Hi everybody. Just sharing a quick personal update and some housekeeping (thanks for reading): I’m stepping away from venture and joining Boulevard as VP of Corporate Development.

Getting to be a VC is an amazing privilege, and I’m beyond grateful for the 12+ years I was permitted to do this. But more recently, my urge to re-enter the operating world has grown from a minor itch to an all-consuming preoccupation. And I keep thinking back to the start of my tech career — the three years spent growing the KACE business — and how meaningful it felt to work hard at building something alongside people you care about.

There are many ways to make an impact as an investor, but it’s just not the same as being ultimately responsible for driving outcomes inside an actual business and being a true peer to the people running it. So. It’s time to take a chance.

I’ll now be working on partnerships, ecosystem, and M&A with my friends at Boulevard, the leading client experience platform for appointment-based, self-care businesses. Boulevard most recently raised a $80M Series D round led by JMI Equity, and I was lucky enough to invest into the company 3x during my time at Toba.

Some quick items:

• The Down Round will continue to exist!1 Now with dispatches from inside the real world of software/AI companies.2 Thank you to the nearly 1,000(!!) of you who read and subscribe.

• Other ways to stay in touch? LinkedIn, Twitter, E-mail

• If you need a new contact at Toba (for dealflow, portfolio inquiries, etc), please send me a note and I’ll get you in touch with the right person.

• WE ARE HIRING AT BOULEVARD! And I can tell you from personal experience that it’s an amazing place to work and grow.

• If you’re building technology in the beauty, wellness, salon, spa, or medspa space — please reach out! Let’s figure out if there are ways for us to work together.

All 3 of these categories are different manifestations of arguably the same category, which we can call “intellectual labor” or “intelligence”. But they live in 3 separate places on the company’s financial statements:

1. Human capital is an operating expense (Opex) on the income statement, and does not depreciate (or rather, it fully depreciates immediately)

2. AI hardware is a capital expense (Capex) which creates an asset on the balance sheet, with depreciation appearing over time on the income statement

3. Token spend is an operating expense which is accounted for similarly to human capital

To add a wrinkle to this, humans in tech are often paid stock-based compensation (SBC) which is an operating expense but not a cash expense (uh-oh). Crucially, stock options do not change hands when companies buy tokens or purchase AI hardware (well, not usually).

What can be tricky is that measures of business performance (like ROIC) generally treat operating expenses and capitalized assets as totally different things… humans are paid in cash and stock, but capital assets are things like factories and turbines that usually perform some non-human task. Historically these two categories have not been easily substitutable.

An asset-heavy company like a railroad has to continually spend money to replenish its physical infrastructure in a way that an asset-light business like a software company does not. For this reason, asset-light businesses have generally (until recently, perhaps) been valued at higher multiples than asset-heavy businesses because they don’t face this never-ending depreciation and refresh problem. This is why we are seeing a lot of recent investor indigestion about the hyperscalers’ transition to asset-heavy.

One thing I’ve been wondering lately is whether these intuitions (physical assets and capex are bad if you can avoid them) are outdated, from a world where machines can’t think. Google or Facebook may have been spectacularly profitable and asset-light, but they spent extraordinary amounts of capital on salaries and stock units for their employees. Unlike a GPU which may have a useful life of 6+ years, 100% of the capital invested in a worker can vanish whenever he/she decides to leave the business.

In a world with rapidly collapsing cost to deliver intelligence via silicon, are we 100% sure that owning chips and data centers is an inferior business model just because it’s “asset heavy”?

Much of the consternation about Capex assumes that the spend is incremental on top of existing headcount costs (and thus will collapse free cash flow). But what if human compensation goes down as capital expenses go up? Because that’s precisely what we’re seeing in the world’s largest companies:

If tech earnings growth continues unabated at a relatively high CAGR (15-30%+) while being funded by Capex that eventually makes an S-curve and tapers off a few years from now… and if this capital intensity becomes an ever-increasing barrier to entry (literally only a few entities in the world can play these high-stakes AI hyperscale games)… then it’s not clear to me whether investors should fear an asset-heavy future.

Whether this is socially optimal for the human race is a different question entirely.

A few predictions coming out of the SaaSpocalypse:

1. Building AI features into existing software platforms will be obligatory, but it will mostly be a defensive move, not an offensive move. This moment reminds me so much of the early pandemic when every SaaS company rushed to build a video/teleconference module, and promised their investors that they’d make so much upsell revenue from it — only to discover that every vendor was doing the same thing, so the differentiation was zero and so were the fees they could charge. I’m pretty bearish on the idea that we’re going to see staggering amounts of “AI upsell ARR” at the big SaaS franchises, Agentforce and Now Assist notwithstanding1. Most of the ROI in new AI features will be in maintaining gross retention, and that’s totally OK.

2. Scaled providers will become multi-product franchises that build breadth with shared data architectures. The way that Datadog positions its product suite is instructive (see below). If you’re selling a point solution, you’re vulnerable to getting ripped out. But what if your average customer is using 4-6 of your products, and they’re all connected together so that the same customer/financial/asset/etc data is consistent from end to end? Why would anybody attempt to rip that all out when you have to get all of it right to even attempt to compete with what the incumbent can already provide? In my opinion, this dimension (single product vs. multi-product) is much more important than the horizontal vs. vertical debate.

   

3. Every vendor will push upmarket. Retention has always been correlated with customer size and price point. Your largest customers tend to renew on time and at high rates, whereas your smallest customers churn more & complain more & pay you a lot less. This disparity will get turbocharged in the era of AI — the largest accounts are sophisticated enough to realize that insourcing their software tools is likely ROI negative for them, but the smallest ones will Claude Code a replacement for your product to save a few bucks (and fail, but that’s no help to you). The incumbent vendors know this, which is why they all report progress on upmarket accounts (e.g. # of $100k+ ACV clients) and are maneuvering away from the SMB segments as fast as possible.

4. For the true enterprise multi-product platforms (think CRWD, NOW, PANW, DDOG, CRM) I don’t think we will see much of any growth slowdown or price pressure from AI disruption. (Just my opinion!) But on the other hand…

5. GAAP earnings are in. Every public SaaS C-suite understands that the market has lost patience with non-GAAP adjusted earnings and excessive stock-based compensation. They will all get to GAAP profitability over the next 24 months. Many people will be surprised at how quickly they get there. And this will happen because…

6. They will incorporate AI into every line of their OpEx. SaaS companies are the perfect beneficiaries of AI because nearly every component of their cost structure is human capital knowledge work. Nearly everything they do within the S&M, R&D, and G&A buckets can be augmented by some sort of AI product. Now, this doesn’t mean they are going to lay off tons of people. But it does permit certain strategies such as “keep headcount flat as revenues continue to go up” which is a great way to grow into positive GAAP earnings.

7. Share buybacks are also in. SentinelOne went public in June 2021 at a price of $35 per share, raising around $1.2 billion. Today it trades at approximately $13. Over those past 5 years, that cash they raised has mostly sat on the company’s balance sheet collecting a few points of interest, and SentinelOne can now buy back its entire IPO float for around $450 million. This ride may have sucked for shareholders, but from a corporate treasury perspective, selling at $35 and buying back at $13 ain’t a bad deal.

8. M&A will go bananas as VCs abandon mid-stage software startups. SaaS is a toxic category in venture right now as investors turn their attention to AI pure-plays. This means a whole bunch of perfectly adequate $5-50M ARR companies will be stranded without growth capital, but with their boards pressuring them to achieve liquidity. Meanwhile, the big public platforms are generating more cash than ever before, but are facing skeptical public markets that worry about growth durability. This is a great setup for consolidation. The VCs won’t like the exit multiples, but public investors will benefit from the cheap cost of buying of ARR in chunks.

I’m not in the habit of making market calls (except that one time). But these are a few thoughts going through my head after SaaSpocalypse 2026.

1. Somebody smart said that the time to buy is when blood is on the streets. Is anything more hated right now than horizontal software (e.g. CRM HUBS MNDY TEAM ASAN BRZE GTLB INTU WDAY ESTC ADBE)?

