Economics on Philipp D. Dubach | Quantitative Finance & AI Strategy

When AI Labs Become Defense Contractors

me@philippdubach.com (Philipp D. Dubach) — Sun, 01 Mar 2026 00:00:00 +0000

Lockheed started by building Amelia Earhart’s favorite plane. Then came a government loan guarantee in 1971 (the L-1011 TriStar nearly killed the company), a Cold War, decades of consolidation, and now a business that earns 92.5% of its revenue from government contracts, with the F-35 alone accounting for 26% of its $71 billion in annual sales. The process took about 50 years. AI labs becoming defense contractors will happen faster.

On February 27, 2026, two things happened within hours of each other. President Trump ordered every federal agency to “IMMEDIATELY CEASE all use of Anthropic’s technology” after CEO Dario Amodei refused to strip safety constraints from Claude’s Pentagon deployment, specifically prohibitions on mass domestic surveillance and fully autonomous weapons. Defense Secretary Pete Hegseth then labeled Anthropic a “Supply-Chain Risk to National Security,” a designation previously reserved for foreign adversaries like Huawei, never before applied to an American company. That evening, Sam Altman announced that OpenAI had signed a deal to deploy its models on the Pentagon’s classified network, posting that the Department of War “displayed a deep respect for safety.” (Whether that reflects the Pentagon’s actual position or Altman’s political optimism, remains unclear for now.)

Most coverage has framed this as an ethics dispute. I think that framing is going to age poorly. What I see is the economics of defense spending doing what they have always done to every company they touch, and the ethics arguments becoming less audible as the financial gravity increases.

The Last Supper and defense industry consolidation

In the summer of 1993, Secretary of Defense Les Aspin and Deputy Secretary William Perry invited the CEOs of America’s defense firms to dinner at the Pentagon and told them, in so many words, that most of them would not survive. Cold War budget cuts meant the government could sustain roughly one prime contractor per equipment category. Norman Augustine, then CEO of Martin Marietta, named it the Last Supper. The message was clear: consolidate or die, and the government would not stop you from consolidating.

The restructuring that followed was fast, even by M&A standards. Within four years, 51 prime defense contractors collapsed into five: Lockheed merged with Martin Marietta in 1995 ($10 billion), Boeing absorbed McDonnell Douglas in 1997 ($13.3 billion), Raytheon folded in Hughes Electronics and Texas Instruments’ defense unit. Between 2011 and 2015, an additional 17,000 U.S. companies exited the defense industry, a contraction that hollowed out the supplier base the Big Five still depend on today.

The revenue dependency data shows what happens to the companies on the inside of that consolidation. Boeing before 1997 was, as Bank of America analyst Ron Epstein put it, “a company where engineers were high church.” Post-merger, Boeing relocated its headquarters from Seattle’s engineering center to Chicago, physically separating leadership from manufacturing. Defense rose to 35.8% of Boeing’s FY2024 revenue ($23.9 billion). The cultural shift that merger carried, financial discipline over engineering judgment, is what most 737 MAX post-mortems eventually trace back to. Companies don’t plan to end up here. They respond to incentives, and the incentives compound.

The AI industry will face the same incentives, just faster, and through a different mechanism: not M&A but access to classified networks and government-funded compute.

How Pentagon AI spending reshapes a company

The FY2025 DoD AI budget was $1.8 billion, a figure that nearly everyone involved described as insufficient. The FY2026 budget request earmarks $13.4 billion for AI and autonomous systems, a roughly 7x increase in a single budget cycle, and the first time these technologies have their own standalone line item inside a total defense request of $892.6 billion. For context: Anthropic’s full annualized revenue as of February 2026 was approximately $14 billion. The Pentagon just made AI a budget category larger than most of the companies selling it.

Anthropic burns an estimated $3–5 billion annually; OpenAI burned approximately $8 billion in 2025. Neither has a clear path to profitability before 2027 at earliest. Government AI contracts offer something consumer businesses cannot: predictable, multi-year, politically protected revenue streams that don’t churn when a competitor releases a better model.

