Chip companies now compete on electricity, not speed

Here's the deal: low-power AI chip design startup Velaura AI announced a $110 million Series A on August 18, led by Seligman Ventures at a valuation above $1 billion.

Becoming a unicorn at Series A is uncommon, and rarer still in semiconductors, where capital requirements are heavy and validation cycles are long. That price tag says investors are sizing the problem, not the product.

The problem is electricity.

What actually constrains AI infrastructure expansion right now isn't GPU supply. It's how much power you can bring to a data center. Building a new site means applying for grid interconnection, and across major US regions that queue runs multiple years. Without a power contract, buying GPUs gets you hardware with nowhere to plug it in. Samsung's interconnect-resistance work covered elsewhere today and Velaura's pitch are both aimed at the same wall.

The cast — Titan Core and the IP business model

Velaura doesn't sell finished chips. What it offers is Titan Core, a proprietary digital chip IP and design platform. The company says it delivers a 2-4x improvement in performance per watt on the mathematical operations inside AI accelerators while maintaining performance.

Worth defining IP here: semiconductor design assets. You don't manufacture and sell chips; you license design blocks to companies that do. Arm is the canonical example — it doesn't build phone chips, it sells designs and collects royalties.

The model's advantage is capital efficiency. Building your own chip means mask sets, wafer prepayments, packaging, test and inventory, easily hundreds of millions of dollars. IP licensing has none of that. A $110 million Series A looks small for a chip company and entirely adequate for an IP company.

The disadvantage is equally clear. Deciding to put someone else's IP into your chip is a heavy, slow commitment — once in, you live with it for years, so validation drags, and adopting IP from a company with no track record is a genuine risk. In IP, the first customer is the hardest to win, and the stickiest once won.

"Performance per watt" deserves unpacking too. Chips used to be compared on operations per second. That number alone means little now. When the power available to a data center is fixed, how much computation you extract per watt determines real throughput — and it sets chips per rack, cooling scale, and the electricity bill.

Why designing for low power is hard is worth a moment. Chip power splits into dynamic power, consumed when circuits switch, and leakage, which drains even when idle. Finer processes raise leakage's share, and lowering voltage cuts leakage but hurts speed and stability. Low-power design is the tightrope of dropping voltage while holding performance and reliability. AI adds its own conditions: matrix multiplication dominates overwhelmingly, and reduced precision often costs little accuracy, permitting far more aggressive optimization than general-purpose arithmetic allows. That's the seam Velaura is working.

The 2-4x range is wide for a reason. Results swing hard depending on which operation, at what precision, under what conditions. The company scopes the claim to "mathematical operations in AI accelerators" — a specific block, not a whole chip. It does not mean whole-chip power efficiency improves 2-4x. That's the most common misreading of announcements like this.

The investor list says more about this company than the product page does. Seligman Ventures led. New investors include Capricorn Investment Group and Prosperity7 Ventures. Existing backers Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund and StepStone Group participated.

Two names stand out. MARA is a bitcoin mining company — an industry that buys electricity, converts it to computation, and sells the result, so performance per watt is literally margin. Prosperity7 is Aramco's venture arm: energy capital investing in compute efficiency. And Samsung Catalyst Fund marks a link into the broader semiconductor ecosystem.

The numbers

Item Detail
Round $110M (Series A)
Valuation $1B+
Led by Seligman Ventures
New investors Capricorn Investment Group, Prosperity7 Ventures
Existing investors Mayfield, Maverick Silicon, MARA, Premji Invest, Samsung Catalyst Fund, StepStone Group
Core product Titan Core — digital chip IP and design platform
Claimed gain 2-4x performance per watt on AI accelerator math
Customers Working with 3 of the 4 largest cloud providers (unnamed)

The last row is the important one. Engagement with three of the four largest cloud providers, if it holds up, means the hardest gate in the IP business is already behind them.

But weigh the word "working with" carefully. In semiconductors that phrase covers everything from signed licenses to evaluation samples to technical review meetings. Hyperscalers evaluate promising IP broadly as routine business. Evaluation and adoption are entirely different stages, and which one applies wasn't disclosed.

The $1 billion valuation probably rests here. For a low-power IP company in conversations with three hyperscalers, a single genuine adoption moves revenue dramatically. Investors bought that probability, not current revenue.

Unicorn status at Series A also merits a second look. Series A normally validates product and early customers; unicorn pricing normally arrives after revenue is on a trajectory. That inversion means one of two things — the team and technology are exceptionally de-risked, or capital crowding into this sector pushed price ahead of substance. The second factor has been demonstrably strong in AI silicon lately.

What each side gets

Cloud providers get power-budget headroom. Google, Amazon and Microsoft all design their own AI silicon, partly to reduce Nvidia dependence but more to tune power efficiency to their own workloads. Dropping in external IP to improve a compute block hits that target while compressing design time.

Velaura gets time and credibility. An IP company's fate is decided by how many chips it ships inside. The $110 million funds growing the design team and building customer-facing organization; the company says it will expand engineering and customer-facing staff and deepen partner collaborations.

Power-intensive operators like MARA get cost structure. Mining or AI inference, converting electricity into computation is the same business. Better performance per watt means more revenue under the same power contract. Having such a company as an early investor also implies a channel for validation in real operating conditions.