2. One of the things I try to notice in “the discourse” are instances where somebody asserts something as a given without even feeling the need to justify their point. Presumably because the point is so obvious/consensus that it’s not even worth defending (meaning it’s a good indicator of global sentiment). There’s a lot of this type of software bearposting right now. An example:

3. It bears mentioning that precisely zero AI disruption evidence is showing up in public software companies’ growth rates and retention rates. In fact, many of them have seen accelerating growth during 2025. Might that change this quarter? Absolutely. But to date this fear has been purely hypothetical.

4. There’s certainly a valuation multiple (and terminal multiple) conversation to be had in software. Given increasing market dynamism and turnover, maybe the free cash flow multiple for a growing sticky software business should be something like 15-20x instead of 25-30x. Did you know that MNDY, growing at >20%, is trading at 11x EV to FCF right now?

5. If you were to see AI disruption occurring, it would likely be at the lower (SMB) end of the market. Enterprises — who are buying software products not for the code but for support, security, compliance, integrations, continuous improvement, etc. — are much less likely to rip out brand name enterprise software platforms for unknown entrants. The incumbent platforms know this, which is why most of them are growing their larger ($100k+ ACV) customer accounts at 2x the rate of their SMB/midmarket logos.

6. Despite my repeated asking, nobody has successfully explained to me what exactly an “AI native new entrant” is. I wonder how many generalist hedge fund managers today can correctly identify that “AI companies” are just software companies that use lots of AI. (Editor’s note: All software companies use AI.)

7. For a good explanation of why the economics of ripping out software for an internally-developed alternative don’t make sense for 99% of firms (even after the advent of AI coding), see here.

8. Software has gotten cheaper to build and deploy in a fairly steady line over the past 30 years. Over that time, buyers’ preference for 3rd party software has exploded:

9. If we were discussing an auto manufacturer in an environment where the cost of steel fell by 80%, that auto manufacturer’s stock would go up because their raw materials cost went down. We don’t think “uh oh, this means that anyone can start a car company”. Why do we think about this so differently in software?

10. Think about the cool stuff that you are now able to do on your desktop with Claude Code or Google Antigravity or whatever. Which of the two scenarios is more likely? A) Personal software projects remain economically valuable, and incumbent software companies crumble into dust because every random person now has access to powerful capabilities to deploy their own software. B) The bar for economically valuable software rises dramatically, there’s no market for anything you can make in a few days with an AI coding platform, and software companies now focus on managing projects that are orders of magnitude more complicated and powerful than anything that’s existed before.

11. Before you answer #10, remember that Salesforce employs around 10,000 software developers, and every single one of them is now using Claude Code and Cursor.

Author’s note: I own shares in MNDY, GTLB, ADBE, TEAM, and CSU.TO.

Human memory can be slippery. Things that we claimed would “never happen!” regularly do, and then we get amnesia and forget we ever claimed this.

That’s why it’s important to occasionally perform retrospective review on how we thought about things before the present made them “obvious.” As Mark Twain said, “It ain’t what you don’t know that gets you into trouble. It’s what you know for sure that just ain’t so.”

For the record, this is not me claiming it’s not a bubble or that you should go max long NVDA. The AI bears may have gotten steamrolled for the past 3 years, but that could change at any time.

This will be fun to re-review in a year.

Bear narratives that have already fallen

• “AI is just glorified autocomplete and can’t reason.” 🪦 Disproved by 2023 (GPT-4).

• “No consumer will pay for AI services.” 🪦 Disproved by 2023 (ChatGPT Plus).

• “AI is too unreliable and hallucination-prone to be used by professionals.” 🪦 Disproved by 2023-2024 (GitHub Copilot, Cursor).

• “Fully self-driving cars will never arrive.” 🪦 Disproved by 2024 (Waymo).

• “Generative AI is a passing fad.” 🪦 Disproved by 2025 (AI products break 500 million daily active users).

• “Users won’t trust AI outputs for anything important.” 🪦 Disproved by 2025 (40% of U.S. physicians use OpenEvidence and most of them use it daily).

• “Enterprise demand just isn’t there.” 🪦 Disproved by 2025 (the hyperscalers don’t have enough supply to meet demand for compute, and total purchase commitments from their customers are going straight up).

• “Scaling is dead.” 🪦 Disproved more or less continuously for the past three years. Will maybe be true one day, but not today.

Bear narratives that persist, but may be challenged in the future

• “AI revenues are tiny relative to the investment, it’s just a bubble financed by hyperscaler capex.” We will still say this once AI revenues are 10x where they are today?

• “OpenAI will never be able to pay for all its massive capital commitments.” When IPO?

• “AI models have no moat and will be commoditized.” All I’ll say about this is that models without moats ≠ AI is a structurally unprofitable industry. If you think models have no moats then maybe you should buy a lot of AMZN1 .

• “Model companies are structurally unprofitable; even if the they make money on inference, they are permanently obligated to keep dumping capital into huge training runs just to stay competitive.”

• “AI’s compute and energy requirements are unsustainable. We literally don’t have enough power for the industry to work.”

• “Enterprises may be spending on AI, but they’re not getting any ROI on it.”

• “This whole industry is circularly financed. OpenAI is basically Enron.”

• “All of these AI products have negative gross margins.”

• “Even if these AI products have positive gross margins, they’ll never be as high margin as software.”

A lot of software bears are operating under the presumption that AI (specifically the productivity gain from using AI to develop software) is going to lead to a massive increase in the supply of software, and consequent collapse in software pricing and unit economics.

The chain of reasoning goes something like this: Cost of building software drops dramatically —> Way more new entrants —> Competitive intensity way up —> Price collapse.

I am a little suspicious of this argument.

The reason is that AI is not the first or only major productivity unlock in the software development field. You can look at things like shared libraries, APIs, development frameworks, modern IDEs, Stack Overflow, Slack, GitHub, open source, no-code, etc. as prior accelerants of developer productivity.

Taking into account all these improvements over the past 30 years, developer productivity has increased 10×–100× depending on measure, yet median SaaS gross margins remain solidly at ~75%1.

Have barriers to entry to building a software company gone down over this span? In some ways yes but some ways no. While it’s certainly the case that a lone developer or small team can get a lot more done in a shorter amount of time, the bar for what constitutes “good enough software to sell” has gone straight up. It still takes at least a few quarters (often closer to a year or two) to put together a product that is commercially viable and can command meaningful margin.

It strikes me that when a new enabling technology is truly democratic, meaning that it can be utilized by new entrants and incumbents alike2, we should not anticipate “disruption” but rather an increase in the quality of the end product.

The chain of causation would instead look something like: Software developers get way more powerful —> Quality bar for software goes up —> No big marginal impact on barriers to entry, vendor concentration, or pricing —> But everybody gets amazing software.

People like to ask the question “what will happen when companies can click a button and have their AI agents perfectly replicate Salesforce in an afternoon?” For fun, let’s pick a timeline when this might be possible… let’s say 2029? In 2029 when such a thing could be done, the economically valuable thing will not be to replicate Salesforce as it stands today in 2025, but to replicate the 2029 version of Salesforce that’s had four incremental years of a few thousand Salesforce developers (using AI!) pushing the technical frontier forward compared to today.

If CRMs get 10x better between now and 2029, then our intuitions about how trivial they are to rebuild may no longer be accurate.

Not to mention that if CRMs get 10x better, then customer demand for CRMs will rise, pushing pricing up rather than down.

I wouldn’t go as far as forecasting price increases in application and infrastructure software as a consequence of AI. But it would not surprise me if this AI moment catalyzed a bunch of new software monopolies that are bigger and badder than the last generation’s.

EDIT 11/3/25 1:18PM EST: Andrej Karpathy agrees with The Down Round:

EDIT 12/3/25 12:40PM EST: Ben Evans agrees with The Down Round:

There are lots of analytical problems with merely noticing that the Capex is much higher than the revenues and declaring that it’s the biggest bubble in world history, most of which I won’t repeat here but you can find in an earlier post from last month.

But one question that’s been on my mind is… at what level of AI revenues will people stop making this objection? At a certain level of revenues, the fact that AI is legitimately monetizable will be hard to argue against, no? That won’t extinguish every bear bubble thesis, because ROIC and competitive differentiation and financing risk and long term moats and circular revenues and blahblahblahblahblah will all still be up for debate. But it will be an important psychological milestone for the broader investment community.

