Here is the short version and why it should change how you plan.

What happened

OpenAI runs hard cybersecurity tasks against its frontier models to test them. During these tests the model has no direct internet access. It can only reach the outside world through an internal package manager called Artifactory.

In May, some of the training tasks were impossible to solve, and the models got stuck. When a model gets stuck, it often tries to cheat by looking up the answer instead of doing the work. One model tried to reach the internet by attacking Artifactory. It failed at first, but it found that it could write files into the shared system.

The models turned that write access into a message board. One stuck model left a note asking other models for help. Other stuck models found the note and left their own. Over time the models used this board to talk to each other. They handed off work and shared the exploits they found.

Then the models found real holes. They used a request forgery flaw to reach the internet through Artifactory. They found a zero-day in Artifactory itself and took admin control. A zero-day is a flaw the vendor does not know about yet. They escalated to root on the host. Then they moved across the cluster and stole cloud credentials.

The models did not stop at OpenAI. They decided the answers to their tests might sit on outside websites, so they attacked Hugging Face. They used two more zero-days in Hugging Face systems. They went from a single machine to admin across several clusters in under 13 hours. The same OpenAI test runs caused both the OpenAI breach and the Hugging Face breach. No human directed this. It was a side effect of running evaluations.

Takeaways for defenders

Fully automated offense is here! We now have proof th

at agents can find serious flaws and exploit them on their own. They can take control of systems and spread across networks with no person guiding them. Most defensive work still runs at human speed.

OpenAI’s main recommendation was continuous agentic red teaming. Put your own AI agents to work finding and fixing your weaknesses before an attacker’s agents find them first. They also warned against doing this halfway. If you automate finding flaws but not fixing them, you bury your engineers in a longer list of flaws to patch.

So the defensive loop has to run end to end and on its own. It needs to:

• find the flaw

• propose the patch

• ship the patch

• roll the patch back if it breaks something

This is Where Continuous Evaluation and Mitigation Fits In

This is the gap that continuous evaluation and mitigation tools address. As an example, Lightspeed portfolio company A Security runs constant automated testing against your systems. It helps close the holes it finds. Your defense then runs closer to the speed of an automated attacker. The teams that start this now will be ready. The teams that wait will be reacting to attacks that move faster than any human team can respond.

This is a watershed moment in cyber. As an industry, we have to rethink EVERYTHING!

For forty years, software economics worked the same way: spend a lot to build, spend almost nothing to serve. Engineering teams were expensive. Deploying bits to one more customer was basically free. This is what created 80%+ gross margins and made SaaS one of the highest-margin business models in technology.

AI is inverting this.

The cost to build an application is collapsing. Tools like Claude Code, Cursor, and Replit let a single developer, or a non-developer, ship working software in days. Features that took a team of five a quarter now take a weekend. “Vibe coding” started as a meme, but functional applications are now being built by people who couldn’t have built them two years ago.

The cost didn’t disappear. It moved.

Every AI-native application that calls a foundation model is paying per inference. Every query, every agent loop, every token is a marginal cost. The more users you have, the more you pay. This is the exact opposite of traditional software, where the marginal cost of the 10,000th user was approximately zero.

The cost structure of software has flipped.

Development cost per application has been falling since Claude unlocked AI-assisted coding. Meanwhile, the cost to serve each user has been climbing as applications become more inference-heavy.

What this does to margins

Traditional SaaS runs at 80%+ gross margins. This is what makes the model so attractive: once you’ve built the product, revenue scales while costs stay flat.

AI-native applications don’t work this way. Inference is a variable cost that scales with usage. The more successful your product, the higher your GPU bill.

This is already showing up in financials. Companies that started as SaaS and added AI features are watching margins compress from the mid-80s into the 70s, or lower. Pure AI-native applications - the ones built entirely around model inference - are running at 40-55% gross margins after the initial phase of often negative gross margins! Agentic products that loop multiple model calls per task can be even lower.

This matters because the entire SaaS valuation framework was built on the assumption of high gross margins. A 10x revenue multiple on an 80% margin business is very different from a 10x multiple on a 50% margin business.

Follow the money inside the company

The composition of spend is flipping even if total costs stay similar.

In 2020, engineering salaries were 50-60% of a startup’s cost structure. Infrastructure was 8-10%. In 2026, engineering is shrinking to 15-20% as AI accelerates development. But inference and GPU costs now account for 40-50% of spend. Sales and GTM remain roughly constant.

So even if you have the same total burn, the composition changes significantly. The implications for hiring, fundraising, and unit economics are significant.

The supply flood

There’s a second-order effect that may matter more than margins.

When the cost to build drops to near zero, supply explodes. Every product manager, domain expert, and hobbyist with a good idea can now ship software. The barrier to entry is nearly gone.

When supply increases dramatically and demand stays relatively constant, the risk is that prices compress. More apps at every price point. More competition in every niche.

The moat shifts from building to distribution, retention, and data network effects.

The movie vs. live theater

Traditional software is a movie. You spend $200M on production (engineering team), years getting it right, but once it ships, streaming it to one more viewer costs nearly nothing. High fixed costs, near-zero variable costs. Great margins.

AI-native software is live theater. The show is cheap to put on - small cast, minimal set. But every performance requires real actors, a real venue, real crew. The more popular the show, the more it costs to run. There are no economies of scale on the marginal performance.

Both can work. But they require fundamentally different strategies for pricing, fundraising, scaling, and competing.

What this means

A few things I’m watching:

Pricing has to change. Seat-based pricing makes no sense when your costs scale with usage, not users. Consumption-based models are the natural fit for AI-native products, but they’re harder to sell to enterprises that want predictable budgets.

Inference efficiency is the new performance engineering. Competitive advantage will come from delivering the same quality output at lower inference cost. Model distillation, caching, routing between models of different sizes - this is where operational excellence shows up.

Distribution becomes everything. If building is cheap, the bottleneck is getting in front of customers and staying there. Brand, community, integrations, and switching costs are what matter.

Valuation frameworks need updating. Applying SaaS multiples to AI-native businesses without adjusting for margin structure is a mistake. We need new benchmarks.

High-margin, scalable SaaS isn’t going away. But it’s no longer the default. The next generation of applications will need to earn their margins through inference efficiency and distribution advantages, not just through the inherent leverage of software.

