Broadcom targets 50% VCF 9 adoption by late 2027 as AI speeds vulnerability discovery

Broadcom expects more than half of VMware’s installed base, measured by processor cores, to be running VMware Cloud Foundation (VCF) 9 by late 2027, as artificial intelligence (AI) models that can quickly uncover software flaws push customers to patch and upgrade more often.

“We have goals more aggressive than that, but I think 50% is a reasonable number for us to target,” Ram Velaga, president of Broadcom’s Infrastructure Software Group, told Computer Weekly on the sidelines of VMware Explore on Tour in Singapore.

The company tracks the move to VCF 9 by cores rather than by customer numbers, which Velaga said can easily be inflated. “Somebody can give you a random high number, and all it is, is they migrated one core from one to the other and declared it as a customer,” he said.

Velaga said the arrival of Anthropic’s Mythos , a frontier AI model that can find vulnerabilities in code, has driven customers to increase their patching frequency. Where they once took a long time to move from one release to the , many now want to patch every few weeks without going through downtime.

According to Velaga, VCF 9 lets IT teams patch and upgrade hosts without shutting down workloads or moving them to another server, something he said was “not very easy for anybody to engineer in hindsight”.

Broadcom is one of the launch partners of Project Glasswing , Anthropic’s programme that gives selected organisations access to Mythos to find and fix flaws in critical software. In his keynote at the event, Krish Prasad, senior vice-president and general manager of Broadcom’s VCF division, said the company had used frontier AI models to harden the latest VCF release and was asking all customers to move to it.

That has led to an “unprecedented ramp” in adoption, with millions of cores and thousands of customers already on the release, said Prasad. “In the era of Mythos and the frontier AI models, you cannot afford to have fragmented operations and fragmented security,” he added.

For flaws that cannot be patched immediately, Broadcom recommends virtual patching , which uses the hypervisor to recognise the network fingerprint of an attack on a known vulnerability and block it, said Chris McCain, global field chief technology officer for application networking and security at Broadcom.

“Nobody loves to patch,” he said. “If you have to patch, you have to test it. But for that period of time it takes to test the application, you are sitting in a vulnerable state. Virtual patching is a compensating control that buys you the time you need to be able to test that patch.”

The push to VCF 9 comes nearly three years after Broadcom completed its $69bn acquisition of VMware and revamped its licensing, scrapping perpetual licences and bundling products into subscriptions. The changes drew complaints from customers and industry groups over steep price rises, and led some organisations to look at alternatives .

Velaga said customers were worried that Broadcom would stop investing in VMware’s roadmap, that support would suffer and that the economics would not work for them. He claimed those concerns had since eased, with customers now seeing VCF 9 as a fully integrated product.

“Before, we had the idea of VCF, but the engineering teams were not all subscribing to the same code base, and some amount of integration needed to be done by the time it got delivered into the field,” he said.

On pricing, he said rising costs across the IT industry had changed how customers view Broadcom’s price rises. “Three years ago, who knew the cost in the IT industry was going to be what it is right now?” he said. “Somebody even laughed and said, ‘You guys were just early to the market’.”

That said, Velaga conceded the transition was not easy: “It happened because the business was not well run, either in terms of how it was priced in the market, how the product was being built and delivered, or the amount of overhead that the business carries.”

There are still customers Broadcom has to “continue to work through”, he said, though not “a massive number”. Customers now want to know what comes , he added, including whether there will be “surprises again”.

Velaga spent most of his 14-plus years in Broadcom’s networking chip business, where his biggest customers were hyperscalers, before taking on his current role nine months ago.

“Every 18 months, I had to go back and win the new silicon design cycle, and to do it, you can’t surprise customers,” he said. “The cloud guys are ruthless. If you don’t deliver the economics, they’ll throw you out.”

Sylvain Cazard, Broadcom’s president for Asia-Pacific, Japan and the Middle East, runs the company’s software business across those regions from Singapore. “I think we are experiencing potentially the fastest growth of the business here,” he said.

Like in Europe, sovereignty concerns are spurring organisations to rebalance workloads between public and private clouds , though Cazard said the drivers vary across markets.

In Australia, New Zealand and Singapore, organisations that went to public cloud are now running repatriation projects with “a lot of pragmatism”, he said. Japan, which is conservative and security-conscious, is moving its public sector onto modern platforms, while banks in India are rolling out private clouds on-premise across the country.

Customers have also gone beyond testing AI use cases and are now making architecture decisions, according to Cazard. He pointed to a large bank that is adding graphics processing units (GPUs) and open-weight language models to its private cloud to run services such as chatbots.