Robotics and drones are potential beneficiaries. The company plans to expand beyond data centers into physical AI, where power constraints bite far harder. On a battery-powered device, compute power is flight time. Whether a drone stays up five minutes longer routinely decides whether a business works.

How much the power constraint actually binds is worth grounding. A large AI data center's demand now rivals a small city's. Site selection leads with grid interconnection timing rather than land price or fiber. A one percent gain in compute efficiency converts to a one percent gain in power contract, which makes performance per watt a permitting variable, not just a technical one.

Someone carries risk here too. A Series A unicorn raises its own bar for the next round: starting at $1 billion means the next round demands a much larger number, and real adoption has to materialize in between. IP validation cycles are long, so that timetable can get tight.

Precedents — how semiconductor IP businesses win and lose

Arm is the canonical success — licensing designs rather than making chips, and effectively owning mobile. What won wasn't performance, it was ecosystem: compilers, operating systems, developer tools and thousands of partners stacked on the instruction set until switching became impossible. In IP, the real moat isn't the technology, it's what accumulates on top.

RISC-V shows another path: an open instruction set usable without license fees, which spread quickly while drawing persistent criticism over fragmentation and software maturity. Velaura sells compute-block IP rather than an instruction set, so it isn't a direct comparison, but open alternatives apply pressure regardless. As open designs at the compute-block level proliferate, paid IP loses pricing power — especially with hyperscalers, who have the staff to take an open design and adapt it, meaning paid IP must win purely on being better.

On the failure side, look at companies that tried building their own AI chips and folded. Wave Computing and others drew attention with promising architectures and broke on production and software support. Velaura's IP model structurally sidesteps that trap — pushing manufacturing risk to the customer and concentrating on design.

But the IP model has its own failure mode. However good the IP, revenue is zero if nobody adopts it. A chip company can at least build something and try to sell it; an IP company depends entirely on someone else's decision. Companies that vanished quietly from the semiconductor IP market usually died of adoption failure, not technical failure.

It also fits the current moment. Etched sells complete racks — the opposite strategy. The Nvidia–Poolside deal bought technology through a license. Across AI silicon right now, "what to own and what to rent" is being answered differently by every company. Velaura picked the most capital-efficient seat.

How competitors respond

Arm is the most direct competitor, having expanded its IP portfolio toward data center and AI accelerators while holding relationships with every customer. The wall a new IP company must clear isn't a performance figure — it's the procurement inertia of buying from the proven vendor.

Synopsys and Cadence, the EDA giants, run large IP businesses whose strength is selling tools and IP together, reducing the customer's integration burden. New IP either enters that integration path or demonstrates an advantage large enough to justify bypassing it.

Hyperscalers' internal design teams are customer and competitor at once. They can design the blocks themselves; buying external IP buys time, not capability. Which means what Velaura actually sells is schedule compression, not performance.

Nvidia looks outside this contest but sets the baseline — every alternative's power efficiency gets measured against Nvidia parts, and Nvidia has raised performance per watt substantially each generation. A fast enough improvement cadence there erodes the alternative's advantage.

What actually changes for you

If you're planning AI infrastructure, the practical implication is power planning. Sizing a three-to-five-year power contract on today's performance per watt risks over-ordering. Conversely, if power procurement is your binding constraint, efficiency gains are equivalent to capacity additions and belong in the same calculation.

If you build robotics or drones, low-power AI compute is directly relevant — compute power is runtime on battery devices, and expanding on-device inference depends on this axis improving. Data-center IP takes time to descend into embedded contexts, and the requirements differ.

If you're in semiconductors, read this round as a re-rating of the IP business model. As the capital required to build your own chip keeps climbing, selling design assets alone looks relatively more attractive. Without adoption wins, though, the valuation can reverse quickly.

If you're a fabless designer in a smaller market, note the alternative. Most domestic low-power inference chip companies build and sell finished silicon, absorbing production capital and software ecosystem burden directly. The IP licensing model reduces that burden but requires building trust in design assets with global customers. Which fits depends on the company; the point is that more than one path exists.

If you're an investor, the thing to watch is how many of those three engagements convert into actual license agreements. The evaluation-to-adoption conversion rate defines this company. And semiconductor IP takes time to produce royalties even after signing — revenue arrives when the customer's chip reaches volume production.

If you follow energy, capital like Prosperity7 investing in compute efficiency is an interesting current: power suppliers investing in power-consumption efficiency, as AI data centers become a dominant customer of electricity markets and both sides' interests entangle.

🥄 Three Things You're Probably Wondering

— Does 2-4x per watt mean my electric bill drops to a quarter? No. That figure covers the math blocks inside an AI accelerator, not a whole chip or a whole data center. Real systems burn substantial power on memory, interconnect, power conversion and cooling. Improving the compute block alone yields a much smaller system-level gain. The direction is still meaningful.

— Why is a company that doesn't make chips worth $1 billion? The IP model is capital-light and high-margin; Arm dominated mobile with it. And this company says it's working with three of the four largest cloud providers, where a single adoption moves revenue substantially. Investors bought that probability, not current results. Whether "working with" means evaluation or contract wasn't disclosed.

— Why did Samsung invest? Samsung Catalyst Fund makes early investments in semiconductors and adjacent technology. The purpose isn't only return — it's visibility into technology trends plus the option to route a company toward foundry business or into internal designs. The optionality behind the check is closer to the point than the check itself.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!