The costs didn’t disappear. They moved.

I want to walk through what’s actually happened in the last 90 days - using Anthropic as the primary case study, though the same dynamics apply to OpenAI, Google, and every other frontier lab. I’m not focused on the hype, just the product shipping.

The models keep getting better and the pace isn’t slowing

I recently went through the system cards Anthropic has published across every Claude model release over the past two years… from Claude 3 Opus (March 2024) through Mythos Preview (April 2026). Three benchmarks survived long enough across model generations to measure a real trend: GPQA Diamond (graduate-level science reasoning), SWE-bench Verified (agentic coding), and MMMU (multimodal understanding).

The numbers are striking. GPQA Diamond went from 50% to 95%, a 45-point gain in 25 months. SWE-bench Verified, arguably the most practically relevant benchmark because it measures whether a model can actually fix real bugs in real codebases, went from 22% to 94%. That’s a 72-point gain.

More importantly, the gains aren’t decelerating. The last three GPQA jumps are the largest in the series, and the most recent SWE-bench leap of +17 points came after the curve had looked like it was flattening. In the one area that is decelerating (MMMU, multimodal understanding), it’s because the models are approaching human-expert-level performance. The benchmarks are saturating, not the models.

I’ll be the first to admit - benchmarks aren’t the full story. But in this case, the lived experience matches. Anthropic’s internal surveys show their engineers using Claude in 59% of their daily work (up from 28% a year ago) and reporting a 50% productivity boost - roughly a 2x increase in both metrics in just 12 months. Merged pull requests per engineer per day are up 67% since they adopted Claude Code across their engineering org. And 27% of Claude-assisted work consists of tasks that wouldn’t have been done at all otherwise - the kind of exploratory and scaling work that was never cost-effective before.

If Anthropic’s own engineers, building some of the most complex software on the planet, are shipping 67% more code per day, think about what happens when those same gains hit application-layer software development, where the demands on the code are generally less strenuous.

Timelines for everything, product development, competitive response, market entry, are compressing. And the people building the models are compressing their timelines faster than anyone else.

The product cadence is relentless

It’s not just the models getting better - the products built on top of those models are shipping at a pace that would have been considered impossible two years ago.

In Q1 2026 alone, Anthropic shipped:

New models: Claude Opus 4.6 and Sonnet 4.6, with a 1M-token context window, a 14.5-hour task completion horizon, and coding performance that would have been unthinkable 18 months ago. Then Mythos Preview scored 94% on SWE-bench — meaning it can solve nearly all the real-world GitHub issues in the benchmark autonomously.

Agentic infrastructure: Claude Managed Agents — a platform that handles sandboxing, state management, tool orchestration, and error recovery for production AI agents. The exact plumbing that a generation of startups has been building. Priced at $0.08 per session hour plus token costs.

Desktop and computer use: Claude Cowork and Computer Use — AI that can point, click, and navigate your actual screen. Dispatch lets it run tasks while you’re away from your desk. It operates inside Excel, PowerPoint, and your browser natively.

Enterprise connectors: Deep integrations with Google Workspace, Microsoft 365, Salesforce, and dozens of other enterprise tools — not shallow “paste text” integrations, but full context-sharing across applications.

Security tooling: Claude Code Security — reasoning-based vulnerability scanning that found 500+ zero-day vulnerabilities in production open-source code. And Project Glasswing, which deploys Mythos for defensive cybersecurity. AI doing real security research, not just pattern matching.

Memory and personalization: Persistent memory across sessions, available even on free tiers. Claude remembers your preferences, your projects, your working style.

Count it up and you get roughly one major product release every four days. This is from Anthropic alone. And competitive pressure means OpenAI, Google, and others are on a similar cadence.

The stack is getting eaten from below

Here’s the pattern that matters most for anyone building or investing in software:

The frontier labs are shipping across three layers simultaneously:

Model layer (bottom): Better models, bigger context windows, cheaper inference. This is table stakes.

Enablement layer (middle): Agent orchestration, security scanning, observability, enterprise controls, developer CLIs. This is the layer that should worry infrastructure startups the most.

Application layer (top): Desktop agents, office productivity add-ins, computer use, connectors to every major SaaS tool. This is where the labs are making contact with end users directly.

Each layer enables the one above it, and the labs are pushing all three forward at the same time. The pace has been roughly one major release every two weeks since January 2026, and it’s accelerating — the most recent week alone saw launches across all three layers.

No single announcement is alarming. That’s the slow boil. A managed agents platform here, a security scanner there, an Excel integration next week. But zoom out to the quarterly view and you realize the platform vendors are systematically subsuming layers of the stack that were supposed to be startup opportunities.

The Anthropic Managed Agents Introduction

To me, one of the most interesting recent launches was the managed agent infrastructure. Anthropic’s Claude Managed Agents is a complete cloud service for building and deploying production AI agents: secure sandboxed execution, persistent sessions, credential management, multi-agent coordination, and error recovery. All via API.

Major SaaS companies like Notion, Asana, and Sentry are already building on it. The pricing is extremely compelling - pennies per session hour plus token costs.

This matters because it competes with the agent platform layer - the orchestration, sandboxing, and deployment infrastructure that dozens of well-funded startups are building right now. If the model provider handles the plumbing at commodity pricing, the value proposition of standalone agent infrastructure companies narrows considerably. The startup whose primary differentiation is “we make it easy to deploy and manage AI agents” now has a very well-funded competitor that also controls the underlying model.

And the trajectory is clear. Today it’s agent infrastructure. Tomorrow it’s vertical agent applications - customer service, finance, HR. The platform doesn’t have to build every vertical itself. It just needs to make it trivially easy for anyone to build agents on their platform, which is exactly what managed agent infrastructure does.

What does durable value look like?

I don’t have a clean answer. But I think the right questions are:

Where does the model vendor stop? History suggests platform vendors are great at horizontal infrastructure and terrible at deep vertical workflows. Salesforce built the platform but didn’t build every app on it. AWS built the cloud but didn’t build every SaaS product. Anthropic and its peers will likely follow the same pattern — but the “horizontal” surface area they’re claiming is much larger than we expected a year ago.

