Cerebras CEO on the Future of Data Centres, Token Costs and Memory & We Are Not in an Infra Bubble | 26 May 2026 | 20VC with Harry Stebbings and Cerebras
"the infrastructure buildout is behind demand we can't build data centers fast enough to keep up with demand we have a $25 billion backlog" - Andrew Feldman [00:03:43]
"let me just ask you this question how big's the market for slow search really how it's zero... there will be zero market for slow" - Andrew Feldman [00:21:47]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"everybody is or most communities are comfortable when their neighbors pay all their own way... our data centers don't need to use a ton of water they can recycle it" - Andrew Feldman [00:32:03]
"if you want to make a lawyer uncomfortable... ask her to work in an area with no precedent they don't know what to do their whole training is about what has everybody else done before" - Andrew Feldman [00:39:37]
"if you don't like delivering for your team you're not a real leader and uh that feels good every day here 800 millionaires" - Andrew Feldman [01:01:43]
"we had an 18-month period where we were spending $8 million a month and we couldn't build it... and nobody else to this day has solved it" - Andrew Feldman [01:05:16]
Speakers & Credentials
Harry Stebbings: Host of 20VC. Renowned venture capital podcaster and investor, recognized for conducting deep-dive interviews with the leading figures in technology and venture capital.
Andrew Feldman: Founder and CEO of Cerebras Systems. A veteran technology executive who recently led Cerebras through a massive public listing. He is a recognized pioneer in building specialized hardware for artificial intelligence, known for pushing the boundaries of wafer-scale chip design to solve the hardest problems in AI compute.
1. Executive Summary
Cerebras Systems recently completed a historic semiconductor IPO, scaling from a $185 to $311 price point and securing a market valuation over $5.5 billion, fundamentally validating the market for specialized, non-GPU AI architecture [00:00:19].
The widespread narrative of an "AI infrastructure bubble" is historically inaccurate; unlike the 1990s fiber optic boom which was built ahead of demand, AI infrastructure is actively trailing real-time, explosive demand, evidenced by a shared industry backlog of $25 billion [00:03:43].
The current bottleneck in AI compute is heavily tethered to memory constraints (specifically HBM), allowing memory manufacturers like Micron to achieve anomalous 80-85% gross margins, a constraint Cerebras avoids entirely through its reliance on SRAM [00:08:14].
Inference speed will be the ultimate competitive moat; there is absolutely zero market for "slow AI," meaning the premium on architectures that can deliver tokens exponentially faster will dictate market dominance [00:21:47].
Enterprise adoption of AI is currently being artificially suppressed not by technology or data cleanliness, but by institutional risk-aversion, specifically internal legal and security teams who are fundamentally incentivized to say "no" to precedent-free technologies [00:38:16].
Geopolitically, there is unequivocal consensus that selling cutting-edge compute to China will directly advance their military and state-backed industrial base; thus, leveraging entities like TSMC and ASML as structural choke points remains a vital US strategic imperative [00:44:34].
The human and financial toll of solving bleeding-edge hardware physics is staggering, requiring immense entrepreneurial grit—exemplified by Cerebras's 18-month period of burning $8 million monthly while suffering daily engineering failures before achieving their architectural breakthrough [01:05:16].
[01:00:40] Creating 800 Millionaires & The Realities of Leadership
[01:03:55] The Toll on Personal Life & Burning $8M/Month
3. Detailed Thematic Summary
The AI Infrastructure Reality & Demand Dynamics
The Inversion of Historical Bubbles: Feldman forcefully rejects the narrative that AI infrastructure is in a bubble, contrasting it directly with the 1880s railroad expansion and the 1990s fiber optic buildout [00:03:03]. In those historical cases, capital was deployed on the assumption that "if we build it, they will come." Today, the infrastructure buildout is radically inverted—it is trailing far behind existing, desperate demand.
The Backlog Metric: This insatiable demand is quantified by a massive $25 billion backlog across the industry (affecting Cerebras, Nvidia, AMD), proving that data centers simply cannot be constructed fast enough to absorb capital [00:03:43].
The Freeway Metering Effect: Rather than being a pure negative, the inability to instantly meet all demand acts as a "meter on a freeway" [00:05:26]. By slowing down the deployment of capital, the market avoids massive indigestion (likened to gorging at a Vegas buffet) and ensures smoother, sustained infrastructural growth [00:05:01].
