Every few weeks, another company announces it is "taking on Nvidia $NVDA." The framing is consistent. The ambitions behind it are not.
Some companies are designing chips built only for inference, the job of answering a question or generating an image once a model is already trained. Others want something more ambitious: control over the entire pipeline, from chip design to the factory that builds it. Plenty of others fall somewhere between those two extremes and chase narrower goals, such as cutting a specific cost or trimming their dependence on one supplier. Almost none are attempting a full-spectrum replacement of Nvidia across training, inference, and the open market.
The distinction matters because companies now trying to cut into Nvidia's dominant position in the chips market aren't all chasing the same prize. The only way to tell them apart is to look at what each one actually wants, not at the headline that lumps them together.
The training-inference divide
Nvidia's hold on training — the computationally intensive process of building AI models from scratch — remains its strongest position. According to Silicon Analysts, which compiles estimates from TrendForce, Morgan Stanley $MS, and TSMC $TSM capacity data, Nvidia's share exceeded 90% in training in 2025. Its inference share sat between 60% and 75% over the same period, undercut by growing competition from custom silicon. Between the two ranges, the gap works out to roughly 15 to 30 percentage points. That's exactly the area most challengers are fighting over.
Groq, an AI chip startup, built its Language Processing Unit specifically for inference, prioritizing fast, predictable responses over the flexibility needed for training. The company was "established in 2016 for inference," according to its own product description, and markets itself as "the only custom-built inference chip" for developers. Groq competes for the workload that happens after a model already exists, not for Nvidia's training contracts.
Google $GOOGL's trajectory tells a similar story. Google describes Ironwood, the company's latest Tensor Processing Unit, as its first TPU "designed specifically for inference." Google still builds TPUs that handle training, but the strategic emphasis has shifted toward the segment where Nvidia's grip is weakest. And Google doesn't sell Ironwood as a standalone chip. The only way to use one is to rent capacity through Google Cloud. Buyers are locked into Google's cloud pricing and infrastructure and can't own the hardware outright.
Internal silicon versus the open market
The most important distinction among Nvidia's challengers may be the simplest: Are they selling chips to others, or using them internally?
OpenAI wants out from under Nvidia's pricing and production schedule, not into the chip business. Its partnership with Broadcom $AVGO, announced in October 2025, involves co-developing custom AI chips that will draw up to 10 gigawatts of power, or as much electricity as 10 large nuclear reactors generate. The stated purpose is to let OpenAI "embed what it's learned from developing frontier models and products directly into the hardware." The chips will be deployed "across OpenAI's facilities and partner data centers," never sold to anyone else.
Amazon $AMZN is the closest any cloud provider has come to building a real open-market rival to Nvidia. AWS, Amazon's cloud division, has made its Trainium chips available to external customers through cloud rental service EC2. As of March, 1.4 million Trainium chips were deployed across three generations, according to TechCrunch, with Anthropic's Claude running on more than one million Trainium2 chips alone. Amazon has also committed two gigawatts of Trainium computing capacity to OpenAI, putting its chips inside two of the industry's most prominent AI labs.
AMD $AMD occupies a distinct category as the only major company selling training and inference GPUs to third parties without also being a cloud platform. Its Instinct MI300 series has won customers including Microsoft $MSFT Azure, Meta $META, Dell $DELL, HPE, and Lenovo. AMD CEO Lisa Su has said the MI350 series offers "the largest generational performance leap in the history of Instinct."
The moat that is not made of silicon
Even where challenger hardware matches or exceeds Nvidia's specifications on paper, a separate barrier remains. Nvidia's CUDA software ecosystem, launched in 2006, has accumulated almost two decades of libraries, tools, and developer expertise. According to Nvidia's own annual report from January 2025, more than 5.9 million developers worldwide use CUDA and related tools, across hundreds of domain-specific libraries.
A December 2024 SemiAnalysis benchmark study found that AMD's MI300X, despite higher marketed performance numbers, delivered 14% slower real-world results than Nvidia's H100 and H200 in key training benchmarks. "The CUDA moat has yet to be crossed by AMD due to AMD's weaker-than-expected software Quality Assurance culture and its challenging out of the box experience," SemiAnalysis concluded.
Beating Nvidia's chip is necessary but not sufficient. The ecosystem around it determines whether customers can use it without rebuilding their entire workflow. That's why Amazon has invested in supporting PyTorch, the software most AI developers already use to build models, on its Neuron developer toolkit, and why AMD launched ROCm 7 and a free developer cloud.
What the spectrum of competition reveals
The big cloud providers building their own chips aren't walking away from Nvidia entirely. Daniel Newman, an analyst at the Futurum Group, told CNBC in November 2025: "They want to have a little bit more control over the workloads that they build. At the same time, they're going to continue to work very closely with Nvidia, with AMD, because they also need the capacity. The demand is so insatiable."
Nvidia's overall AI chip market share, estimated by Mizuho Securities at between 70% and 95% depending on segment, is projected by Silicon Analysts to settle near 75% as the total market expands past $200 billion. The market isn't zero-sum. Nvidia can lose share in percentage terms while selling more chips than ever.
No two companies here are running the same strategy. Where Groq picked inference and stuck with it, Google is chasing something narrower: cutting its own cloud costs. OpenAI wants the opposite of narrow. It's locking down its own supply chain so Nvidia's pricing and production schedule can't touch it. Amazon went further and built an entire ecosystem it can actually sell to others. AMD is the one holdout willing to fight Nvidia head-on in the open market.