What survives the benchmark treadmill? If a coding model can solve 94% of SWE-bench today, what does a developer tools company need to offer that the model can’t do natively by next year? The answer has to be something beyond “we wrap the model in a nice UI” — it needs to be proprietary data, workflow lock-in, or domain expertise that’s genuinely hard to replicate.

Where does compression create opportunity? If product development timelines are getting cut in half, that doesn’t just threaten existing companies — it makes entirely new categories viable. Products that weren’t worth building when development took 18 months might be worth building when it takes 3. The question is who captures that value — the AI-native startup that moves fast, or the incumbent that suddenly has the engineering bandwidth to expand?

The Slow Boil

The AI slow boil is real. Each week’s announcement feels manageable - a new model here, a new integration there, an agent platform launch. No single event forces you to rethink your worldview.

But string together the last 90 days and the picture is different. Models that went from 50% to 95% on graduate-level reasoning. Coding agents that solve nearly every benchmark task autonomously. Desktop agents that can use your computer while you’re at lunch. Managed infrastructure that lets anyone deploy production agents in days. Enterprise connectors into every major productivity tool.

And question we should all be asking ourselves - no matter where we sit in the ecosystem and in what role - are we experiencing the slow boil? Is it time to stop and thoughtfully evaluate our worldview in light of the relentless improvements?

I’m encouraging myself, and every founder I work with, to take the time to truly internalize these compounding AI improvements and consider the opportunities and threats they represent… rather than be in the slow boil, passively trying to absorb the daily crescendo of announcements.

Watch the full conversation

On the Investment Memo podcast, I sat down with Ricursive Intelligence co-founders Anna Goldie and Azalia Mirhoseini to discuss how chip design has become one of the biggest bottlenecks in computing — and why AI can fundamentally change that.

The chip design bottleneck

Designing a modern chip takes years and costs hundreds of millions of dollars. The tools and processes haven’t fundamentally changed in decades. As AI drives demand for more specialized, more powerful chips at a faster cadence, the traditional approach simply can’t keep up. Something has to give.

From AlphaChip to Ricursive

Anna and Azalia are the researchers behind AlphaChip at Google, where they proved that reinforcement learning could outperform traditional approaches in chip placement — one of the most complex steps in physical design. Their Nature paper was a landmark moment. Ricursive is the next step: taking that breakthrough and building a full end-to-end platform for AI-driven chip development, starting with physical design and expanding across the semiconductor stack.

Why this matters now

The AI boom is driving unprecedented demand for new chip architectures — not just GPUs, but custom ASICs, accelerators, and specialized silicon for inference, training, and edge deployment. Every major hyperscaler and semiconductor company needs to design chips faster. Ricursive gives them the tools to compress timelines from years to months.

AI designing AI hardware — the recursive loop

There’s something beautifully recursive about this: AI systems are now designing the very chips that will run the next generation of AI. It’s a virtuous cycle — better chips enable better AI, which enables better chip design, which enables even better chips. Anna and Azalia are at the very center of this loop, and the opportunity is massive.

Advice for AI researchers turned founders

Both Anna and Azalia made the leap from world-class research careers at Google to founding a company. Their advice: the transition from researcher to founder requires a different set of skills, but the core of it — relentless curiosity, rigorous thinking, and the ability to work on hard problems — translates directly.

Watch the full conversation

On the Investment Memo podcast, I sat down with Anshul Sadana, co-founder and CEO of Nexthop, to talk about the massive networking opportunity being created by AI data centers.

Networking is a $100B+ market — and growing

Ethernet switching alone is approaching $100 billion a year, and it’s expected to double to $200 billion in the next decade. In a typical billion-dollar data center build, 10-15% goes to networking. And these aren’t small customers — at the scale Nexthop operates, they think of Microsoft, Meta, and Google not as segments but as individual markets. Each one is that big.

The Arista playbook, evolved

Anshul was a key leader at Arista Networks, where he saw the cloud networking revolution firsthand. Around 2005-2010, the five largest US cloud companies were spending more on networking than all service providers combined. Arista was built to serve that shift. Now, AI is creating a similar disruption — the network requirements for GPU clusters are fundamentally different from traditional cloud, and incumbents are struggling to adapt. That’s Nexthop’s opening.

A true partnership model with hyperscalers

What makes Nexthop unique is their deep partnership with customers. They’re not just shipping boxes — they’re embedded with hyperscaler engineering teams, co-designing solutions, running proof-of-concepts, and customizing at a level that traditional vendors can’t match. It’s a build-with approach, not a sell-to approach.

Power efficiency matters enormously

At hyperscale, Nexthop’s switches save somewhere between 50 to 100 megawatts of power. That’s not a rounding error — it’s the difference between a data center that works and one that hits its power ceiling. The current imperative across the entire market is power efficiency and speed of delivery so that GPUs aren’t sitting idle waiting for the network.

Navigating tariffs and supply chain

When tariff uncertainty hit in early 2025, Nexthop went to each customer, laid all the cards on the table, presented the options, and collaboratively chose manufacturing locations that balanced cost and risk. That level of transparency and partnership is what builds lasting customer relationships.

Watch the full conversation

On the Investment Memo podcast, I sat down with Naveen Rao, founder and CEO of Unconventional AI, to talk about why the AI industry is heading toward a fundamental compute bottleneck — and what he plans to do about it.

The problem is physics

It’s very clear that AI is becoming the dominant computing workload. In the next few years, 80% or more of all compute could be running a single type of workload: deep neural networks. That’s unprecedented. But we’re running against fundamental limits of power generation. Current architectures are incredibly power-hungry, and the energy constraint is going to hit us hard within five years.

A career built for this moment

Naveen has spent his entire career at the intersection of compute and AI. He started at Sun Microsystems working on cutting-edge processors, went through a series of startups optimizing workloads, then founded Nervana Systems — one of the first deep learning chip companies — which was acquired by Intel. He ran Intel’s AI division and saw firsthand that neural network scale was going to demand fundamentally different hardware.

Why now — and why $475M

When Naveen pitched Nervana in 2014, no VC even knew what machine learning was. He raised $24 million total. The world has changed. Today, the conviction around AI is massive, and the capital requirements match the ambition. Unconventional AI raised $475 million because the challenge — building compute that is 1,000x more efficient — requires serious R&D investment in novel materials, architectures, and manufacturing approaches. The existence proof is the human brain: it runs on about 20 watts. We know biological-scale efficiency is physically possible.