The defense procurement structures deepen that dependency over time. IDIQ contracts (Indefinite Delivery, Indefinite Quantity), which now account for roughly 56% of DoD contract award dollars, run five years with extension options. Palantir’s Maven Smart System contract started at $480 million and expanded to nearly $1.3 billion through 2029. The JWCC cloud contract, which replaced the cancelled $10 billion JEDI contract, placed over $3.9 billion in task orders within three years of award to AWS, Google, Microsoft, and Oracle. Once embedded in classified systems, switching costs become close to prohibitive. A competitor cannot simply offer better inference speed.

Security clearances are maybe the most underappreciated asset in the defense tech ecosystem. Processing a clearance takes an average of 243 days end-to-end, up to a year for TS/SCI with polygraph. Only around 4.2 million Americans hold active clearances, roughly 2.5% of the labor force, and an estimated 500,000 to 700,000 cleared positions currently sit unfilled. Average cleared professional compensation hit $119,131 in 2025; full-scope-polygraph holders averaged $141,299. For AI labs accustomed to hiring from MIT, Cambridge, and ETH Zürich, the cleared talent pool is thin and gets more expensive every year.

Any lab serious about classified work has to build a parallel organizational structure: separate hiring pipeline, separate facilities, separate operational security requirements. The lab that builds that structure first has a moat no competitor can cross quickly.

Palantir’s trajectory as the defense tech blueprint

The clearest view of where this ends is Palantir, which has been running the experiment at scale for a decade. It posted $4.48 billion in FY2025 revenue, up 56% year-over-year, with government comprising 53.7% of the total, down from a peak of 58.2% in 2021 as its commercial AIP platform gained traction. Its $10 billion U.S. Army Enterprise Agreement in July 2025 consolidated 75 existing software contracts into a single framework. Its market capitalization reached roughly $320 billion by late February 2026, making it worth nearly twice Boeing. The model, government as the client that funds and validates the technology, commercial as the client that justifies the valuation, is what the AI labs are now building toward.

OpenAI at an $840 billion valuation with a classified Pentagon network deal is already further down that road than most coverage acknowledges. It has appointed retired General Paul Nakasone, former NSA director, to its board. It hired Dane Stuckey, who spent a decade at Palantir and served as its CISO for six of those years, as its own CISO. It has active job postings for Government Account Directors in Defense requiring Top Secret clearance and defense revenue targets exceeding $2 million per year.

The publishing record is moving the same way. OpenAI’s 2015 founding post promised researchers “will be strongly encouraged to publish their work.” GPT-1 shipped with open-sourced code. GPT-2 was partially withheld in 2019, GPT-3 fully closed in 2020, GPT-4’s architecture undisclosed in 2023. OpenAI released smaller open-source models in August 2025 (its first since GPT-2, six years later) but they were text-only, trained on synthetic data, not frontier systems. Google removed the “AI applications we will not pursue” section from its principles in February 2025, including the explicit weapons prohibition. Meta opened Llama to defense agencies and contractors including Lockheed Martin and Anduril in November 2024. Anthropic has never open-sourced a Claude model. Every major lab is moving in the same direction.

The counterargument, and it’s a real one, is that defense R&D has historically generated civilian spillovers: ARPANET, GPS, jet engines, the semiconductor supply chain. Moretti, Steinwender, and Van Reenen, writing in the Review of Economics and Statistics (2025), found that a 10% increase in government-funded defense R&D generates a 5–6% increase in privately funded R&D in the same industry: crowding-in, not crowding-out. The estimated total effect: U.S. private R&D investment is $85 billion higher than it would be without government defense spending.

But there’s a difference between how much research gets done and what it gets pointed at. Lockheed’s R&D is now probably almost entirely classified hypersonics and directed-energy weapons. What it learns there does not flow back to commercial applications in any useful timeframe. The research volume expands; the scope narrows. Bell Labs devoted a substantial share of its personnel to government contracts at its Cold War peak; the 1956 AT&T Consent Decree forced royalty-free patent licensing on the transistor, which accidentally accelerated the civilian semiconductor industry by giving Texas Instruments and Fairchild Semiconductor access to the core technology. AI labs operating under classification will not be forced to open-license anything. That mechanism does not exist for software under ITAR.