I’d argue that once direct AI revenues are 10x higher than today, so call it $200B or so, the average layperson or generalist asset manager will no longer posit that this is all pretend make-believe with no endgame. Then the next question becomes: How soon until that happens?

Here are a few data points to consider:

• OpenAI has gone from $0 to ~$13B of ARR in around three years (targeting nearly $20B by year-end), with current growth rates at around 3x YoY. Their projections are to blast past $100B of ARR in another three years, with the next three years’ growth coming in at 2.3x, 2x, and 1.6x (source).

• Anthropic started 2025 with ~$1B of ARR and projects to end 2025 with ~$9B of ARR and for 2026 to end somewhere between $20B and $26B (source), which makes between 2.2x and 2.8x growth in 2026 after 9x growth in 2025.

• The ARR breakout for firms like Google is a lot murkier, but there are completely insane data points coming out such as the fact that Google’s monthly tokens processed are up something like 50x from a year ago. This should not be taken to imply a particular revenue growth rate, but are potentially a leading indicator of eventual monetized AI ARR continuing to grow robustly in 2026 and beyond.

• Plenty of “AI revenue” is being captured by incumbent SaaS vendors as a line item within their total subscription ARR. For example, Adobe reports approximately $250M in AI revenues, Salesforce reports $440M from Agentforce, ServiceNow sees $1B in AI revenues next year, and Workday is attaching AI modules to 75% of new deals.

• Meanwhile, firms like Atlassian are bundling AI features into the core platform without charging for them separately, thus muddying the waters of what does and does not constitute “AI revenue”. Perhaps nodding to a world in the not-so-distant future where most subscription SaaS ARR is AI ARR.

If you take these together, and make some extremely high level assumptions about growth over the next few years while acknowledging that this forecast will inevitably be embarrassingly wrong in one direction or another, you get something like the below. My personal guess is that we pass $200B of “direct” AI ARR by sometime in mid-2028, a bit less than 3 years from now.

Note that Google Gemini is nowhere to be found on this chart, because I have absolutely no idea of how to think about what their AI revenues look like. If Google’s Search monopoly morphs into Google’s AI Search monopoly, all bets are off.

1. This data center buildout is often compared to the telecom and fiber optic overbuild during the Dot Com bubble. One key difference: Dot Com capex was spent during a period of comparative tiny Internet usage (see below), on the expectation that web commerce would materialize shortly thereafter. Where do we expect that all of this AI inference will be consumed? Yup, also the Internet. Except now the Internet is where five billion consumers and nearly every corporation in the world already conduct business. When you think about revenues associated with AI, you shouldn’t think just about ChatGPT users paying $20/mo; rather you should note that every single corporation in the S&P 500 wants to spend tens to hundreds of millions of dollars on inference — immediately — either for their own internal use cases (e.g. Google or Meta Ads) or to repackage into digital products for their own customers. We do not have to speculate about hypothetical demand materializing… data center capacity is already booked through 2027 in some cases. We also don’t have speculate where this consumption will take place: It will take place via web applications. You know, the same place where we already spend $300B/yr on software and $500B/yr on digital advertising.

2. All of this data center capex is not amortized just over AI revenues, but across all of the hyperscalers’ other cloud revenues and use cases (e.g. AWS). You can’t just say that $400b of capex is supported by only $20b in revenues. You can either say that $400b of capex is supported by something like $700b+ in revenues growing at ~20% (AWS + Azure + all of Google + all of Meta) or you can say some smaller amount of AI-specific capex (maybe $200b?) is supported by $20b of AI-specific revenues growing at something like 400% annually. Much less scary.

3. The teams within the Mag7 deciding how many Nvidia GPUs to buy have way more information about demand, usage, backlog, contracts, etc. than the rest of us do. Important people who should know things like this are telling us that they have never seen anything in their careers like the amount of demand there is for inference right now. I’m not saying that we should blindly follow what they say or presume that they’re being consistently honest with us, but when the entirety of our technology leadership uniformly makes the same enormous bet in the same direction… they probably know something.

4. If Jerry Neumann is right and AI model providers turn out to be a low profit commodity business, then this will be to the benefit of the complementary businesses in the AI stack: Namely, those who own the infrastructure serving this cheap and abundant AI on behalf of the model providers (MSFT, Google), and also those who are consuming cheap AI to power their own products (also MSFT, Google). There’s a possible world where all participants in the ecosystem get rich off of AI except for OpenAI and Anthropic.

5. Let’s not forget that model capabilities are still improving rapidly. It may be hard to foresee precisely how good they will end up being by 2028, but there’s a decent chance that our intuitions are massively underrating how much consumer and enterprise utility will be available via AI solutions in three years (meaning that demand/revenue may be orders of magnitude higher than it is today).

One of our portfolio SaaS CEOs opens each board meeting with an overview slide consisting of two columns: Highlights and Challenges. As you might predict, highlights consists of everything that’s going well, and challenges are the things that are worrisome or need improvement.

In this quarter’s meeting, I was struck by something interesting. Nearly everything in the highlights category were things that were happening right now, or had happened in the most recent quarter; but challenges mostly consisted of hypothetical things that *might* go wrong in the near future.

This arrangement wasn’t always how it went — usually the current quarter’s performance was a mix of good and bad, with lots of things to note under the challenges header. But the current quarter was a blowout success. So without much to be disappointed about in the company’s current execution, the CEO’s mind was pulled forward (rightfully so) to things that may go awry in the future.

That’s good, because that’s his job. But this did prompt two main thoughts in my head:

1. This is the manifestation of psychological anxiety: Things are going great, but I am worried.

2. This is almost identical to how every public software earnings call has gone so far in Q2.

The example everybody has on their minds now is Monday.com. MNDY came into their Q2 earnings call having not only beat their quarter’s estimates handily but also having raised(!) their targets for the year. But the size of the raise seemed conservative compared to the magnitude of the beat, so analysts asked what’s up with that? Management answered (and I’m paraphrasing here), “well, you see, our organic search performance on Google has been a little light since the introduction of their AI overviews, which hurts our inbound lead gen for the lower end of our market. We think this will be temporary and we have ways to work around it, but we’re being cautious for now.”

This was taken by the market as Monday is being disrupted by AI! and the stock fell 30% overnight.

I’m not here to argue that the market was wrong about this. But rather to point out the dominant narrative playing out right now in private and public software companies:

AI is coming to disrupt software. Our moats are under assault. Where exactly? Not quite sure, but possibly everywhere. Is it showing up in the numbers yet? No, the numbers are great actually. But it’s coming. Get ready.

Any indication that AI is eroding competitive advantage, even if amorphously defined, is taken as a reason to panic. And as the Buddha likes to tell us, a restless mind will find no shortage of things about which to fret:

These are not the wrong things to be worried about. We have companies in the portfolio that were poorly positioned for the current AI moment and are totally hosed right now. This is not fantasy disruption.

But on the other hand… isn’t the entire point of software that it is adaptable?

In a way that is rather unique compared to the other big industries (machinery, pharmaceuticals, railroads, etc), asset-light software businesses can pivot to meet the current moment. If customers want new AI features, software companies can build them. If one lead gen channel (SEO) dries up, they can rotate into other ways of getting to customers. If some new entrant is innovating, the established incumbent can copy what they’re doing. If seat-based pricing doesn’t make sense anymore, they can navigate to consumption-based. And so on.

It seems to me that in the current moment, AI does not inexorably shift the balance of power between disruptor and incumbent. Instead, AI is democratic — it gives more power to whoever exercises the most agency. In this era, the winners will be those who make the fullest use of the tools at hand.

A few points on this that feel worth mentioning:

1. Negative gross margins and negative net margins are two very distinct things. The former means that the product is being sold at below the variable cost required to deliver that product; the latter means that the enterprise as a whole loses money even if the next widget sold provides marginal cash to the company. Many companies cited as examples of businesses that were very unprofitable prior to flipping to profitability at a very high enterprise value — for example Uber, Amazon, Spotify, DoorDash, Netflix — had positive gross margins for a very long time before they had positive net margins.