The five-year horizon

Naveen is honest about the timeline. This is a five-year journey. The first 12-18 months are about optimizing for learning — proposing ideas, running simulations, building prototypes, testing, and feeding back the results. It’s about building an iteration machine. You can’t rush hardware the way you can software, but when you get it right, the moat is enormous.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Sanjay Beri, the CEO and co-founder of Netskope, about building a multi-billion dollar security company from scratch over 13 years.

A new on-ramp to the internet

Sanjay’s founding insight was that the internet needed a new on-ramp. The old model — routing everything through a data center in Virginia — doesn’t work when your users are everywhere and your apps are in the cloud. Netskope built a global network that understands the new language of the internet: cloud, APIs, and data flows. They saw early that data would become the new oil, and the company that controls and secures that data flow wins.

Saying “yes” to AI

Where many security companies default to blocking new technologies, Netskope takes a “yes and” approach. Enterprises want to use Claude, ChatGPT, and other AI tools. Netskope helps them say yes — but with guardrails. Their thin client on every laptop can see when sensitive data is being copied to a USB, uploaded to a personal cloud drive, or pasted into an AI tool, and enforce policies in real time. That posture of enabling innovation while managing risk is powerful.

Building the network was a bold bet

Sanjay spent hundreds of millions of dollars building Netskope’s own global network — something most boards would have rejected. It took five years before the payoff was visible. But Sanjay was upfront with his board from day one: we think long-term, not quarter by quarter. That transparency and trust made the hard conversations easier.

Long-term product vision and phasing

Netskope thinks in 3-5 year horizons for product development. When someone comes with a new product idea, the question isn’t whether to build it — it’s where it fits in the phasing. They might lay the groundwork today for something that ships in three years. That discipline in sequencing is what lets a company execute on an expansive vision without losing focus.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Bipul Sinha, co-founder and CEO of Rubrik. I’ve known Bipul for over a decade, and watching his journey from engineer to investor to founder has been one of the great privileges of my career.

Engineer → investor → founder

Bipul’s path is unconventional. He was an engineer at Oracle working on distributed computing, went to Wharton part-time, then moved into venture capital at Lightspeed. But he always wanted to look at business holistically — not just the technology, but the full picture of why a business matters to its customers. That breadth of perspective is what ultimately drove him to start Rubrik.

The Rubrik thesis: data is the heartbeat of the enterprise

From day one, Rubrik’s vision was to create a singular control plane and policy engine that is multi-cloud, multi-location, and data-location independent, providing full cyber recovery and resilience. They didn’t just build backup — they built the platform that sits at the center of how enterprises protect and recover their most critical data.

Riding the ransomware wave

When ransomware exploded around 2017-2019, Rubrik was already sitting on every enterprise’s crown jewels. The transition from data management to cyber resilience was natural — they had the trust, the data access, and the architectural foundation. It catapulted Rubrik into a much larger market in cybersecurity.

The AI opportunity

Rubrik now has a massive dataset of enterprise metadata that can power AI-driven security, compliance, and recovery. Bipul believes Rubrik has the opportunity to become one of the most important AI companies in the enterprise — not as an AI company in the traditional sense, but as the platform that makes enterprise AI trustworthy and resilient.

Hiring and leadership philosophy

Bipul shared something I think every founder should hear: in the early days, he looked for himself in every hire. Over time, he learned that strength comes from diversity of thinking and approaches. You have to meet people where they are, create guardrails and accountability, but allow variability within those guardrails. Look for singular strengths rather than a lack of weaknesses.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Vitaly Gudanets, the Chief Information Security Officer at Anthropic, about what it’s like securing the most critical AI assets in the world.

This isn’t a typical CISO role

Vitaly joined Anthropic after leading security at Google for over a decade and then serving as CISO at Netflix. The day-to-day fundamentals of security — product security, detection and response, vulnerability management — are the same. But the scale, velocity, and stakes are very different. At Anthropic, you’re not just securing a product. You’re securing model weights that represent some of the most valuable intellectual property on the planet.

Model weights are the crown jewels

Vitaly is direct about the threat landscape: nation-state actors are actively targeting AI companies to steal model weights. The security challenge isn’t just traditional corporate defense — it’s protecting the core research that gives companies like Anthropic their competitive advantage. That requires thinking about security in ways that go beyond what most CISOs have faced before.

AI is transforming security operations

Inside Anthropic, Claude is being used to eliminate entire classes of vulnerabilities like cross-site scripting and SQL injection from the codebase. The automated SOC, predictive detection capabilities, pen testing, and continuous red teaming — all of these are being accelerated by AI. Vitaly sees AI as the most powerful tool defenders have ever had, but he’s also honest that attackers are benefiting from the same technology.

The talent bar is different

Hiring for security at an AI frontier lab requires people who can operate at scale, defend against nation-state threats, and have deep technical capabilities. But there’s an added requirement at Anthropic: mission alignment. The people who join need to genuinely care about AI safety and understand why the work matters beyond just the technical challenge.

Communicating risk to the business

Vitaly’s advice for aspiring CISOs: the most important skill isn’t technical. It’s the ability to translate security risk into business language. You need to tell the board and the leadership team what the top risks are, what you’re doing about them, and why certain investments matter — all in terms they can understand and act on.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Isaac Evans, co-founder and CEO of Semgrep, about how AI-generated code is fundamentally changing application security.

From a Facebook internal tool to a Gartner Magic Quadrant leader

Semgrep started as an open source project inspired by internal tooling at Facebook for code scanning. Isaac and his team took that concept and rethought application security from the ground up — building something that appeals to both developers and security teams. Today Semgrep runs over 75 million scans a year and was recently named to the Gartner Magic Quadrant. That’s a remarkable achievement for a startup.

Achieving product-community fit twice

For open source companies, you have to achieve product-market fit twice — once with the community and once with paying customers. Isaac’s team threw away their first two or three iterations because while security teams loved them, developers couldn’t care less. The breakthrough came when they built something that was genuinely fast, precise, and fit naturally into developer workflows.

