Jensen Huang Told Investors to Buy at a Discount Two Months Ago. Here's What the Data Actually Says About the AI Infrastructure Cycle.

Jensen Huang stood in Seoul two months ago, amid a selloff that sent the Kospi tumbling and dragged US tech stocks lower on overheating fears and Fed rate-hike anxiety, and told investors: "We're at the beginning of it, and whatever happened to the stock market, you should be very happy because now you can buy at a discount."

The headline this week isn't whether that call paid off. At $223.96, Nvidia is up roughly 30% from where it sat around mid-June, up 20% year-to-date, and its $5.42 trillion market cap sits just below the 52-week high of $236.54. The discount has been more than recovered.

The question that actually matters is the one Huang was pointing to underneath the discount comment — and the one the market consistently misses when it fixates on price levels. We're two months past that selloff, but still at the beginning of the infrastructure buildout. The evidence stack from product architecture to hyperscaler capex commitments supports that claim. The capital allocation question is whether that long runway is priced into the stock already, or whether the return curve is still front-loaded.

The market fixates on the selloff. The signal is in the architecture cycle.

What happened in June was a liquidity event, not a thesis event. Fear of overheating in the AI trade and macro anxiety around rates triggered a broad tech sell-off. The Korea angle — where SK Hynix parried losses after the South Korean president called the domestic market undervalued — was theater around a global risk-off moment.

I deal with facts, not opinions about whether a 10-15% pullback in a parabolic stock is "normal." What I check when the market panics is whether the product cycle has changed, whether the supply chain still has room to grow, and whether the architecture that's driving revenue still has a generation ahead of the competition.

On all three, the answer is yes. And it's more than a yes — it's a structural advantage that gets wider with every architecture generation.

Rubin is in full production. Volume shipments begin in the second half of 2026 to AWS, Azure, Google Cloud, Oracle, CoreWeave, Lambda, Nebius, and Nscale. This isn't sampling. It isn't qualification. Full production. And the architecture that Rubin delivers is specifically designed for the workloads that are eating the AI market right now — inference, not training.

That distinction is critical. Nvidia's CUDA moat was built on training dominance, where the ecosystem lock-in of software, libraries, and developer workflows made switching costs prohibitive. Inference is different. Inference doesn't need CUDA's parallel training orchestration. Inference needs lower latency, lower cost per token, better memory bandwidth for KV cache management, and architectures optimized for the long-context, reasoning-heavy workloads that are driving actual AI product revenue.

Rubin delivers on that exact transition. The Rubin GPU offers 50 petaflops of NVFP4 inference compute, with 288 GB of HBM4 memory at 22 TB/s bandwidth per chip — nearly tripling Blackwell's memory bandwidth. The Vera CPU, with 88 custom ARM cores and 1.5 TB of LPDDR5X memory, handles the orchestration layer that matters as agentic AI systems multiply tool calls and context windows. The full Vera Rubin NVL72 rack — 72 GPUs, 36 CPUs, all liquid-cooled — claims up to 10x reduction in inference token cost compared to the Blackwell NVL72.

Put plainly: Nvidia didn't just ship another GPU. They shipped the architecture that the market is transitioning into. Training dominated the 2023-2025 AI cycle. Inference will dominate 2026-2028. Nvidia is on the right side of that transition, not on the wrong side.

This is what separates Nvidia from the narrative that inference will erode the CUDA moat. The moat wasn't just CUDA — it was the full-stack architecture of GPU, CPU, interconnect, network, and DPU designed to work as a system. Rubin has six chips, not one. That's the moat. AMD ships a GPU. Nvidia ships a data center.

The supply chain doesn't lie: hyperscaler capex is still accelerating

Management quotes from earnings calls are where I start, because Jensen Huang has visibility into hyperscaler procurement pipelines that investors don't. On the Q2 FY2026 call, he said the top four hyperscalers — Amazon, Google, Microsoft, and Meta — were spending roughly $600 billion annually on capex. That figure drew skepticism because it exceeded analyst composites, which placed 2025 spending closer to $400-450 billion across the Big 4 plus Oracle.

Here's what the numbers say now. For 2026, Goldman Sachs and Futurum Group estimates put combined hyperscaler capex at $660-725 billion. That's a near-doubling from 2025 levels, up 77% year-over-year. Amazon is projecting $200 billion. Alphabet is targeting $175-185 billion — revised upward three times from initial estimates. Microsoft is tracking toward $120 billion or more, with an $80 billion Azure order backlog that remains unfulfilled primarily due to power constraints. Meta is guiding $115-135 billion. Oracle is targeting $50 billion, a 136% increase over 2025.

That's not a slowing demand story. That's the opposite. The market was pricing in a capex peak in mid-2026, and the data says the peak hasn't arrived.

What's constraining supply isn't demand. All five major hyperscalers report their markets are supply-constrained — power availability, permitting, physical infrastructure, and chip supply are the bottlenecks. Jensen's $600 billion figure from two quarters ago now looks more like an understatement than an exaggeration.