I’m more confident in the direction of this analysis than in the timeline. The Anthropic supply-chain-risk designation may not survive legal challenge. The $13.4 billion FY2026 AI budget might not survive unchanged. Amodei might find a compromise that others in the industry treat as a ceiling rather than a floor. What I don’t think reverses is the structural pull. The defense budget is the largest single purchaser of advanced technology on earth, it’s growing, it operates on multi-year contract cycles that reward incumbents, and it is willing to use blunt regulatory tools against companies that don’t cooperate, as Anthropic learned in about six hours on February 27.

The Last Supper logic applies here too: the government will not block consolidation, and it will not save the AI defense contractors that don’t participate. It will just find a different partner who will.

Economics of a Super Bowl Ad

me@philippdubach.com (Philipp D. Dubach) — Fri, 20 Feb 2026 00:00:00 +0000

A 30-second Super Bowl ad costs $8 million. That’s $267,000 per second, roughly the median U.S. home price for every tick of the clock. Super Bowl LX drew 124.9 million average viewers with a peak of 137.8 million, the highest peak audience in American television history. The NFL accounted for 84 of the top 100 most-watched U.S. telecasts in 2025. The Oscars, by comparison, managed 19.7 million.

Ro (that’s the name of the direct-to-patient telehealth company) CEO Zachariah Reitano, writing from direct experience as a 2026 Super Bowl advertiser, published a detailed cost breakdown based on his own spending and interviews with 10+ brands. The picture that emerges is considerably more expensive than the headline number. Production runs $1–4 million for studio, crew, and post-production before any famous face enters the frame. Celebrity endorsement talent adds $1–5 million, with the current A-list sweet spot at $3–5 million according to WME agent Tim Curtis. Then comes the companion buy: for every 30-second slot, advertisers are generally required to commit to spending an equivalent amount on other programs broadcast by the same network. For NBC’s 2026 Super Bowl, that meant additional inventory across the Winter Olympics and NBA All-Star Game, adding another $7–10 million to the tab.

Total committed spend: $16–23 million for a single 30-second spot. CFO.com’s Jason Hershman brackets the full range at $15–50 million depending on ambition.

For companies already spending nine figures annually on marketing, the framing of a Super Bowl ad as a “portfolio bet with capped downside” applies to virtually any marketing investment at that scale. It’s whether that $10 million generates more value here than in the other places you’ve been spending $10 million. The observation is reductive but directionally useful: the special-ness of the Super Bowl needs to be demonstrated in the data, not assumed from the vibes. But then on the other hand, as Matt Levin puts it, it’s comparably cheap:

One thing that the ads made me think about is how cheap Super Bowl advertising is, for an AI company. A Super Bowl spot costs something like $10 million for airtime plus another few million to produce, for a total at the high end of maybe $20 or $30 million, or roughly the cost of paying one employee for one month at a leading AI lab. Mark Zuckerberg carries around $30 million in his wallet in case he runs into an OpenAI engineer at Starbucks. The cost of creating a cutting-edge AI model — in compute and researcher pay — is astronomical in a way that makes the cost of any advertising, even Super Bowl advertising, look like nothing.

But let’s look at the data.

The CPM looks reasonable. Everything else is complicated.

At $8 million reaching roughly 125 million viewers, the Super Bowl’s effective CPM lands around $63–65 per thousand impressions. Standard primetime TV runs $20–30. Streaming TV sits at $15–35. TikTok charges $5–10. Digiday calculated that for the same $8 million media buy, an advertiser could purchase 1.6 billion TikTok impressions, 267 million Google search impressions, or a primetime network TV spot every night for four months.

But CPM comparisons are misleading here because they treat all impressions as equivalent. They aren’t. The Super Bowl is the last true monoculture event in American media, and the only advertising environment where the ads are the product. People rewatch them, rank them, discuss them at work Monday morning. The Today Show airs them as content. EDO, a TV outcomes measurement company, found that a single Super Bowl ad generates the same brand-search engagement as 1,056 primetime ads.

There is academic evidence on Super Bowl ad ROI

The cleanest causal evidence comes from Wesley Hartmann at Stanford GSB and Daniel Klapper at Humboldt University, published in Marketing Science. Using Nielsen data across 55 media markets and six years of Super Bowls, they exploited exogenous variation in viewership (specifically, ratings spikes caused by local team participation) to estimate causal effects. Their results: Budweiser earned an extra $96 million from Super Bowl advertising, a 172% return on investment. Budweiser’s short-run sales revenue ran 15.75% higher per household than competitors in the weeks following the game.