In his keynote, Velaga recounted how one bank told him that the 25% of its workloads running in public cloud cost as much as the 75% it ran on-premise.

“Before you go down the cost path, you have to think strategically, which means thinking two steps ahead, not just one step ahead,” Velaga told Computer Weekly. “One step ahead is, ‘Do I save on cost?’ Two steps ahead is, ‘Do I own this infrastructure?’.”

Owning the infrastructure lets enterprises decide later whether a given inference job goes to a public AI service or is routed through an internal token gateway to an open-weight model running on-premise, said Velaga. He also expects cloud providers to raise prices to recoup what they are spending on AI infrastructure.

Some consultants have suggested that enterprises rent on-premise infrastructure through consumption-based models to avoid the upfront cost of GPUs. “If you want it on-premise and you are going to eventually own it, why would you rent it?” he said. “The guy in the renting business is charging a margin over the cost of capital. It’s not like the guy who’s renting to you is able to get a GPU any cheaper. The only time you rent is because you are cash poor.”

Hardware costs feed into the argument, too. Velaga said server prices are now two to three times what they were a year ago, with long lead times. Even Broadcom, which runs less than 20% of its workloads in public clouds, can wait six months to a year for the compute capacity it wants from cloud providers, he said.

How far enterprises have got with AI has “very little to do with the infrastructure capabilities”, and more to do with the strategic choices they make, according to Velaga. Broadcom itself is using AI to speed up product development and improve quality, even though that is costing it more.

“When I talk to other large enterprises who are spending a lot of money on AI right now, their focus is not cost cutting,” he said. “They actually think in the near term their costs will go up.”

As inference workloads grow, Velaga expects GPUs to be spread across the datacentre alongside central processing units (CPUs) rather than concentrated in large, monolithic systems.

“Anything that looks like a monolithic scale-up mainframe kind of stuff over time doesn’t survive the economics,” he said. “Most customers don’t even have 10kW, 20kW or 40kW racks. They have 5kW racks, 6kW racks, sometimes 2kW racks.”

With GPUs spread out, the software that decides where workloads run and how they talk to each other matters more, according to Velaga. “Our software has always been an abstraction layer between the workloads and the hardware,” he said. “We don’t care whether the underlying silicon is a GPU or a CPU. We don’t care whether the instruction set is x86 or Arm. We don’t care if the compiler is a Cuda compiler or a PyTorch compiler.”

That is the basis of VMware Private AI Cloud, which Broadcom announced at VMware Explore in Las Vegas in August as the latest step in its private AI strategy . Its VMware AI Factory component includes the drivers and open source software needed to stand up models as a service, tested and supported, for more than 150 models and accelerators from Nvidia and AMD, plus servers from Dell, Lenovo, Supermicro and Cisco, Prasad said.

AMD is the newer of the two accelerator suppliers. VMware has long virtualised Nvidia GPUs, and Prasad said VCF 9.1 , launched in May, and the 9.1.1 maintenance update released last month extend that to AMD’s latest GPUs. The updates also bring vMotion live migration to AI workloads.

In a demonstration built around a fictitious insurer, Broadcom showed applications calling models through a fixed address, so the model behind it can be swapped without touching the application. Guardrails mask personal data in prompts and responses, and token quotas are set by department. The model gallery on screen included open-weight models from Chinese developers DeepSeek, Alibaba, Moonshot AI and Zhipu AI alongside those from Nvidia, Google, Meta and Microsoft.

Digital Realty’s CTO talks up the pace of AI silicon innovation , the growth of inferencing workloads, and why boasting datacentre megawatts misses the point.

Rajiv Ramaswami talks up Nutanix’s agentic AI play, the growing demand for sovereign cloud capabilities, and why decoupling storage from HCI is hastening migrations from VMware .

Neocloud and sovereign cloud providers offer alternatives to hyperscalers for AI infrastructure and data sovereignty, but availability gaps and a lack of managed AI services can pose challenges to enterprise customers.

Vast Data is targeting banks, government agencies and other regulated organisations in APAC with DataEnclave, a confidential computing capability that enables them to run third-AI models on data that cannot leave their own environments.

AI security, VKS boost VCF 9.1 appeal as support deadline looms

With VCF 9 'bedrock' in place, Broadcom fleshes out private AI

Nationwide bets on VCF as Broadcom responds to Tesco By: Cliff Saran

Broadcom updates VCF to address on-premise AI By: Cliff Saran