Nvidia deal paves the way for Intel buy
Cette publication existe aussi en Français
A key deal announced today paves the way for Nvidia to buy the design business of struggling chip maker Intel.
The deal is significant in many ways. Nvidia will work with Intel to create a custom x86 processor with the NVLink interface for the AI datacentre market, and Nvidia will also supply GPU chiplets to Intel for consumer AI PCs.
Nvidia is also investing $5bn in Intel shares, which gives it a key negotiating position in any future negotiations and the ability to push for a merger. TSMC and Broadcom have both been linked to a potential deal to buy the company, which has been selling off Altera and its RealSense business and closing its automotive chip unit.
The move is significant in many different ways, not least driving Nvidia’s NVLink interconnect technology into the wider datacentre market. Intel has historically developed leading edge standards such as PCI Express and the UCIe standard for chiplet interconnect. Now Nvidia is driving NVLInk, which evolved from its acquisition of Mellanox in 2019, as a key standard for datacentres.
“General purpose computing has reached its limits and today we are taking the next great step with a historic partnership to jointly develop multiple generation of x86 CPUs,” said Jensen Huang, CEO of Nvidia, which is the world’s most valuable company with a valuation of over $3 trillion. Huang was in London for the state visit of President Trump to the UK and to announce a $2bn investment.
The core of the Intel datacentre deal focusses on the NV72 racks that are currently based on ARM-based processors.
“NV72 has a custom CPU called Vera with the Blackwell GPUs to scale up the entire rack but we had to customise the CPU to do that and this architecture is only available for the Vera ARM CPU and its really unavailable for x86 and PCIexpress. Now with the x86 we can integrate it and create rack scale supercomputers,” said Huang.
The deal also includes designing new x86 processor for the mobile market.
“Our partnership is 100% focussed on custom CPUs for the datacentre and completely focussed on mobile SoC for consumer PCs,” he said. “We are creating an SoC that fuses the CPU and RTX GPU using NVLink into one virtual SoC and that would become a new class of integrated graphics laptop.”
Neither company would comment on the process technology that would be used for the next generation x86 CPUs, whether on TSMC’s 2nm or Intel Foundry’s 14A process.
“This announcement is about the product collaboration,” said Lip-Bu Tan, recently appointed CEO of Intel. “TSMC has been a long term partner for Nvidia and Intel and we will continue doing this, and more announcements on process down the road when the product is ready.”
The deal will not impact on the ARM side of Nvidia’s business, says Huang.
“We are fully committed to the ARM roadmap and we are building the next generation of Grace called Vera and we have the Thor chip for autonomous driving and robotics. Today should have no impact on ARM.”
Future direction
The elephant in the room is what this means for the future of Intel. Nvidia has $57bn in cash, and Intel is currently valued at $85bn for both the design business and the foundry business. The valuation of the foundry business is not public, but the fabs and advanced packaging business are surely worth well over $30bn with advanced fabs in Oregon and Arizona, and other fabs in Ireland and Israel. This puts the x86 business at $50bn, which is also the value of the market that this deal addresses.
These are clear hints. The AI PC chips and the AI datacentre chips will be the major growth businesses for Intel and these markets will be inextricably linked to Nvidia.
“This will address $25 to $50bn of annual opportunity,” said Huang. “The return on that investment is going to be fantastic with a $50bn annual market,” which gives a clear price point for Intel design business.
“Intel has the Faveros packaging technology that is really enabling here,” said Huang. “Nvidia’s technology is based on the TSMC foundry, and connecting the chiplet in a multi technology packaging is a fabulous way of mixing and matching technology which is one of the ways we can build these incredibly complex systems.”
Nvidia also does not want to be a manufacturing company and would be happy to buy x86 chips from Intel Foundry in the same way it buys its ARM-based Grace and Vera chips from TSMC to add to its Blackwell and forthcoming Rubin GPUs.
- US government buys stake in Intel with Chips Act cash
- Nvidia pushes Blackwell GPU as start of a whole new industry
- AI bubble heads for crash
The moves are afoot already, with three architecture teams at Intel and Nvidia working closely together for the last year. “We want to match the fast moving culture of Jensen,” says Lip-Bu Tan. That seems to be a very clear prelude to a ‘merger’ of the design business with Nvidia.
Such a deal would meet with political approval, as the Trump administration had been pushing the sale of Intel.
“The US administration is happy to see US technology companies working together,” says Huang. The deal would also avoid the risk of anti-trust regulations that would hit an approach by the other x86 supplier AMD, whose Radeon based AI chips also compete directly with Nvidia in both the consumer PC and the datacentre markets
But time is short. Nvidia has more to lose than Intel from the collapse of the AI bubble, and the timing is right for a strategic merger of two key technology companies.
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
