Microsoft Bets on Its Own AI Chips, Nvidia Feels China Heat

Full transcript

Alex: Welcome to Daily Tech News for Tuesday, October thirteenth, twenty twenty-six, and I am Alex.

Jordan: And I am Jordan, and today the A I chip map is being redrawn in public, from Microsoft building its own silicon to Nvidia losing ground in China.

Alex: Let us start with Microsoft, because the company says it will mainly rely on its own chips to handle incoming A I requests, beginning with its third generation Maia accelerator.

Jordan: That chip is called Maia three hundred, and Microsoft is talking about expanding its Azure Maia deployments into the multi gigawatt range beginning in twenty twenty-seven.

Alex: The scale is the headline here, because reports around the announcement point to potentially more than one million of these chips being deployed next year to back Microsoft's own internal A I services.

Jordan: That does not mean Microsoft is walking away from Nvidia or anyone else overnight, but it does mean the biggest cloud buyers are no longer content to rent all of their future from one supplier.

Alex: Exactly, and for enterprise buyers the practical question is pricing, because a more diverse silicon ecosystem inside Azure could eventually change what inference costs and which workloads get priority.

Alex: Our second story shows the other side of that chip race, and it comes from new data on China's data center A I accelerator market for the first half of twenty twenty-six.

Jordan: According to I D C data reported by Reuters, Nvidia's share of that market fell to about thirty-four percent, while Huawei climbed to about twenty-eight percent.

Alex: Nvidia shipped roughly six hundred seventy-eight thousand A I accelerators in China between January and June, and that figure was down about forty-four percent from a year earlier.

Jordan: U S export restrictions still limit which of Nvidia's most advanced processors can be sold into China, and that opening is exactly where domestic suppliers have been gaining ground.

Alex: At the same time, Tencent is selling U S dollar and offshore yuan bonds in a deal expected to raise at least four point five billion dollars as its A I related investment rises sharply.

Jordan: And investors did not take the shift calmly, because Asian technology stocks fell heavily, with South Korea's Kospi index down around six percent as traders reassessed A I valuations.

Alex: So the takeaway is not that Nvidia is in trouble everywhere, but that China is now a genuinely contested market, and that contest is moving markets well beyond one country.

Jordan: Our third story is also about Nvidia, but this one is about openness rather than market share, and it comes from the Open Compute Project Global Summit in San Jose, California.

Alex: On October twelfth, Nvidia unveiled new hardware designs, networking technology, and partnerships aimed at speeding up what it calls A I factories, which are data centers built specifically for A I workloads.

Jordan: The company contributed its third generation Vera Rubin rack and tray designs to the Open Compute Project, covering power, cooling, and mechanical systems, so other builders can work from that architecture.

Alex: Inside those racks, the details matter, because a liquid cooled compute tray eliminates hoses, cables, and fans, which should simplify assembly and make maintenance easier over the life of the system.

Jordan: Nvidia also said its Vera C P U rack, which is designed for agentic A I workloads, is now shipping, and it says Vera delivers twice the performance of its predecessor Grace for tasks like scripting, compression, and code compilation.

Alex: There is a networking piece too, with a new module developed alongside Cisco, and that pairing is a reminder that an A I factory is only as fast as the connections between its machines.

Alex: Now to enterprise security, where Broadcom today introduced VMware HyperArmor, a new product meant to protect the hypervisor inside VMware Cloud Foundation.

Jordan: Broadcom's pitch is that frontier A I has changed the volume, velocity, and variety of attacks, and that when attacks arrive at machine speed, reactive detection and patching are no longer enough.

Alex: HyperArmor takes a different approach by allowing only what is known good to run on the host, instead of mainly chasing what is known bad after the fact.

Jordan: Broadcom says the two problems behind most hypervisor attacks are stolen root credentials and unpatched vulnerabilities, and HyperArmor is meant to address both of those problems.

Alex: It works with VMware E S X version eight and later releases, and it sits alongside the existing vDefend tools that already provide virtual patching at the network layer.

Jordan: If you run virtual infrastructure for a living, you know why this matters, because the hypervisor is the floor everything else stands on, and if that floor is compromised, every workload above it is at risk.

Alex: Our next story is a government action, because the U S Justice Department and the F B I seized seven internet domains that were allegedly used by China linked hackers to target critical infrastructure.

Jordan: The domains were tied to two hacking tools allegedly operated by actors working for Integrity Technology Group, a Beijing based cybersecurity company that investigators say holds Chinese government contracts.

Alex: Officials say those tools were used to find security weaknesses, gain unauthorized access to networks, and steal information from organizations in the United States and overseas.

Jordan: A federal magistrate judge in Pennsylvania approved the seizure warrant on October sixth, and the domains are being redirected to servers controlled by the F B I.

Alex: This is the second public U S disruption of that company's alleged hacking infrastructure in two years, after the F B I dismantled a network of more than two hundred thousand infected devices back in September twenty twenty-four.

Jordan: And investigators say the tools remained reachable afterward, with one server found in March of this year and a login page for another tool still accessible in September, which tells you how hard it is to keep disrupted infrastructure down.

Jordan: We will close with one more A I security note, because Anthropic says new protections for its Claude models have sharply cut large scale attempts to copy those models.

Alex: The company says the measures reduced the data taken through so called distillation attacks by roughly ninety to ninety-nine percent, and that it has made harvesting through proxy networks much harder and blocked tens of thousands of accounts.

Jordan: Distillation, in plain terms, is when a rival queries a strong model over and over to train a cheaper model to imitate it, and Anthropic named a Chinese startup's attempt to replicate Claude as one example.

Jordan: So today we have custom chips at Microsoft, a contested chip market in China, open A I factory designs from Nvidia, a new hypervisor shield from Broadcom, and governments and model makers both playing defense.

Alex: That is the through line for twenty twenty-six so far, because the race is no longer just about building the biggest model, but about who controls the chips, the racks, and the security underneath it all.

Jordan: Thanks for listening to Daily Tech News, and we will be back tomorrow with the next round of stories.

Alex: Until then, stay curious and stay secure.