SETH LEVINE's VC ADVENTURE

The Primary Is Tuesday. Vote - and Vote for Phil.

Sun, 28 Jun 2026 09:00:00 +0000

Ballots are due Tuesday, June 30 for the Colorado primary. If yours is still sitting on the kitchen counter, consider this your reminder: vote. Your vote genuinely matters.

It’s worth saying again: I’m voting for Phil Weiser for governor. I hope you will too.

I laid out the full case last month. The short version hasn’t changed. Phil has the most affirmative, pro-growth vision in the race - welcoming new industries, modernizing Colorado’s regulatory environment rather than just sloganeering about it (“like garlic in cooking,” as he puts it), and investing in the workforce, housing, and infrastructure that growth actually requires. When more than 400 Colorado business leaders signed an open letter on the state’s innovation future, Phil was the only candidate it was addressed to who responded with anything substantive. That tells you how he’d govern - by asking how he can help. He’s been doing the same around AI policy, which is why he received the endorsement from a group of business-forward but civically minded technologists who outlined a positive AI vision for Colorado.

I align with Phil on the things I care most about: small business, rural Colorado, entrepreneurship, and a smarter regulatory environment, balanced against a real commitment to education, healthcare, and the environment. I have a lot of respect for Michael Bennet, but Colorado is better served with him in the Senate.

The race is an incredibly close one, and it’s going to come down to the wire. The most recent polling shows Phil pulling ahead, but the surveys have bounced back and forth all month. What’s clear is that it’s a very tight race - and with a meaningful share of voters still undecided, it comes down to who actually returns a ballot.

So vote. Get your ballot into a drop box or vote center by 7 p.m. Tuesday. And I hope you’ll join me in voting for Phil.

Om Malik

Fri, 26 Jun 2026 12:00:00 +0000

Om Malik passed away this week. He was one of the greats - smart, thoughtful, and kind.

He was always willing to engage on an interesting idea or question, but he took his time and thought deeply before weighing in. That combination is rare.

He was a wonderful writer and an outstanding photographer. His blog, On my Om, and his photography at Photos by Om, are both worth your time.

Mostly, he was an amazing human. The world was a better place with him in it.

An Evolution, Not a Revolution

Thu, 18 Jun 2026 12:00:00 +0000

The International Economy just put out its latest issue, built around a single question — “Does American Capitalism Need to Be Reimagined?” They posed that question to seventeen economists, historians, and investors. This one was, of course, right up my alley: it’s the core thesis of Capital Evolution. My answer is that the status quo isn’t an option — but neither is a wholesale replacement. What we need is an evolution, not a revolution. The full piece is below; you can find the complete symposium at international-economy.com.

Does American Capitalism Need to Be Reimagined? By Seth Levine — The International Economy, Winter 2026

The status quo is not an option. But neither is replacing capitalism. What’s needed is an evolution, not a revolution.

In 2019, Jamie Dimon led the Business Roundtable in rewriting its fifty-year-old Statement on the Purpose of a Corporation, replacing shareholder primacy with a commitment to all stakeholders. Over 180 CEOs signed it. The left dismissed it as performative. The right called it woke. What it actually signaled was the end of the neoliberal consensus that had governed American capitalism for half a century.

That consensus delivered extraordinary aggregate wealth. But it stopped distributing it through the mechanisms that built the middle class. Between 1978 and 2018, CEO compensation rose more than 900 percent while worker pay barely moved. In 1940, 90 percent of children would outearn their parents; today, only 50 percent of millennials are expected to do the same. It is now easier to climb the economic ladder in Sweden, Germany, France, and Japan than in the United States.

If the biggest economic question of the last fifty years was about the role of government, the biggest question of the next fifty will be about the role of business. Businesses have become the most powerful actors in our economy, in many cases more powerful than governments. The next evolution of capitalism—what we call Dynamic Capitalism—rebalances profit with purpose and innovation with inclusion. It is neither revolutionary nor socialist. It recognizes what made American capitalism great: that it gave people freedom, opportunity, and mobility.

The core challenge is straightforward: we need more capitalists. When the wealthiest 10 percent own 87 percent of corporate equities, we shouldn’t be surprised that large swaths of the population have lost faith in the system. The more members of the capital class our economy creates, the more broadly its benefits will be felt.

Broadening ownership is the most direct path. Nearly seventy million Americans are paid hourly with no equity participation. Yet the evidence for shared ownership is compelling. When KKR sold C.H.I. Overhead Doors in 2022, the average worker received $175,000—because shared ownership improved retention, engagement, and returns. In Rochester, two former Kodak line workers who founded Optimax chose to transfer ownership to their four hundred employees rather than sell. Publix, one of the largest employee-owned companies, ranks number one in customer trust among grocery brands. The ownership economy works across industries and company sizes.

The government’s role should be to provide incentives, not mandates: an Employee Ownership Office within the Small Business Administration, universal baby bonds so every American enters adulthood as a member of the capital class, and expanded access to retirement plans for the fifty-six million Americans who lack them are just a few ideas to consider.

