Nvidia is down roughly 21% from its 52-week high, the chip sector is flat on the year, and the fear trades are back. So is Nvidia stock done? According to semiconductor analyst Tae Kim, the answer is a clear no — and if anything, this selloff rhymes almost perfectly with what happened a year ago. "It's like Groundhog Day," Kim said in a recent conversation. "Everyone freaked out about DeepSeek, everyone freaked out about tariffs — and then the stock ripped." The fundamentals, he argues, have never been stronger.
Is Nvidia Stock Over After Its 21% Drop?
The short answer: no. The longer answer involves understanding what's actually driving the selloff versus what's driving the underlying business. Kim points out that a year ago, Nvidia suffered a 30% drawdown while its business was "actually flying" on a fundamental level. The macro fear — DeepSeek efficiency gains, Trump tariff wars, now Iran tensions and oil prices approaching $100 — created a sentiment vacuum that had nothing to do with GPU demand.
02:15
Tae Kim comparing the current Nvidia selloff to the DeepSeek panic from a year ago — calling it 'Groundhog Day' for the chip sector
Watch at 02:15 →
"Oil can't be $100 forever, and Trump will probably backpedal in the next few weeks," Kim noted. The fear is real but the cause is external. Meanwhile, inside the walls of Meta, Google, and Nvidia itself, engineers are reporting something very different: they can't get enough compute. People are hitting rate limits. Vibe coders on X are running multiple subscriptions across model providers just to keep up with their token needs. That's not a picture of a market that's peaked.
Why Is AI Chip Demand Still Exploding in 2025?
The driver right now is inference — specifically, the inference demands created by AI coding assistants and agentic workflows. Kim spoke with dozens of engineers at Meta, Google, and Nvidia during GTC and came back with the same message from all of them: crazy inference demand and AI compute shortages across the board.
This isn't just anecdote. OpenAI is pivoting hard toward coding agents. Anthropic is doing billions in ARR with growth that seems to accelerate every few weeks. And the gold-rush dynamic is visible in wild ways — people are literally running bots to snap up any available B200 GPUs the moment they come online, like sneaker bots for NeoClouds.
07:40
Discussion of Jensen Huang's Groq acquisition and how it fits into Nvidia's inference strategy alongside Vera Rubin
Watch at 07:40 →
Jensen Huang, characteristically, saw this coming. He locked up supply agreements for memory, co-packaged optics, and connectors ahead of the curve. "He's very prescient," Kim said. "He probably saw this demand months away."
Beyond coding, the next wave is broader still. At GTC, a session between Google's Jeff Dean and Nvidia's Bill Dally outlined several incoming vectors: context window innovations that can focus intelligently on relevant documents, stacked memory directly on top of GPUs and TPUs, and synthetic data generation for audio and video that opens entirely new training frontiers. AI models, both chief scientists agreed, have an enormous runway ahead.
What Is the Nvidia-Groq Deal and Why Does It Matter?
Nvidia's acquisition of Groq's assets and team is, in Kim's view, one of the most strategically elegant moves Jensen has made since acquiring Mellanox in 2019. The logic of the Mellanox deal was that Jensen saw the world shifting to massive 10,000-to-100,000 GPU clusters and understood that networking would be the bottleneck. With Groq, the logic is similar: he saw low-latency inference becoming critical for agentic AI and moved to own that capability.
The combination works like this: roughly 75% of inference workloads run on Vera Rubin (Nvidia's next-generation architecture), while the remaining 25% — the ultra-low-latency, time-sensitive tasks — runs on Groq's silicon. Together, they cover the full spectrum of inference demand economically and efficiently. "It's like the perfect combination to take advantage of this," Kim said. Ian Buck and Jensen himself highlighted the pairing at GTC as central to Nvidia's inference strategy.
18:22
Kim breaking down the hidden CPU shortage story and why AI agents are driving a 4x increase in CPU core demand
Watch at 18:22 →
Is a Semiconductor Supply Shortage Coming?
Yes — and it may already be here. TSMC is not ramping capital expenditure fast enough to meet what could be another 10x increase in compute demand over the coming years. Every major hyperscaler with custom silicon (Google's TPUs, Amazon's Trainium, Microsoft's Maia) is fighting for more wafer capacity. The queue is brutal.
