For three years, the defining question of the artificial intelligence boom has been how many chips a company could get its hands on. On Tuesday, a group of investors wagered $110 million that the more urgent question is how many of those chips the electrical grid can afford to keep running.
Velaura AI, a Santa Clara semiconductor startup, said it had closed a $110 million Series A at a valuation of more than $1 billion, vaulting the company into unicorn territory on the strength of an unfashionably narrow claim: that it can wring two to four times more useful work out of every watt an AI accelerator consumes.
Seligman Ventures led the round, joined by new investors Capricorn Investment Group and Prosperity7 Ventures, the diversified growth fund backed by Saudi Aramco. Existing backers Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund and StepStone Group all participated. The company said the money will fund development and commercialization of its AI compute portfolio, expand its engineering and customer-facing teams, and deepen work with partners building AI infrastructure and physical AI systems.
Selling watts, not FLOPs
At the center of the pitch is Titan Core, the proprietary digital chip IP and design platform Velaura announced earlier this year. The company says Titan Core delivers a 2x to 4x improvement in performance per watt on the mathematical operations that dominate AI accelerator workloads, without giving up raw performance — a combination that, if it holds, lets a data center operator fit more computation inside a fixed power envelope rather than negotiating for a bigger one.
Velaura is not selling a finished accelerator. It is selling the arithmetic blocks that go inside somebody else's. That is a less glamorous position than competing with Nvidia head-on, but it comes with an unusual credential for a young chip company: Velaura says the underlying technology has already shipped in more than 30 million ASICs across leading manufacturing processes. Most silicon startups raising at unicorn prices are still arguing from simulation results. Velaura is arguing from yield data.
The company says it is already working with leading hyperscalers to fold the technology into future XPU roadmaps, and is extending the same architecture toward physical AI — robots, drones and autonomous machines that must run demanding models inside brutal power and thermal budgets. Its leadership and engineering ranks are drawn from Apple, Nvidia, Google, Qualcomm and Marvell, and the company is led by co-founder and chief executive Rajiv Khemani.
“Every advance in AI, from reasoning models to embodied intelligence, creates demand for more compute, and ultimately more power,” Khemani said in a statement accompanying the round. “The next era of AI will be defined not only by better models, but also by fundamentally better compute economics. Velaura is building the ultra-low-power silicon and software foundation needed to scale AI from hyperscale data centers to intelligent machines operating in the physical world.”
Umesh Padval, managing partner at Seligman Ventures, framed the thesis around the second market. “Physical AI represents one of the next major frontiers for AI, and it will require a fundamentally different approach to compute centered on extreme power efficiency,” he said. Seligman Ventures launched with a $500 million allocation aimed at AI infrastructure, cloud computing and modern data center hardware, and Padval has been an active backer of chip-design startups.
Why the bottleneck moved
The numbers behind the deal are less about Velaura than about the grid. Gartner forecasts global data center electricity consumption will reach 565 terawatt-hours in 2026, up from 447 TWh in 2025, a 26 percent jump in a single year. Measured in capacity rather than energy, worldwide data center power demand is expected to climb roughly 27 percent this year to about 132 gigawatts, from 104 GW in 2025. The International Energy Agency projects data center electricity demand will more than double to roughly 945 TWh by 2030, driven overwhelmingly by AI.
Those figures describe a supply problem that money cannot immediately solve. Hyperscalers have committed hundreds of billions of dollars to AI buildouts, but interconnection queues, generation capacity, transformer lead times and cooling constraints all run on timelines measured in years. New campuses are now designed at 100 to 300 megawatts, and in some cases a full gigawatt. When power, not real estate or even silicon allocation, is the binding constraint, efficiency stops being an environmental talking point and becomes the cheapest available source of additional capacity. A 2x improvement in performance per watt is, functionally, a second data center you did not have to build.
That is the arbitrage Velaura is selling, and it explains why a company with no accelerator of its own can command a billion-dollar price. It also explains the risk. Performance-per-watt claims scoped to specific math operations do not automatically translate into whole-chip gains once memory bandwidth, interconnect and software overheads are counted. Hyperscaler qualification cycles are long and unforgiving. And Velaura is not the only party working this problem: Nvidia, AMD, Broadcom and the in-house silicon teams at Google, Amazon and Microsoft are all chasing the same joules, with vastly larger budgets.
The tell over the next twelve months will be design wins. Velaura has said it is engaged with hyperscalers on future XPU roadmaps; whether any of those engagements converts into a disclosed production socket, and on what timeline, is the number that matters more than the valuation. Watch also for independent benchmarks on full workloads rather than isolated operations, and for whether the physical AI side of the story — the part Seligman appears most excited about — produces a named robotics or autonomy customer. The AI industry spent the last cycle asking how many GPUs it could buy. The next one may be spent asking how much power each of them justifies.
“The next era of AI will be defined not only by better models, but also by fundamentally better compute economics.”— Rajiv Khemani, Co-founder and CEO, Velaura AI