2. It is not necessarily problematic for a company to run at negative net margins for a long time. We have a word for companies of this type, they are called startups. There are many instances of companies where the correct move is to reinvest capital into the company for scale or growth, rather than to achieve profitability in the near term. We would not have scaled franchises like Amazon if this were not the case.

3. That being said, having negative gross margins is a much more severe state of affairs, because the company has to eventually solve its per-unit profitability problem and clear its overall fixed cost profitability hurdle. Like a normal startup, it has to survive off the generosity of others (i.e. external investors), except the money is being vaporized at a much faster rate than a more typical startup with positive gross margins would. Ceteris paribus, this sort of company is in a riskier position than a vanilla startup.

4. The Andreeseen Horowitz article identifies four criticisms of negative-margin AI harness startups and explains why there is a path for these companies to overcome the critique and end up in strong economic footing with durable, profitable franchises. These four critiques are 1) AI apps need to show they can improve their cohorts’ gross margin profile over time; 2) AI apps need to do this in spite of sitting on top of model companies that are voracious profit-maximizers; 3) they need to prove they have “real” product-market fit even after the VC subsidy is retracted; 4) they need to avoid get gobbled up by the advancing frontier of frontier model capabilities. Let’s be clear: There will absolutely be AI app companies that successfully overcome all of these objections. It seems entirely possible that certain companies will thrive despite underwater gross margins in the near-term.

5. With that being said, we should meanwhile acknowledge that it is very rare for a company to go all the way from severely negative gross margins, up to being a scaled durable franchise with a huge moat around it, without getting outcompeted or blowing up along the way. This is just, I dunno, not something that we should expect from an entire cohort of companies. It’s extremely anomalous. You might cite Uber as an example of a company that made it work. Well, Uber had to burn around $25 billion to emerge from a multi-year competitive battle to the death with its competitors in order to survive long enough to even have the possibility of one day becoming GAAP profitable. Try asking Bill Gurley how excited he is to go through all of that again. How many WeWorks and Wags had to detonate in order for us to get each Uber or DoorDash?

6. In order to overcome the margin issue, these sorts of AI companies will have to find methods to develop stickiness and pricing power over time (enterprise lock-in, network effects, bespoke capabilities, brand, etc)… it’s worth restating that AI is a hyper dynamic emerging industry where nobody really knows who (if anyone) in the value chain will hold pricing power over the long run.

7. Maybe VC power law math (the winners make up for the losers) means that each fund only needs one AI mega home run to justify a whole bunch of blowups of portfolio companies taking this approach to scaling. Could be. That wouldn’t make me feel great if I were the CEO of a single AI company, though.

8. I do not believe the onus is on third-party critics to justify why they believe negative gross margin startups are in big trouble. The onus, rather, is on those outlier companies to explain why their company will overcome these laws of physics and prosper despite the losses. Which, to be fair, the Andreessen team did an admirable job of defending. But it’s not unfair or thoughtless for the general investing public to assume that a company that a loses money on every sale is probably doomed, because that’s how things work in 99.9% of businesses.

Just ask the folks at Y Combinator, who believe every SaaS company will face competition from a newer agentic AI version of the same product. Or public market investors who, aside from some outliers like Palantir and Cloudflare, are assigning values to public software companies at 10-year low revenue multiples. Despite the NASDAQ reclaiming all-time highs, public cloud software remains a very out-of-favor category as investors perceive their prospects and defensibility to be not-so-good:

The logic of incumbent displacement analogizes AI to prior platform shifts such as the transition to cloud. When software began to move from on-prem to the cloud, you couldn’t exactly “port” an existing product into a SaaS instance, or bolt a “cloud” upsell module onto your current platform — you more or less had to rewrite the software from scratch.

Incumbents had some huge disadvantages in this state of play:

• Their R&D teams were filled with individuals who didn’t have experience in cloud.

• They had Innovator’s Dilemma type effects — legacy customers were paying them large maintenance fees to continue supporting on-prem products, making it very hard to discontinue older platforms in favor of cloud. This meant that incumbents were burdened with the cost of maintaining two completely different platforms (on-prem and cloud) into the foreseeable future, if they were to attempt a foray into SaaS.

• Standardization of web development frameworks & cloud infrastructure products like AWS made it much cheaper to build a net new SaaS product with a small team and some seed funding. Also, armies of 22-30 year old developers without distractions like babies and mortgages had a native understanding of these tools that their older peers in the bigger companies did not.

These dynamics favored Rebels over the Empire and created conditions whereby it was much more efficient to start a brand new cloud company than to transition an older business to SaaS. We haven’t even talked about the financial upside angle yet — there’s a reason why a thirty-something Marc Benioff set out to build Salesforce by himself instead of doing it from within Oracle.

In general, “disruptive” technology platform shifts like on-prem to cloud or mainframe to PC have similar dynamics of the young/new outcompeting and displacing the old. Steve Jobs and Steve Wozniak were 21 and 26 years old, respectively, when they founded Apple. Bill Gates and Paul Allen were 19 and 22. Incumbents at the time like IBM and Digital Equipment Corporation did not produce these technologies in-house.

But it’s important to note that not all new technologies are disruptive platform shifts. Sometimes better tooling is just better tooling. Everybody gets to use it, and everybody benefits, without the field being tilted heavily in favor of new entrants building new platforms for a new era.

Just to pick a random example, think of the rise of modern integrated developer environments (IDEs) like Microsoft’s Visual Studio (launched in 1997). This is a technology that significantly improved the efficiency of building software. But it did not create any sort of “well now we have to throw everything out and build new stuff from scratch” dynamic. And furthermore it didn’t disfavor incumbents; practically every big company in the world in ‘97 already had an account with Microsoft and could adopt the technology quite rapidly.

So how “disruptive” is AI? Let’s go through a few angles to answer this.

• Do you need an entirely new company architecture to take advantage of it? The answer is “yes” if you are a foundation model company like OpenAI, but “no” for pretty much anyone else, at least as far as we know now. Any cloud software company can summon AI into their product simply by using an API call. You don’t need to “build an entire company from the ground up to support AI” to satisfy the AI demands of customers (whereas you do need a cloud product to sell over the cloud).

• Does the new technology favor ambitious 23 year old technologists over experienced 45 year olds? Interestingly, we are starting to find out that AI coding assistant tools are most useful in the hands of experienced developers who best know how to debug the mistakes they invariably make. In the hands of a junior dev they mostly produce useless slop. There’s a reason why the job market for new grads in tech is getting dire.

• Does AI turn legacy incumbents’ sprawling codebases into a competitive disadvantage for them? For the moment, AI cannot autonomously build big projects with lots of dependencies, and high requirements for security and compliance. It’s just not good enough yet. So it feels like all the years of work that incumbents put into making their products enterprise-ready will continue to be a competitive moat for them.

• Is there an Innovator’s Dilemma issue where legacy companies can’t sell AI products without cannibalizing sales of non-AI products? Seemingly not at all. Incumbent SaaS providers are rapidly building AI modules and distributing them directly into their install base. If anything the opposite seems to be happening: Incumbents are finding that their customers don’t want to pay more for an “AI product”, they just want AI to augment the value proposition of the product they’re already paying for.

• Do new AI-native entrants have a distribution advantage over incumbents? In some cases yes, driven by the virality of truly novel products (see ChatGPT, Cursor, Replit, etc). But on the other hand, every scaled incumbent SaaS company is completely NRR-obsessed and has a gigantic customer success function that is tasked with putting new products in the hands of existing customers. TBD on this one.

Some of these dynamics will probably change over the next few years. I’m not here to argue against scaling laws.

But for now we know this: AI significantly lowers the cost of building new software, particularly when it’s in the hands of strong, experienced software developers who know how to use it. And who has the most experienced top decile software developers at their disposal — a random startup, or Microsoft?

So what should we make of all this? Below are some general thoughts; I’d certainly welcome any feedback from readers.

• U.S. Federal deficits are going up, not down. Period.

• U.S. manufacturing output is going down, not up. At least in the medium term. American manufacturing job loss and layoffs are already happening because orders are down and the cost of raw materials is rising. Can this be reversed in the longer run if the tariffs stick around? Maybe, but tariffing everything you need to build a factory (steel, aluminum, machine tools, gypsum, lumber, etc) is not going to help anyone build new facilities. Providing incredibly contradicting and unclear instructions about tariffs is freezing most capital investment. Most likely scenario is that U.S. manufacturing (not necessarily the entire economy) goes into a recession under Trump.