OpenAI's Forecasting Superpower: Sam Altman’s core advantage was trusting the math of exponential growth; while others suffered cognitive dissonance trying to fathom gigawatt-scale compute needs, OpenAI aggressively contracted for hardware and power years in advance, even if it meant taking down-rev H100 chips from Elon Musk out of sheer necessity [00:06:06].
Supply Chain Dynamics & Silicon Economics
The Memory Squeeze (HBM vs SRAM): The explosion in AI demand has heavily stressed the High Bandwidth Memory (HBM) supply chain (dominated by Samsung, Micron, and SK Hynix). Because fabs cost roughly $40 billion and take five years to construct, memory constraints will persist for years [00:08:58].
Software Margins on Hardware: Due to this choke point, memory makers like Micron are pulling 80-85% gross margins—margins typically reserved for SaaS companies, not physical component manufacturers [00:08:14].
Cerebras's Architectural Moat: Cerebras avoids this crippling HBM bottleneck completely because their wafer-scale design utilizes SRAM, etched directly onto the logic chip by TSMC [00:14:16]. This insulates them from the 4-5x price inflation and the CoWoS packaging constraints hamstringing traditional GPU manufacturers.
Speed, Inference, & The Neocloud Ecosystem
The Negative Market for Latency: Feldman emphasizes that speed is the ultimate differentiator in AI inference. He analogizes it to dial-up internet—there is a negative market for slow compute [00:21:47]. If Cerebras can run Kimi K2 6.7x faster than the next fastest GPU cloud, it creates compounding advantages in coding and agentic workflows [00:19:50].
Nvidia's Proxy War Strategy: Nvidia has strategically funded, backstopped, and overallocated GPUs to "Neoclouds" (like CoreWeave) to intentionally foster competitors against traditional hyperscalers (AWS, Azure, Google) [00:11:38]. While this creates an unhealthy dependence on Nvidia, it serves a market segment that simply wants "cheap compute" without the high-cost security and enterprise software layers (the "leather seats") offered by AWS [00:13:08].
Google's Vertical Integration Risk: While Google owning the full stack (TPU to data center) theoretically makes them the lowest-cost producer of tokens, historically, limiting hardware sales exclusively to one's own internal demand restricts total market volume and economies of scale [00:17:03].
Historical Paradigms & Deep-Time Context
The 1990s Telecom Crash vs 2026 AI: Feldman explicitly contrasts the current AI buildout with his experience during the late 90s fiber optic boom [00:02:55]. Telecom companies dug millions of miles of trenches based on speculative future demand, leading to a catastrophic bust. AI data center expansion is chasing existing, unfulfilled daily usage across demographics (from 11-year-olds to 85-year-olds) [00:10:40].
The Disappearance of the VP of Telco Infrastructure: Just as the mid-1990s rise of Cisco birthed the CIO role and killed the PBX hardware industry, AI will spawn Chief AI Officers while eradicating traditional middle-management "information gatherer" roles and drastically reshaping Human Resources [00:36:39].
The Shadow of William Shockley: Feldman grounds his relentless pursuit of intellectual horsepower in his childhood on the Stanford campus, where his father played tennis with Nobel laureates and William Shockley (co-inventor of the transistor) lived next door [01:01:09]. This environment normalized tackling globally impactful physics problems over chasing incremental wealth.
The Absolute Consensus on China: Feldman states there is zero debate among security professionals: if the US sells leading-edge chips to China, the Chinese military will use them, and the Chinese state will subsidize their industry to compete unfairly against American businesses [00:44:34].
The Choke Point Strategy: To prevent China from replicating American technological supremacy, the US must weaponize structural choke points. TSMC and ASML serve as these bottlenecks, allowing the West to keep industrial adversaries significantly "down-rev" on technological capability [00:46:51].
The Onshoring Mandate & Regulatory Relief: Rebuilding TSMC-level capabilities (and the surrounding packaging ecosystem) on US soil is an existential necessity [00:48:23]. Feldman proposes a radical policy: granting TSMC and Samsung a 20-year absolute exemption from all local US ordinances to build fabs seamlessly, treating these fabs as the "modern pyramids" they are [00:49:24].