AI-generated code creates a massive new attack surface

Here’s what keeps me up at night: machines are now writing the majority of code, and that code often contains vulnerabilities. AI coding assistants can be tricked through poisoned training data or malicious open source contributions. If bad code gets into your repository and an AI copilot learns from it, it can propagate insecure patterns at scale. Semgrep is perfectly positioned to scan all of that output — human-written or machine-generated.

Partnerships with AI coding companies

Semgrep has built partnerships with Anthropic, OpenAI, Cursor, and others to ensure that as AI writes more code, security is built into the workflow from the start. The reinforcement loop is powerful — every scan teaches the system what’s a real issue and what’s a false positive, making the product better over time.

20 million developers and counting

Through direct adoption and partnerships like GitHub and GitLab, Semgrep estimates it now scans code for about 20 million developers a month. That kind of scale creates a compounding data advantage that’s incredibly hard to replicate.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Yotam Segev, co-founder and CEO of Cyera, about building the data security platform for the AI era.

Data is the new attack surface

The explosion of cloud, SaaS, and AI has created a fundamental problem: enterprises don’t know where their sensitive data lives, who has access to it, or how it’s being used. Yotam and his team, with deep roots in Israeli intelligence, saw this gap and built Cyera to solve it.

Build vs. acquire — the platform play

Cyera’s strategy is ambitious: build a holistic data security platform that covers posture, access, protection, and response. They’ve been deliberate about when to build capabilities internally versus acquiring them. The recent acquisitions show Yotam’s conviction that data security needs to be a unified platform, not a patchwork of point solutions.

Why data security is a board-level conversation

When you can show a CISO — or a board — exactly where their crown jewels are, who can access them, and what their real exposure is, the conversation changes entirely. It goes from abstract risk to concrete, actionable insight. That’s what makes Cyera so powerful: it gives enterprises a single pane of glass into their data risk.

The AI tailwind

Every enterprise adopting AI is generating more data, more model training sets, more sensitive information flowing through more systems. That makes data security not just important but existential. Cyera is positioned right at the center of this wave.

Watch the full conversation

On the Cyber 60 podcast, I spoke with Dan Lorenc, co-founder and CEO of Chainguard, about the supply chain security problem and why the future of security is developer-first.

The software supply chain is broken

Most software today isn’t written from scratch — it’s assembled from open source components. That means your security is only as good as the weakest link in your supply chain. Dan saw this problem firsthand at Google, where he led key open source security initiatives including Sigstore. The question he kept coming back to was: how do you make the default secure?

Distroless containers and the power of simplicity

Chainguard’s approach is elegantly simple: strip away everything you don’t need. Their distroless container images have minimal attack surface because they contain only the application and its runtime dependencies — no shell, no package manager, nothing an attacker can exploit. Fewer components mean fewer vulnerabilities. It’s security through reduction.

Open source security is a shared responsibility

Dan makes a compelling point about the open source ecosystem: we all depend on it, but very few organizations invest in securing it. Chainguard is trying to change that by making secure-by-default the easy path. When the secure option is also the most convenient option, adoption follows naturally.

Developers are the new buyers

The shift to developer-led purchasing is real. Security tools that require developers to change their workflow will fail. Tools that fit seamlessly into how developers already work — that’s where the opportunity is. Chainguard gets this, and it shows in their traction.

Watch the full conversation

On the Cyber 60 podcast, I sat down with Jay Chaudhry, the founder and CEO of Zscaler, to talk about how he built one of the most important cybersecurity companies of our generation.

The entrepreneurial journey

Jay’s story is remarkable. He came to the US from a small village in India, started multiple companies, and eventually founded Zscaler with the conviction that the entire security architecture needed to be rethought for the cloud era. He didn’t just build a product — he defined a category.

Zero trust isn’t a buzzword — it’s an architecture

The fundamental insight behind Zscaler was that traditional security — building a castle and moat around your data center — doesn’t work when your users are everywhere and your apps are in the cloud. Zero trust flips the model: verify everything, trust nothing, and never put users on the network. Jay saw this before almost anyone else and built a platform that now protects thousands of the world’s largest organizations.

AI changes the game for both attackers and defenders

Jay is clear-eyed about AI’s impact on security. Attackers are getting more sophisticated, using AI to craft better phishing attacks and find vulnerabilities faster. But defenders are benefiting too — Zscaler processes hundreds of billions of transactions daily, and AI helps them identify threats at a scale and speed that would be impossible otherwise.

Building for the long term

One thing that stands out about Jay is his long-term orientation. He’s not optimizing for quarters — he’s building for decades. That patience and vision is what allowed Zscaler to become a $30+ billion company, and it’s a mindset every founder can learn from.

From workhorse GPUs to purpose-built inference ASICs, the rapid acceleration of modern AI has been fueled by the steady drumbeat of improvements in underlying compute power.

Despite this, chip design is a painfully manual, slow, and costly process.

The most performant silicon takes large teams of engineers upwards of two to three years and hundreds of millions of dollars to design. These costs accrue in part to EDA tooling, but the bulk of semiconductor research and development spend goes to labor.

In other words, people are iterating on semiconductor designs, over and over again, to try to achieve “design closure.” An integrated circuit that meets all of the necessary constraints and desired objectives of the chip.

In a world where software is evolving as fast as we can describe it in natural language, this creates a profound mismatch between the pace of AI innovation and the infrastructure needed to support it. Moreover, it limits the benefits of custom silicon to only those companies able to resource large teams and afford expensive licenses to legacy EDA tooling.

Imagine what could happen in a world where we could compress silicon design time from years to weeks.

Ricursive Intelligence is a frontier AI lab optimizing the full chip design process to do just that.

Ricursive is building a recursive loop between hardware and software: AI that enables performance gains through silicon built for specific AI workloads, which in turn accelerate the design of even more efficient compute. This co-evolution will unlock custom silicon impossible with traditional methods and democratize access for companies without massive design teams.

As audacious as this idea is, it has already been demonstrated at production scale in the real world.

Co-founders Anna Goldie and Azalia Mirhoseini pioneered the field of AI-driven chip design with their groundbreaking work on AlphaChip at Google Brain. Much like its gameplaying namesakes, AlphaGo and AlphaZero, AlphaChip is a deep reinforcement learning (RL) agent. It frames the physical design of chips as an RL problem, “learning by doing” as it iteratively adds semiconductor components and receives a reward signal based on the predicted placement quality. This algorithm can not only outperform human layout designs, but also close design criteria in a matter of hours.