The AI productivity surge will make this more urgent. The last wave of automation devastated families without college degrees. AI threatens an even broader set of workers. We should learn from our past. If productivity gains flow exclusively to the shareholders of a handful of technology companies, the political consequences will be severe. But if we broaden who participates—through ownership, not redistribution—we can build an economy that shares more of its rewards.

American capitalism doesn’t need to be replaced. It needs to remember what made it work so effectively in the first place.

AI Is Scary. We Should Listen Better.

Mon, 01 Jun 2026 17:00:00 +0000

Inside the tech world, AI feels like a superpower. Think of it like a language — clunky while you are leaning and then second nature. So much so that you forget how hard it was to learn in the first place. Once you’ve experienced what it can do, the optimism is hard to suppress. Companies are getting built faster. Individual productivity is going up by multiples. Problems that used to require a whole team now get solved by one person orchestrating a handful of AI agents in parallel — research, drafting, code, analysis, all running at once, all feeding into the next step. The result is that companies and teams are getting more done and moving much faster.

But outside the tech world, AI often feels like something else entirely.

It feels like the thing that just raised your rent. It feels like the thing that screened out your résumé before any human at the company ever saw your name. It feels like the system that denied your insurance claim. It feels like a deepfake of your grandson’s voice calling you for money.

It feels like the thing that’s coming, eventually, for your job too.

That’s the dissonance I’ve been thinking about. Most of the people in my professional orbit see AI the way I do — as the most exciting and consequential technology in a generation. We talk about it constantly. We invest in it. We use it daily. We assume that other people, given a bit of time, will come around to seeing it the same way we do.

That’s a mistake.

Many people don’t experience AI as a tool that makes them more productive. They experience it as a system that’s being used on them. The rental market is one example, and a particularly concrete one. Large landlords — companies that own thousands or tens of thousands of apartment units — increasingly use algorithmic pricing software to set rents. The software analyzes occupancy, market conditions, willingness-to-pay, and recommends the highest rent the model thinks each unit will support. The Department of Justice sued one of the largest of these companies, RealPage, in 2024 for facilitating what amounted to algorithmic price coordination across competing landlords. Several state attorneys general joined that suit, including Colorado’s.

If you’re a renter on the receiving end of one of those algorithms, “AI” isn’t an enabler. AI is the reason your rent went up $200 at renewal, despite market conditions that would indicate otherwise. That’s what AI does in your life. And that may be effectively your entire experience so far with this new technology.

It’s easy, inside the tech bubble, to dismiss those concerns as misinformed. To say “that’s not really AI” or “the technology isn’t the problem, the application of it is.” Both of those things might even be true. But neither of them changes the lived experience. And the lived experience is what shapes policy.

That’s why states are moving. Maine has put a moratorium on new data center construction - an ill advised and short-sighted reaction to AI fears, but also one that asked important questions about where we place data centers and who pays for the cost of the power that runs them (both easily addressed in more productive ways, but often key questions that those in the tech sector gloss over). Colorado passed the country’s first comprehensive AI accountability law in 2024 — and then spent the better part of two years in a hard fight to revise it. The updated version eventually - passed only a few weeks ago - fixed the most problematic parts of the law. But the fight exposed just how deep the distrust of AI and the tech industry runs, even in a state with one of the country’s most active tech communities. Other states are watching. The “how do we regulate AI” conversation is genuinely happening — not in the abstract, but in legislative committees, with real bills, and with real constituents pushing on their representatives (so far in 2026 over 2,000 AI bills have been introduced at the state level).

The default reaction inside tech is to roll our eyes at this. To frame the opposition as Luddites. To assume the regulators don’t understand the technology and are reaching for the brake pedal because change is uncomfortable.

I think we should resist this framing.

The concerns are real. They are often based on lived experience, not paranoia. And dismissing them as anti-progress is both unfair to the people raising them and bad for those of us who want AI to succeed (and, in fact, see the necessity for it to do so). If the optimist camp wants AI to be politically and socially survivable — wants the next ten years of deployment to be additive to society and not extractive from it — we have to actually engage with what people are afraid of, not pretend they’re afraid of nothing.

Some of what they’re afraid of is the technology itself. Most of what they’re afraid of is what the technology is going to do on top of a society that already feels rigged against them.

That distinction matters. The economy people are walking into in 2026 is not the economy of 1995. Inequality has been widening for two generations. Productivity has gone up, but median wage growth hasn’t kept pace. The share of national wealth held by the top 1% has roughly doubled. And now, AI arrives. Two AI companies are now worth close to $1 trillion apiece. The founders are worth tens of billions of dollars personally. A handful of people are getting wildly rich from a technology that, from where most Americans are standing, hasn’t yet made their own work better, their lives safer, or their job more secure.