But Hartmann and Klapper’s most important finding on ad effectiveness is that when two brands in the same product category both advertise, neither gains incremental profit. The effects cancel out. Coca-Cola and Pepsi have both advertised annually in the Super Bowl for years. The researchers found no statistically significant volume increase for Coca-Cola regardless of whether it advertised, and the direction of the coefficients, if anything, suggested a negative relationship. The entire soda category’s Super Bowl spending appears to be a value-destroying exercise that neither side can unilaterally exit.

This is a textbook prisoner’s dilemma. Game theory applied to advertising predicts exactly this outcome: if Bud Light and Coors Light both spend $50 million on ads, they each profit $200 million. If both spend only $10 million, they each profit $240 million. Both rationally choose $50 million.

Anheuser-Busch understood this and paid to avoid it. The company held exclusive beer advertising rights for 33 consecutive years (1989–2022), spending $278 million over a decade partly to prevent competitive neutralization. When exclusivity ended in 2023, the Super Bowl immediately featured nine beer ads from multiple brands. Budweiser’s ROI almost certainly declined.

Stock price studies paint a muddier picture. An MDPI Sustainability study examining 272 ads from 142 firms (2010–2019) found positive cumulative abnormal returns of 2.35% over 10 days post-game. Bridgewise, covering 2021–2024, found the opposite: a portfolio of Super Bowl advertisers underperformed the S&P 500 by 9.2% after six months, with only 25% of individual advertisers outperforming. Kantar’s analysis reports an average ROI of $4.60 per dollar spent, a figure broadly consistent with their multi-year tracking. A Georgia Southern study by Eastman and Iyer found that USA Today Ad Meter likeability scores, the industry’s most-cited metric for judging Super Bowl ads, had no significant relationship with financial effectiveness.

Attribution is difficult

I always wondered how well attribution works. It seems mostly guesswork to me. The evidence suggests this is more right than wrong, though “guesswork” understates the sophistication of modern marketing attribution tools while overstating their accuracy. A 2024 Ascend2 survey found that only 29% of marketers are “extremely confident” in their attribution accuracy. More than a third of CMOs do not fully trust their own marketing data. The problems are real: privacy signal loss from GDPR, CCPA, and iOS opt-outs has degraded observable data. Cross-device fragmentation means customers touch 3–5+ devices before converting. Platform self-reporting creates systematic overcounting, with Google, Meta, and Amazon each claiming credit for the same sale.

For Super Bowl ads specifically, the attribution challenge is amplified by confounding. Brands run concurrent promotions, digital retargeting campaigns, influencer activations, and PR blitzes. Many release ads days before the game. Academic research suggests pricing relative to competition has 20–25x greater impact on sales than total advertising across all channels, which means a coincidental price change during Super Bowl week can wash out the advertising signal entirely.

Cloudflare’s DNS data showed TurboTax saw a 24,875% traffic increase above baseline after its 2024 Super Bowl ad. e.l.f. Cosmetics saw 8,118%. Poppi saw 7,329%. But a Similarweb analysis of 28-day post-game traffic found an average increase of only ~1% across all advertisers. The spike is enormous and ephemeral. YouGov BrandIndex found that only 10 of roughly 50+ advertisers saw positive buzz lift above the margin of error, with a maximum duration of two weeks.

CFO.com’s Hershman had the clearest framing for anyone trying to evaluate this honestly: marketing will come back with impressions, social mentions, and “earned media value,” which he described as Wall Street’s least favorite made-up metric. The only meaningful number is incremental contribution profit. At 40% gross margin, a $13 million all-in Super Bowl investment needs $32.5 million in incremental revenue just to break even on pure acquisition economics.

The NFL’s advertising pricing machine

The NFL operates a price-discrimination machine that has outpaced inflation for 60 years. Super Bowl ad prices have increased from $37,500 in 1967 to $8 million in 2026, a 213x nominal increase and roughly 22–23x in real terms. The compound annual growth rate of approximately 9.6% is more than double average CPI inflation over the same period. Only three year-over-year price decreases have occurred in the entire 60-year history.

