Andrew Feldman

Co-Founder and CEO of Cerebras Systems

171

Claims

As speaker

Mentioned in

Key positions and views

NVIDIA's dominance is built on market share inertia, not a defensible technical moat like CUDA, and its GPU architecture is a fundamental weakness for AI workloads.

The primary bottleneck for AI's future growth has shifted from the availability of silicon to the constraints of physical power generation and grid infrastructure.

Cerebras's wafer-scale architecture, with its massive on-chip SRAM, solves the memory bandwidth problem that plagues GPUs, making it superior for both AI training and inference.

The AI hardware market is on the cusp of a major shift where enterprise value will accrue to chip providers over model providers, and NVIDIA's market share will significantly decline within five years.

The U.S. is at risk of falling behind in the global AI race due to misguided government policies on immigration, chip exports, and a decades-long failure to invest in its electrical grid.

Podcast consensus on Feldman

Points of consensus

▶Feldman consistently argues across multiple appearances that NVIDIA's GPU architecture is fundamentally flawed for AI, especially for inference, due to memory bandwidth bottlenecks that lead to massive underutilization.Jun 2026

▶He repeatedly identifies the primary constraint for large-scale AI deployment as physical power availability and grid infrastructure, rather than the supply of silicon chips.

▶Feldman maintains that NVIDIA's competitive moat is its dominant market share and default status, not a technical 'lock-in' from its CUDA software, which he claims is irrelevant for inference workloads.Jun 2026

▶He asserts that Cerebras's wafer-scale architecture provides a fundamental advantage by co-locating massive amounts of fast SRAM with compute, directly addressing the memory bottlenecks of GPUs.May–Jun 2026

Points of debate

▶Feldman directly debates the industry's perception of 'CUDA lock-in,' arguing it is a myth for inference tasks where simple APIs are sufficient, though he concedes software porting for training remains a challenge.

▶He challenges the conventional wisdom of a rapid, two-year depreciation cycle for AI hardware, pointing to the continued utility of older NVIDIA chips like the A100 and H100 for less demanding tasks.

▶Feldman refutes near-term predictions of massive AI-driven labor shortages, arguing that while productivity gains will be enormous, the organizational restructuring required means significant economic dislocation is more than five years away, possibly closer to fifteen.

▶He contests the idea that vertical integration (building both models and chips) is necessary for success in AI, citing the success of non-integrated players like OpenAI (using Azure) and Anthropic (using AWS/Google) as evidence.

Key themes

▶The Architectural War on NVIDIAJun 2026

Feldman consistently frames the AI hardware market as a battle of architectures. He argues that NVIDIA's GPU architecture, a legacy of graphics processing, is ill-suited for modern AI workloads due to memory bottlenecks, while positioning Cerebras's wafer-scale engine as a purpose-built, superior solution.

This narrative suggests the AI hardware market is not a settled matter of incremental improvement but is ripe for a fundamental architectural disruption, creating a high-risk, high-reward environment for investors betting against the incumbent.

▶The Geopolitics of AI Compute

Feldman highlights the global race for AI dominance, criticizing US policies on chip sales and immigration while noting China's aggressive state-backed investment and the UAE's ambitious nuclear-powered data center projects. He underscores the strategic importance of compute resources on a national level.

This theme indicates that market success for AI hardware companies is increasingly dependent on navigating complex international relations, government policies, and national security interests, not just technological superiority.

▶The Physical Constraints of a Digital Boom

Feldman repeatedly emphasizes that the exponential growth of AI is colliding with real-world physical limits, primarily power availability and an aging electrical grid. He argues that the key bottleneck is no longer the number of chips but the ability to power and connect them.

Investors should look beyond chip performance metrics and evaluate companies based on their strategies for power efficiency, data center location, and access to energy resources, as these will become critical long-term differentiators.

▶AI's Economic Transformation

Feldman predicts AI will drive a massive increase in labor productivity and economic growth, but only if organizations fundamentally restructure their workflows around its capabilities. He also foresees a long-term shift in enterprise value from AI model providers to the underlying chip providers.

This suggests the true economic impact of AI will be a lagging indicator, and the most durable long-term investments may be in the foundational infrastructure (hardware) rather than the more visible and currently hyped application layer (models).

Source episodes

Timeline

Founding Period

Feldman's discourse focused on the historical challenge of wafer-scale computing, explaining how Cerebras solved the 75-year-old yield problem by using redundant tiles, a technique adapted from memory manufacturing.

August 26 (Year Unspecified)

Cerebras launched its inference product, marking a strategic pivot in Feldman's public narrative to address the larger inference market and emphasize the irrelevance of CUDA for these workloads.

Recent Period (Davos)

Discourse shifted to the macro and geopolitical scale of AI, with Feldman highlighting massive international deals (G42, OpenAI), the critical bottleneck of power infrastructure, and national AI strategies.

Current Fundraising Period

Following a $1 billion pre-IPO funding round, Feldman's messaging centers on aggressive scaling, including expanding manufacturing and adding new data centers to meet what he describes as underestimated demand for AI compute.

Suggested prompts

How does Feldman's critique of NVIDIA's memory bandwidth limitations align with the technical roadmaps of other major competitors like AMD and Intel? &nearr;What geopolitical risks and opportunities arise from Feldman's observation that nations like the UAE are building sovereign, nuclear-powered AI infrastructure? &nearr;If Feldman's prediction about the dominance of synthetic data is correct, which adjacent industries (e.g., data labeling, content moderation) face existential threats or opportunities? &nearr;Analyze the potential contradiction in Feldman's argument that CUDA lock-in is weak for inference while also acknowledging the significant software porting effort required for training on Cerebras hardware. &nearr;

Key concepts

NVIDIA 3 ep GPU Architecture 3 ep Wafer-Scale Engine (WSE) 3 ep AI Inference 3 ep Market Dynamics 3 ep Geopolitics & Government Policy 3 ep Power & Infrastructure Constraints 2 ep CUDA 2 ep AI Training 2 ep Synthetic Data 1 ep

Notable quotes

“In a GPU, most of the time it's doing inference, it's 5% or 7% utilized. That means it's 95% or 93% wasted.”

Andrew Feldman · Andrew Feldman, Cerebras Co-Founder and CEO: The AI Chip Wars & The Plan to Break Nvidia's Dominance

“The limiting constraint for large AI deployments and data centers is now power availability, not square footage or number of chips.”

Andrew Feldman · Coinbase CEO's Top 3 Crypto Trends for 2026 + More from Davos!

“You don't talk about the field failure rate of your products, which are massive.”

Andrew Feldman, Cerebras CEO · Cerebras CEO, Andrew Feldman on Why Raise $1BN and Delay the IPO & Why NVIDIA’s Worried About Growth

“I think in five years from now, NVIDIA is going to have 60%, somewhere between 50% and 60% of the market.”

Andrew Feldman · Andrew Feldman, Cerebras Co-Founder and CEO: The AI Chip Wars & The Plan to Break Nvidia's Dominance

Report last updated: Apr 21, 2026

Get started free

Back to Entities Intelligence Report