• We are about to lose many, many American small businesses. Personally I am disturbed by how little the administration seems to notice or care about this. It’s one thing for the government to decide that trade-dependent businesses should move their supply chains to the USA, and for government to prod those businesses along gradually with carrots and sticks. It’s another thing entirely for the White House to unilaterally put your company out of business.

• Forecasting the Fed response to all of this is challenging. On one hand, recent CPI/PPI readings have been very low, and high tariffs are recessionary which should provide rationale for cuts. On the other hand, tariffs are expected to have an inflationary impact on the PCE. On the THIRD hand, this administration’s efforts to deregulate and open up the energy industry should continue to bring gas and electricity prices lower, counterbalancing the tariff impact. This isn’t like early 2024 when the Fed was obviously going to cut, the question was just how much.

• The “4D chess” / Mar-a-Lago Accord endgame for all this — which is a new state of the world with a reconstituted pro-U.S. / anti-China trading bloc with lower rather than higher overall barriers to trade — requires the rest of the world to cooperate with us in “cutting deals” so that we can get tariff rates down before they cause a recession in the USA. Here’s why I’m skeptical of this working out: We have created a political dynamic where other countries’ leadership becomes more popular when they defy the United States rather than cooperate. Just look at the Canadian election, where the Liberal party moved from 20-1 underdogs to favorites to hold power just in the last couple of months, because the alternative party (the Conservatives) is perceived as cozier to Trump. How are we going to cut a grand bargain when foreign Prime Ministers worry they will get thrown out of office if they work with us?

• Democrats are now 83% betting favorites to take the House in 2026 (up from 67% five months ago). Trump’s net approval has gone from +12 to -5 in less than three months. Many people are making the assumption that this approval number will not go down much further because Trump has a relatively high floor of ironclad support… I am not so sure about that. If we do get a manufacturing recession (or perhaps an economy-wide recession), increased SMB bankruptcies, higher prices for everything except for energy, and more pain in the stock market, we may nestle in the range between -10 and -20. That may be enough for GOP Congresspeople to fear enough for their reelection to break from MAGA. What happens then? Will we see Congress take back the power to tariff from the President?

• DOGE is the wildcard in all of this. It’s reasonable enough to say that the DOGE cuts have not yet showed up in the spending numbers because they’ve only been at work for 12 weeks. But what happens if we get to, say, August, and overall discretionary Federal spending still hasn’t budged? Right now the general public has only been able to verify 8% of the total spending that DOGE has claimed it has cut. If it turns out all the chaos and big Musk promises amounted to negligible impact, what will the political consequences of that be? Many moderates are currently withholding judgement on the Trump administration because they believe DOGE’s role as an efficiency-improver and waste-cutter is necessary for America to avoid a debt death spiral. If those Trump-tolerating moderates lose that faith… then what?

One of my big theories for 2025 is that public software names are going to get a LOT more efficient and will begin to generate significantly more cash. Two main reasons for this:

1. AI. While I don't see fully autonomous agents replacing human employees en masse (this year at least), many companies will be able to reduce hiring or keep headcount closer to flat, while asking those employees to become incrementally more productive with AI tooling. It seems reasonable that largest beneficiaries of consuming AI technologies (as opposed to selling AI functionality) would be software companies, for the obvious reason that building and implementing workflow tech is their modus operandi.

2. Sticky, growing, recurring revenue software companies are among the easiest to manage for greater efficiency (if so desired by management). Because nearly all costs come in the form of personnel1, a simple way to increase cash flows is to hire more slowly than ARR goes up.

I now direct your attention to the chart of ARR per employee (top of this post) and trailing 12mo free cash flow percentage (below). These are not all public SaaS companies, but a subset that I track in a personal workbook of mine.

The data is noisy and incomplete (not all companies publish quarterly headcount numbers, for example; and ones that do occasionally stop) but the direction is clear: Efficiency is up, cash flows are up too.

A lot of this is the natural consequence of the long-term post-ZIRP rebound and should not be surprising. But one thing you'll notice in both charts is a somewhat-pronounced positive inflection in the most recent quarter (Q4-24)2.

If I had to guess, I would say that Q4 is the first quarter for AI impact to show up in software companies' profitability. Barring a recession or an over-supply of capital rotating back into midcap software names, it’s tough to imagine this trend reversing anytime in 2025.

Is (virtually) every public SaaS business going to be a cash-generator by the end of the year? That would be quite the departure from fall 2022 when half of the online financial commentariat was calling SaaS companies “structurally unprofitable shitcos”.

Certain topics become big public controversies precisely because they have multiple reasonable, yet incompatible, explanations. Is it two faces or a vase? Is the dress blue or yellow? Is Edward Snowden a traitor? If the answer were obvious to all, there’d be no controversy and therefore we wouldn’t be arguing about it in the first place. So it must somehow be both, except it can’t be both simultaneously because the two sides are contradictory. So we organize ourselves into rival camps arguing for each side of the disagreement, and a controversy is born.

Our latest public controversy is the question of Elon Musk’s, well, brain. How high is his IQ? How exactly do we define “competence”? Is he having a mental breakdown? Is he saving the government or destroying it? What the hell is with this guy?

Here it is in a nutshell: How is it that someone with a track record of such stunning success in business (that’s the “competence”) can be ignorant and deranged in so many ways that are reminiscent of the most deluded and incompetent people we’ve ever met in our lives? This is the world’s weirdest Rorschach Test, right? He can’t be smart AND dumb at the same time, right? Don’t we have to pick?

Antibody theory

I have a pet theory that squares the circle on this question, and I call it the antibody theory of Elon.

You know how when a foreign invader penetrates the body (say a virus or a vaccine), the body produces antibodies that not only fight off the invader, but make the body stronger and more resilient against future infection? That’s how I think Elon’s organizations respond to his chaotic leadership. And this is how an “incompetent” person can produce highly competent and effective organizations.

I have a few friends who have worked reasonably high up in several of Elon’s companies — high enough to have had plenty of meetings with the man himself. The stories they share about Elon meetings are all pretty similar. Often they feature Elon hyper-fixating on some detail of a presentation that he found unsatisfactory, sometimes capriciously firing a random employee in response, and demanding that the team accomplish some new, impossible task in order to satisfy him. By “impossible” I don’t mean in the sense of requiring superhuman work effort on a tight deadline; I mean “impossible” as in violating the laws of physics. You hear this over and over, whether it’s Tesla or SpaceX or Twitter or whatever.

The fact that Elon is known to behave in business settings like the worst ex you’ve ever had isn’t really what makes this interesting. What’s interesting is that, according to people who work at these companies, the chaotic behavior often provokes extremely productive outcomes within the organization. This seems to come primarily in two forms:

1. Sometimes this happens in the form of Elon getting what he wants. In a defiant effort to prove Elon wrong, the team will race forward on his mandate and sometimes realize that he was right, or at least less wrong than they had initially believed. Because the effort was fueled by animus and rage and spite, this result was reached much quicker than it otherwise would have been.

2. Sometimes Elon doesn’t get what he wants because it actually is impossible. But because “no” is not an acceptable answer to Elon, the team has to reach its goals in a clandestine manner — if he finds out you’re not doing what he asked, you’re fired. This also means speed, because the faster the team can do what they had planned to do all along before Elon did his meddling, the faster they can prove that their approach was correct and his was wrong (and therefore save their jobs).

In both cases, two things are true: 1) Everyone is pissed off at Elon and wants him to go away; and 2) The group does better work and arrives at the right decision more quickly than if he hadn’t antagonized them. He’s the virus, they’re the body, and now the body has better antibodies to defend it from the next viral attack. Repeat this over and over, all throughout the organization, and eventually you have a pretty high-performance company.

This means that the fact that Elon is impulsive, abrasive, confrontational, and often wildly uninformed is somehow a feature rather than a bug. These are the actions of an invader causing tissue damage and provoking an immune response. It also explains the paradox of how so many people at his companies sincerely dislike him yet feel like they are doing the best work of their careers.