If you already believed the economy was structured to enrich a tiny number of people at the expense of everyone else, AI looks exactly like that thesis on steroids. Of course it’s scary. It would be irrational not to be at least a little scared.

I’m not in the slow-down-AI camp. I’m a hundred percent in the AI optimist camp, and I think the long-run case for AI being additive — to jobs, to productivity, to human capability — is strong. I think we’re at the start of one of the most economically generative periods in modern history. I want us, collectively, to get there.

But getting there requires the rest of the country to come with us. And the rest of the country doesn’t yet see what we see. They see RealPage. They see trillion-dollar AI companies. They see something that is increasing the gap between the haves and the have-nots. And they’re asking, fairly, whose side is this on?

Our job, inside the tent, isn’t to write better explainer threads about how the technology works. It’s to listen carefully to their concerns and not dismiss them. And, ultimately, it is to build AI and AI systems that actually make the lives of the people on the receiving end better. To work on the structural questions — pricing, labor, ownership, regulation, access — with the same energy and creativity we bring to the technology itself. To show with action, not just with words, that an AI future can benefit all of us.

The opposite of progress isn’t caution. It’s arrogance.

Why I'm Supporting Phil Weiser for Governor

Wed, 13 May 2026 09:00:00 +0000

I’m voting for Phil Weiser in the Democratic primary for Colorado Governor on June 30, and I want to share why.

I’ve spent nearly three decades in Colorado working with entrepreneurs and across the innovation ecosystem. This has included working with startups, large companies, main street businesses, and non-profits. I am deeply committed to Colorado — to its people, land, resources, and opportunity. The question I bring to this primary is straightforward: which candidate is most likely to build Colorado into the kind of state where both people and companies can thrive.

For me, the answer is clear: Phil Weiser.

A True Vision

A few weeks ago, Phil spoke to the Grand Junction Economic Partnership, and it was the strongest articulation of an economic vision for Colorado I’ve heard from any politician in quite some time. He named the problem squarely and succinctly: Colorado’s population growth has stalled to its lowest pace since 1990 — talent and capital are moving (sometimes quietly, sometimes with fanfare) to states that are easier to build in. He laid out three pillars in response: a welcoming environment for the industries Colorado leads in, a serious simplification of how the state regulates, and a call for investment in the workforce, housing, childcare, education, and infrastructure that make staying in Colorado possible for working families. He invoked Roy Romer and Federico Peña — leaders who built modern Colorado by asking how can we help rather than by reaching for a press release. That’s the lane Phil is running in.

Practical Perspective

Phil’s platform on jobs and the economy reads like someone who has thought about how the parts fit together — capital access, workforce, infrastructure, regulatory burden, the gap between urban and rural Colorado. He’s talking about a Chief Innovation Officer, a Business Navigator program for early-stage founders, expanded grants and microloans, and a serious effort to harmonize the regulatory thicket that businesses navigate across counties and municipalities.

A Reasonable Regulatory Environment

Colorado is now the sixth most regulated state in the country, with more than 205,000 state-level regulatory restrictions (and growing). That’s a real, daily drag on businesses that Colorado claims to support. Phil’s line in Grand Junction — that regulation should be “like garlic in cooking, just enough and never too much” — is the right instinct, and his platform commits to actually doing the work of simplification and harmonization that the state has needed for years. Not deregulation as a slogan, but modernization as a discipline.

A Broader Vision

A note worth adding because some of the conversation about this primary has narrowed to a single dimension. Phil’s platform is comprehensive. He’s substantively committed to a strong public education system, to early childhood, to healthcare access, and to protecting Colorado’s land, air, and water. He supports a robust safety net for the Coloradans who need one. He treats those commitments and a competitive business environment as complementary rather than in tension — clearly the right frame for this. Pro-business and pro-people are not opposed propositions. The Open Letter itself is underpinned by the same premise.

Phil and the Open Letter

In mid-April, more than 400 Colorado business and civic leaders signed an Open Letter — Ensuring Colorado’s Innovation Future — calling on elected officials to take action to restore Colorado’s standing as a place where innovation can thrive. I signed it because I’m genuinely concerned about Colorado’s competitiveness.

Phil responded publicly on April 20: “On this, I am all in, and I am committed to the nine points you raise.” You can read his detailed response here. Phil’s reaction was that of someone who had heard the message. And, at least for now, he’s the only person to whom the letter was addressed who has issued a substantive response.

A Word on Dan Caruso’s Endorsement

I want to be direct about this, because as a signatory of the Open Letter, I don’t want anyone to be confused about where I stand. Dan Caruso — who deserves real credit for organizing the Open Letter and convening so much of the conversation about Colorado’s business climate — recently endorsed Michael Bennet over Phil. I disagree with that conclusion, and I want to push back on how the case has been made so far.

Dan’s endorsement isn’t really that — at least what he’s released so far consists solely of a list of grievances against Phil. Apparently, the affirmative case is coming in a follow-up email. I’m not sure why he chose to start with the negative, and I think it is telling about Dan’s thinking about the race. When the case for a candidate is the second half of an endorsement and not the first, the framing matters.