Goldman Sachs' broader view puts a combined $5.3 trillion in capex for the four largest hyperscalers from fiscal 2025 through fiscal 2030. The aggregate baseline across compute, data centers, and power through 2031 sits at $7.6 trillion. That's not a cycle that's winding down. That's an industrial-scale buildout that's accelerating.

The financials: growth still accelerating, not decelerating

A CEO telling you to buy at a discount only carries weight if the underlying business can justify the claim. Nvidia's latest financial data doesn't just justify it — it suggests the selloff was a dislocation between sentiment and reality.

Q1 FY2027 revenue hit $81.6 billion, up 20% sequentially and 85% year-over-year. Gross margins held at 74.9%. Q2 FY2027 revenue came in at $46.7 billion, up 6% sequentially and 56% year-over-year. Revenue growth across the trailing twelve months is running at 70.68% year-over-year. Free cash flow for the trailing twelve months is $119.1 billion, with a 46.97% FCF margin. Return on invested capital sits at 89.42%. Return on equity is 114.3%. Debt-to-equity is 4.3%. The company carries $13.24 billion in cash against $64 billion in total debt, for a net cash position of $72.1 billion.

The stock trades at roughly 34 times trailing earnings, which looks elevated until you factor in that 70% revenue growth rate — the PEG ratio comes out to roughly 0.31. Forward P/E is around 60x, which reflects the market pricing in continued acceleration. The question isn't whether the multiples are high. The question is whether the growth trajectory that justifies them has a floor.

Consensus estimates for the coming quarters — Q3 FY2027 at $0.75 EPS and $33.2 billion in revenue, Q4 FY2027 at $0.85 EPS and $38.1 billion — represent the bar the market has set. Actuals have beaten consensus in every quarter going back to Q3 FY2024. The most recent Q2 FY2027 actuals were $1.05 EPS and $46.7 billion in revenue, versus consensus of $1.01 EPS and $46.0 billion.

The next earnings report drops after market close on August 26, for FY2027 Q3. The forward-looking revenue consensus for that quarter sits at roughly $44 billion on a quarterly basis. The bar is high, but the pattern of beats suggests the bar may not be the real constraint — execution is.

The risk that changes the risk/reward

Demand is not the issue. The issue is whether timing and opportunity cost still justify the same allocation today.

I've said repeatedly that I still believe Nvidia can reach $20 trillion by 2030 — but much of that return is likely back-half weighted. The stock has rallied from the Seoul selloff levels, and at a $5.42 trillion market cap with a forward P/E near 60x, the market has already done serious work pricing in Rubin ramp, inference demand acceleration, and continued hyperscaler spending growth.

What would change the thesis is not a macro shock or a competitive scare — it's execution failure on the Rubin production ramp. Rubin's supply ceiling in 2026 is estimated at 200,000-300,000 units, constrained by TSMC's N3 process competition with Apple and AMD, and HBM4 yields from SK Hynix and Samsung that remain below mature HBM3e levels. If Rubin ships late or volumes fall short of the hyperscaler pipeline, the growth acceleration narrative loses its engine.

There's also the longer-term question of inference economics. Rubin claims 10x lower cost per million tokens versus Blackwell. If inference costs drop that fast, it could trigger Jevons Paradox — where efficiency drives higher total usage volumes, necessitating more infrastructure. Or it could compress the revenue duration of each architecture generation, forcing Nvidia to rely on faster cycle cadence to maintain growth rates. The annual cadence (Rubin now, Rubin Ultra in H2 2027, Feynman in 2028, Rosa Feynman in 2029-2030) suggests Nvidia is banking on speed to offset efficiency. I believe that cadence is sustainable, but it's a risk worth monitoring.

So what? Where does the capital go?

The debate is not whether Nvidia remains the dominant AI infrastructure company. The evidence is overwhelming that it does — in training, in inference, in the full-stack architecture race, in the hyperscaler pipeline, in developer adoption. Jensen Huang's Seoul comment was not a timing call. It was a statement that the infrastructure cycle has more room to run than the stock price pullback implied.

What the debate actually is: is the return profile from here still as compelling as what can be found elsewhere in the AI trade? Nvidia is up 23% on a rolling annual basis. The stock has recovered the selloff and is approaching its highs. At 60x forward earnings with the next earnings report in 17 days, the near-term risk/reward is less asymmetric than it was in June.

I believe Nvidia remains a core holding for any AI infrastructure portfolio. The Rubin architecture, the hyperscaler capex pipeline, the inference transition, and the annual cadence through 2030 give me high conviction in the long-term thesis. But conviction and allocation are different things. A stock that's up 30% from its low in two months is a different entry point than a stock that's down 30%.

If you entered at the Seoul discount, the thesis has already delivered. If you're looking at Nvidia now, the question is whether you're comfortable buying into a stock that's approaching its highs, priced at 60x forward earnings, with a binary earnings event in less than three weeks — or whether the return profile is better deployed elsewhere in the AI infrastructure chain, where the market hasn't yet caught up to the same demand signals.

That's the capital allocation question Jensen Huang was actually asking in Seoul. Not "buy now." "The cycle has more room to run. Decide where to stand."