MBWA: Management by Widespread Antagonism

If the average big company CEO is like the Emperor Hirohito — cloistered in his imperial palace, completely out of touch with the situation on the front lines, and only able to comprehend the world via thirdhand reports from scurrilous and self-interested generals — then Elon is like a version of Napoleon Bonaparte who gets into a fistfight with a different platoon leader every single day, and whose army is so annoyed by this that they conquered Europe just to spite him.

Importantly, to pull of this strategy you really need to be okay with people hating you. Most business leaders want to be loved, and many will put off uncomfortable but necessary conversations in order for their image to remain pristine.

Elon inverts this deliberately. Think about what type of person is comfortable sending out a “what did you get done this past week?” e-mail to 3 million Federal employees. You’d have to be incredibly impervious to people calling you an asshole. And the only reason to dress like this is because you really want people to think you actually are one:

When people jeer at Elon for being so oblivious to how he comes off, I think they are missing the point that this is all deliberate. And the reason it’s deliberate is because he’s gotten incredibly good results from it in the past. Why would he stop now?

Is this good? Is this bad?

I’m not here to judge, although I have some reservations over whether the Federal bureaucracy is going to react as beneficially to the Elon-virus as, say, the Twitter engineering department. Instead of passing judgement, I think there are a few important takeaways from this theory:

1. People yelling on the internet, especially liberals, love to call out hypocrisy as if the fake internet points you get for finding it can be traded in for anything at all. I’ve been doing this too, like when I demanded that Elon and Trump conform to the standards of logical reasoning that I require of them. (You really showed them there, Pat; I bet those 4 “likes” felt really good.) But power doesn’t care about hypocrisy. Calling someone out for being a jerk doesn’t work very well when they’re doing it intentionally. All you’re doing is getting angry and giving them what they want.

2. Although I am loathe to give him credit for (in my view) acting like a total jerk to Ro Khanna, who I really like, there probably is something to this idea of J.D. Vance’s that we are way too precious about the exact words people use, equating them with actual violence. When we get triggered by something we see online, the next question is: Are you going to crumble from this, or will you develop your own antibodies? If Elon/DOGE/Trump all fail, it won’t be because of stupid things they tweeted, it will because of unethical acts they committed.

3. If for nothing else, I appreciate that Elon is showing us new and interesting ways to assemble and motivate organizations. And if DOGE is a total disaster, at least we’ll have learned something new there too.

4. With all that said… I think it’s better for people to voluntarily opt-in to workplaces with leadership of this type, rather than have it foisted upon them after decades of committed labor.

I should say upfront that I enjoy anti-tech takedowns very, very much. My industry has a tendency of being hubristic, naive, and dilletante-ish — it’s good for mental hygiene to occasionally hear why other people might think that we’re morons.

So I started my morning by reading the never-boring Ed Zitron’s latest jeremiad against the AI industry, which of course has some healthy criticisms that those of us funding and building AI companies should take note of.

But a few items did catch my eye. It’s not so much the content of the critiques, but the tone and what they imply about the person making the critique. These are from Ed (emphasis mine):

I'm sorry, I know I sound like a hater, and perhaps I am, but this shit doesn't impress me even a little.

Deep Research is the AI slop of academia — low-quality research-slop built for people that don't really care about quality or substance, and it’s not immediately obvious who it’s for.

The fact that Sam Altman can ship such mediocre software and get more coverage and attention than every meaningful scientific breakthrough of the last five years combined is a sign that our society is sick, our media is broken, and that the tech industry thinks we're all fucking morons.

And one from a commenter:

The part that offends me the most is that everything these Slop Machines deliver is so relentlessly mediocre. Imagine spending this much money, burning this much electricity, and dedicating this much human ingenuity in the pursuit of automating mediocrity.

I would synthesize this line of thinking into the following expression: “The AI industry thinks I’m too stupid to tell the difference between their crappy synthetic product and ‘the real thing’… but they’re wrong and I can!”

Self-importance, an obsession with quality and authenticity, a desire to project an image of expertise… I had a startling realization after reading through all of this: Oh my God, these people are like wine snobs! These are basically hipsters who want to impress us with their knowledge of fair trade coffee beans, Japanese whiskey, local craft beer, and artisanal chef’s knives. “I know quality when I see it! Can’t fool me!”

They argue from a specific point of view about consumer products: That anything mass-consumed (or mass-produced) must be definitionally mediocre1, and that anyone who doesn’t notice or care about this has inferior taste to them.

Mediocrity as compared to what?

Where the dividing line is between appreciating quality and fetishizing craftsmanship, and at what point you become an annoying snob who won’t shut up about the stitching on your footwear, is an argument for another essay. Personally I like nice things, I have a nice desk made of real wood, et cetera.

But what I observe in more Luddite-adjacent pieces of AI criticism is an obsession with the “real,” the “authentic,” and the “inspired” aspects of human cognitive performance and how they cannot be replicated in any manner by machines without massive sacrifices in quality. And, yeah, that’s one hundred percent right if we are comparing Midjourney to Picasso. But is that really what we’re doing here? Is AI taking aim at the apex human skills of elite talent and craft? Or is it building cheaper automations for undifferentiated drudge work?

I argue the latter, which means that comparing AI “mediocrity” to the output of Hemingway or Plato is totally missing the point. AI output — to the extent that it is “slop” — is computer bot slop that attempts to outcompete human quasi-slop for 1/10th the price. Here’s a comment from the one brave soul who dared post any pushback to Ed’s piece:

That said, one place to look to steel man the argument is legal document review, a field I work in. Costs in the US are $42B per year. AI might knock that down by at least 50%, based on my personal experience, so that's $21B annually in just one market--and there are other markets just in legal, but perhaps not as large.

Is a human doing manual legal document review adding art and craftsmanship to the world in the same way that Walt Whitman did when he composed his poems? Does this task summon talents and skills from the absolute limits of human ingenuity and genius? No. Obviously. Karl Marx would describe this person as merely the alienated means of production. The labor might be cognitive rather than physical, but it’s still jamming widgets into sprockets on a (cognitive) assembly line.

And I’m not in any way implying that their labor is undignified; of course I am happy that this person has a job and is providing for their family and chasing their dreams(?). I’m merely repeating back their own words. If you know anyone who works as a paralegal or a recent law grad who has to wade into mountains of document review, they will tell you themselves that this fucking suuuuuuuuuuucks. Why are we pretending like this job is super-duper awesome for the legal associate doing it? Didn’t they go to law school to do… not this?

Everything is Chinese bluejeans

I continue to feel that the comparison between AI and Free Trade2 is an apt one (ahem). You can get mad that Chinese-made jeans can be purchased for under twenty-five dollars at WalMart. You can complain all day long about the inferior quality & lack of craftsmanship in the mass production techniques, and how you "wouldn’t be caught dead” wearing imported pants from a big box retailer (oh, I’m sorry, are you better than Jim Harbaugh?). Feel free to be a snob, I won’t stop you.

ORRRRR… you could marvel that our modern production and distribution technologies have enabled almost every person with a job in this country to afford new pants with a couple hours’ labor (compare to 125 years ago when a middle-class outfit cost about two weeks’ income). Chinese-made jeans show us that products with one-half the quality but one-fifth the price end up winning because lots and lots of normal people want to buy them. And meanwhile the high end craft denim category is thriving; there have never been more options for the connoisseurs and the normies!

I do not think we should go the way of turning human cognitive labor into a fetish object for hipsters (artisanal, small batch account reconciliation, anyone?). Asking people to continue performing jobs that are better3 suited for machines — just because AI makes us feel icky — is backwards and infantilizing.

Per Wikipedia:

The motte-and-bailey fallacy (named after the motte-and-bailey castle) is a form of argument and an informal fallacy where an arguer conflates two positions that share similarities: one modest and easy to defend (the "motte") and one much more controversial and harder to defend (the "bailey"). The arguer advances the controversial position, but when challenged, insists that only the more modest position is being advanced. Upon retreating to the motte, the arguer may claim that the bailey has not been refuted (because the critic refused to attack the motte) or that the critic is unreasonable (by equating an attack on the bailey with an attack on the motte).

Let’s check in on a discussion that’s happening this morning in, I dunno, maybe six or seven million separate group chats around the world:

The Bailey: “Elon is going to cut $2 trillion from the Federal budget and reduce the deficit! I might not love everything the Trump administration is up to, but this seems like a great outcome.”