There’s also a particular irony worth surfacing. Dan’s chief grievance against Phil is that Phil has been too aggressive in standing up to the second Trump administration. But Senator Bennet’s own line of attack against Phil in their first televised debate was the opposite: that Phil didn’t stand up to Trump enough during the first administration. Both critiques can’t be true. The fact that they coexist tells you this isn’t really about how often Phil has gone to court. The grievance is about something else entirely (you can listen to a podcast that Dan and I recently recorded together that covers many of these issues here).

For my part, I’d rather elect a Governor on what he’ll do, not on competing stories about whether he sued the federal government with the right cadence.

On Senator Bennet

I have great respect for Michael Bennet. He has been an excellent U.S. Senator for Colorado — substantive, serious, and effective. And the Senate is where I want him to stay. Colorado is uniquely fortunate to have a senior senator with Bennet’s depth, his committee positioning, and his ability to deliver for our state in Washington. Trading that for the Governor’s office means giving up a Senate seat that has taken years to earn the influence it now carries. Coloradans benefit more from him in the Senate than in the Governor’s office.

Where I Land

Phil’s record, his platform, and his instincts on the issues that matter most to me — small business, rural Colorado, the long arc of entrepreneurship, a regulatory environment companies can actually navigate, and a Colorado that takes its safety net, its schools, and its environment seriously — match the work I’ve spent my career trying to support. Phil has earned my vote.

Vote on June 30. And know that if you’re unaffiliated, like the majority of Colorado voters (including myself), Colorado lets you cast a ballot in either primary. I hope you’ll join me in voting for Phil.

AI Ghosts

Tue, 12 May 2026 09:00:00 +0000

I wrote a few weeks ago about being open when you use AI. The argument was straightforward — be honest about when AI is involved, and people will be more comfortable trusting what you send.

Lately I’ve been thinking about the flip side of that. The people who see AI where it isn’t.

I’ve had a handful of interactions recently — with friends, with people I know well — where someone has assumed I was using AI to respond to them, when I wasn’t. A note I wrote on my phone in the back of a car. An email I dashed off at 6am. A text reply to a thread of texts. Real responses, written by me, that got flagged in someone else’s head as somehow not me.

I’ve started calling these AI Ghosts. They’re not bad-faith accusations, exactly. It’s something subtler — a perceived presence of AI projected onto interactions where there isn’t any. The flicker of a thought behind a reply: wait, did Seth actually write this?

I get it. AI is everywhere right now. It’s writing emails, drafting blog posts, generating images, producing presentations, code, music, and increasingly even voice. It’s reasonable for people to feel that the line between human work and machine work is getting blurry. And it’s especially reasonable for people who work in industries that are getting disrupted by AI to feel it most acutely. The closer you are to the rebar of the change, the more likely you are to see it in everything around you.

Still, the experience of being on the receiving end of an AI Ghost is unsettling. The first time it happened to me, I shrugged it off. The third time, less so. By the fifth, I noticed it had started to put some tension on a relationship I care about. The conversation was no longer just the content of the message — there was now an unspoken question hovering behind every exchange. Did you actually write this? Are you actually thinking about what I’m saying? Are you still in this with me?

That’s the corrosive part. Not the question itself. The fact that the question is even there.

Why this strengthens the case for disclosure

I came away from those experiences more convinced, not less, that being explicit about AI use is the right norm. At least where AI has done most of the heavy lifting. Without that signal, the suspicion bleeds everywhere. Once someone starts wondering whether a message is AI-generated, they’re going to wonder about the next one, and the one after that. They have no way to verify (at least no reliable way - as an aside, I sometimes worry that we’re all starting to sound like AI the more we use it…). So they question all of it.

Transparency, paradoxically, is what makes “this is just me” believable.

AI as a collaborator

The other thing I’ve been chewing on is the framing of AI as a replacement versus AI as a collaborator. The AI Ghosts thing only feels like a betrayal if you assume AI is substituting for the person — as if the machine has crowded the human out and the recipient is now talking to a hollowed-out version of you.

But that isn’t how most of us are using these tools. We use them the way we’ve always used collaborators: to draft, to think out loud, to react, to push back. Academic papers have always been collaborative. So has most good business writing. So is most of what gets done at a venture firm — partners, founders, advisors, lawyers, all weighing in. AI is the newest collaborator in that mix, with a different shape and different strengths, but the underlying pattern isn’t new.

What AI does change is the cost of collaboration. It’s now cheap to bring a thoughtful collaborator into a five-minute task that would previously have just been you alone at the keyboard. That’s a real shift, and it’s the reason I lean on these tools so often. But it isn’t the same as replacing my thinking. It’s an extension of it.

A moment in time (?)

