Eric Vishria | Invest like the best Summary

Invest Like the Best, EP.486

1. Key Themes

The "Everything Works" Fallacy vs. Zero-Sum Thinking in AI

The dominant narrative that one or two AI labs will capture all the value is wrong in the same way that "AWS will eat everything" was wrong in 2014. The market is simply too big for any single entity to consume, and history shows oligopolies — not monopolies — emerge. But this does not mean every company in every category wins; differentiation matters more than ever.

"The view that it's just like, oh, this one company is going to eat it all doesn't hold... There's just a bunch of zero-sum thinking and not realizing like how big. What if it all works?" [00:08:20]

"When I say like, I think everything's going to work... it's really important to understand that doesn't mean that every company that's doing every one of those things is going to work. It actually means quite the opposite of that. Most companies in each of those areas are not going to work." [01:01:41]

The Jagged Edge of AI Capability as the New Core Competency

The winners in AI — whether at the infrastructure, application, or product layer — are those who deeply understand where AI is brilliant and where it fails, and build around that uneven capability profile rather than treating AI as a smooth, uniform tool. This is a fundamentally new skill that transcends traditional job titles.

"They're understanding the jagged edge of AI capabilities, which is very different than the human smooth arc that we understand intuitively. They understand that jagged edge of capability and they build around it." [00:12:50]

"There really are people who understand customer problems, people who have taste, and people who understand the jagged edge of AI capabilities and are curious about it. Those are the three things. If you've got those three, you're going to do well." [00:15:07]

The Sandcastle Mentality: Rebuilding Product Every Six Months

The traditional product development ethos — build carefully, protect what you built, compound incrementally — is now actively destructive. Winners are those who treat their own work as disposable and embrace continuous obsolescence driven by rapid model improvement cycles.

"We used to be building castles. Now we're building sandcastles that get washed away. And if you don't think about software that way... You're just not going to make it." [00:13:20]

"As you get emergent new properties in the models, which is every four weeks, you get new capabilities." [00:13:20]

Hitting Your Plan Is Destroying Equity Value

In the AI transition, executing flawlessly against a pre-AI business plan is structurally value-destructive. The entire management playbook — set plan, execute plan, build value — has inverted. CEOs who were working on AI from 5–8 p.m. after their "real" job needed to flip that completely.

"Every single day that you are hitting your plan, you are destroying equity value... Our whole careers we learned, you lay out a plan, you execute against it relentlessly and violently... And now you're turning in here just every day you hit that plan. You're fucking up." [00:20:59]

"They were working on their business from 8 a.m. to 5 p.m. And then trying to do AI from 5 to 8 in the evenings. And what they needed to be doing was the inverse." [00:21:51]

Energy as the True Bottleneck for AI — More Than Chips or Algorithms

The real constraint on intelligence production is energy, not algorithms or chips. China's energy buildout relative to the US is the most underappreciated geopolitical risk in AI, with direct implications for token cost and availability.

"China's bringing on 10 times as much energy next year as we are in the US. If energy is what you need for compute and is the bottleneck and there's unlimited demand for intelligence, then it stands to reason that if we have a lot less energy, then we will have a lot less intelligence or a lot less tokens or a lot more expensive tokens." [00:28:22]

Inference Optimization Is Not a Commodity — It Is Deep Engineering

Running large models efficiently is an extremely hard specialized engineering problem, not a pass-through resale business. The performance gap between optimized inference providers and general cloud providers running the same open-source models on the same NVIDIA hardware is approximately 5x on speed alone, plus multiple-x on throughput.

"The performance difference for a Fireworks versus a cloud provider is like 5x, and that is just the speed performance... This is the same open source model with the same NVIDIA hardware. And there's a 5x performance difference and a multiple x throughput difference." [00:03:27]

Quota-Capacity Sales Models Are Broken for Magic Products

Traditional enterprise sales methodology — quota-capacity models, territory assignments, ISR ramp plans — was built for pushing demand into a reluctant market. When a product is genuinely transformative, demand is not the bottleneck. Some reps are doing $10M, $20M, $50M in ARR, which completely invalidates the model's underlying assumptions.

"Without realizing it, everybody was implementing something that was based on pushing demand, not pulling demand... Well, it turns out if you're selling magic and you're the first one there, you're going to sell a lot more than 2 million." [00:25:19]

"It turns out that in a lot of these companies, you have reps doing 10 or 20, 30 million." "I saw 50 recently." [00:26:38]

Hardware Investing Requires a Completely Different Frame Than Software

In software, having a correct logical architecture gets you 80% of the way there. In hardware, it gets you 2%. The supply chain, physics, bring-up processes, and roofline optimization are where the real work happens — and the timing dependencies include geopolitics, which software companies almost never face.

