As the world's appetite for computing power threatens to outpace what human hands and conventional methods can build, two researchers have stepped forward with a quiet but radical proposition: let artificial intelligence design the chips themselves. At TechCrunch Disrupt, the founders of Ricursive Intelligence argued that a projected 10,000-fold surge in compute demand over five years cannot be met by building more infrastructure alone — it must be met by building smarter silicon. Their work raises an older question in new form: when machines become capable of doing what only experts once coul
AI-Designed Chips Could Solve Computing's Efficiency Crisis, Ricursive Intelligence Argues
Humans bring ideas, AI handles execution.
Mark
So the core claim is that we need more efficient chips because we can't build enough data centers to meet demand. How confident are we in that 10,000x figure?
Mimi
Goldie herself said it's potentially conservative, so she's not overstating it. The demand for AI compute is genuinely accelerating. But you're right to push—that's a projection, not a measured fact.
Luke
Exactly. A 10,000x increase over five years is extraordinary. That's a compound growth rate of roughly 60% per year. For context, semiconductor demand has never grown that fast. So either the AI boom is genuinely unprecedented, or that number is an upper-bound scenario. We should know which.
Mark
Okay, but let's say the demand is real. Can AI-designed chips actually solve it?
Mimi
The evidence so far is the AlphaChip donut design—it worked better than human designs. And the idea of optimizing the whole chip instead of pieces makes intuitive sense. But that's one example.
Luke
One example from a system they built themselves. We don't have independent verification yet. We don't know how much better these designs are, or whether the improvement holds across different types of chips, or whether the AI can handle the constraints of actual manufacturing at scale.
Mark
What about the speed claim—designing a chip in a week instead of years?
Mimi
That's where the real disruption would be. If true, it opens up custom silicon for applications that can't justify the current timeline and cost.
Luke
But again, that's a claim about what's possible, not what's proven. We don't know if Ricursive has actually achieved that speed, or if it's a theoretical best case. The article doesn't say they've shipped a chip designed this way.
Mark
So we're looking at a promising direction, not a solved problem.
Mimi
Right. The technology is real—they have working AI systems. But scaling it, proving it works across the industry, getting manufacturers to trust it—that's all still ahead.
Luke
And the biggest unknown: can AI-designed chips actually be manufactured reliably? Design is one thing. Turning that design into millions of working units is another.
O Pulso
- A 10,000x surge in compute demand is coming, and the semiconductor industry has no conventional answer ready for it.
- Ricursive Intelligence's AI is already producing chip layouts — including circular 'donut' geometries — that outperform designs crafted by experienced human engineers.
- By optimizing entire chips as unified systems rather than isolated blocks, the approach breaks a fundamental assumption that has governed chip design for decades.
- Compressing design timelines from years to weeks could unlock custom silicon for space, AR/VR, and biological computing — markets previously too niche to justify the cost.
- The founders frame the disruption not as replacement but as redirection: AI absorbs the execution, freeing human minds for the harder work of imagining what computing could become.
As the world's appetite for computing power threatens to outpace what human hands and conventional methods can build, two researchers have stepped forward with a quiet but radical proposition: let artificial intelligence design the chips themselves. At TechCrunch Disrupt, the founders of Ricursive Intelligence argued that a projected 10,000-fold surge in compute demand over five years cannot be met by building more infrastructure alone — it must be met by building smarter silicon. Their work raises an older question in new form: when machines become capable of doing what only experts once could, what becomes of expertise itself?
On Wednesday at TechCrunch Disrupt, Anna Goldie and Azalia Mirhoseini made a case that was simple in its logic and staggering in its implications. Over the next five years, they argued, global demand for computing power will grow by at least 10,000 times. Building enough data centers to match that demand isn't feasible. The only realistic path, they said, is to make chips dramatically more efficient — and to do that, the design process itself must be transformed.