Leadership, Hard Problems, and The Human Toll
The $8 Million Monthly Desert: Hardware is ruthlessly unforgiving. Cerebras spent 18 months burning $8 million every single month, failing to solve the physics of wafer-scale architecture every single day [01:05:16]. The board held steady, and Feldman's team learned from micro-failures until they achieved an architectural leap no one else has replicated to this day.
Creating 800 Millionaires: For Feldman, the true metric of leadership isn't personal wealth (which didn't change his lifestyle after his last exit). It is wealth distribution. His last company minted 100 millionaires; the Cerebras IPO minted 800 millionaires out of the engineers who dedicated a decade of their prime working years to the mission [01:01:43].
The CFIUS Battle & IPO Grit: Cerebras's successful $5.5B IPO was not luck; they attempted to go public 18 months earlier but were stymied by opaque, obstructionist CFIUS (Committee on Foreign Investment in the United States) inquiries regarding foreign customers [00:53:59]. Instead of quitting, they kept building through the delay, emerging as a radically stronger business when they finally listed [00:57:27].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Cerebras IPO Share Price Range
$185 to $311
The stock price appreciation verified during the opening public market listing.
The Infrastructure-Demand Inversion: Historically, physical tech booms (1880s rail, 1990s fiber optics) were characterized by a speculative "build it and they will come" methodology, where capital deployment massively outpaced actual consumer utility, leading to dramatic market bubbles [00:03:03]. The modern AI buildout flips this model. Because AI crossed a critical utility threshold in early 2025, data center construction is actively chasing relentless, multi-demographic daily usage. It is not a bubble; it is a desperate game of supply-chain catch-up quantified by a $25 billion industry backlog [00:03:43].
The Negative Market for Latency: In computing, there is no upper bound on the value of speed. Feldman asserts that just as there is zero market for a "slow search engine" or "dial-up internet" (even if heavily subsidized), the future of AI inference will mercilessly punish latency [00:21:47]. If a hard problem can be solved in 3 minutes on Cerebras versus 20 minutes on a legacy GPU cloud, the compounding advantage in agentic workflows and developer velocity means the slower technology will be entirely priced out of existence, regardless of marginal cost savings [00:21:02].
The Institutional Bottleneck (The "Saying No" Business): The primary drag on enterprise AI adoption is not technological capability or data cleanliness, but institutional risk architecture [00:38:16]. Internal lawyers and Chief Information Security Officers (CISOs) operate in environments where they receive no credit for innovation but suffer total career destruction for failure. Therefore, when confronted with precedent-free AI tools, their structural incentive is to default to "no." Massive productivity gains remain locked behind an administrative wall of fear until executive leadership forces adoption via decree [00:41:13].
The Geopolitical Supply Chain Choke Point: To contain an industrial and military adversary like China, physical supply chains must be weaponized as strategic choke points [00:46:51]. By strictly controlling the output of TSMC (wafer fabrication) and ASML (lithography systems), the United States ensures its adversaries remain at least a generation "down-rev." This framework dictates that exporting leading-edge silicon is indistinguishable from exporting sovereign military supremacy [00:44:34].
The "Pay Your Own Way" Good Neighbor Principle: Data center expansion is increasingly blocked by local municipalities angry over resource drain [00:27:32]. The framework to solve this is total financial autonomy and transparency. Instead of using legacy loopholes to amortize the infrastructure costs of new power lines onto local residents over 30 years, tech giants must pay upfront to upgrade the local grid, recycle their own water, and build community infrastructure [00:32:03]. Self-funding the externalities of compute removes the political friction of expansion.
