Nvidia stands alone at the top of the AI hardware world. Its GPUs power the vast majority of model training and inference. CUDA, the company’s software platform, keeps developers locked in. Yet that grip shows signs of loosening. Rivals have spotted an opening. They target the biggest customers with chips built for one job only.
Hyperscalers spend billions each year on compute. Training frontier models demands tens of thousands of accelerators running nonstop. Inference adds another massive bill. A 20 to 30 percent cut in those expenses changes the math. It frees up cash. It lowers risk. So Amazon, Google and Microsoft design their own silicon or partner with specialists who do. These application-specific integrated circuits trade flexibility for efficiency. They run specific AI tasks faster and cheaper than general-purpose GPUs. Yahoo Finance laid out the case two days ago.
But Nvidia does not sit idle. The company struck a $20 billion licensing deal with Groq late last year. It brought in top talent from the inference specialist. Groq’s Language Processing Units deliver blazing speeds on token generation. They shine in real-time applications where latency matters most. Nvidia now prepares versions of that technology for restricted markets, according to Reuters reports from earlier this year.
Startups raised record sums in 2026. AI chip newcomers pulled in $8.3 billion globally. Cerebras Systems landed $1 billion in February. MatX, Ayar Labs and Etched each closed $500 million rounds. European firms Axelera and Olix secured more than $200 million apiece. The money flows because hyperscalers want options. They refuse to depend on one supplier. CNBC tracked the funding surge in April.
Cerebras took a different path. Its wafer-scale engine measures 58 times larger than a standard Nvidia die. That size packs enormous on-chip memory. Bandwidth jumps. Inference runs at tremendous speeds. The company filed to go public in recent months. It signed deals with OpenAI and Amazon Web Services. One OpenAI agreement exceeds $20 billion and covers 750 megawatts of compute. AWS will distribute Cerebras chips worldwide. Shares jumped after the IPO. Yahoo Finance covered those moves in April.
AMD fights on two fronts. Its Instinct MI300 and upcoming MI400 series GPUs chase Nvidia in training. The company also partners with Cerebras. The duo announced a disaggregated compute platform this week. It combines AMD GPUs with Cerebras SRAM-powered accelerators. The target? Ultra-low latency inference for agentic AI workloads. AMD CEO Lisa Su unveiled the collaboration during her Advancing AI keynote. It fills a gap that once cost Nvidia $20 billion to close through Groq. The Register reported the partnership Thursday.
Broadcom plays a quieter but critical role. It designs custom ASICs for Google and others. The firm also supplies networking silicon that ties everything together. Its AI revenue grew 32.3 percent last year. Analysts still see 13.4 percent upside. Investors who look past the GPU wars often land here. Memory makers like Micron ride the same wave. Their high-bandwidth memory sells into every major system. Investing.com highlighted six stocks beyond Nvidia just two days ago.
Intel pushes its Gaudi 3 accelerators. The company holds strong enterprise relationships. Yet it trails in raw performance and software maturity. Market share data from May showed Nvidia at roughly 86 percent of AI GPUs. AMD and Intel sit in single digits for training accelerators. Hyperscalers’ custom chips erode that number further. CompaniesHistory.com compiled the figures in late May.
Google’s TPUs power its own models and serve external customers. Anthropic expanded its deal for multiple gigawatts of TPU capacity. Amazon’s Trainium chips handle training inside AWS. Microsoft works on Maia accelerators. These internal projects reduce dependency. They also force Nvidia to compete on price and availability. OpenAI reportedly collaborates with Broadcom on custom silicon. Anthropic weighs its own designs.
China adds pressure from outside. Export controls block Nvidia’s newest chips there. Local firms step up. Some U.S. companies explore workarounds. Nvidia itself prepares Groq-based chips for the Chinese market, Reuters noted in March. Geopolitics complicates every forecast.
Performance claims fly in every direction. Cerebras says its giant chip beats Nvidia on memory bandwidth and inference speed. Groq touts tokens per second that outpace GPUs in certain benchmarks. AMD promises its Venice CPU will hold a 20 percent edge over Nvidia’s Vera in AI server tasks. Real-world results vary by workload. Software compatibility often decides adoption. CUDA remains the moat. Competitors scramble to match the ecosystem.
Yet the war extends beyond chips. Networking matters. Power delivery matters. Cooling systems matter. Every accelerator, whether from Nvidia or a rival, runs through the same handful of suppliers further down the stack. Applied Materials equips the fabs. TSMC manufactures most advanced dies. Those names rarely grab headlines. Their revenues rise with total AI spend regardless of who wins the GPU battle.
Recent X conversations capture the tension. Users note Jensen Huang’s new push for open models. More models mean more demand for compute. Nvidia benefits. Meta, OpenAI and others balance open and closed strategies for their own reasons. One post summed it up: everyone loves openness that opens doors for their business. Another highlighted AMD’s latest claims against Nvidia’s upcoming CPUs.
Analysts split on the outcome. Some see Nvidia’s lead holding for years. CUDA’s installed base proves sticky. Others predict custom silicon will claim 30 percent or more of hyperscaler spend by decade’s end. Inference workloads fragment fastest. Training still favors flexible GPUs. The split creates room for multiple winners.
Public market reactions tell part of the story. Broadcom and Micron outperformed Nvidia over the past year in some periods. Cerebras stock soared on its debut. Startup valuations climbed on each big funding round. Private capital chases the next architecture breakthrough: photonic computing, analog approaches, novel memory tech. None have displaced GPUs yet. All keep pressure high.
Enterprise buyers watch closely. They want choice. They want predictable costs. They also need talent that knows how to program these systems. Nvidia’s developer community still dominates. Rivals invest heavily in tools and frameworks. Progress shows. It rarely matches the depth of CUDA overnight.
The smartest money may sit in the middle. Companies that enable the entire stack avoid picking one horse. They ride them all. As AI matures, inference takes a larger share of spend. Specialized chips gain ground there. Training clusters still scale with Nvidia iron. Both trends can coexist.
Supply chain realities constrain everyone. Advanced packaging, high-bandwidth memory and rare materials create bottlenecks. Geopolitical friction adds cost and delay. No single vendor escapes those limits.
Watch the hyperscalers’ capex mix. Rising percentages for custom silicon signal market share loss for Nvidia. Stable or growing GPU purchases suggest the moat holds. Early data points to a hybrid future. Nvidia retains the crown for now. The challengers grow stronger. The race stretches years ahead. And the real winners could hide where few investors look.
WebProNews is an iEntry Publication