Whoever buys AI chips has historically bought Nvidia. The Santa Clara company held approximately 85 percent of the AI accelerator market early 2026, and its revenue recently surged 85 percent to $81.6 billion. Yet AMD's stock rose 142 percent this year, while Nvidia gained only four percent. The market is pricing in something not yet visible in quarterly earnings: a structural shift away from universal graphics processors toward specialized chips optimized for specific AI tasks.
What is the difference?
A GPU is a universal parallel processor. Originally developed for computer graphics, it has become the workhorse for machine learning because it executes thousands of operations simultaneously. Nvidia sells primarily these GPUs: flexible, programmable chips that work across many different tasks.
An ASIC is the opposite: a chip optimized for a single function. Google builds ASICs for its AI service (Tensor Processing Units, TPUs); Amazon for its AWS infrastructure (Trainium chips). They are faster and more energy-efficient for their specific purposes but inflexible for anything else.
What changed in 2026: The AI industry is splitting into two phases. Training a large model—learning once from massive datasets—is a task where Nvidia GPUs remain unmatched. Inference—answering millions of daily user queries—is a repetitive, standardized task where an ASIC can be vastly more efficient. The training market remains Nvidia-dominated. The inference market is opening up.
Why are tech companies buying AMD now?
Market research firm TrendForce forecasted in January 2026 that shipments of customer-specific ASICs will grow 44.6 percent this year, while merchant GPUs grow only 16.1 percent. The ASIC share of the AI server market will rise from roughly 15 to nearly 28 percent.
Simultaneously, alternatives are catching up on training chips. Meta signed a partnership agreement with AMD in February 2026 worth up to 100 billion dollars, providing for delivery of AMD's Instinct MI450 processors. TechCrunch reported that Meta aims to reduce dependence on a single chip supplier while working on "Personal Superintelligence." OpenAI signed a separate contract in October 2025 for six gigawatts of AMD Instinct GPUs across multiple generations; the first gigawatt is scheduled for delivery in the second half of 2026.
AMD's Instinct MI400 series, unveiled at the Consumer Electronics Show in January, uses CDNA-5 architecture with 432 gigabytes of HBM4 memory, 19.6 terabytes per second memory bandwidth, and 40 petaflops in FP4 precision. Currently available variants include the MI455X for training and inference, and MI430X for high-performance computing.
What are startups doing differently?
Alongside AMD, specialized startups are targeting the inference market. SambaNova Systems unveiled the SN50 chip in February 2026, based on a proprietary Reconfigurable Dataflow Unit architecture, claimed to be five times faster than Nvidia's B200 GPU for agentic AI inference, with three times lower total cost of ownership. SambaNova simultaneously raised 350 million dollars in fresh capital; external benchmarks are absent since the chip hasn't shipped yet.
d-Matrix from Santa Clara, California, entered full production in June 2026 with its Corsair platform. The company combines Corsair accelerators with Nvidia GPUs in a hybrid system. CNBC reported tests at Gimlet Labs: A language model's response time dropped from 24 seconds to under two seconds—a tenfold improvement. The gain lies in inference operations, where GPU capacity often sat idle while awaiting user queries.
What does this mean for Nvidia's position?
Nvidia is not passive. The company develops its own inference-optimized products and markets both Hopper and Blackwell architectures, each covering training and inference. Nvidia's strength is its ecosystem: The CUDA stack, which developers have used for nearly 20 years, is deeply embedded in development pipelines. Switching to AMD or an ASIC means substantial effort, engineering hours, and risk for any company.
Yet AMD's 142-percent stock gain versus Nvidia's four percent shows investors believe the shift is coming. The GPU share of the AI server market will fall from 83 to 70 percent according to TrendForce. This is no disaster for Nvidia, but the end of monopoly comfort.
AMD stock as a leading indicator: Deliveries starting H2 2026
The first major test for AMD comes in the second half of 2026 when MI400 deliveries to OpenAI ramp up. Analysts expect AMD's AI market share could grow to ten to twenty percent by 2030. Whether SambaNova and d-Matrix make the jump from pilot installations to mass production decides in the same twelve months. For cloud customers, it's good news: more competition in training and inference will eventually drive API costs down.