Velaura AI (formerly Auradine) announced Titan Core, a breakthrough silicon design and IP platform engineered to redefine power efficiency for AI accelerators.
Titan Core enables up to two times lower overall chip power consumption for AI accelerators, delivering up to 500 watts of power savings on a typical 1,000-watt GPU or XPU.
For hyperscaler data centers, this translates into significant reductions in electricity, cooling, and infrastructure costs, often equating to tens to hundreds of millions of dollars in savings. At Nvidia’s GTC event last week, I spoke with Sanjay Gupta, cofounder and president of Velaura AI, in an interview about the significance of the technology.
Additionally, this unlocks the ability for XPU deployment growth that would otherwise be constrained by power availability.
Built on Velaura AI’s patented, ultra-energy-efficient digital design technology and validated across over thirty million leading-edge ASICs in production deployments over
multiple years, Titan Core delivers fundamental innovations to optimize AI compute
power.
Velaura AI has ongoing engagements with several leading hyperscaler XPU partners
across 3nm and 2nm process nodes and is in discussions to integrate Titan Core into
next-generation AI accelerators. The company is now expanding its engagement with
additional GPU and XPU vendors seeking transformational power-efficiency gains.
“Our initial engagements with top hyperscalers are validating Titan Core’s capability to
dramatically reduce global AI data center power costs and improve AI sustainability,” said Rajiv Khemani, CEO of Velaura AI, in a statement. “The future of AI will be defined by who can deliver the most meaningful performance within real-world power limits. We are excited to unveil Titan Core, our ultra-low power technology perfected over several years.”
Titan Core: Redefining power efficiency for AI accelerators
As AI workloads scale across data centers worldwide, power, not compute capability, is
becoming the limiting constraint. Global AI-related electricity demand is projected to
more than double to over 945 TWh/year by 2030 (IEA Report), placing enormous
pressure on infrastructure expansion.
Velaura AI is addressing this challenge with its breakthrough silicon technology that
dramatically reduces the energy used by AI accelerators. A significant amount of power
in modern AI accelerators is consumed by mathematical operations that perform matrix
multiplications (MATMUL), the fundamental building blocks of AI training and inference.
Titan Core reduces the energy required for these operations by 2-4X using proprietary
circuit and library technology at advanced semiconductor process nodes. This equates
to approximately $1,300 in electricity savings over three years per XPU, in addition to
meaningful reductions in cooling and infrastructure costs, ultimately driving substantial
improvements in total cost of ownership (TCO).
Integration of Titan Core into customers’ existing SoC architectures and design flows is
seamless, maintaining full functional equivalence. Velaura starts from the customer’s
RTL design and applies its proprietary libraries and physical design methodology to
deliver an optimized physical layout. The resulting ultra-low-power design operates at a
lower voltage and integrates easily into the customer’s SOC.
“At SemiAnalysis, we track 200+ neoclouds and 6,000+ datacenters globally. Power
availability is a key constraint driving half of datacenters expected to go behind-the-meter by 2027. By reducing the energy needed to generate AI tokens, Titan Core is a pragmatic solution that Velaura AI’s customers can use to drive efficiency, lower OPEX, and allow more of the total datacenter BOM to be spent on compute,” said Dylan Patel, CEO of SemiAnalysis, in a statement.
A proven technology stack for low-voltage silicon
Delivering ultra-low-power compute requires far more than incremental optimization. It
requires deep expertise in custom circuit design and silicon behavior at voltages
significantly below nominal GPU and XPU operating levels.
Velaura AI has developed comprehensive, production-proven technology perfected over
multiple years that consists of upfront design service and licensable IP, including:
● Proprietary low-voltage digital design libraries at advanced semiconductor
nodes, including 3nm, 2nm, and beyond
● Custom Electronic Design Automation (EDA) flows and advanced circuit
methodologies optimized for low-voltage design, that deliver world-class yield,
reliability, and performance
● Potential silicon area reduction depending on the design targets
● Seamless integration into existing SoC architectures and design flows
The company has demonstrated large-scale production capability by shipping over 30
million energy-efficient ASICs at leading process nodes over several years. This
field-proven silicon foundation underpins Titan CoreTM for AI accelerators, providing
customers with a multi-year time-to-volume advantage versus do-it-yourself (DIY)
approaches that carry significant yield, reliability, and schedule risk.