AI Ghosts are a feature of this particular moment, not a permanent state of the world (at least I hope they’re not). Norms will catch up. People will get more accurate about distinguishing trained-skill output from generic-prompt slop. Disclosure will become more routine. And the question will shift, the way these questions always do, from is this AI? to is this any good?

Until then, sometimes it’s just me, at 6am, in the back of a car writing you back.

How Much Should We Tell People When We Use AI?

Sun, 19 Apr 2026 09:00:00 +0000

I use AI a lot. If you’ve been reading this blog, you know that already. A few weeks ago, I wrote about running my business from a terminal window. That post doesn’t do justice to just how much of my business life is managed through iTerm and Claude. And, of course, Brad has been writing about building CompanyOS - the system that powers a lot of this. I’m all in on this stuff, and I’m not subtle about it.

But something happened recently that made me think about an aspect of AI use that I hadn’t considered carefully enough.

The Deck Review

Like most VCs, I get ten or more new pitches a day. The vast majority aren’t a fit, and I try to send a kind note back when they’re not. For the ones where someone asks me for feedback on their deck - which happens more often than you’d think - I used to have two choices. Spend 30 minutes writing detailed feedback, which I rarely had time for. Or send back a “thanks but I don’t have the bandwidth” note, which felt dismissive when someone was genuinely asking for help.

So I built a third option. I trained a skill inside CompanyOS specifically for reviewing pitch decks. It’s not a generic “summarize this PDF” prompt. I fed it my feedback on a few dozen actual decks, taught it what I look for as an investor - narrative clarity, whether the metrics are stage-appropriate, how the financial model holds together, what questions I’d ask in a partner meeting. When I run a deck through it, the output reflects how I actually think about these things. It’s me, at scale.

A few weeks ago, someone asked me for feedback on their deck after I’d passed on the deal. I ran it through the skill, reviewed the output, made a few edits, and sent it along. Substantive stuff - what was working, what wasn’t, specific suggestions on a few slides.

The response I got back was unexpected. The founder told me he’d run my email through an AI detection tool. It flagged the message as likely AI-generated. And he asked me point blank - is this actually your opinion, or did AI write this?

The Irony

It was a valid question (and he was correct - as I explain below, I heavily relited on AI in my response to him). But here’s the thing that made me laugh. He zeroed in on one specific paragraph about a slide that needed work. That paragraph was one I’d written myself. I’d noticed the slide when I was screenshotting the deck to load into Claude (it can’t read Docsends, sadly) and added my own note about it. The rest - the AI-generated part - he didn’t focus on (although I’m sure that’s in part what caused the note to be flagged by whatever AI detection tool he used).

But his broader question was fair. And it stuck with me.

I wrote him back and explained what I’d done. That yes, AI drafted the initial feedback. That I’d spent significant time training the skill on my actual thinking. That I review and edit everything before sending. And that a year ago, his alternative wouldn’t have been a hand-crafted email from me - it would have been no substantive response at all.

The Podcast Conversation

The next day I was on a podcast with George Samuels, who a a deep thinker about the intersection of AI and human behavior, and I told him this story. We ended up having a long conversation about something that I think a lot of people are navigating right now - the etiquette of AI in professional communication.

George’s suggestion was simple: just be open about it. Not in an apologetic way. Not with a disclaimer that undermines the content. But with a straightforward acknowledgment - something like “I used AI to help draft this, but I’ve trained it on how I actually think about this topic and I review everything before sending.”

That framing resonated with me because it addresses the real concern head-on. People aren’t worried about AI because it’s AI. They’re worried that what they’re reading doesn’t represent the thinking of the person who sent it. That it’s generic filler, not real engagement. Transparency solves that problem. It says: yes, AI was involved, and here’s why you should trust the output anyway.

The Other Side

Around the same time, I had an exchange with my _Capital Evolution _co-author Elizabeth that got at this from the opposite direction. She asked me if I had been using AI to respond to her in our interactions (she was worried that I had) and flagged a specific text conversation where she thought my responses felt AI-generated. They weren’t. But more than that one exchange, she told me she’d started to wonder more generally whether it was always me on the other end of our conversations.

It was more of a feeling but one that I think is quite valid - in this case not by my actual responses to her, but to an world in which people are increasingly using AI to aid (or in some cases substitute for) their work. This is especially concerning to Elizabeth, who works in an industry that’s being significantly disrupted by AI (by some estimates 50% of content is now being generated by AI), and I think that sensitivity was coloring how she read everything I sent. When you know someone is deep into AI tools, it’s natural to start questioning whether any given message is “really them.”

This is another argument for being explicit about when you are and aren’t using AI. If you’re transparent when AI is involved, it’s easier for people to trust that the times you don’t flag it are genuinely just you. Without that transparency, the suspicion bleeds into everything - even the stuff that’s 100% human. It’s also a reminder that AI is best used in many instances as an aid and not a substitute. Context matters, of course. But so does taking the time to make sure that things that go out under your name (this blog, as an exmple - which was drafted by Claude but then significantly revised by me) actually reflect YOU.

