“The days of millions of GPU data centers are coming,” said Nvidia CEO Jensen Huang during a keynote at Computex. Huang’s predictions are becoming bolder and bolder; he recently said programmers won’t need to learn coding because of AI. But Huang can afford a few wild guesses, given that he is presiding over a multitrillion-dollar company.

By comparison, underdogs Intel’s Pat Gelsinger and AMD’s Lisa Su were more grounded in reality during their Computex keynotes. The companies are in catch-up mode to Nvidia.

Su focused on new products that would destroy Nvidia’s kingdom. Gelsinger unexpectedly focused on x86 power efficiency and products that would fend off ARM, which is threatening Intel’s dominance in PCs and servers.

For Nvidia and AMD, AI is a goldmine, and the companies expanded their GPU and CPU roadmaps by introducing products into 2026. Intel’s GPU roadmap is on life support, with Intel’s Ponte Vecchio GPU officially dead and its Falcon Shores GPU being revived with a redesign.

Intel instead focused on its new Xeon 6 E-core chips, which have opened up to rave reviews for the unexpected power-efficiency gains.

Nvidia GPUs

Nvidia added new GPUs, CPUs, and networking chips to its roadmap for 2026. The “Rubin” platform’s GPU in 2026 will support HBM4 memory, which requires dense memory configurations. For 2027, the company is planning Rubin Ultra, an incremental upgrade, also with HBM4.

The Rubin GPUs will support 12-socket servers, compared to eight-socket servers currently.

“We push everything to technology limits, whatever TSMC process … whatever memory technology … SERDES technology, optics technology, everything is pushed to the limit,“ Nvidia CEO Jensen Huang said in his keynote.

Huang is upgrading the entire chip stack with Rubin. In 2026, Nvidia will introduce a new CPU called Vera, which will pair alongside Rubin. Vera will succeed the current Grace CPU, which is paired with the Hopper and Blackwell GPUs.

Nvidia also introduced the NVLink 6 interconnect to link GPUs in racks and data centers. The roadmap indicated it will have a speed of 3600 GB/second, about two times faster than the current NVLink 5 speed of 1800GB/second, which is in Blackwell. NVLink is a proprietary chip-to-chip interconnect.

The Rubin platform will also double the networking speed to scale out GPU communications across data centers to 1600GB/second, which is twice the speed of predecessors. The CX9 SuperNIC networking interface and x1600 Ethernet/Infiniband networking interface will allow millions of GPUs to communicate.

Rubin will follow the Blackwell Ultra GPU in 2025, which will be an upgrade from Blackwell GPUs coming this year. Blackwell Ultra includes HBM3e support and supports 800GB/second X800 networking interfaces.

“The X800 Ultra is designed for hundreds of thousands of GPUs, and the X1600 is designed for millions of GPUs. The days of millions of GPU data centers are coming,“ Huang said.

Blackwell is close to becoming available and will be available in Nvidia’s DGX air-cooled and liquid-cooled systems. The air-cooled system will be available with x86 CPUs.

All major cloud providers have already announced the availability of Blackwell GPUs and systems. Microsoft said the first instances of Blackwell in Azure cloud will go up later this year.

AMD GPUs

AMD CEO Lisa Su took clean shots at Nvidia in announcing new GPUs. AMD has struck gold with MI300, which is being adopted by Microsoft, Meta, and Oracle Cloud. Microsoft CEO Satya Nadella said MI300X offers the best price-performance for GPT-4 workloads.

The next major GPU will be the MI325X, an upgrade from the MI300X that will be out by the end of this year. It will include 288GB of HBM3E memory—which Nvidia has put in H200—and six terabytes per second of memory bandwidth.

“A single server with eight MI325 accelerators can run advanced models up to 1 trillion parameters. That’s double the size supported by an H200 server,“ AMD CEO Su said.

In 2025, AMD will release the MI350 series, which will also be based on HBM3e. It will be made using the 3-nm process. The GPU will be based on the new CDNA-4 architecture, which is tuned for AI.

“When we launched CDNA-3, we had eight times more AI performance compared to our prior generation. And with CDNA-4, we’re on track to deliver a 35 times increase,“ Su said.

The MI325X and MI350X series will support the universal baseboard OCP server design, and “what that means is that our customers can very quickly adopt this new technology,“ Su said.

By comparison, Nvidia’s GPUs aren’t easily available to regular customers off the shelf.

In 2026, AMD will launch the MI400 GPU, though Su didn’t talk much about it except to say it will be based on the CDNA-Next architecture.

AMD also launched new 5th Generation Epyc CPUs code-named Turin, which will feature 3-nm and 6-nm chiplets and will be available later this year. It will have up to 192 cores and 384 threads.

Intel Chips

In the keynote, Intel’s CEO Pat Gelsinger was his usual honest self, which focused on introducing Xeon 6 E server chips code-named Sierra Forest and Lunar Lake PC chips. There were no GPUs and no grand proclamations.

Gelsinger gave a sense that Intel is getting back in the CPU game with its new generation of server and PC chips, which are becoming leaner on power consumption while remaining performance leaders.

The Xeon 6 E-core chips are targeted at power-efficient servers with a large concentration of CPU cores, a market that fits ARM’s profile. Microsoft, Google, and Amazon have already introduced power-efficient CPUs based on ARM architecture.

The Xeon 6 E-core chips are available in several configurations up to 288 cores. The chips have opened up to great reviews on Serve The Home  and Phoronix . Both sites have seasoned and reliable reviewers with extensive experience testing hardware on multiple workloads and OSes.

The general takeaway of reviews was of surprise –that  x86 chips could deliver performance with decent power efficiency, which was the calling card of ARM.

“We see this as an essential upgrade for the modern data center, with a high core count, high density, and exceptional performance per watt. It’s also important to note that this is our first product on Intel 3, and Intel 3 is the third of our five nodes in four years,“ Gelsinger said.

Xeon 3 chips with P-cores (performance-cores), code-named Granite Rapids, are also on their way, Gelsinger said.

Gelsinger also said Intel was busting the myth of x86 being power-hungry with its Lunar Lake PC chip, which is 40 percent more power-efficient than its previous-generation Meteor Lake PC chip.

Apple’s Macs, built on ARM technology, deliver nearly twice the battery life of Windows PCs with x86 chips. Qualcomm recently introduced Snapdragon X Elite chips, which also touted multiple-day battery life for Windows PCs.

“Some say x86 can’t win on power efficiency. Lunar Lake busts this myth as well. This radical new SOC architecture and design delivers unprecedented power efficiency, up to 40% lower SOC performance than Meteor Lake, which was already very good,“ Gelsinger said.