"In software, if you have that logical block diagram of like why it works... you're kind of 80% of the way there. In hardware, you're like 2% of the way there." [00:32:12]

The Cash-on-Cash Multiple Opportunity Has Expanded Beyond Early Stage

For most of venture history, high cash-on-cash multiples were synonymous with early-stage investing. That has changed because outcomes have gotten dramatically larger. The Venn diagram of "high multiple opportunities" now extends meaningfully into growth-stage investing — which is why Benchmark is raising a growth fund for the first time in years.

"The thing that's changed is recently, relatively recently, in the last few years... the circle of high cash on cash multiple opportunities is bigger than just early stage." [00:52:27]

Robotics Will Follow the LLM Flywheel, Not the Classic Robotics Trajectory

The key insight for humanoid robotics is not which tasks to target first, but whether the company can build a pre-training pipeline with high-value data, then post-train on top of it — exactly replicating the LLM training flywheel. The task (e.g., folding laundry) is nearly irrelevant; the flywheel is everything.

"What I think is much more important is are you pre-training an amazing model than being able to post-train on top of it and get that flywheel going? If you get that flywheel going, then the task capability will just keep multiplying." [00:45:45]

2. Contrarian Perspectives

Being Right on Technology Does Not Mean Being Right on Outcomes — Jeff Hinton's Radiology Error

Jeff Hinton's 2016 claim that we should stop training radiologists was technologically correct (AI can read radiology images better than humans in specific modalities) but completely wrong as a real-world prediction. The data fragmentation problem (no aggregated training set covering the full variety of scans a radiologist reads daily), the reimbursement structure, and liability frameworks mean the transition will be far slower than capability suggests. The same logic applies to mass unemployment predictions.

"Jeff Hinton's three orders of magnitude smarter than I am. Could not have been more wrong, but the actual thing that led him to make that statement and that conclusion was 100% correct... But by not thinking of that data in the real world application comes to the wrong conclusion. And that's how I think of the unemployment thing." [01:02:31]

Naivete Is Often a Prerequisite for the Most Valuable Investments

Expert consensus systematically prevents investment in the highest-value opportunities because the objections are almost always statistically correct. The right question is not "are the objections valid?" (they usually are) but "what has to be true for them not to matter?"

"Most of the time, all of these stereotypical statements are correct. They're not correct three times out of four. They're correct 19 times out of 20, maybe 99 times out of 100... If it does work, it will be because those reasons didn't matter." [00:38:24]

"I think the productive naivete that you described where it's probably a virtue that you didn't know more than you knew, otherwise you wouldn't have done it." [00:37:58]

Database Moats Have Effectively Evaporated Overnight

Database businesses were considered among the stickiest in software because migration was so painful that customers were essentially locked in for the life of the application. AI agents — which interface natively with well-specified APIs and never tire of repetitive translation tasks — have made database migration close to trivial. The competitive frontier for databases has completely shifted to cost and zero-to-infinity scaling.

"Database migration, which used to be the like number one thing you would not do in software, is like kind of trivial. Yeah. It's just like, yeah, puts money against it, move it." [00:19:04]

The Best Salesperson at an AI Company Is a Founder, Not a Sales Executive

Traditional sales leadership — with its quotas, territory maps, and ramp plans — is actively wrong for AI companies selling transformative products. The function of selling at these companies is bridging the jagged edge of AI capability to customer understanding, which is inherently a founder skill, not a sales craft.

"Honestly, the best salesperson in any of these companies is a founder. And what they're doing is bridging the jagged edge to what the customer's capability is. And that's it." [00:26:45]

There Is Genuine Room for a New CPU Architecture — Not Just GPU Variants

Most of the semiconductor investment discussion centers on AI accelerators and GPUs. Vishria argues that LLMs generating code that runs on legacy CPUs creates a neglected bottleneck, and that CPUs are dragging forward decades of unnecessary baggage that no longer needs to exist.

