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

The capability, expected to roll out in the coming months, extends Vast’s protection for data at rest and in transit to model weights and enterprise data while they are being processed. This way, neither the data owner nor the model developer has to expose their intellectual property to the other or to whoever runs the infrastructure.

At a media briefing in Singapore, Sunil Chavan, vice-president for Asia-Pacific at Vast Data, noted that regulated sectors would lead adoption in the city-state. Local banks were already talking to Vast and cloud partners such as Sharon AI about how they could use the technology, he said, though he did not name any as those talks are still ongoing.

What differentiates enterprises’ use of AI is their own data, said John Mao, vice-president of business development at Vast Data, and sometimes regulations prevent it from leaving the country. Model developers have the opposite concern.

“The models oftentimes are the IP [intellectual property], and yet they can’t distribute that like a free piece of software, because if I gave it to somebody at an enterprise, it can be easily replicated and reproduced,” said Mao. “Both sides have a trust issue.”

Vast developed DataEnclave with Nvidia, which burns a hardware root of trust into its newer graphics processing units (GPUs) and central processing units (CPUs) at the time of manufacture. In trusted execution environments, CPUs, such as those from Intel and AMD, can connect with Nvidia GPUs running in confidential computing mode, protecting GPU memory as well as traffic between GPUs.

Nothing will be decrypted until the platform proves itself. Model owners hold their keys in their own vaults and verify the environment cryptographically before releasing them. Enterprises can do the same for their data and models they have fine-tuned, which have corporate data “effectively baked in”, Mao noted. Attestation can be run on-premise for air-gapped datacentres.

Vast is also extending the tamper-proof audit logging capability on its storage platform to compute, recording every call to an application inside the enclave in the Vast DataBase, where the audit trail can be queried at high speed.

“Sometimes bad actors can come in, and they will go in and edit the record book, and then your audit is no longer trusted either,” said Mao, adding that as Vast is not a security company, it passes the logs to partners such as CrowdStrike for anomaly detection and incident response.

Mao stressed that neither Vast nor the cloud provider has the encryption keys, though independent third-party audits of the compute layer would come eventually.

The launch also brings model management to the Vast AI Operating System as the company moves beyond its storage origins. Models will become a logical resource, much like data, with the operating system matching them to tasks based on purpose and cost, and on which data each can see.

“People think about AI as if there’s one single model that will run everything, and that’s just simply not true,” said Mao. Enterprises are increasingly using a smaller model to drive a workflow and calling on frontier models for specific tasks, he said.

Over 20 launch partners have signed up for DataEnclave. Model developers such as Cohere, TwelveLabs, Factory, CrowdStrike, Deepgram and Fundamental were each chosen to cover a different modality. Cisco and Supermicro will supply rack-scale confidential AI appliances, while Nscale, G42 and Sharon AI serve as AI cloud partners.

Sharon AI, a US-listed neocloud founded in Australia, sees DataEnclave as a way to win over regulated firms that have been hesitant to move their data for AI processing. James Manning, its co-founder and CEO, said the product would help the company pursue sovereign compute, which has become a “very important part of our growth profile”.

He cited the example of a tax office using a large language model (LLM) to handle initial phone calls from taxpayers, who might need to hand over tax file numbers and identity details, as a potential use case.

Then, there are also Australia’s strict privacy rules, which include mandatory breach reporting. “How do we verify things like we’ve had a breach? How do we qualify if that breach has occurred? What data’s been compromised? These are complex problems and things that AI hasn’t really been able to answer up until today,” Manning said.

“It’s not a solution that every customer is going to need, but it is a solution in highly regulated industries that enables data sovereignty to be properly managed, deployed and delivered,” he added.

Asked by Computer Weekly if DataEnclave was aimed at training or inference workloads, Mao said most GPU capacity still goes to model training at large organisations, but the market is moving towards inference, where DataEnclave will matter most.

“The training is usually a curated set of data,” he said. “When you start to get into inferencing, that’s where you’re feeding it real-time business data. You’re feeding your other systems of record.”

Vast has about 1,200 employees worldwide and is cashflow positive. It runs the region from its Singapore headquarters and has offices in China, Japan, South Korea, India and Australia. About 30 staff are in Singapore, a number it plans to increase to between 60 and 75 over the next few years.

Chavan said that Vast is partnering with Singapore’s Economic Development Board to establish an engineering and support centre in the city-state. Vast’s regional customers include SK Telecom, which has implemented a sovereign AI strategy for Korean government projects, as well as Coupang, Ola and Bitdeer AI.

Sharon AI, which followed Vast into Singapore, has Asian customers processing data in Australia, which is “close enough to be relevant, but far enough to be safe and secure enough to be trusted”, Manning said. The continental-sized country’s small population, energy resources and good connectivity put it in a good position to export AI compute to Asia, he added.

“Whatever we build in Australia, we can’t consume enough,” said Manning. “It’s far easier to send you the compute output than it is to try to ship you the power.”

When Sharon AI releases new GPU capacity, it is usually three to eight times oversubscribed, Manning noted, an imbalance he expects to continue over the next two years. At the same time, storage needs are rising, with the company going from modelling 3PB (petabytes) of storage per 1,000-GPU cluster to signing a customer contract for 10PB per cluster in about nine months.

“Customers are realising they need to store more – they’re processing more,” said Manning.

• Chronosphere details mechanics of modern approach to observability

  

  By: Adrian Bridgwater

• Career cure for AI phobia: Be a beekeeper, not a worker bee

  

  By: David Essex

• KubeCon EU 2026: AI tightens the dev-prod loop

  

  By: Torsten Volk

• What Western companies misunderstand about China’s AI strategy