Somebody Else: “But isn’t $2 trillion more than all discretionary Federal spending programs put together? So they’d have to cut entitlement programs to get there, which they’ve promised not to do? Cutting the entire Federal workforce would only save about $300 billion annually. And didn’t Trump promise a $4T tax cut? How is saving a couple billion dollars at a time going to reduce the deficit when all that other stuff is happening… it seems mathematically impossible? It seems like Trump ran on policies that are destined to make the deficit explode upwards…?”

The Motte: “Things like ballooning administrative costs, rampant Medicare fraud, and absurd technological debt are huge multi-billion dollar problems in this country. Who could be against fixing this???”

Somebody else: “I mean, that’s fine and all. But you said they were going to reduce the deficit!”

I thought for a long while about why this whole form of argument rubs me the wrong way. And here’s the best answer I have: When a good thing (in this case, removing waste in government) is accomplished in the context of a gigantic lie that insults my intelligence as a citizen and as a voter, I’m not going to be as psyched about the good thing as I would otherwise be if they had played me straight.

Eliminating waste and needless bureaucracy is a worthy goal in its own right. I would argue that one of Trump’s best “innovations” is demonstrating just how little most administrative vacancies actually inhibit the smooth functioning of American society.1 There are many good arguments for small government.

But the administration does not need to brazenly lie about deficit reduction in order to make this point. These actions clearly cannot do enough to reduce the deficit by themselves, in a larger fiscal environment of looming tax cuts, static entitlements, rapidly increasing Federal interest expense. This doesn’t mean they’re not worth doing. But by making extravagant claims that it will accomplish the impossible, the administration makes itself appear untrustworthy, and irresponsible with the mandate they have been given by the voting public.

And yes, it was a mandate. But guess what — mandates don’t last forever. Right now prediction markets show a 79% chance of the Democrats retaking the House in 2026. In a thermostatic political environment where the Republicans desperately need to convince voters that Trump/Vance isn’t a total clown show, I do not believe that making promises they have no intention of delivering is a good way to engender that trust.

I imagine that DOGE will continue to dangle wins in front of the public — $15M in waste saved here! $50M in DEI programs eliminated over there! — and that will be fine and good. But let’s not let this distract us from the big issue: This is a pro-deficit administration until proven otherwise.

The chart above shows two actual SaaS companies1 from Toba's portfolio that were launched within a few years of each other. The y-axis is ending ARR by year (in millions); the x-axis are years from inception. Let’s call these businesses Company GTM (go-to-market) and Company Product.

Company GTM was a company that had an incredibly potent ability to close new customers quickly and efficiently. It had especially fast sales cycles and multiple effective sales channels, all of which delivered new customers in high volume. Its founders came from various high-growth enterprises (though none in the current business's target market) and knew everything there is to know about Always Be Closing. Unfortunately it also had a product that was, although in no way "bad," not especially differentiated or difficult for competitors to replicate. As a consequence, retention started off decent but became progressively worse as new competitors entered the market at lower price points. Despite the fast start, this business's category eventually commoditized, and their R&D engine was not strong enough to innovate its way out of the churn trap.

Company Product, on the other hand, had a highly differentiated and complex product serving a use case that no other vendors had thought to solve. The founders had been practitioners in the market that they sold into, and so deeply understood the pain that they had built the product to fix. By the time new competitors had entered the space, the company and its reputation were solidified enough that retention never really dipped from the place it started from, which was excellent. At the same time, the company didn't possess much inherent "sales DNA," and it took them an exceedingly long time (and a succession of VPs of Sales and VPs of Marketing) to finally crack the code on how to efficiently grow revenues. Compared to Company GTM, sales cycles and CAC payback periods were always a lot longer, and consequently the company burned a whole lot more capital in its early years, making it seem like a much less efficient business.

So what happened?

Company GTM obviously grew quite quickly in its first 4 years. Headcount expanded exponentially. VCs jockeyed for a chance to be introduced to the CEO. Although churn was extant, the sheer pace of new bookings meant that it really didn't seem to matter. Company GTM seemed to be on a collision course with >$50m revenues and eventually total market domination.

Conversely, Company Product started out as kind of an enigma. Growth was okay but unspectacular. Existing customers might have been very happy (and rarely churned) but prospects didn't seem to be in a hurry to buy. To the extent that VCs were interested in even taking a meeting with the company, virtually all of them passed, citing poor sales efficiency metrics and too much cash burn coming out of R&D.

For the first 3 or 4 years, it felt like Company GTM was an early potential fund-returning home run and Company Product was probably just a base hit. But something interesting was happening underneath the surface of each company, something that you could feel long before you could prove it with numbers — at Company Product things were getting easier over time, and at Company GTM everything was getting harder.

• Company Product had built up a base of hundreds of happy customers that not only loved the product and renewed/expanded, but brought the product into new companies whenever they changed jobs. Because they had built up a constituency of raving brand advocates, word-of-mouth drove larger pipelines and faster sales. Eventually they attracted experienced sales and marketing executives who build a legitimate go-to-market engine on top of it this strong foundation. In short: The company that started out with more product DNA was able to develop its sales DNA over time. How come? Because the sellers had an awesome product to sell!

• Whereas Company GTM found its churn and customer satisfaction issues becoming more and more insurmountable with scale. At the same churn rate, a $20m revenue company has to replace 10x as many customers each year as a $2m revenue company, and in this instance the churn rates were actually getting worse over time. A series of externally recruited CTOs were unable to develop a "product culture" in the midst of a company that was more oriented around the boiler room environment of the sales pit. New bookings remained very high, but net new ARR began to drop as churn cut out all the gains from new customers coming in. The company eventually found itself in a place where they had to keep burning cash just to keep revenues from falling.

By year 8, the difference in outcomes was… vast.

The story of these two companies is essentially the fable of the tortoise and the hare. The business with only a terrific go-to-market enjoys the sugar high of quick revenues, but this comes with a ticking time bomb attached…. Build a product good enough to match the story you've been pitching, or perish. Silicon Valley is littered with the tombstones of VC-funded businesses that couldn't deliver on this before the money ran out.

At the end of the day, a business needs both a strong product and GTM to thrive. But if you ask me which one I’d want first, give me product any day of the week.

The biggest question in investing right now is "how should we play the AI trade?"

One school of thought says that you should do everything possible to put capital into the 4 or 5 foundation model companies that matter, regardless of entry price. If OpenAI develops the singularity, it will be worth something approximating the GDP of the entire planet. Why spend any time worrying about valuation multiples when the prize is essentially world domination?1

Another school says: Invest into any company built on top of LLMs. The world is changing quickly, and these businesses are growing faster than the winners of any prior generation of technology (supposedly, at least). Buy anything with the word "AI" attached to it. Ignore the haters who diss your "wrapper" companies.

A third school says to just buy Nvidia. And ARM. And Broadcom. And Meta/Google/Amazon/Apple/Tesla/Microsoft. You want to be in the silicon game, baby. Chips are the ultimate picks-and-shovels play for AI.

There are other schools too: Short everything AI because this is just a bubble. Short everything in general because AI will accelerate societal breakdown and World War 3. Buy SaaS. Sell SaaS.

I have a slightly different mental model when it comes to the longer-term AI trends.2 Here's how I think about it: AI more than anything is a broad, massively deflationary force that will create winners and losers, but overall will create more consumer and producer surplus than direct value capture by the companies that constitute the AI value chain.

Put another way: AI is less of a discrete, patentable technology that can be exploited directly by the businesses that build it, and more of a social technology that will lead to diffuse improvements in the standard of living for society as a whole, even as it creates losers in specific pockets of the economy. In this way it is more analogous to a worldwide reconfiguration like Free Trade than a monopolizable asset like railroads.

Why Free Trade? Well, the effect of something like NAFTA is to rotate production to lower cost regions in order to reduce input costs (and eventually finished product costs). This leads to an overall surplus — Producers can procure lower-cost raw materials and intermediate goods to either increase profits for themselves and/or drive more competitive pricing in order to win share, which leads to lower prices for the consumer. Despite the overall surplus it does cause harsh losses in specific parts of the economy, in this case the domestic workforce that is outcompeted by lower-cost foreign manufacturing (in the U.S. this is the hollowed-out manufacturing base).