Their startup, Ricursive Intelligence, is building AI systems that design semiconductors from the ground up. Where traditional chip design requires human engineers to partition a chip into smaller blocks and optimize each one separately, Ricursive's AI treats the entire chip as a single unified problem. The results have been striking. An earlier AI system the founders developed, called AlphaChip, generated a circular, donut-shaped component arrangement that defied conventional engineering intuition — and outperformed it. The team continues to explore wave-like and other unconventional geometries that emerge organically from the AI's search, designs that human designers might never have thought to attempt.
The speed gains are as significant as the design gains. Chip development that once took six months to a year could be compressed to a week or two. That shift changes the economics of custom silicon entirely, making it viable to build specialized processors for space systems, augmented reality, or biological computing platforms — each tailored precisely to its application and more energy-efficient than any general-purpose alternative.
Goldie and Mirhoseini are careful to frame this not as the end of human expertise, but as its elevation. As AI takes on the intricate, iterative labor of layout and optimization, engineers are freed to focus on what new capabilities are worth pursuing. Whether Ricursive can deliver this vision at scale remains an open question — but the pressure driving it is already here.
Anna Goldie and Azalia Mirhoseini took the stage at TechCrunch Disrupt on Wednesday with a straightforward argument: the world is about to need vastly more computing power than we can reasonably build, and the only way out is to make the chips themselves smarter.
Their startup, Ricursive Intelligence, is betting that artificial intelligence can do the work of designing semiconductors better and faster than humans ever could. The premise rests on a stark projection. Over the next five years, Goldie said, demand for compute will increase by at least 10,000 times—a figure she acknowledged might even be conservative. Building 10,000 times more data centers is not a viable path forward. The alternative, she argued, is to engineer chips that extract more performance from the same infrastructure: chips that use bandwidth more efficiently, that squeeze more capability from the wafers and fabrication plants already in operation.
The company's approach inverts the traditional hierarchy of chip design. Rather than having humans partition a chip into dozens of smaller blocks and optimize each one in isolation, Ricursive's AI systems optimize the entire chip as a unified whole. The results have surprised even experienced engineers. An AI system the founders previously developed, called AlphaChip, produced a circular, donut-shaped arrangement of components that defied conventional wisdom but ultimately outperformed designs created through traditional methods. The team is also exploring wave-like patterns and other intricate geometries that emerge from the AI's search process—designs that humans might never have conceived to try.
Goldie framed the shift as part of a broader reordering of work. As AI becomes capable of handling execution, humans can redirect their energy toward ideation and application development. In chip design specifically, this means engineers and architects can focus on what new capabilities are possible rather than spending months or years on the mechanics of layout and optimization. The same principle, she suggested, is already reshaping software development and other fields.
The speed advantage compounds the efficiency gains. Traditional chip design can take six months, a year, or longer. If that timeline compresses to a week or two, the economics of custom silicon change entirely. Applications that once seemed too niche to justify a dedicated chip—specialized processors for space missions, augmented and virtual reality systems, biological computing platforms—suddenly become financially sensible. Each could have its own optimized design, tailored to its specific constraints and requirements, all while consuming less power than a general-purpose alternative.
The founders are careful not to dismiss human expertise or suggest that AI will simply replace chip designers. Rather, they see automation as liberation—freeing the most creative minds in the field to ask bigger questions about what computing could become, while AI handles the intricate, iterative work of translating those ambitions into silicon. Whether the semiconductor industry embraces this shift, and whether Ricursive's technology can deliver on its promise at scale, remains to be seen. But the pressure is real, and the timeline is urgent.
Citações Notáveis
There is this understanding that in the next five years we are going to have at least 10,000x more demand for compute, and that's somewhat potentially conservative.— Anna Goldie, Ricursive Intelligence
If you can design a chip in like a week, then you can design custom chips for all sorts of different types of applications in space and AR, VR, and biological computing, and it makes sense financially to do that.— Anna Goldie, Ricursive Intelligence
Quer a matéria completa?
Leia o original em TechCrunch ↗
Technology