Big Tech was only beginning to move out of the house Nvidia built. Led by Jensen Huang, the GPU giant cornered the datacenter market and swept to a peak valuation of around $5.7 trillion in a dizzying rise. Now, the hyperscalers are positioned to benefit from any Nvidia misstep, just as technology and circumstance are making exactly that a likely future. Anthropic, less than a week out from a long-awaited IPO, may suffer in the short term.
Friction is the name of the game in the AI infrastructure buildout. Apart from getting enough power, permits and processors, tech companies building their datacenters couldn’t afford to think twice about their procurements. Nvidia’s GPUs continue to be dominant for AI training and carry an enormous premium. So much so that, under Jensen Huang’s leadership, GPU generations have accelerated at an astounding rate. The goal was clear: nobody’s getting past Nvidia’s lead.
In inferencing, that lead has essentially become a moot point. Firstly, there’s far more demand than Nvidia hardware could ever hope to satisfy. Secondly, efficiency is key for those running AI workloads at scale, and optimized, specialized processors are all around us nowadays. Google has TPUs, Amazon has Trainium, Microsoft has Maia. Even OpenAI, an AI model maker, is getting in on the act with Jalapeño, an inference chip it designed and brought to production with Broadcom. These are all indications that Nvidia’s omnipresence is far from certain in the day-to-day running of AI.
That isn’t to say the hyperscalers are where they want to be. All signs point to Nvidia being key for major training runs, with only the occasional mention of a successful model being created without it. This is where Nvidia has some time to find a CEO that can continue the frantic pace of development. Needless to say, the engineers actually getting the technology in place haven’t changed. That is until they, perhaps, someday soon, find themselves unwilling to rise to whatever demands a different CEO may place on them.
CUDA’s moat is drying up
Remember ZLUDA, the translation layer that got CUDA workloads running somewhat competently on non-Nvidia hardware? AMD funded its development for a time, then dropped it in early 2024. Nvidia’s CUDA EULA forbids translation layers of this kind, and in August 2024 AMD’s lawyers had the open-sourced ZLUDA code taken down, although the developer said Nvidia itself never contacted him. The upshot was that CUDA retained its exclusive edge, and AMD went back to the drawing board with ROCm and its many trials and tribulations.
Think of how trivial it now seems to poke holes in that CUDA dominance. AI agents, while fallible and inconsistent, are helping developers find ways to optimize workloads and essentially port entire codebases within hours. These may not be perfectly optimized or indeed maintainable in the long run. No software layer, regardless of the legalese, is impenetrable any longer.
At the same time, Nvidia’s then-CEO said at GTC Berlin in October that running all of Nvidia’s software for AI agents costs about $2,000 per GPU per year. This was a bit odd given that Nvidia AI Enterprise already costs $4,500 per GPU per year. At any rate, it was just an example, albeit one to take seriously. It’s just that the CEO who proposed it has now exited stage left. What are Nvidia’s plans going forward now that the software it is increasingly turning to for added profits is more vulnerable than ever?
China won’t be hooked
The China angle to the Jensen Huang scandal is obviously well-trodden ground. Huang admitted to an affair with a former executive of Megaspeed, a Singapore-based Nvidia customer that is under U.S. criminal investigation for allegedly smuggling Nvidia chips into China. However, in all the controversy around that alleged smuggling scheme and potential skirting of export controls at a vast scale, one key element often gets lost. China already uses a bunch of lesser Nvidia GPUs, judging by troves of leaked videos, images and reports from multiple media outlets. It is repurposing consumer chips, soldering on extra memory, all clear signs that China is eager to gain access to what Nvidia would very much like to sell it.
Without China, one of the two most significant AI buildouts is lost to Nvidia. It is clear why Huang went to such lengths to argue against restrictions, and why China hawks in Washington have seen those restrictions as important in recent years. But China will not be hooked, and just like the hyperscalers, it is finding ways not to use Nvidia. In 2025, Beijing told its tech giants to stop buying the H20 and RTX Pro 6000D, chips Nvidia had designed specifically for China. Even after Washington approved H200 exports, Beijing has withheld approval for large-scale imports. What followed was news of Huawei’s AI chip progress. Those chips are held back by the lack of EUV equipment and reportedly suffer from poor yields, but Huawei expects around $12 billion in AI chip revenue this year, and Nvidia’s own filings describe it as effectively foreclosed from China’s datacenter market.
Anthropic faces incredible uncertainty
Essentially, Nvidia was facing pressure from all sides. It was always likely to see a tail-off to its growth rate, with or without Huang at the helm. But the specific circumstances we find ourselves in today are having one especially bad effect on the prospects of one of its most prolific customers: Anthropic. With its Nasdaq debut set for November 19, the roughly six times oversubscribed offering is a bellwether for AI stocks in general. Several of those stocks, Nvidia among them, have given up all their gains for 2026. While signs of recovery exist, as we have seen time and again with stock turbulence over the past three years, the general sentiment around AI has turned skeptical in a decisive way.
Once again, much of this predates Huang’s departure. However, the exit of the AI industry’s most fervent proponent, a close ally of President Trump and a key player in accelerating Nvidia’s roadmaps, will cause short-term worries that may leave Anthropic’s once-delirious valuations looking far removed from reality. Nvidia itself has committed $10 billion to a private placement alongside the IPO. And you know why Anthropic needs that IPO to go well? To pay for compute, much of which runs on Nvidia GPUs.