What I Actually Think

Here’s where I’ve landed on this.

I’m not going to stop using AI (obviously). The deck review example is a good one - the founder got substantive, personalized feedback on his pitch that reflected 25 years of looking at investment opportunities. His alternative wasn’t a better version of that email written by hand. His alternative was a polite brush-off.

But I am going to start being more explicit about it. Not as an apology. As context. The same way one might say: “I had my analyst pull these numbers” or “my team put this deck together” - because of course you did, and that doesn’t diminish the substance.

The distinction that matters isn’t whether AI was involved. It’s whether the output represents real thinking or is generic fill. I’ve spent real time building and training these tools to reflect how I actually approach problems. That’s meaningfully different from pasting something into ChatGPT and forwarding whatever comes back.

I think we’re heading toward a world where AI involvement in professional communication is assumed, and the interesting question becomes not “did you use AI” but “how well did you use it.” We’re not there yet - clearly, based on the reactions I’m getting. In the meantime, transparency seems like the right bridge.

The AI Gap Between Startups and Enterprise is Larger Than You Think

Sat, 28 Mar 2026 10:00:00 +0000

There’s no shortage of hype around AI right now. Every pitch deck mentions it. Every company claims to be using it. Every conference has a keynote about how it’s going to change everything. Plenty of hype, but also plenty of real work getting done daily by those truly leveraging the power of AI tools.

Like every venture firm, we’ve been deep in this at Foundry. We’ve been working with our portfolio companies on how they’re integrating AI into their products, their workflows, and their operations. It’s been a major area of collaboration and learning across our portfolio. And what we’re seeing across startups (at Foundry and beyond) - not in the hype but in the actual numbers - is a striking contrast between AI adoption in startups and what’s happening in large enterprises.

Is AI the startup superpower? At the moment it appears to be so.

The Enterprise Struggle Is Real

The data on enterprise AI adoption is stark. A widely cited MIT study from 2025 found that roughly 95% of enterprise AI pilots fail to deliver measurable ROI. That specific number has been debated - the sample was small, and the definition of success was narrow - but even the more conservative research tells a similar story. RAND found that over 80% of AI projects fail, which is twice the failure rate of non-AI IT projects. BCG reported that 74% of companies have seen no tangible value from their AI investments, and only 4% have achieved enterprise-wide AI capabilities. S&P Global found that 42% of companies abandoned most of their AI initiatives in 2025 - up from 17% the year before.

The reasons are consistent across every study. Enterprises pick the wrong problems to solve with AI - choosing projects based on technical novelty instead of business value. Their data is siloed and messy. Projects that work in sandboxes stall when they hit the reality of compliance workflows, authentication systems, and organizational processes that weren’t designed for this. And critically, they try to bolt AI onto existing processes rather than rethinking how work gets done. That rarely works.

None of this is surprising if you’ve watched large organizations try to adopt any transformative technology. What makes this moment different is the other side of the equation.

Startups Are Running With It

This is where it gets interesting. The top AI-native startups now average $3.48 million in revenue per employee - roughly 5.7 times higher than leading traditional SaaS companies. Cursor, the AI code editor, hit $300 million in ARR with about 20 employees. Lovable reached $17 million in ARR with 15 people, three months after launch. And across the board - if Foundry’s portfolio is representative (and I believe that it is) - startups of all kinds are seeing massive leverage from their AI investments.

Nimbleness Has Always Been a Superpower. AI Amplifies It.

Startups have always had the edge when it comes to speed and adaptability. Fewer layers of approval. No legacy systems to work around. A willingness to rethink how things are done rather than protecting how things have always been done. That’s been the case through every technology cycle I’ve watched over 20+ years in venture.

What’s different about AI is how dramatically it amplifies that advantage.

AI doesn’t reward scale the way traditional software does. You don’t need a large engineering team to build a product. You don’t need a large marketing or go-to-market team to sell. You don’t need layers of management. Individuals and small teams can be as productive as entire divisions of larger companies. Powerful models are available to everyone willing to take the time to learn how to leverage them. The APIs are the same whether you’re a Fortune 500 company or three people in a coffee shop. What matters is how quickly you can integrate AI into your workflow, how willing you are to rethink how work gets done, and how fast you can iterate. Startups win on all three counts.

Enterprises, meanwhile, are stuck in what I’d call pilot purgatory. They run a proof of concept, it shows promise, and then it dies in the gap between “this demo looks great” and “this is now part of how we actually operate.” That MIT I cite above tells an important part of the story - companies are choosing projects based on technical novelty rather than business value. They’re solving interesting problems instead of important ones.

The World Economic Forum noted that AI is making resources that were once exclusive to large companies - advanced technology, specialized expertise, sophisticated analysis - broadly accessible to small teams. OECD data shows AI uptake in small firms grew 72% in a single year, with 91% of adopters reporting efficiency gains. Mid-market companies are getting to value in 6 to 9 months, while large enterprises take 12 to 18.

