Itay Inbar

Partner @ Greenfield Partners

[Itay Inbar @itayinbar_]

Models learned from data. Agents increasingly learn from environments.

As AI moves from answering questions to taking actions, simulated worlds become the training ground - places to act, fail and improve before touching the real world.

Great chat with @rebeccatqian @ @PatronusAI .

[Itay Inbar @itayinbar_]

@elonmusk Mind reading is no longer science fiction

[Itay Inbar @itayinbar_]

@katiekirsch Yes! 🙋‍♂️

[Itay Inbar @itayinbar_]

America is building complex hardware at a pace it hasn’t attempted in decades. But the supply chain underneath it isn’t ready.

Rockets, drones, robots, and defense systems are iterating at software speed, while critical components can still take months to source.

That gap is creating a massive opportunity: to rebuild the industrial stack itself - from contract manufacturing and new fabrication technologies to critical components and raw materials - using AI, robotics, and software.

@oliviaalevine, @curious_jt , and I went deep on the subject to imagine the future supply chain for Physical AI.

https://t.co/6VV7k9PoXf

[Itay Inbar @itayinbar_]

@GavinSBaker Crazy how fast narratives change. Just months ago Anthropic was crowned as the undisputed champion in the enterprise.

Model performance matters more than ever, and customers have built the right harnesses to let them easily switch over to whatever performs best right now.

[Itay Inbar @itayinbar_]

This is forming one of the largest upcoming bottlenecks in AI infra, and subsequently an opportunity for innovators. Cooling has historically been a relatively commoditized part of the compute stack, but as these power densities increase, it is becoming a much more advanced engineering problem. Something that was largely infrastructure plumbing is increasingly becoming part of the core system architecture. (3/3)

[Itay Inbar @itayinbar_]

AI datacenters have only recently made the transition to liquid cooling, but the most widely adopted solutions are already approaching their limits. NVIDIA’s Rubin generation is 100% liquid cooled, while skived cold plates - the dominant approach today - start running into practical limits around ~2-2.5kW per chip. The problem is that the next generation of chips is already being designed well beyond that envelope. (1/3)

[Itay Inbar @itayinbar_]

The trajectory is fairly remarkable: H100 was ~700W, while Rubin Ultra is expected around ~3.5kW and Feynman could move beyond 4kW. In other words, chips being designed today and launching just a few years from now are moving beyond what today’s widely deployed cooling architectures can handle at full power. And the power curve is still moving up. (2/3)

[Itay Inbar @itayinbar_]

@oliviaalevine @_jeff_liu @assort_health has quickly become a leader in transforming patient access to healthcare through AI.

Proud to have featured them on last year’s AI Disruptors 60.

Fantastic conversation between @_jeff_liu and my partner @oliviaalevine.

[Itay Inbar @itayinbar_]

@DIU_x What looks like a simple demo day photo will define billions of dollars in spend in the coming years.

Will be fascinating to see who comes out on top.

[Itay Inbar @itayinbar_]

Dozens of chip startups are trying to take on Nvidia in the datacenter.

Almost nobody is going after Jetson, Nvidia's chip for the edge.

Every humanoid, drone and autonomous machine being built today runs on that same brain, and the only serious challenger is Qualcomm.

Edge compute for Physical AI will be one of the largest markets of the next decade. Who is building here?

[Itay Inbar @itayinbar_]

@anandnk24 Aside from the fact that my kids are already hard at work trying to beat the agents on this, there have always been deep insights on agent advancement we can glean from how they play games.

Great job as always by @anandnk24 and team @PatronusAI.

[Itay Inbar @itayinbar_]

@elonmusk Innovation in power and cooling will create some of the largest new winners of the next decade. The scale and demanding performance needed at the frontier is hard to grasp.

Few see the emerging bottlenecks as much as @elonmusk and @SpaceXAI.

[Itay Inbar @itayinbar_]

Vega is reimagining one of the largest markets in cybersecurity - loved chatting with Shay about the future of cyber in the age of AI.

[Itay Inbar @itayinbar_]

@anandnk24 @PatronusAI Amazing work from the @PatronusAI team, quietly pushing forward the frontier across complex domains.

[Anand Kannappan @anandnk24]

Today, we're open sourcing the first computer-use dataset for design.

We captured 200+ hours and 3400+ trajectories of real design work in Figma, recording step by step how designers execute complex, long-horizon tasks.

FigmaTrace dataset, paper, and blog below :)
@PatronusAI

[Itay Inbar @itayinbar_]

The convergence of AI and services is one of the largest opportunities right now. AI-natives aren't selling software to accountants and lawyers - they're becoming the firm itself.

But owning the work ≠ keeping the gain. When automation makes work cheaper, savings usually flow to customers, not operators.

New blog post w/ Jason Cohen on control points - the license, network, or workflow position that decides who keeps the gain, and whether the right play is to buy the operator, build it, or arm it.

https://t.co/cOXg89sL6x