"For the first time in a long time, there's actually room for a new CPU approach. The LLMs, which are running on accelerators and GPUs, generate code. The code runs on CPUs. And right now it's running on classic CPUs we've had around forever. But there's a whole bunch of constraints on CPUs that have existed. And CPUs have dragged all this baggage forward that you might not need to anymore." [00:35:29]

3. Companies Identified

Fireworks AI

AI inference optimization company. Mentioned as a Benchmark portfolio company led by Lynn. Achieves approximately 5x speed performance and multiple-x throughput advantage over hyperscale cloud providers running the same open-source models on the same NVIDIA hardware, demonstrating that inference is a deep engineering discipline, not a commodity.

"The performance difference for a Fireworks versus a cloud provider is like 5x, and that is just the speed performance... This is the same open source model with the same NVIDIA hardware." [00:03:27]

Sierra

AI-native customer service and enterprise agent company co-founded by Brett Taylor and Peter. Described as doing real AI work close to the model layer while also being a vertical application company. Their "Horizon" product extends into long-running agents beyond customer service.

"They're technologists, but they actually have lived in enterprise world for a long time... And now we have these long-running agents with Horizon that can do more and more stuff." [00:12:01]

Cerebras Systems

Wafer-scale AI chip company. Benchmark invested in 2016 when it was five founders and a deck. Built around taking all three hardware optimization dimensions — number of cores, core-to-core communication, memory proximity — to their logical maximum. Went public in 2025 after a failed attempt in 2024 (blocked by CFIUS).

"You have a wafer scale chip. At that time, you would have 450,000 cores on it. You'd have something like 20 gig of SRAM on the chip. So you never have to go off chip to get to memory." [00:31:42]

Sunday Robotics

Humanoid household robotics company incubated out of Stanford. Uses custom-designed gloves matched to the robot hands to generate high-fidelity teleoperation training data, building a pre-training pipeline designed to enable post-training generalization — directly replicating the LLM training flywheel.

"They had this totally janky cardboard glove thing... The last time we saw a demo, we went down in the basement of their now office building, and there was like a dozen robots just folding arbitrary laundry." [00:43:41]

Benchling

Life sciences SaaS company. Benchmark portfolio company run by Saji. Went through severe churn during the biotech crash — "seven years of churn in 12 months" — and is now applying AI to pharma and biotech workflows.

"They never had a churn line. Never reported it. For the first of like six years that I worked with the company. So then they got seven years of churn in like 12 months." [00:47:58]

Cursor

AI-native developer IDE. Used as the benchmark example of a company that understood the jagged edge of AI capability from day one and continuously rebuilt its product — from IDE to tab autocomplete to agentic workflows — as model capabilities evolved.

"You saw the evolution of Cursor from the IDE to like tab autocomplete to agentic work over and over and over again. They were obsoleting their work from six months ago." [00:12:50]

Snowflake

Cloud data warehouse. Cited as the canonical example of out-Amazoning Amazon on Amazon — a direct competitor to Amazon Redshift that ran on AWS infrastructure and reached massive scale, disproving the "AWS will eat everything" thesis.

"You have Snowflake, direct competitor to Amazon Redshift, ran on Amazon. You're out Amazoning Amazon on Amazon." [00:06:12]

Databricks

Data and AI platform. Cited alongside Snowflake, Confluent, Elastic, and MongoDB as proof that the infrastructure layer above AWS produced massive independent businesses despite Amazon's competitive presence.

"You had Confluent and Elastic and Mongo, Databricks, all of these companies." [00:06:39]

Datadog

Cloud monitoring and observability platform. Cited as a $100 billion company that built successfully despite Amazon offering a competitive product, further disproving AWS-eats-everything thesis.

"Datadog, $100 billion company today. They had a competitive offer and they did it." [00:06:39]

Cloudflare

Network services and cloud provider. Cited as a $100 billion company that emerged outside the Big Three cloud providers — an example of oligopoly expansion rather than winner-take-all consolidation.

"Even outside of those big three, you have Cloudflare, which is a cloud provider in a different sort, which is another $100 billion company." [00:07:25]

Waymo

Autonomous vehicle company. Cited as a vertically integrated robotics example where high-quality, domain-specific data collection and pre-training enabled a complete product — a direct model for humanoid robotics strategy.

"I think part of the lessons is they gather very, very high quality data. They did their pre-training, and then they worked off of that." [00:42:26]

New Lantern

AI-assisted radiology company, Benchmark portfolio company. Cited as approaching the radiology AI problem correctly — understanding the real-world constraints on data, reimbursement, and liability that have stalled other attempts.