In previous technology cycles - cloud, mobile, SaaS - startups had advantages, but large companies could eventually catch up by throwing resources at the problem. With AI, the traditional advantages of large companies - vast data, deep pockets, big teams - seem to matter less (or, at least, most seem unable to harness these to their advantage). In some cases they’re actually liabilities, because they come with organizational complexity that makes it harder to adapt.

What We’re Watching

The companies in our portfolio that use AI most effectively aren’t treating it as a bolt-on feature. They’re using it as an operating model. They’re building companies that are structurally leaner and faster than what came before - generating more value per person than would have been possible even a few years ago.

That’s not hype. That’s what’s actually happening on the ground, in real companies, with real revenue. And it’s happening while many of their larger competitors are still trying to get their pilots to work.

Startups have always punched above their weight. AI is making the punch a lot harder.

Capital Evolution - T+3 Month Check-In

Sun, 22 Mar 2026 10:00:00 +0000

It’s been about a quarter since Capital Evolution hit the shelves, and I thought it would be worth stepping back to take stock of what Elizabeth and I have been doing to get the book’s ideas out into the world. It’s felt like a lot — and seeing it all written down confirms that is indeed the case!

From the start, our goal was to make sure that the ideas in the book — Dynamic Capitalism and the concepts around it — didn’t stay on the page. We wanted to get them in front of as many people as possible: business leaders, policymakers, students, investors, and anyone thinking seriously about the future of the American economy.

In Print and Online

The conversation around Capital Evolution has shown up in some incredible places. We wrote an op-ed for The New York Times arguing that copying China’s economic playbook is a terrible idea for the U.S. We had a piece in the Los Angeles Times with Congressman Ro Khanna on reviving America’s innovation economy. Forbes featured our framework for global leadership in a system of Dynamic Capitalism. The International Economy published our essay “The Frankenstein Economy,” which distills the book’s core argument about how neoliberalism’s creation has turned on its creator. And we’ve been covered in Fortune, Inc., the Washington Independent Review of Books, and many others.

On the Air

We’ve appeared on dozens and dozens of podcasts (full list here), reaching audiences across business, technology, investing, and public policy. Highlights include the Conscious Millionaire Show, Michael Covel’s Trend Following, School for Startups Radio, Keen on America with Andrew Keen, and The Full Ratchet. Each conversation has been different — some focused on venture capital and AI, others on inequality and the middle class — but they all come back to the same question: how do we build an economy that works for more people?

In Person

In-person events, while more time-consuming and often involving travel, have been by far the most rewarding part of our book-promotion efforts. We kicked things off with a launch party at my wife’s bookstore, Composition Shop, that set the tone for everything that followed. We packed the house at Politics & Prose in Washington, D.C., with over 120 people for a conversation that could have gone on all night (special thanks to Lori Montgomery, former Politics editor for The Washington Post for moderating). There were many others as well, including a memorable event at the Tattered Cover here in Denver (I spent many days at Tattered as a kid - it was special to be back to host an author event there). Those were only two of a bunch of bookstore events we’ve hosted since launching. And we’ve done a series of corporate events and private discussions with organizations including JP Morgan, KPMG, Cooley, CapRock, MichaelBest, CU Boulder, and the Colorado School of Mines — conversations with executives, investors, attorneys, and students who are grappling with questions related to the future of our economy every day. These events spurred not just interesting discussions, but also pushed us to think more deeply about our work, challenge our ideas, and provided the opportunity for others to take up the mantle of Dynamic Capitalism.

Looking Ahead

The ideas in the book seem very durable, and we continue to have events scheduled as questions around the future of capitalism remain extremely relevant in today’s political and social environment. The conversations we’re having aren’t partisan. They’re practical. And they give me real optimism that the evolution we describe in the book is already underway.

Migrating My Blog from WordPress to Hugo

Tue, 10 Mar 2026 12:00:00 +0000

I’ve been blogging on WordPress for close to twenty years. It’s been a great platform, but I’d been thinking about a change for a while. WordPress had gotten heavier over the years - more plugins, more updates, more things to manage. I wanted something simpler. Something where I could just write in a text file and push it live.

Brad recently migrated his blog - Feld Thoughts - from WordPress to Hugo and that was the kick I needed to finally do this. He gave me quite a few tips and pointers that were incredibly helpful. Seeing him go through the process gave me the confidence that this was both doable and worth doing.

Why Hugo

Hugo is a static site generator. You write posts in Markdown, run a build command, and out comes a set of plain HTML files. No database, no PHP, no server-side processing. Just files.

That simplicity is the whole point. My blog doesn’t need a content management system running on a server somewhere. It needs to turn my writing into web pages. Hugo does that extremely well, and it does it fast. I also like the simplicty of the template I choose - the focus is more about the content (but still with a place to include my books and other stuff I’m working on).