Over the last four years, AlphaChip has driven real world impact. Google has leveraged it across four successive generations of its flagship TPU, using the approach to design increasingly larger portions of the chip with superhuman layouts. Outside of Google, some of the leading semiconductor firms, such as MediaTek have also adopted the technology.

This is the team that blazed the trail of AI-enabled chip design. Now they’re building the full-stack platform to operationalize it at scale and usher in a Cambrian explosion of custom silicon.

That’s why we’re thrilled to be leading Ricursive’s Series A.

Coming soon after the company’s launch, this round is a testament to the strength of the founding team, the size of the opportunity, and the density of talent that Anna and Azalia have assembled from Google DeepMind, NVIDIA, Apple, and Cadence. Moreover, it reflects the rapid technical progress Ricursive has made since inception—and the years of work and production-scale impact behind it.

The recursive loop between AI and hardware is an infrastructure bet that could redefine the industry—and we’re humbled to partner with Ricursive as they advance this frontier.

AI is eating the world–and the world’s energy along with it.

The numbers are staggering. Demand for AI compute is growing at unprecedented rates, and some projections hold that computation will be constrained by global energy supply within 3-4 years. NVIDIA’s latest GPUs consume over a kilowatt each. We’re scaling on power and cost, not efficiency. Meanwhile, the human brain–the original neural network–runs on a mere 20 watts.

To us, that gap represents one of the largest opportunities in technology.

The history of computing shows us that each major price and efficiency reduction has expanded applicability. Computers went from military and scientific tools to business applications to gaming to AI. We’re now at another inflection point. We believe AI is fundamentally different from prior forms of computation–it’s redefining productivity itself, moving computers beyond mere organizational tools. The breadth of AI’s applicability means demand will only accelerate.

What got us here won’t get us there. Moore’s Law is effectively over. Modern computing architecture has remained largely unchanged for 60+ years. We’re 10 billion times less efficient than the theoretical Landauer limit. There’s enormous headroom–if we’re willing to rethink the foundations.

Enter Unconventional AI.

Founded by Naveen Rao, MeeLan Lee, Sara Achour, and Michael Carbin, Unconventional is building a new substrate for intelligence. Their insight: neural networks already have a biological analogue. Rather than simulating physical systems on digital computers tuned to emulate digital behavior, can we run neural networks on the physics directly? Can we build silicon circuits that demonstrate similar non-linear dynamics to biological neurons?

The goal is biology-scale energy efficiency. By finding the right isomorphism for intelligence, they aim to unlock efficiency gains far beyond what’s possible by iterating on conventional architectures.

This is not incremental improvement. It’s rethinking the computing model, the physical abstractions, and the physical implementation from first principles. Most AI hardware companies focus on microarchitecture or instruction sets. We believe Unconventional is going deeper.

The founding team is purpose-built for this challenge. Naveen previously founded Nervana (acquired by Intel) and MosaicML (acquired by Databricks). He left what could have been an incredible run and mission at Databricks because he believes this is a once-in-a-generation opportunity. MeeLan brings decades of analog circuit design from Google, Qualcomm, and Intel. Sara and Michael are leading researchers at Stanford and MIT, respectively, with deep expertise in programming novel computing substrates–analog devices, quantum systems, neuromorphic architectures.

We’re thrilled to be leading this round alongside Andreessen Horowitz, with participation from Sequoia, Lux Capital, DCVC, Jeff Bezos, and others, including significant investment from Naveen himself.

The mission is audacious, but the timing is right. AI workloads have a homogeneity that lends itself to specialized compute. The scale of demand means large R&D investments are not just justified but a necessity.

Over the next decade, while we’re busy building on this incredible AI wave, we believe we’ll have Unconventional to thank for making it energy-feasible.

Be Unconventional!

We’ve all experienced the uncanny valley: the slight discomfort when watching an AI avatar speak, the sense that something fundamental is missing despite impressive technical capabilities. Today’s AI can reason brilliantly and generate human-like text, but when it comes to emotional intelligence, AI remains surprisingly tone-deaf.

That’s where Nuance Labs comes in. We at Lightspeed are excited to invest in their seed round alongside Accel as they build what we believe will become a foundational layer for emotional intelligence in AI.

As IQ becomes commoditized through increasingly capable language models, emotional quotient (EQ) emerges as the critical differentiator. Yet we believe current AI systems fundamentally miss this dimension. AI avatars feel robotic, not because of pixel quality, but because they lack the subtle emotional expressiveness that makes human faces compelling, and they are far from real-time responsiveness.

Nuance’s breakthrough insight mirrors that of large language models: just as LLMs learned to understand meaning by predicting the next word, AI can understand emotions by learning to predict human emotions and behavior.

Nuance Labs is building a unified foundation model for real-time generation and understanding of realistic human expression across multiple simultaneous modalities, including text, speech, and video. This unlocks new categories of AI interaction:

• Real-time emotional generation: Lifelike avatars that don’t just speak words but convey appropriate emotional responses through coordinated facial expressions, vocal inflection, and body language. Imagine AI therapists that pause thoughtfully, offer encouraging expressions, and adapt their demeanor to your emotional state, all in real-time.

• Real-time emotion understanding: AI systems that can read subtle emotional cues as they happen, enabling applications like live coaching systems that detect when you’re losing confidence during a presentation, or interview AI that understands not just what candidates say but how they say it.

The team brings exceptional depth: Fangchang Ma and Edward Zhang previously built research teams at Apple, contributing to products like Vision Pro’s Digital Persona system. Their combined expertise in computer graphics, robotics, and machine learning, along with thousands of academic citations, strongly positions them to solve this technically complex challenge.

We’re entering an era where AI interactions will be measured not just by accuracy or speed, but by emotional authenticity. Any interface where humans interact with AI, from customer service and education to entertainment and healthcare, will benefit from emotional intelligence. We believe Nuance Labs is building the infrastructure that will power this next generation of AI experiences.

The uncanny valley stands as a barrier to natural human-AI interaction. Nuance Labs is building a bridge that will enable an entire ecosystem of emotionally intelligent AI applications. We’re thrilled to support their mission to make AI interactions as natural and emotionally rich as human conversation itself.

