From DeepSeek to Alibaba: False Narratives Haunt Nvidia
Mainstream media chases headlines and scant facts, analysis, and distinctions.
Okay, that’s no surprise! But it’s annoying when it affects our investments as it has done repeatedly to Nvidia (NASDAQ: NVDA). Every time a new player enters the conversation, whether a hyperscaler, a custom chipmaker, or a Chinese tech giant, the story is told as if Nvidia’s entire franchise is under existential threat. Investors panic, Nvidia’s stock falls, and the financial press chalks it up as evidence of “competition” finally catching up.
When the details are scrutinized, the headline story is revealed as nonsense. These AI entrants are never competitors with Nvidia. Nvidia’s high-performance training GPUs are not the same thing as lower-performance inference chips. Yet the headlines have done their damage. When was the last time the New York Times or Wall Street Journal ran the headline “We Wuz Wrong! (Again)”?
DeepSeek and the Panic of January 2025
The first major misstep of this year came in January with the arrival of DeepSeek, a Chinese startup that claimed to have dramatically improved the efficiency of AI training through software. The headlines wrote themselves: “Nvidia’s Moat Under Siege.” CNBC ran segments on whether DeepSeek would “crater GPU demand.” Bloomberg speculated that “Nvidia’s era of unchecked dominance may be ending.” Nvidia’s stock price dropped sharply, wiping out billions in market value.
The claim collapsed under the simplest scrutiny. DeepSeek had not invented a new chip. It had not built a rival platform. It had devised a software layer that made existing Nvidia GPUs run more efficiently by managing memory, load balancing, and computation across large clusters. The DeepSeek solution was dependent on Nvidia’s CUDA software platform running on top of Nvidia’s GPUs. DeepSeek’s innovation made Nvidia’s hardware more valuable.
As I explained in my March 7, 2025, Substack article entitled Nvidia Benefiting from DeepSeek in AI Training While Broadcom and Marvell Gain from AI Inference, the panic was unwarranted. DeepSeek’s software still depends entirely on Nvidia GPUs.
Yet Nvidia CEO, Jensen Huang, must take some responsibility for the PR disaster. A few simple clarifications could have calmed the market: that training GPUs remain essential, that CUDA is the foundation, that software innovations deepen rather than replace Nvidia’s value. Instead, Huang said little, and the false analogy — software efficiency as hardware replacement — spread. Investors bore the cost of that silence. History is perfectly clear on the point: the more customers can do with software, the better the hardware sells.
Broadcom and Marvell: The Mid-Year Overreaction
In mid-2025, another round of headlines emerged, this time off the back of strong earnings at Broadcom (NASDAQ: AVGO) and Marvell (NASDAQ: MRVL). Both companies are deeply respected in the semiconductor industry. Broadcom has long been the dominant provider of custom ASICs for hyperscalers, especially in networking. Marvell has a strong footprint in DPUs and storage accelerators. When both firms reported better-than-expected quarters, the press quickly drew a line to Nvidia. “Are hyperscalers moving away from Nvidia GPUs?” was the headline.
Nvidia’s share price sagged again, even though nothing fundamental had changed about its training GPU business. Broadcom and Marvell are not in the same competitive lane. Their chips serve inference and specific networking functions. They may save hyperscalers money in certain workloads, but they cannot replace Nvidia’s A100, H100, or the new B100 and B200 GPUs that sit at the heart of AI training clusters.
I laid this out in my May 18, 2025, Substack article Marvell and the AI Infrastructure Curve: What to Expect from Earnings, and Why It Isn’t a Nvidia Competitor, and my June 30, 2025, Substack article Broadcom and the Rise of Custom AI Silicon, these companies make excellent inference and networking chips. But they do not compete head-to-head with Nvidia’s training silicon.
Yet again, Huang left the distinction unspoken. The silence allowed the false narrative — that strong quarters for Broadcom and Marvell must come at Nvidia’s expense — to persist.
Alibaba: China’s Political Substitute
Most recently, in late August and in light of U.S. export controls, Alibaba (BCBA: BABA) announced a new inference chip designed to serve Chinese data centers. The Wall Street Journal framed it as “filling the Nvidia void.” CNBC asked whether “China’s hyperscalers are turning the page on Nvidia.” Investors watched Nvidia’s stock fall more than 3%, its steepest weekly dip in months.
Yet, once again: Alibaba’s chip is for inference, not training. It is not built to train GPT-5, Gemini, or the next generation of multimodal models. It is designed to run inference workloads within a geopolitically constrained environment. It exists not because Alibaba can match Nvidia’s performance, but because U.S. restrictions on the H20 chip (now lifted, conditionally) forced Chinese companies to develop substitutes.
Nvidia’s valuation and earnings growth are powered by training GPUs — the highest-performance chips on the planet, designed for memory bandwidth, interconnect speed, and massive parallelism. Inference chips are cheaper, narrower in scope, and politically driven. Yet the press missed the distinction, and Nvidia once again paid the price. Huang’s silence was deafening when a single clear statement might have reset the narrative.
Training Versus Inference Blind Spot
The common thread in these episodes is the media’s failure to distinguish between training and inference. Training makes Nvidia’s moat. It requires the fastest GPUs, terabytes per second of HBM memory bandwidth, and interconnects like NVLink and NVSwitch that stitch thousands of chips together into AI factories. This is where Nvidia’s CUDA platform is unrivaled, where its software and hardware integration create lock-in, and where its margins are fat.
Once a model is trained, it can be run on a wide range of hardware: CPUs, ASICs, DPUs, and cheaper GPUs. Confusing inference hardware with training hardware is like confusing scooters with building Formula 1 cars because they both have wheels and engines.
The Responsibility of Leadership
Huang has repeatedly failed to make distinctions that matter deeply to investors. He is quick to highlight Nvidia’s triumphs when earnings are strong. He has no hesitation in presenting the company’s dominance in glowing terms when speaking at Nvidia’s GTC Technology Conference (GTC). But when misinformation drags the company’s stock down, he too often says nothing.
A CEO’s job is not only to build a company but to explain it to the world and especially to capital markets. Nvidia’s biggest risks are geopolitical, not technological. (Most of that risk comes from Washington.) Investors pay the price for Huang’s silence.
Investors must recognize two truths:
- Media headlines will continue to conflate inference with training. Each time a new competitor emerges, expect the “Nvidia killer” storyline to resurface.
- Jensen Huang is failing to clarify these distinctions. Until he does, Nvidia stock will remain prone to sharp selloffs on superficial news.
The press will keep repeating its mistakes. But Huang’s silence is its own mistake and should be factored into investors’ expectations. Nvidia remains the gold standard in training, but it does not have a gold standard communicator at its helm.