AI Infra Summit 2026 Amid a history lesson on chip advances since 1965, Nvidia and Intel outlined their plans for AI computing.
Akul Saxena
SANTA CLARA, Calif., Sept. 16, 2026 — In the Tuesday kickoff keynote at the AI Infra Summit here, Dave Patterson, a distinguished engineer and fellow at search giant Google, traced the role that chip advances have played in artificial intelligence computing.
The push to make AI complete tasks autonomously, Patterson said, is increasing demand for central processing units, or CPUs, the chips that run programs on everyday computers.
When frontier lab OpenAI brought the ChatGPT to mass attention in late 2022, much of the hardware discussion focused on graphics processing units, or GPUs. Those chips perform many calculations simultaneously, providing the computing power to train AI models and generate answers.
AI agents go further by using software tools to act on those answers, Patterson said. A model might write a program to analyze a company’s finances, then use a CPU to run the program and return the results.
AI chipmaker Nvidia, best known for its GPUs, is promoting its Vera CPU for that additional work. Slow execution of software tools can delay an agent even when its AI processors are ready, said Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, also speaking on the main stage at the conference here following Patterson.
Agents can repeat those steps without waiting for another human instruction. While a chatbot user might to read an answer, an agent can immediately request another calculation or run another tool, increasing the computing required to complete an assignment.
Electricity use, Moore’s law and Dennard scaling
The growing demand for CPUs follows decades of investment in making chips faster and more efficient, according to the history outlined by Patterson.
In 1965, semiconductor researcher Gordon Moore observed that the number of electronic components on a chip was doubling annually. His prediction, later revised to doubling every two years, became known as Moore’s Law and helped guide chipmakers’ investment for decades.
Moore co-founded Intel with Robert Noyce in 1968.
Putting more components to work also required controlling electricity use. In 1974, researchers at computer manufacturer IBM explained how shrinking transistors, the switches that perform computing operations, and lowering their voltage could keep power consumption steady per unit of chip area.
That relationship became known as Dennard scaling. Smaller switches needed less electricity, allowing manufacturers to fit more onto a chip and run them faster without a corresponding increase in heat.
By the mid-2000s, the tiny switches were increasingly leaking electricity even when turned off. The leakage produced heat and limited further voltage reductions, making additional speed harder to achieve within the same power limit.
Patterson said chip technology continued improving, but much of the gains manufacturers once relied on had slowed. Designing processors for particular calculations and use cases became an increasingly important way to improve performance.
The rise of the Tensor Processing Unit (TPU)
Google faced a concrete reason to invest. In 2013, its researchers projected that three minutes of daily voice search per user would require doubling the company’s data center computing resources using conventional CPUs.
In response, Google developed its tensor processing unit, or TPU, a chip designed for machine learning calculations, and deployed the first version in 2015. The specialized chip could perform those calculations more efficiently, reducing the need to buy additional CPUs.
Researchers also found that some AI calculations required less precision than conventional scientific computing. Using 16 bits rather than 32 bits to store each value, for example, reduced memory requirements while preserving sufficient accuracy for certain tasks.
Patterson said he had not expected renewed innovation in floating-point computation, the method computers use to calculate with numerical approximations, during his lifetime.
Agents need more computing
By 2026, AI agents had increased demand for CPUs, the processors that run ordinary software, alongside the GPUs that run the models themselves. Nvidia designed Vera to execute the code and the software tools an agent calls, Buck said, so the model can read the results and continue its task.
An agent writing software, for example, might use a GPU to generate code and a CPU to run tests. The model then uses the test results to decide whether the code needs revision, so slow CPU execution can delay the assignment even when the GPU is ready.
The requests themselves got larger. The text a model processes before answering has grown from roughly 1,000 tokens in the 2023 benchmarks to about 142,000, Buck said, and each length requires its own software optimization.
An agent testing software must keep track of the code it wrote and the errors its tests found. The model uses that information to decide what to change, increasing the amount of data the system must store and retrieve as the assignment continues.
Meeting that demand requires more memory, the hardware that holds data for processors to use. Patterson warned that manufacturers could not quickly increase production since building additional factories takes years.
Processors also need fast connections to exchange data as they complete different parts of an assignment. Nvidia sells that connecting technology through NVLink Fusion, which allows chips designed by other companies to operate within Nvidia’s computing systems.
It includes a small connecting chip that companies can package with their processor, plus the designs needed to make the connection work. Customers can therefore build their own AI processors and still buy Nvidia technology to use them together.
The other Santa Clara chipmaker
Nvidia is developing CPUs while partnering with longtime CPU supplier Intel. Also based in Santa Clara, Intel agreed in September 2025 to build custom data center processors for Nvidia, which announced a $5 billion investment in the struggling company
Intel CEO Lip-Bu Tan said Tuesday, also at the AI Infra Summit, that the partnership combined Intel’s CPUs with Nvidia’s GPUs and technology connecting the chips. He said Intel sought the investment to “strengthen its balance sheet first.”
The U.S. government also has a financial interest in Intel’s recovery. In August 2025, it agreed to acquire a 9.9 percent stake for $8.9 billion, using previously awarded funding to support domestic chip manufacturing.
Tan said chip manufacturing and advanced packaging, the process of assembling and connecting chips, required substantial investment. The Nvidia agreement covered Intel-built CPUs but did not announce that Intel would manufacture Nvidia’s GPUs.