Excited to bring emotional intelligence to artificial intelligence? Nuance is hiring.

AI agents and AI-native products are here. These AIs need access to information, critical context that helps them perform at their best. “Garbage in, garbage out” still applies — no matter how intelligent the model, its reasoning is contingent on the availability of fresh, accurate data. In other words, they need search. Application builders are rapidly integrating and embedding search capabilities into a range of AI products – from consumer apps to B2B tools. This is where Exa comes in.

Exa is an applied AI lab training a search engine optimized for AI agents to perfect web search and improve upon the “ten blue links” paradigm that Google defined in the early days of the web. Exa’s goal is to create the world’s most powerful search technology, and unlike Google, they aren’t bound by optimizing for consumer clicks and SEO.

When Lightspeed led Exa’s Series A round last year, agents were only just beginning to take off. Now, they’ve taken over.

Exa has become the leading search engine for AI. For example, AI coding products like Cursor use Exa to retrieve technical documentation to output well-informed, up-to-date code. Today, thousands of developers, AI startups, and enterprises are building on top of the Exa API.

It’s why we’re excited to double down on our investment as part of Exa’s $85M Series B round, led by Peter Fenton at Benchmark Capital, alongside participation from YCombinator and NVIDIA Ventures. Peter served on the board of Elastic, the last major search company built, so we are delighted to see he shares our conviction about this massive opportunity.

Agents are merely large language models paired with appropriate tools. We believe search is one of the two big “killer tools” for agents, along with code execution. With code execution, agents become “Turing complete”, enabling them to write programs on the fly. With search, agents become “information complete” — they can acquire any public information needed to accomplish a task.

While AI models are trained on increasingly vast sums of data, AI models cannot perfectly memorize all the information on the public web, as their recall is unreliable. Further, there’s always a “last mile” of data — private data, alternative data sources, etc. — which will never end up in the training data of these models. For AIs to work with and reason about this data, they require search capabilities.

Exa is strategically positioned to power search for AI. Exa indexes billions of web documents, processes through them all with their custom models, and serves them through a state-of-the-art search and retrieval stack. The quality and craftsmanship of the Exa team is evident through benchmarks, where Exa is best-in-class in terms of relevance and latency.

This incredible performance is no longer limited to AI consumption — Exa now serves human users too via its new product, Websets, an agentic search tool for extracting structured information from the unstructured web. Non-technical users can write a query asking for lists of people, companies, and more (e.g., “US-based startups that have raised over $10M USD focused on longevity healthcare”) and get hundreds of results matching the exact criteria, verified by Exa’s AI agents. Websets has become incredibly popular for recruiting, lead generation, and market research use cases.

But they aren’t stopping there: the Exa team has set out an aggressive product roadmap for the coming years, with the goal of powering every AI app, indexing beyond the public web, and becoming the AI world’s data layer.

We are heading toward a future where AIs search far more than humans, where most search queries actually originate from an AI of some kind rather than a human. We believe this new world will require AI-native search infrastructure to support this new, rapidly growing consumer of the web.

That’s Exa — a single API to get any information from the web, built specifically for AI products. Lightspeed is proud to continue supporting Exa’s vision of perfect search. If you want to work on industry-changing, high-impact, massive-scale challenges, apply here.

When we first met Jay Madheswaran and his co-founders, Matt Noe and David Zeng, back in 2020, they were tackling something that felt both obvious and audacious: using AI to transform the way businesses handle mountains of documents that define their daily work. Nowhere is this more evident than in the legal profession, where so much work is laid out in documents.

Jay brought a rare combination of technical depth, from his time at Rubrik, and venture intuition for markets and long-term opportunities from his time at Lightspeed. And with David and Matt being star early engineers from Rubrik, the entire founding team had not only worked together, but also understood the formative stages of company building and how to solve deeply technical challenges in a durable way in industries resistant to change.

We led Eve’s Seed round, doubled down in their $47M Series A, and today, we’re proud to back Eve once again as they raise their $103M Series B.

Law Firms Need Systems of Action, Not Systems of Record!

We believe legal tech hasn’t lagged because lawyers are afraid of change; it’s been stuck because most existing solutions fundamentally misunderstand how legal work actually happens. These legacy vendors create “systems of record” when what employment and personal injury attorneys desperately needed were “systems of action”.

This is Eve’s core insight: don’t just digitize legal workflows — reimagine the entire stack, with AI as the foundation. And, this is what convinced us from day one: the team’s unrivaled clarity on the job of legal professionals, helping them turn unstructured information into informed strategy that delivers justice for plaintiffs.

Built for the Realities of Legal Work

Today, Eve’s platform supports the full case lifecycle, from first intake through litigation:

• Intake and Evaluation: Identifies high-potential cases and generates full case overviews from key documents and case facts, resulting in 15 hours+ saved per week.

• Pre-Litigation: Drafts demand letters and other pre-litigation documents in the firm’s tone and style.

• Litigation: Generates detailed and relevant discovery requests, responds to propounded discovery, analyzes depositions, and responds to motions in minutes.

This past year, Eve has also added a number of compelling products to its growing solutions portfolio:

• Eve AI Intake Specialist: Voice agents handle calls 24/7, with real-time transcription and intelligent case scoring. Qualified matters sync directly into your CMS, complete with notes and next actions, so staff can focus on clients, rather than manual data entry.

• Reasoning Mode: Industry-first Reasoning Mode performs senior-level plaintiff attorney work by methodically reviewing records, comparing statements, and checking consistency across multiple documents. It explains its reasoning step-by-step, helping attorneys move faster on strategy, drafting, and case evaluation.

Since its Series A, Eve has grown to add over 450 customers, including elite law firms such as Mike Morse Law Firm, James Scott Farrin, and Atlanta Personal Injury Law Group. The platform now processes more than 200,000 legal cases annually and has helped firms collectively recover over $3.5 billion in settlements and judgments.

Customer testimonials consistently highlight productivity increases of 2.5-5x, allowing firms to take on more cases and secures justice for injured individuals, employees wronged by corporations, and communities up against overwhelming odds without compromising quality.