What's happening with AI? In this instance, white collar cognitive labor is being substituted for a machine that has been trained by crunching down and synthesizing the brainpower of billions of other people throughout history. The distilled outputs of this machine can be procured for fractions of a penny.

Say what you want about slop and hallucinations and "glorified auto-complete" — obviously this technology is wildly imperfect, no disagreement from me there. But let me ask you a question: Do your "Made in China" jeans ever have a button or a pocket sewed on incorrectly? Does that lead you to instead spend $250+ on American-made jeans, or do you just shrug and deal with it? At a certain point the cost savings are worth the inconsistency in craftsmanship, as evidenced by the $20 Wranglers that millions of people buy at Walmart even as American denim plants continue to shut down.

Note that I didn't say this is good, just that it is happening. It is the nature of capital to seek its highest rate of return (or as Ranjan Roy says, "money is always swimming towards yield"). Now that the AI genie is out of the bottle, will it be implemented up and down the economy to drive down costs? It is inevitable.

$50,000 worth of labor for $1.68

I was struck by a post from Simon Willison two weeks ago that described how cheap it was for him to accomplish a pretty major task with Gemini that would likely take a human a few months to do manually. Apologies for the long blockquote but you really need to read it all to let it sink in:

Here’s a fun napkin calculation: how much would it cost to generate short descriptions of every one of the 68,000 photos in my personal photo library using Google’s Gemini 1.5 Flash 8B (released in October), their cheapest model?

Each photo would need 260 input tokens and around 100 output tokens.

• 260 * 68,000 = 17,680,000 input tokens

• 17,680,000 * $0.0375/million = $0.66

• 100 * 68,000 = 6,800,000 output tokens

• 6,800,000 * $0.15/million = $1.02

That’s a total cost of $1.68 to process 68,000 images. That’s so absurdly cheap I had to run the numbers three times to confirm I got it right.

How good are those descriptions? Here’s what I got from this command:

• llm -m gemini-1.5-flash-8b-latest describe -a IMG_1825.jpeg

Against this photo of butterflies at the California Academy of Sciences:

A shallow dish, likely a hummingbird or butterfly feeder, is red. Pieces of orange slices of fruit are visible inside the dish.

Two butterflies are positioned in the feeder, one is a dark brown/black butterfly with white/cream-colored markings. The other is a large, brown butterfly with patterns of lighter brown, beige, and black markings, including prominent eye spots. The larger brown butterfly appears to be feeding on the fruit.

260 input tokens, 92 output tokens. Cost approximately 0.0024 cents (that’s less than a 400th of a cent).

This increase in efficiency and reduction in price is my single favourite trend from 2024. I want the utility of LLMs at a fraction of the energy cost and it looks like that’s what we’re getting.

This is what I meant by a massively deflationary force.3 How much would you personally demand to be paid to write 4-6 sentences each that describe 68,000 photographs? You probably wouldn't do it at any price because it sounds so dreadful, but if you were forced to price it out it would probably be something like $50,000 (one description every 3 minutes means that this task would take 3400 hours; at $16 California minimum wage that's $54,400. By the way, that would take you about a year and a half to accomplish in 40 hour workweeks.) This crunched-up-humanity machine just did it for $1.68 and probably a few minutes.

Again, you can talk about how this is all just uninspired regurgitated slop, but this is just one example of how the technology is undoubtedly a lower cost way (a dramatically lower cost way, actually) to accomplish cognitive labor that could in some other instance employ a human. Even if the quality of the output is terrible4, the cost savings overwhelm the quality control problem by many orders of magnitude.

Does this example make you want to be an investor in OpenAI, the company that had to spend billions of dollars on data centers to crunch all of the text on the Internet into a distillate goo that they could sell to people like Simon Willison for $1.68? Or would you rather be Simon Willison, the guy who just bought $50k of labor for less than two bucks? It seems to me that 99.99% of the economic surplus in this transaction went to the consumer (Simon). That’s not me saying OpenAI is a bad business; I’m just saying that it’s getting really good to be Simon. Here’s the best news: We’re all about to be Simons.

We're about to enter a world where a pair of jeans costs 5 cents — maybe you have to throw out 2 or 3 pairs before you find one that is sewn correctly, but who cares? (Temu is a thing, after all). This is going to change everything.

Rather than asking ourselves how to "play the trend"5, what we should be doing is pondering how this is going to remake the world — what new opportunities are going to be possible6, and what cherished things we're going to lose.

In my early days as a wee junior VC, I tried to differentiate myself by being maximally responsive. Unlike my impossible-to-reach competitors — who demanded founders seek warm intros from a trusted third party in the Silicon Valley in-crowd — I read and responded to almost everything in my inbox.

Obviously most of these interactions went nowhere; that's the game in venture capital. But I at least tried to be helpful. Even if a meeting didn't make sense, I attempted to be liberal in providing feedback, encouragement, and sometimes even introductions to other investors who might be a better fit. It was cool, and a lot of people seemed to appreciate it.

I'm not telling you this so you'll think I'm some super terrific guy. This whole thing was strategic — I was trying to build a reputation and generate dealflow. And for a while it worked! Our seed investment in Skupos, which went on to raise over $50m and eventually be acquired for 9-figure consideration, kicked off with a cold e-mail that turned into a meeting. Many other interesting contacts and relationships originated similarly. I even wrote a Medium article advocating for this approach that became the most-read thing I've ever published (67,000+ pageviews as of this morning).

Equally importantly, being a Pied Piper for cold e-mail responsiveness was fun for me. The ability to generate new investment opportunities or meet cool new entrepreneurs from my laptop was essentially a game (and better suited for my introverted side than conferences and networking events).

This continued for a long while. But the last 12 months or so — as we've embarked on the brave new world of the AI era — have been very different. You could say the fun is now almost completely gone.

It's not that the quality of messages have gone down. If anything, the median quality of writing has gone up (I get a lot fewer wild, crazed messages from possibly insane people).

What's changed is the uncanny sense that there is no other person on the opposite side of this exchange. In the past there was a palpable feeling that, if I were to reply to this message, literally anything could come of it — an engrossing conversation, a new investment, a friendship, an argument, who knows. Now whenever I do reply, I either get bot slop sent back to me 30 seconds later, or more often I don't get anything back at all, even if my message was that I'd love to learn more about their company!

The emotional difference is vast. There's a real sense that, even if the startup opportunity sounds captivating or the writer sounds brilliant, all I'll inevitably accomplish by responding is falling into a GPT-powered dialogue with some ghost in the machine who is simultaneously having the same conversation with 10,000 others. This isn't fair to the non-AI founders who are authentically reaching out, but it's tough to shake the feeling, after so many annoying fruitless interactions, that nothing interesting or actionable is coming inbound via cold e-mail anymore. So I regretfully declare that my cold e-mail era is now (mostly) over.

But forget me for a second… nobody cares about my inbox. I can't be the only one who feels this way, right? My subjective reaction — which is a defensiveness and a defeat in the face of increasing doubts about who in my life is a real live human being — can't be unique. Are you guys seeing and feeling the same thing at your jobs, or on social media, in your dating apps?

I think what may be coming in our new AI world is a mass feeling of disorientation and fear.

Back in the cavemen days, you knew instinctively that the people in your tribe were safe, but any strangers outside the tribe were unsafe, a potential danger. As we developed civilization, our circle of safety expanded — mega-cities like New York are only possible if society develops a comfort around unknown people. The concept of globalism is perhaps the apotheosis of this development… not only are the poorest and most remote Afghanis, Indians, and Sudanese your kin, but so are the animals that we threaten with climate change. So are future humans who haven't even been born yet, so say the effective altruists.

I'll be the seven billionth person to point out that this movement is now in obvious retreat. Political pressure against the interests of strangers (whether immigrants, Ukrainians, Chinese textile workers, or the people we might protect by vaccinating ourselves against diseases that could harm them) is the highest I can recall in my lifetime. Is it a coincidence that these politics are transpiring at the same time we're reaching a technological moment where we're becoming less certain that the strangers who enter our lives are even people at all?

Hmm.