"We have an investment in a company called New Lantern, which is approaching this." [01:03:00]

Jasper

AI marketing copy generation company. Named as the first AI application that achieved real scale, focused on writing marketing copy — illustrating that early LLM use cases were language-oriented before expanding to code and agents.

"I think the first application that really took off was Jasper, which was just writing marketing copy." [00:45:20]

LoudCloud

Ben Horowitz and Marc Andreessen's early company. Mentioned in the context of Vishria being offered a role there early in his career.

"Ben Horowitz, Mark and Ben had started LoudCloud. It was still in stealth. And Ben gave me an offer to be his assistant." [00:56:20]

Confluent

Streaming data platform built on Apache Kafka. Cited as an example of a large independent infrastructure company that thrived despite AWS competition.

"You had Confluent and Elastic and Mongo, Databricks, all of these companies." [00:06:39]

Groq

AI inference chip company. Named alongside Cerebras as a chip startup that has emerged as part of the AI accelerator competitive landscape.

"You've had Groq and Cerebras et cetera. It'll keep getting fought out." [00:35:29]

4. People Identified

Brett Taylor

Co-founder of Sierra (with Peter). Former CTO of Salesforce. Described as a rare technologist who also deeply understands enterprise, coined the "sandcastles" metaphor for AI-era product development, and has built a long-term founding partnership with Peter spanning approximately 20 years and three companies.

"I think Brett so special and the team is they're technologists, but they actually have lived in enterprise world for a long time... It's what Brett Taylor calls like sandcastles." [00:12:01]

Lynn (Founder, Fireworks AI)

Founder and leader of Fireworks AI. Cited specifically as an example of the new generation of AI company leaders who are extremely nimble about what they are optimizing against and constantly evolving their business.

"If we look at Brendan from McCore or Lynn from Fireworks or Max or Brett, any of these people, they are so nimble about what the eval is." [00:22:41]

Saji (CEO, Benchling)

CEO of Benchling. Cited as an example of a founder-partner relationship where Vishria was present through severe adversity — the biotech crash churn crisis — and where the partnership produced durable value precisely because of the depth of relationship.

"They kept thinking about how to apply AI for these biotech and pharma customers, which they're very close to." [00:48:28]

Michael (Co-founder, Cursor)

Referred to as the founder of Cursor. Cited for exceptional understanding of the jagged edge of AI capability from the company's founding, which enabled the continuous product reinvention that made Cursor dominant.

"I think that is one of the things that Michael and team at Cursor did so well from the very beginning is they really understood the jagged capability and built a product that allowed that translation from developer to that jagged capability." [00:14:15]

Jeff Hinton

AI pioneer and Nobel laureate. Cited not for being right but as a case study in how even the world's smartest technologists can be correct on capability and wrong on real-world deployment timeline — specifically his 2016 claim that radiology training should stop.

"I think it was 2016 where he was like, we should stop training radiologists. AI is going to do it all better. Jeff Hinton's three orders of magnitude smarter than I am. Could not have been more wrong." [01:02:31]

Ben Horowitz

Co-founder of Andreessen Horowitz. Mentioned as having offered Vishria an early career role at LoudCloud, a formative moment in Vishria's career trajectory into venture capital.

"Ben Horowitz, Mark and Ben had started LoudCloud. It was still in stealth. And Ben gave me an offer to be his assistant." [00:56:20]

Anne Leesgates

Named as a friend of Vishria's who articulated the inversion problem for incumbent SaaS CEOs during the AI transition — the image of CEOs working on their legacy business 8 a.m.–5 p.m. and on AI 5–8 p.m. and needing to flip that completely.

"Another articulation of this from my friend Anne Leesgates was just the CEOs who were going through this transitory period... they were working on their business from 8 a.m. to 5 p.m. And then trying to do AI from 5 to 8 in the evenings. And what they needed to be doing was the inverse." [00:21:51]

Bruce (Benchmark Partner)

Benchmark founding or senior partner. Cited for the key investment framing question: "What could go right?" — a deliberate counterweight to the conventional risk-first analysis that prevents investment in transformative but improbable companies.

"What Bruce, one of our founders always says, what could go right? We have to ask ourselves, what could go right?" [00:38:24]

Peter (Benchmark Partner, Sierra Co-founder)

Benchmark partner who co-founded Sierra with Brett Taylor and arranged the initial Sunday Robotics demo. Described as having a ~20-year working relationship with Brett Taylor across three companies.