The NFL’s leverage comes from a structural scarcity that it actively maintains. Super Bowl ad inventory sells out months in advance. NBC sold out its 2026 inventory before the NFL season even started, with some companies paying $10 million or more due to what NBCUniversal’s Mike Marshall called “the marketplace demand.” Fox reported $800+ million in gross ad revenue from Super Bowl LIX in 2025, a record that industry analysts expect to become a billion dollars within two to three years. The mandatory companion buys force advertisers into additional network inventory they might not otherwise purchase, extracting surplus beyond the headline slot price.

Viewership has cooperated. The Super Bowl drew 51.2 million viewers in 1967 and 127.7 million in 2025. Streaming hasn’t fragmented the audience; it’s expanded it. Tubi alone delivered 13.6 million streaming viewers for Super Bowl LIX, a 94% increase over Fox’s previous Super Bowl. AdImpact data showed streaming at 49% of total viewership, up from 41.5% in 2024. The audience skews younger on streaming: Tubi’s Super Bowl audience was 38% more likely to be 18–34 than the overall game audience, which is exactly the demographic advertisers pay premiums to reach.

The result is a market where the seller has near-monopoly pricing power, the buyers face a prisoner’s dilemma that prevents collective resistance, and the audience keeps growing. The NFL has essentially created a Veblen good in advertising: the price itself signals legitimacy, which makes the price self-sustaining. The 2000 “Dot-Com Bowl” saw 14+ internet companies advertise, using the Super Bowl as a credibility play. At least eight went bust within a decade. The 2022 “Crypto Bowl” featured Coinbase, FTX, Crypto.com, and eToro spending a collective $54 million. FTX collapsed into bankruptcy within nine months. The pattern repeats because the mechanism works: bubble industries pay the premium precisely because appearing in the Super Bowl signals they belong among established brands. That this signal is often false doesn’t reduce its price.

The cases that define the genre

Apple’s “1984” ad cost approximately $750,000–$900,000 to produce plus $800,000 in airtime, roughly $4 million in today’s dollars. Apple’s board hated it and ordered the time sold back. Steve Jobs intervened. The ad generated $155 million in Macintosh sales within three months. Apple sold 250,000 Macs in the first year against a 30,000-unit break-even target. It sits in the Smithsonian.

Coinbase’s 2022 QR code ad cost $14 million for 60 seconds of a bouncing QR code on a black screen. The landing page received 20+ million hits in one minute, crashing the app. Downloads surged 309% week-over-week. The ad won the Clio “Super Clio” and finished dead last in USA Today’s Ad Meter consumer rankings simultaneously. Then the crypto market collapsed, Coinbase laid off 18% of staff, and the massive awareness evaporated. A reminder that advertising cannot fix a product’s relationship to reality.

GoDaddy advertised in every Super Bowl from 2005 to 2015, deliberately courting controversy with provocative ads. Their first appearance generated a 378% website traffic spike and 51.4% share of voice among all advertisers, largely because Fox pulled the second airing and created a news cycle. Today over 60% of visitors go to GoDaddy.com directly rather than through search. The company grew to 21 million customers before going public. Provocation as a launch strategy worked, until the brand matured and pivoted away.

System1 data offers a sobering counterpoint to these highlights: 21% of viewers in 2025 couldn’t recall which brand was behind the ad they’d just watched. That means roughly one in five Super Bowl ads converts millions in ad spend into brandless entertainment. The audience enjoyed the show. They just have no idea who paid for it.

What the economics of a Super Bowl ad tell us

I keep coming back to Hartmann and Klapper’s central result because it’s the one that reshapes how you think about the entire exercise. The Super Bowl ad works brilliantly as an investment, but only when the advertiser has category exclusivity. The moment a competitor shows up, the gains evaporate. What looks like an advertising problem is actually a competitive strategy problem.

Anheuser-Busch paid for exclusivity for 33 years because the company understood this. The $278 million over a decade wasn’t a media buy. It was an entry barrier. The moment that barrier fell in 2023, the category filled with nine competing brands and the collective value of Super Bowl beer advertising almost certainly declined. The NFL captured the difference.

This means the honest answer to “should you buy a Super Bowl ad?” isn’t about CPMs or brand lift or even ROI in the traditional sense. It’s about whether your competitive position allows you to capture the value or whether you’re paying for a prisoner’s dilemma that the NFL designed.

