

"I'm gonna give half a trillion dollars of orders to these suppliers." And I bet you they come to United States. Boom.

“Subsidy-only” US fab thesis underweights anchor offtake from a single orchestrator spanning chips → computers → AI factories.

Trigger: TSMC / Samsung-class US fab tool pull-ins; domestic subfab / substrate milestones tied to Nvidia-dense buildouts.

Names: Applied Materials (AMAT) / Lam Research (LRCX) / ASML (ASML) (WFE absorption); Entegris (ENTG) (materials purity tail).

Investment Thesis #5: Neo-Cloud “Anchor Investor” Model — Signaling Value > Nominal Check Size

The argument: Early $1 of Nvidia capital into neoclouds required $9+ of co-investors but de-risked the category; Huang claims CoreWeave, Nebius, Nscale co-investors are “extremely happy / incredibly happy” because Nvidia sees pipeline visibility others lack.

"…we invest some amount and our anchor investment gave all of the investors confidence that we're behind this company."

“Circular financing” critiques miss information asymmetry: strategic anchor as forward demand signal, not passive PE.

Trigger: Neo-cloud EBITDA inflection vs. depreciation fear; elongating multi-year GPU offtakes.

Names: CoreWeave (CRWV) (named); Nebius, Nscale (named, not US-list watchlist slots—track via IPO/ADR overhang).

Investment Thesis #6: Open-Source Cyber “Swarms” vs Monolithic Frontier Models — Asymmetric Defense Stack

The argument: On frontier coding models (Mythos named) dual-use for cyber, Huang argues symmetric super-model duels lose; defenders need abundant cheap open checkpoints trained for telemetry + response—swarms of white blood cells.

"The answer to that, as it turns out, is not another mythos. The way you defend against a super force is not with another super force. It's with an abundance of cheap force… The best answer for mythos is actually open source."

Export control / hoarding narratives overstate closed-model moats in security; OSS distribution + fine-tune economics may capture enterprise SOC budgets.

Trigger: Enterprise security teams standardizing on open-weight SOC copilots; rising threat-hunt automation attach.

Names: Alphabet (GOOG) (open-weight Gemma lineage cited elsewhere at Google DeepMind); defense primes & security ISVs expected to bundle fine-tuned OSS checkpoints for SOC automation (thematic).

The Ecosystem Map (Huang’s Current Focus)

Silicon → system integrators: Claims largest supply chain, ~7 chip roles per AI computer, relentless bottleneck rotation.
Neo-cloud anchors: Public print on CoreWeave + named Nebius / Nscale as prior strategic supports; highlights $1 → $9 co-investment structure.
Model customers: Notes OpenAI & Anthropic gross margins “extremely positive” recently—capacity race implication for accelerator pull.
National scale: Positions AI as job-creator ($100B startup funding → jobs cited) and grid modernizer; open to backstopping land/power/shell financing domestically.
Geopolitics: US should lead on best SKUs while still exporting where legal; references H200 class as example of tiered China access framing in discussion.
Anthropic: Praises culture & Claude Code / agent contributions; hopes USG resolves contract frictions—national competitiveness angle.

Key Risks

Perception risk: Huang warns US fear could forfeit adoption lead even if tech leadership persists.
Regulatory / dual-use: Cyber-oriented frontier models invite narrowing deployment—could slow certain closed API monetization arcs.
Energy & interconnection: AI factories still gated by site power / land / shell timelines—projects stall even with GPU allocation.
Circular deal optics: Neo-cloud financing scrutiny resurfaces whenever GPU lessor multiples compress.
Concentrated customer capex: If frontier labs throttle spend post-IPO (Huang hints OpenAI / Anthropic may need less Nvidia balance-sheet help later), leasing mix could shift.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	Nvidia (NVDA)	Thousand-X agentic compute + integrator-of-$4–5M racks + anchor neocloud signaling	Data-center GM resilience as model customers hit positive gross margins
2	CoreWeave (CRWV)	Validated strategic neocloud cohort; Huang cites co-investor outcomes explicitly	Forward ASP / backlog vs. renewal waves on aging Hopper
2	TSMC (TSM)	3D packaging + CoWoS-like flows underpin Vera Rubin-class integration	Packaging revenue mix vs. leading-edge logic
2	SK Hynix (HXSCF) / Micron (MU)	“Most advanced memories” inside photonic, liquid-cooled racks	HBM share & pricing vs. smartphone DRAM rationing
2	Lumentum (LITE) / Coherent (COHR)	Silicon photonics penetrates rack scale-up domains per Huang system description	Datacom laser backlog & cloud AI buildouts
3	Vertiv (VRT)	Liquid cooling + facility power alignment to “land power and shell” financing theme	Thermal content per kW deployed in new AI factories
3	Synopsys (SNPS)	Agentic automation of codification expands EDA / verification seat intensity (agent-as-tool-user arc)	SNPS AI-attach growth vs. legacy EDA multiples

Monitoring Checklist

Agentic inference share-of-wallet — Validates ~1000× compute step-up vs single-shot chat.
Vera Rubin / rack ASP & bill-of-materials tear-downs — Confirms photonics + memory + cooling $ content slope.
Hopper / legacy GPU secondary pricing indices — Tests “good wine” depreciation thesis.
Neo-cloud renewal & co-investor returns — Cross-checks anchor-investor signaling story.
Open-weight cyber playbooks at Fortune 500 SOCs — Tests Mythos → OSS swarm defense narrative.
US grid + site PPA announcements tied to AI factories — Tracks energy as swinging constraint vs silicon.

Bottom Line

Agentic workflows are not marginal; they reshape the compute multiplier—plan for orders-of-magnitude more per user outcome than generative demos, not incremental token adds.
The investable edge is migrating from “GPU die” to “photonics + thermals + packaging + megasystem integration”—Huang’s $4–5M rack description is the fundamental unit of analysis.
Balance-sheet scale can move physical supply chains faster than subsidy alone—the $500B orders anecdote reframes US reshoring as demand-led, not grant-led.
Investor skepticism on neocloud “circularity” underweights information value of strategic anchors—pipeline visibility may justify repeat equity/debt appetite if secondary GPU pricing stays firm.
Open-weight cyber stacks may be the asymmetric answer to frontier coding models—a tailwind for OSS distribution + security automation franchises, not only the closed API duopoly narrative.