6. Anecdotes
Beating the "Bozo Analyst" on Live TV: While watching a prominent financial analyst on television boldly claim that Cerebras could not achieve specific processing performance metrics, Feldman and his team simultaneously published their benchmarks proving they ran Kimi K2 6.7x faster than the fastest GPU cloud. Feldman tells this story to highlight the sheer joy of empirical victory over speculative punditry, cementing why execution always trumps narrative in Silicon Valley. [00:19:50]
The Vegas Buffet vs. The Freeway Meter: Feldman compares the current metering of AI compute (due to supply chain delays) to his 20s in Vegas at a buffet, where having everything available at once causes you to gorge and feel sick. He tells this to illustrate Gavin Baker's point: if data centers could be built instantly, demand would gorge on supply and cause market indigestion. The friction is actually saving the market from itself. [00:05:01]
Raising Capital in the Ashes of Lehman Brothers: When discussing the frustration of IPO delays due to regulatory pushback, Feldman recounts the summer of 2008. While Bear Stearns collapsed in March and Lehman Brothers exploded in September, he was out raising venture capital. He tells this to enforce a vital lesson for founders: macroeconomic catastrophes are out of your control. The only rational response is to keep building product and create distance from competitors. [00:58:00]
William Shockley and the Stanford Halloween: Growing up on the Stanford campus surrounded by academic giants, Feldman's next-door neighbor was William Shockley, co-inventor of the transistor. Yet, to the kids, Shockley was just the eccentric guy who handed out full-size candy bars on Halloween. This anecdote grounds Feldman's perspective on wealth; he was raised in an environment where intellectual horsepower, not money, was the only respected currency. [01:01:09]
The 18-Month, $8 Million-a-Month Desert: To illustrate the brutal reality of hardware engineering, Feldman shares that for 18 agonizing months, Cerebras burned $8 million every month without successfully building their architecture. He faced his board every six weeks to admit failure. He shares this to demystify CEO confidence—he had immense self-doubt—but it proves that rigorous failure analysis (failing differently every time) eventually yields breakthroughs that no competitor can copy. [01:05:16]
7. References & Recommendations
People
Andrew Feldman: Brought up as the central guest to share his decade-long journey founding Cerebras and navigating a successful public offering [00:00:12].
Harry Stebbings: The host, steering the conversation around macroeconomic trends, data center physical constraints, and venture dynamics [00:00:12].
Eleanor Roosevelt: Referenced incorrectly by Harry as "Alan Roosevelt" when citing the famous framework that great minds discuss ideas, average minds discuss events, and small minds discuss people [00:01:57].
Jensen Huang: Brought up regarding his macro-prediction that global AI infrastructure spending will cross $3–4 trillion by 2030 [00:02:27].
Gavin Baker: Cited for his thesis that local permitting delays and grid constraints act as helpful traffic meters preventing infrastructure hyper-indigestion [00:04:35].
Sam Altman: Mentioned for his unique ability to accurately trust exponential curves and pre-contract multi-gigawatt power grids before incumbents understood the need [00:05:36].
Elon Musk: Referenced as the source from whom OpenAI had to buy legacy-generation H100 hardware nodes during supply crunches [00:06:28].
Sarah Friar: Mentioned (phonetically as Sarah Fry) regarding her comments on cloud commoditization and the underlying economic reality of frontier model competition [00:11:15].
Marc Benioff: Cited (phonetically as Benov) regarding Salesforce’s internal financial allocation of $300M annually spent on Anthropic tokens to boost internal developer velocity [00:34:24].
Brad Smith: Praised for creating Microsoft’s corporate playbook on how data center operators must proactively engage local communities transparency-first [00:29:22].
Mark Zuckerberg: Noted (as Zuck) regarding the human reality behind Meta's broad organizational layoffs delivered via automated early morning messaging [00:32:46].
Lisa Su: Mentioned (as Lisa) along with Jensen Huang to argue that semiconductor executive self-interest should be ignored when establishing ironclad national security export controls against China [00:44:14].
William Shockley: Noted as Feldman's childhood neighbor on the Stanford campus who famously handed out full-size candy bars on Halloween, completely unaware of his monumental historic status to the children [01:01:09].
Companies
Cerebras Systems: The primary corporate entity discussed, highlighting its specialized on-wafer SRAM architecture and recent historic public market listing [00:00:12].
Nvidia: Discussed for its massive chip backlog, its multi-trillion market target, and its proxy war strategy of allocations to Neocloud startups [00:00:35].
AMD, Qualcomm, ARM: Grouped together as silicon architecture giants who will collectively iterate chip designs to dramatically lower the cost of compute over time [00:15:44].
OpenAI: Discussed as a principal driver of structural compute demand, scaling aggressively from its early model classes into hyper-scale infrastructure consumers [00:05:36].
TSMC: Brought up as the single most critical corporate manufacturing choke point in the global tech ecosystem, etching SRAM logic components directly [00:07:36].
Samsung, Micron, SK Hynix: Identified as the memory fabrication triopoly driving massive profits by command-pricing HBM units amid structural supply deficits [00:08:04].
AWS & Microsoft Azure: Mentioned for their core enterprise utility, offering robust baseline data protection layers ("leather seats") that justify premium hosting costs [00:11:55].
