Gavin Uberti, mentioned 26 times across podcast episodes and expert conversations analyzed by Sonic.

Gavin Uberti

Co-founder of Etched, an AI hardware company specializing in chips for inference.

Claims

As speaker

Mentioned in

Key positions and views

A vertically integrated stack, from chip to model, is necessary to win in AI hardware, as opposed to providing general-purpose components [9].

Minimizing chip-to-chip latency is the most critical factor for scaling AI inference performance, more so than maximizing single-chip computational density [15, 18].

Winning AI hardware solutions must exhibit economies of scale, where the cost to serve a 10x increase in demand is less than 10x the original cost [19].

Future hardware development tools should be designed for AI models as the primary users, not human engineers, to achieve optimal performance [12].

System-level innovations, such as custom interconnects and unified cluster memory pools, are the key to unlocking the next level of AI performance [24, 25].

Timeline

Early Career

Worked on kernels at Xnor.ai, a company that was later acquired by Apple for $200 million [2].

Pre-Etched Career

Worked on kernels at OmniML, which was subsequently acquired by NVIDIA for hundreds of millions of dollars [1].

Etched Founding & Partnership Building

Co-founded Etched and secured crucial early support from industry giants TSMC and Synopsys, who provided access to manufacturing and emulation tools on favorable terms before major funding was raised [10, 11, 26].

Series A Funding

Etched raised a $103 million Series A round, capitalizing on strong investor interest by accepting soft commits [8].

First-Generation Product Development

Led the development of Etched's first chip on a 4nm process, tackling extreme technical challenges like 50-picosecond clock alignment [3, 7] and resolving vendor delays by deploying a team to India for six months [22].

Leadership Team Expansion

Recruited Mark Ross, former CTO of the $9 billion Cypress Semiconductor, to join Etched as its full-time CTO, strengthening the company's executive leadership [23].

Podcast consensus on Uberti

Points of consensus

▶Multiple claims emphasize that Etched's core technical strategy is to drastically reduce chip-to-chip latency, positioning its custom interconnect as a more than 5x improvement over competitors like NVIDIA's Blackwell [15, 25].Jun 2026

▶Uberti's background is consistently rooted in developing low-level software ('kernels') at startups like Xnor.ai and OmniML, both of which were successfully acquired by major tech companies (Apple and NVIDIA, respectively) [1, 2].Jun 2026

▶The narrative highlights the critical importance of securing strategic partnerships with semiconductor giants. Etched successfully engaged TSMC and Synopsys, receiving collaborative support and favorable financial terms before it had raised its Series A [10, 11, 13, 26].Jun 2026

▶NVIDIA's Blackwell GPU is repeatedly used as a key performance benchmark, particularly its point-to-point latency of approximately 4,000 nanoseconds, which Etched aims to significantly undercut [4, 14, 15, 16].Jun 2026

Points of debate

▶Uberti's strategy of building a specialized chip for inference [9, 18] presents a direct challenge to the dominant industry paradigm of using general-purpose GPUs for all AI workloads.Jun 2026

▶His claim that traditional semiconductor venture funds largely missed the AI chip wave [5] suggests a debate over whether established investment theses in the sector are misaligned with the needs of modern AI hardware companies.Jun 2026

▶Etched's philosophy of designing debugging and profiling tools for an AI model as the primary user, rather than a human engineer [12], contrasts with conventional hardware development and tooling methodologies.Jun 2026

▶The assertion that a startup can co-develop and improve manufacturing processes with a foundry like TSMC [13] challenges the typical power dynamic between large-scale manufacturers and their smaller customers.Jun 2026

Key themes

▶Vertical Integration for Specialized WorkloadsJun 2026

Uberti advocates for a deeply integrated approach, controlling the entire stack from the chip fabrication layer up to the AI model. This strategy is built around the belief that specializing hardware for a single, dominant workload—AI inference—will yield performance and efficiency gains that general-purpose hardware cannot match.

This all-in strategy on inference for transformer models is a high-risk, high-reward bet on the stability of the current AI paradigm; a fundamental architectural shift in AI could render Etched's specialized hardware obsolete.

▶Latency as the Fundamental Scalability BottleneckJun 2026

A core theme is the identification of chip-to-chip communication latency as the primary barrier to scaling AI performance. Uberti repeatedly contrasts the thousands of nanoseconds of latency in current systems with Etched's goal of a >5x reduction, enabled by a custom interconnect and 'cluster-scale memory'.

This focus suggests a belief that future performance gains will come from system-level architectural innovation and efficient data movement, rather than simply increasing computational density on a single chip.

▶Navigating the Semiconductor Supply Chain

The claims illustrate the critical importance of building strategic, collaborative relationships with industry titans like TSMC and Synopsys. Uberti details how Etched secured access to essential manufacturing and emulation tools on favorable terms, even before raising significant capital, by demonstrating deep technical credibility.

For capital-intensive hardware startups, this theme underscores that technical innovation alone is insufficient; success is equally dependent on the business acumen required to persuade incumbent supply chain partners to bet on an unproven venture.

▶Aggressive Execution and Talent Strategy

Uberti portrays a culture of intense, hands-on problem-solving to overcome immense logistical challenges. This is exemplified by relocating a dozen top engineers to India for six months to fix a vendor delay and attracting veteran leadership like Mark Ross, the former CTO of a multi-billion dollar semiconductor company.

This narrative highlights that competing with established giants requires not only a superior technical vision but also an operational tempo and commitment to talent acquisition that can navigate the brutal realities of hardware development.

Source episodes

Suggested prompts

How does Etched's specialized hardware strategy mitigate the risk of being outcompeted by the massive R&D budgets of general-purpose GPU manufacturers like NVIDIA? &nearr;What specific technical and business milestones did Etched have to demonstrate to convince TSMC to collaborate on process improvements at the foundry's own expense? &nearr;Beyond latency, what are the other key performance metrics and trade-offs involved in Etched's 'cluster-scale memory' architecture compared to traditional memory hierarchies? &nearr;How does Uberti's vision for AI-driven debugging tools change the skill set required for hardware verification and performance engineering teams? &nearr;

Key concepts

Etched 20 ep Chip-to-Chip Latency 6 ep NVIDIA / Blackwell 7 ep Semiconductor Supply Chain (TSMC, Synopsys) 5 ep AI Inference Scaling 4 ep Venture Capital & Funding 4 ep Custom Interconnect 3 ep Kernels 2 ep Cluster-Scale Memory 1 ep Hardware Debugging & Profiling 1 ep

Notable quotes

“We can go ahead and cut this by more than a factor of 5x, and that allows us to then use the memory of other chips much more effectively.”

Gavin · The Two Harvard Dropouts Who raised $800M to take on NVIDIA

Report last updated: Jul 28, 2026

Create a free account to see Gavin Uberti's full intelligence report - every claim, the relationship network, and AI Q&A across all sources. No card needed.

Get started free

Back to Entities Intelligence Report