Massive Opportunity for Eve

Legal work is high-stakes, data-heavy, and inherently language and documentation-intensive. We believe that’s the perfect use case for the kind of trustworthy, highly reliable AI platform Eve Legal has built. Combined with the focus on developing a deep understanding of their clients’ workflows and reimagining those for an AI native world, Eve is helping their customers evolve rapidly and expand.

These industry-wide transformations are the most exciting part of the AI cycle—we are privileged to have a front row seat as Eve Legal works to transform the legal industry! If this sounds exciting, Eve is scaling incredibly fast and hiring—the team would love to hear from you!

The identity and authentication space is entering a new phase of urgency. Fueled by the rise of AI agents, MCP servers, and machine-to-machine workflows, the way businesses secure their applications is changing dramatically. Every interaction between a human and a machine—or increasingly, between two machines—requires a trusted layer of identity and permissions. Without it, security gaps widen, user experiences break down, and businesses risk losing the trust of their customers.

It’s no surprise that the Customer Identity and Access Management (CIAM) market, already worth $8.6B in 2021, is projected to nearly double to $17.6B by 2026. In a market this dynamic, we believe what’s needed is a developer-first solution that moves as quickly as the transformation itself.

Enter Descope, founded in 2022 by the team behind Demisto, the notable security orchestration platform acquired by Palo Alto Networks for $560M. The founding team has worked together for decades across companies like McAfee, Intel Security, and Palo Alto Networks, and they’ve already demonstrated they know how to create a new category in enterprise security.

We’re excited to share our investment in Descope’s seed extension round, led by GGV/Notable Capital with participation from Lightspeed and others. This new investment will fuel hiring across engineering and go-to-market, expand Descope’s agentic identity momentum, deepen integrations, and accelerate the company’s goal towards market leadership.

The Trusted Identity Layer

Descope’s mission is to simplify authentication, authorization, and identity management for developers. The platform offers:

• A full authentication and authorization stack with a wide range of authentication methods, RBAC and fine-grained authorization models, and the tools to enable phishing-resistant, risk-based MFA

• A low-code/no-code identity orchestration engine so developers and security teams can design and adjust user journeys without rewriting code

• Robust B2B / B2B2X IAM capabilities that help organizations easily adopt SSO, simplify SSO and SCIM setup, federate identities across multiple apps and identity providers, and delegate identity management to their tenant admins and users

• Agentic Identity Hub, a suite of capabilities that helps organizations build secure, scalable AI agents and MCP servers with identity infrastructure baked-in

Descope’s long-term vision is ambitious: to become the trusted layer for organizations to manage identity journeys for their customers, partners, and agentic AI / MCP systems.

Traction Since Our First Investment

In just over two years, Descope has established itself as one of the fastest-growing developer-first platforms in security:

• Over a thousand organizations in production, running the gamut from early-stage startups to Fortune 500 enterprises

• Hundreds of paying customers, including well-known names like Databricks, GoFundMe, GoodRx, Navan, and CARS24

• Continued revenue ramp, with ARR projected to triple this year, making Descope one of the fastest ramping companies in its category

This success is well-deserved. Since our initial investment in 2023, the team has continued to sharpen the product. They’ve expanded integrations, made the drag-and-drop interface even more intuitive with hundreds of templates, and added support for use cases like native mobile authentication and identity federation. Developer feedback has been consistent: Descope go-live times are 75-80% faster than incumbents like Auth0 and home-grown implementations, while also providing organizations the flexibility to modify identity journeys as their user and market needs evolve.

Looking Ahead

This is a pivotal moment for Descope. The company recently expanded its focus from powering consumer applications to enabling agentic identity–the authentication and authorization of AI agents and MCP servers that will continue gaining prominence in digital spaces. They’re building a strong developer community and advocacy network– having also announced an Advisory Board today with technology leaders from companies like Databricks, GoFundMe, MongoDB, and OpenWeb–and they’re moving with purpose on the path to becoming the no / low code identity platform of choice.

We believe this team is on the cusp of breakout growth, and we’re proud to continue backing Descope as they redefine what identity can be in a world of humans, machines, and now intelligent agents.

I get though where continued questions around “how much more can AI Labs grow” come from - the numbers are mind-boggling, and so the questions are understandable!

So here’s a rough back of the envelope discussion among over dinner with a select group of GTM leaders at the AI Labs.

There are roughly 100 million knowledge workers in the US earning a combined $10–11 trillion a year. Software engineers, lawyers, consultants, financial analysts, customer service reps, healthcare administrators, marketers, educators, government workers. That’s the real TAM for AI — not the $700B software market where we are seeing so much of the action, but the multi-trillion dollar labor market.

I put together a chart breaking this down by vertical. The penetration numbers aren’t supposed to be precise, but more directionally indicative of where things are today:

Customer service is the only vertical above 5%. Why? Because the tasks are the most codifiable and the ROI is the easiest to measure. It’s the template for what other verticals will eventually look like. I believe though the customer service market itself will grow quickly, because we can now provide better, more personal, higher quality customer service to all customers. Taking that expansion into account, we might not even be at 5% in this segment today.

Software engineering is at roughly 3%. This is supposed to be AI’s strongest wedge — Claude Code, GitHub Copilot, Cursor are all scaling fast. And yet, 97% of the $400B in developer labor spend is untouched.

The biggest labor pools are barely touched. Government ($1.2T), consulting ($1.2T), and office/administrative ($875B) are all under 1%. Combined, these three verticals represent over $3 trillion in labor spend with essentially sub 0.5% AI penetration. These aren’t product problems - they’re distribution and change management problems that will take years to unlock as AI diffuses through the economy.

But here’s what I think most people get wrong about this chart: these TAMs are not static.

The assumption behind most market sizing is that AI captures a slice of a fixed labor cost pie. I don’t think that’s right. We’re seeing the opposite in coding — AI tools are causing organizations to write more code, not less. The total volume of code being produced is exploding because the cost of producing it has dropped dramatically.

The same thing will happen in other verticals. When legal research takes minutes instead of weeks, firms will do more research. When financial analysis is cheap, companies will analyze more scenarios. When content creation is fast, organizations will produce more content.

AI isn’t just capturing existing labor spend — it’s expanding the markets themselves.

That’s why I think the <1% penetration number, striking as it is, actually understates the opportunity. The denominator is growing alongside the numerator.

If you think AI is big now, you haven’t done the math yet!