Google: Discussed regarding the economic risks and benefits of its vertically integrated TPU ecosystem, which limits hardware supply liquidity solely to its internal operations [00:16:11].
Nebius & CoreWeave: Brought up as prototypical "Neocloud" models optimized for hyper-fast GPU provisioning and capital restructuring via creative debt financing [00:18:42].
G42: Mentioned as the sovereign tech entity with whom Cerebras initially executed a milestone $1 billion foundational compute deployment deal [00:22:58].
Crusoe Energy Systems & SoftBank: Noted as forward-looking infrastructure operators driving massive multi-gigawatt data center expansion designs worldwide [00:25:08].
Meta: Noted for its organizational shifts and automated corporate downsizing following post-pandemic engineering adjustments [00:32:46].
Cisco: Brought up historically to explain how its mid-1990s rise directly established the modern enterprise CIO position out of necessity [00:36:16].
Palo Alto Networks: Cited to demonstrate how major security transformations historically birthed entirely new executive roles like the CISO [00:36:58].
Rolm: Cited as a historical titan of legacy PBX telecom architecture whose primary business completely evaporated as mobile cell networks emerged [00:36:52].
Anthropic: Referenced as a leading enterprise model builder capturing substantial subscription volume from major platform providers [00:34:24].
Mayo Clinic & Glaxosmithkline: Referenced to demonstrate the massive competitive advantage organizations possess if they maintained disciplined multi-decade data structuring practices [00:40:20].
Harvey & Agora: Cited as frontline legal vertical applications pioneering the direct automation and acceleration of legacy regulatory research [00:40:54].
Baidu, Tencent, Didi: Mentioned as elite Chinese technology platforms led by world-class builders with whom Feldman happily partnered before geopolitical shifts occurred [00:46:04].
Intel: Noted in connection with national defense, emphasizing that simply funding legacy logic foundries is insufficient without establishing core domestic packaging lines [00:48:47].
Mistral AI & SAP: Grouped as rare European technical standouts fighting against a regional landscape heavily bogged down by defensive regulatory policies [00:50:20].
Lovable, 11Labs, Synthesia, DeepMind: Mentioned as important examples of European-founded software and application talent counter-balancing infrastructural deficiencies [00:51:16].
SpaceX: Cited as a generation-defining private tech company whose pending public capital events alter late-stage market liquidity considerations [00:53:30].
Bear Stearns & Lehman Brothers: Brought up to vividly illustrate the sheer terror of attempting to secure venture funding rounds during the peak of the 2008 banking collapse [00:58:00].
Emirates Airline: Mentioned humorously because their corporate office tracks Feldman’s extreme travel metrics closely enough to send him holiday baskets [01:04:06].
Geopolitical Institutions
CFIUS: Cited as the primary regulatory body responsible for stalling Cerebras's public listing path due to protracted, opaque national security reviews [00:53:59].
Trump Administration: Mentioned as the executive shift under which Cerebras finally secured clear regulatory resolution to execute its delayed IPO path [00:57:01].
Media/Pop Culture
John Wick / Baba Yaga: Invoked in a joking analogy to describe how corporate attorneys view unseen risk phantoms similarly to the legendary cinematic boogeyman [00:41:49].
Pokémon: Used by Stebbings as a humorous generational baseline, noting he was an 11-year-old child playing video games while Feldman was out navigating the global financial crisis [00:58:18].
8. The Bottomline (by AI)
The prevailing narrative of an AI infrastructure bubble fundamentally misunderstands the physics of current market demand; the capital being deployed is frantically chasing a $25 billion backlog of real-time usage, not speculative futures [00:03:43]. As hardware economics hit a wall of crippling memory (HBM) costs and protracted fab construction timelines, non-standard architectures utilizing SRAM (like Cerebras) will hold a violent, asymmetric speed advantage [00:14:16]. Investors and operators must monitor the rapid shift from cost-per-compute to pure inference velocity, recognizing that in the next 36 months, there will be absolutely zero market tolerance for latency in agentic and enterprise workflows [00:21:47].
"Brookfield's the largest infrastructure owner in the world... We drew a pipeline and we showed all the different components of the payments ecosystem on a pipeline and said it's like a pipe that moves any commodity except what it's moving…
Memory (HBM) Cost Increase
4x to 5x
Extreme price inflation recorded for High Bandwidth Memory due to absolute factory oversubscription.