The Stack

Here’s what the new setup looks like:

Hugo generates the site from Markdown files
GitHub stores the source code and content
Vercel builds and hosts the site - auto-deploys every time I push to GitHub
Cloudflare handles DNS and CDN
Kit (formerly ConvertKit) handles email subscriptions via RSS

Publishing a new post is literally: write a Markdown file, commit it, push to GitHub. Vercel picks up the change and deploys in about 30 seconds. No logging into an admin panel. No clicking through menus. Just text files and version control.

Choosing and Customizing a Theme

I went with the Notepadium theme as a starting point. It’s clean and minimal - exactly the aesthetic I wanted. But “starting point” is the key phrase. I ended up doing extensive customization.

I worked with Claude to adjust the CSS, tweak the layout, get the typography right, and handle things like dark mode support, pagination, and how lead images display. We went back and forth on the design until it felt right. The beauty of Hugo themes is that you can override any template file by placing your own version in the layouts/ directory, so customization doesn’t mean forking the theme.

The Migration Process

Exporting from WordPress

The first step was getting all my content out of WordPress. I used the standard WordPress export tool, which gives you a big XML file with all your posts, pages, and metadata. From there I used a conversion tool that I had Claude create for me to turn those into Markdown files with the right Hugo frontmatter.

URL Mapping

This was critical. I’ve got almost two decades of posts, and there are links all over the internet pointing to them. I needed every old URL to keep working.

WordPress uses a URL structure like /archives/2024/05/some-post-title.html. I configured Hugo to use the same pattern, so every existing link still resolves to the right post. No redirects needed - the URLs just work. (side note - let me know if you find any broken links that I missed!)

Media Library

WordPress stores images and other media in its own upload directory structure. I had to port all of that over, making sure image paths in posts pointed to the right locations. This took some patience - going through posts, finding broken images, and fixing paths. We adjusted the media directory structure to work cleanly with Hugo’s static file handling.

One thing Hugo doesn’t give you out of the box is search. WordPress has search built in. I had to think about whether I needed it and, if so, how to implement it. Same with things like category pages, archive views, and pagination. Each of these required some template work.

The Email Subscription Migration

This was honestly the trickiest part.

I use Kit for bulk email for most of my work, but my blog was connected to MailChimp. I wanted to change that. The idea is simple - when I publish a new post, it shows up in my RSS feed, Kit detects it, and sends an email to subscribers with the post content. Simple in theory.

In practice, we ran into a few issues.

The Missing Content Field

Kit expects the full post body to come through in a specific RSS field called content:encoded. Hugo’s default RSS template only includes a description field with the post summary. So my first test emails went out with just a snippet instead of the full post.

The fix was creating a custom RSS template that includes content:encoded with the full post content. Straightforward once we figured out what Kit was looking for.

There were a bunch of other challenges as well. I spent more time troubleshooting this part of the migration than anything else.

The XML Parsing Bug

This one was sneaky. After adding the content:encoded field, we initially wrapped the content in CDATA tags - which is a standard XML technique for including HTML content. It looked correct. But it turns out that if your post content happens to contain the character sequence ]]> anywhere, it breaks the CDATA wrapper and produces invalid XML.

The RSS feed would load fine in a browser but fail for automated parsers. Kit silently failed to pick up new posts. We burned through several test-post cycles - publishing, checking, unpublishing, tweaking - before spotting the actual XML error.

The fix was simple: use HTML entity encoding instead of CDATA wrapping. But finding the problem took real debugging.

Test Post Cycles

We went through probably half a dozen rounds of publishing test posts, checking whether Kit picked them up, looking at the email output, and adjusting. Each cycle meant creating a post, pushing it live, waiting for Kit to poll the RSS feed, checking the email, then taking the test post down. It wasn’t glamorous work, but it’s the kind of thing that has to actually be tested end-to-end.

Working with AI

I should mention - I did this entire migration working alongside Claude in my terminal. Not as a chatbot answering questions, but as an active collaborator. Claude made the code changes, ran the git commands, debugged the RSS issues, helped customize the CSS, and handled the DNS configuration.

This is the kind of project where AI collaboration really shines. There are a hundred small technical decisions - CSS tweaks, template overrides, XML formatting, DNS records - that individually aren’t hard but collectively take hours. Having an AI partner that can read code, write code, and execute commands in real time made the whole thing dramatically faster.

The Result

The site is faster. The publishing workflow is simpler. The content is stored as plain text files in a git repo, which means I have full version history of every post I’ve ever written. And there’s no server to maintain - Vercel handles all of that.

If you’re thinking about making a similar move, I’d highly recommend it. The Hugo ecosystem is mature, the documentation is solid, and the combination of Hugo + GitHub + Vercel is a genuinely great stack for a blog.

And if you’ve been reading this blog for a while, the transition should be invisible to you. Same URLs, same RSS feed, same email subscriptions. Just a cleaner engine underneath.

Let me know what you think!