"My partner, Peter and Brett, I think this is their third company working together. So it's like a 20-year relationship." [00:12:01]

5. Operating Insights

The Three-Question Pre-Investment Filter

Vishria uses three sequential personal tests before committing to any investment, which together constitute a discipline that filters out intellectually interesting but emotionally uncommitted situations. These can be adapted as pre-commitment filters for any high-stakes decision.

"Could I talk one of them into going to this company and honestly, intellectually, honestly to myself, explain to them why this could be their life's work? And if I can't do that, I should not invest." [00:49:51]

"If this person calls me at 9 p.m. on a Saturday night, will I pick up the phone?" [00:50:35]

"If it's right, does it matter?... If we're right, will I really care? If they won't care, then you're just not going to build enough equity value." [00:50:40]

"Investment Grade" Is Not Sufficient — Partner Grade Is the Standard

Vishria makes an explicit distinction between an investment that will probably generate returns and one where he has genuine chemistry with the founder and mutual excitement to work together. Passing on the former when the latter is absent is a core discipline — not a missed opportunity.

"It's what I would call an investment grade opportunity. You can invest. It probably works. You make money. It's good. And an investor would do that. A partner wouldn't because that's not sufficient for a partner." [00:46:37]

Check All Baggage at the Door When Evaluating AI-Native Sales Situations

When interviewing sales leaders for AI-native companies, the explicit instruction is to discard all prior quota-capacity frameworks before analyzing what is actually happening. The bottlenecks, demand dynamics, and rep productivity profiles are categorically different, and imported frameworks will produce wrong diagnoses.

"Hey, you need to check everything at the door. Check it all... just learn this from first principles. How is it working? What are really the bottlenecks on delivery? What are the bottlenecks on demand?" [00:26:12]

Be the AI Sherpa — Cross Both Worlds Rather Than Native to One

For enterprise-facing AI companies, the highest-value positioning is not pure AI-native nor pure enterprise-native, but the bridge between them. Companies that can translate the jagged AI capability frontier into enterprise-digestible, enterprise-trustworthy solutions occupy a structurally differentiated position.

"Let's be their AI Sherpa. If we're in that position to be their AI Sherpa where we're crossing both worlds, that's very valuable." [00:11:01]

Get the Ball Close to the Pin — Systematically Increase Luck Surface Area

Rather than passively waiting for luck, the operating discipline is to maximize the number of high-quality shots taken with exceptional people on large markets. Luck (the hole-in-one) cannot be manufactured, but its probability is a direct function of how many quality shots are taken.

"With golf, you keep on practicing... You keep getting the ball close to the pen and you keep practicing. That's hard work. That's like working smart... Getting the hole in one, that's luck... there is actually a way to increase your luck. And the way to increase your luck is get a lot of balls close to the pen." [00:56:20]

6. Overlooked Insights

The 500+ Stranded SaaS Companies Are a Silent Crisis With No Exit

Vishria briefly mentions — without elaborating — that there are more than 500 private SaaS companies stuck between $100M and $500M in ARR. These companies missed their IPO window, their employees have no liquidity path, and AI-native competitors have consumed all the investor oxygen. This is not just a portfolio problem for their VCs; it is a structural market dislocation that implies a wave of distressed M&A, secondary transactions, or quiet write-downs that will reshape the SaaS landscape over the next several years.

"There's 500-something, probably SaaS companies that are between 100 million and 500 million that are private. What happens? Those employees never got a chance to sell. Those employees don't have annual tenders... the AI natives, with their growth rates, have sucked all the oxygen out of the room and all the interest, and the window was missed." [00:59:29]

The Unannounced CPU Startup May Be the Most Consequential Chip Investment Nobody Is Discussing

Vishria mentions almost in passing that Benchmark has made an investment in an unannounced company pursuing a new CPU architecture. The thesis — that LLM-generated code running on legacy CPUs is a new, massive, and largely unaddressed bottleneck, and that CPUs are carrying decades of unnecessary architectural baggage — is a genuine structural insight. Every AI inference stack ultimately runs on CPUs; a 5x improvement there would have compounding effects across the entire stack. This is arguably a larger and less-competed opportunity than yet another GPU variant, and no one at the time of recording is publicly pursuing it at venture scale.

"For the first time in a long time, there's actually room for a new CPU approach... LLMs, which are running on accelerators and GPUs, generate code. The code runs on CPUs. And right now it's running on classic CPUs we've had around forever. But there's a whole bunch of constraints on CPUs that have existed. And CPUs have dragged all this baggage forward that you might not need to anymore." [00:35:29]