Delivering Real Economic and Deployment Impact
Ultra-low power compute translates directly into measurable infrastructure benefits:
● Expanded AI deployment capacity in power-constrained environments
● Reduced electricity consumption and operating costs
● Decreased cooling requirements, allowing for simpler thermal management
● Increased rack-level compute density within existing power envelopes
Backstory
The company has 85 people and its headquarters is in Santa Clara, California. To date, it has raised $300 million for two very different products. One of its board members is Lip-Bu Tan, the current CEO of Intel.
Gupta himself has been in the tech and chip industries for 35 years. He was an electrical engineering grad who went to business school and then became a management consultant at McKinsey.
He was successful in a number of roles in payments and fintech. Then he met with a team of founders — Rajiv Khemani, Barun Kar, Barun Na and Saptadeep Pal. They cofounded Auradine about three years ago and Gupta joined them. He is now president of the newly renamed company Velaura AI, formerly Auradine.
The story is complex, so here’s more of the history. Khemani had a few successful exits for his startups before. He sold a network processing company to Intel and ran Intel’s networking business for a time. Khemani was also part of Cavium Networks, which Marvell bought for $6 billion. And Khemani also ran Innovium, a security processor firm that Marvell bought for $1.1 billion in 2021. Then the group started Auradine, which focused on Bitcoin processing as well as AI.
Back then, the blockchain processing market was mature, AI was just emerging. The team raised more than $300 million across three rounds and they incubated two other companies along the way — Aurascape AI, where Khemani is a cofounder, and Upscale AI, where Kar is CEO.
Auradine succeeded in launching tech that could mine Bitcoin using less energy. In 2025, that company generated close to $70 million in revenue. The company has shipped something like 30 million chips around the world.
Bitcoin or AI?
And it turns out that the ultra-low power processing for Bitcoin also has applications in AI computing.
Speaking about AI compute, Gupta said, “We are able to reduce power consumed for mathematical functions by an order of magnitude of two to four times. If you take AI computing, a lot of that is matrix multiplications and other arithmetic functions. So we can take the same logic equivalents and apply special techniques, including reducing the overall voltage, adjusting it.”
The company creates a customized layout and a specific low-voltage methodology that brings down the overall power consumed. We’re perfecting it in the area of Bitcoin mining. And then we can apply it to the area of AI computing.”
Since the Bitcoin market has taken a significant downturn, the team has pivoted the company with a new name, Velaura AI, to focus on the AI applications. They also brought aboard Manu Gulati (who designed chips at Apple, Google and Nuvia — the latter was sold to Qualcomm for $1.4 billion).
“With this ultra-low power technology, we’ve been in discussions with some of the top hyperscalers in the marketplace. And they got very excited” about the chance to reduce the amount of power consumed in mathematical operations, which consumes around 50% to 70% of the power of an AI XPU chip.
“We are actually working with two of the top hyperscalers on the concept,” he said. “And we are showing that you can get a two to 4x reduction in power with the same functionality.”
How it works
The tech works by modifying the voltage domain of the matrix multiplication. The control voltage is typically around 750 millivolts. With blockchain mining, the team brought that down to around threshold voltage, or below 300 millivolts. The aim for AI chips is to bring down the voltage while ensuring the chip’s reliability is not compromised, he said.
For the past four years, the team has worked with TSMC on ultra-low-power libraries and EDA flows as well. The net result is TitanCore. The core is an intellectual property platform, and both hyperscaler firms want to use it in their mainstream XPUs.
“Power is the binding constraint on AI scale, and the industry has been slow to address
it at the silicon level, where it matters most,” said Patrick Moorhead, CEO and chief analyst at Moor Insights & Strategy, in a statement. “Velaura AI’s approach targets the single largest power consumer inside an AI accelerator – matrix multiplication – using low-voltage design techniques proven across tens of millions of production ASICs. That production track record at 3nm is what separates this from a whiteboard exercise. If the efficiency gains hold up under independent validation, the implications for data center TCO and deployable compute capacity are substantial.”
Because of the lower power, Gupta believes the tech is in a class by itself. Customers will work with Velaura AI’s design methodology platform and the output is a design that can be sent to a foundry for manufacturing.
Not every part of the chip has to operate at the lowest voltage. Gupta said that some parts of the chip will operate on a different voltage island. One of the possible business options might be to license the tech to a company like Nvidia and dramatically lower Nvidia’s power consumption.
“It’s an enormous savings on the power, like maybe $1,000 to $1,500 per chip,” he said. “You don’t have to come up with new software.”
Velaura AI focuses on eliminating the error rates that typically happen at lower voltages.
He added, “At a very simplistic level, one of the techniques is around the ability to maintain the same functionality while reducing the voltage applied to the chip. That’s one of the key techniques. It’s not easy.”