Nvidia's advantage here is relationship-driven. Jensen visits TSMC five or six times a year, speaks at their employee events, and has cultivated a partnership deep enough that Nvidia gets priority wafer allocations — and can prepay tens of billions to lock them in. "They're the biggest dog in the house," Kim said.
Samsung and Intel remain the only credible alternative fabs at the leading edge, and there's a scenario — particularly relevant now that the US government holds a stake in Intel — where a coordinated deal brings major AI buyers to the table to guarantee offtake agreements that justify Intel's fab investment. It's not a sure thing, but the incentive structure is there.
Why Do AI Agents Need So Many More CPUs?
This might be the most underappreciated story in semiconductors right now. The ARM CEO recently noted that AI infrastructure is requiring four times more CPU cores compared to last year's models. Dell, AMD, and Intel's CFO have all flagged hyperscalers locking in three-to-five year CPU supply contracts.
The reason is architectural. AI agents don't just run on GPUs — they orchestrate. Every tool call, database query, web search, and workflow routing decision runs through a CPU. As agentic systems become more complex, with dozens of parallel agents operating simultaneously, CPU demand scales with it. Kim frames it plainly: "The whole thing requires orchestration, and that's all handled by the CPU."
This CPU shortage narrative isn't consensus yet, which is exactly what makes it interesting from an investment standpoint.
Will GPU Prices Crash as Newer Chips Arrive?
"Depreciation gate" — the fear that H100s and older GPUs would become worthless within months as Blackwell flooded the market — has so far proven to be a non-issue. CoreWeave has reported that GPU assets are lasting five to six years and still commanding 90-95% of original pricing. Rental prices for hardware that's six years old are still selling out.
Could it become a problem? Yes — but only in a scenario where AI demand stalls and a genuine compute glut emerges. Kim doesn't think that's happening. "The AI compute demand outpacing supply is so large that this is not an issue right now." In a bubble scenario, all bets are off. But the evidence on the ground points the other direction.
Is Meta Stock a Buy Despite Massive AI Spending?
Kim has a complicated history with Meta — he's called it cheap before, it dropped another 30%, then he was right in the long run. But his conviction on the core thesis hasn't wavered: nobody is replacing Instagram, nobody is replacing Facebook, and billions of people will keep using those platforms regardless of what happens in AI.
More interesting is his counter-intuitive take on Meta's competitive position in the AI era. Google faces existential search risk as AI chatbots eat into query volume. Meta doesn't. If anything, Meta's ad targeting and recommendation engine only gets more powerful as AI improves it. Even if Meta's frontier AI spending is partially wasted — and Kim acknowledges Zuckerberg may burn $100 billion chasing a frontier model position — the core ad engine that generates the cash is untouched. "That money-making engine is not going to be affected by this."
The open-source strategy is also working in Meta's favor. Llama models are driving developer adoption, ecosystem lock-in, and GPU utilization — all without requiring Meta to win the frontier race outright.
Can Elon Musk Actually Build a Competitive Chip Fab?
Kim is skeptical, and the reasons are structural. TSMC and Intel represent the only realistic leading-edge alternatives, and TSMC is already supply-constrained. Acquiring ASML extreme ultraviolet equipment, hiring technicians with the right expertise, and building up the institutional knowledge required for cutting-edge fabrication takes decades — not years.
"Chip fabs are almost like cooking," Kim said. "It takes a lot of trial and error accumulated over decades. It's not something you can just jump right in and do." The more credible XAI play, he suggests, might be something closer to a neocloud model: Starlink satellites embedded with GPUs, serving as distributed AI compute infrastructure that other companies can rent. That model — infrastructure as a railroad — fits SpaceX's existing playbook far better than trying to out-TSMC TSMC.
The bottom line across all of these threads is the same: the AI compute boom is not over, it's not even close to over, and the current fear cycle looks almost identical to the one that resolved sharply upward twelve months ago. The fundamentals are intact. The demand is real. And Jensen Huang, for better or worse, seems to be about three moves ahead of everyone else on the board.