Reitano’s asymmetric upside thesis is logically sound for a company like Ro, which was advertising in a healthcare category without heavy Super Bowl competition and used the spot as a genuine brand awareness play. But the framework breaks down when applied generally. The 2026 Super Bowl featured Novo Nordisk, Ro, Hims & Hers, Novartis, Boehringer Ingelheim, and Eli Lilly all running health-related ads. Northwestern’s Tim Calkins called it the “GLP-1 Super Bowl.” If the Hartmann-Klapper result holds across categories, those brands collectively spent north of $100 million on ads whose effects substantially cancelled each other out.

Ozempic is Reshaping the Fast Food Industry

me@philippdubach.com (Philipp D. Dubach) — Fri, 16 Jan 2026 00:00:00 +0000

Something strange is happening in the food industry. New US dietary guidelines call for more protein and less sugar. Greggs, the UK bakery chain, just warned of “flatlining profits” in the food-to-go market. Food companies are racing to overhaul their brands, ditching artificial dyes and packing protein into products. Earnings calls across the sector blame “inflation” and “subdued consumer confidence.” Nobody mentions the elephant in the room: GLP-1 medications.

New research from Cornell finally puts numbers to what the food industry doesn’t want to discuss. Using transaction data from 150,000 households linked to survey responses on medication adoption, Sylvia Hristakeva, Jūra Liaukonytė, and Leo Feler tracked exactly how Ozempic and Wegovy users change their spending. The results deserve attention from anyone holding food stocks.

The headline: households with a GLP-1 user cut grocery spending by 5.3% within six months. For high-income households, that figure jumps to 8.2%. Fast food takes an even harder hit, with spending at limited-service restaurants falling 8.0%. These aren’t people switching brands or trading down. They’re simply eating less.

The category-level data tells the real story. Savory snacks see the largest decline at 10.1%. Sweets, baked goods, cookies, all down. Even staples like meat, eggs, and bread decline. In the entire grocery basket, only one category shows a statistically significant increase: yogurt. Fresh fruit and nutrition bars trend up slightly, but yogurt is the lone winner with statistical confidence.

As of July 2024, 16.3% of U.S. households have at least one GLP-1 user. The adoption curve is steepening. Nearly half of adopters report taking the medication specifically for weight loss rather than diabetes management. These weight-loss users tend to be younger, higher income, and more willing to pay out of pocket. They’re also the most profitable customers for fast food chains, the ones who don’t flinch at price increases.

This creates what the researchers call a “double whammy” for the food industry. Companies are losing their highest-margin customers to a biological shift in appetite while being left with a more price-sensitive demographic that actually does respond to inflation. When McDonald’s CEO Chris Kempczinski talks about losing lower-income customers to home cooking, he’s describing the wrong problem.

The research also suggests why food executives might be keeping quiet. About 34% of GLP-1 users discontinue within the sample period. When they stop, their spending doesn’t just return to baseline. It becomes less healthy. Candy and chocolate purchases rise 11.4% above pre-adoption levels after stopping the medication.

If you’re running a snack company, the math might look survivable: lose customers to Ozempic for a year, then welcome them back once they quit. The drugs suppress appetite biologically; they don’t teach new habits. When the biology reverts, so does the behavior.

Scott Galloway has called the food industry an “obesity index” and predicted a “tsunami of shareholder destruction.” The Cornell data suggests he’s directionally right but possibly too aggressive on timing. The industry has a built-in buffer: medication discontinuation. The question is whether that buffer lasts as drugs get cheaper, side effects improve, and insurance coverage expands.

The deeper issue is about the persistence of dietary change. Previous studies found that even major life events, a diabetes diagnosis, job loss, childbirth, produce only modest and short-lived changes in diet. Information campaigns and price nudges have mixed results at best. GLP-1 medications work differently because they alter the biological reward system directly. Users describe the experience as “silencing food noise,” a constant background hum of cravings that simply disappears.

But this biological dependence cuts both ways. The changes don’t stick without the drug. Stopping medication means losing both the appetite suppression and whatever habits might have formed during treatment. The Cornell team notes that “GLP-1s could complement existing nutritional interventions” but cautions that “their broader public health relevance ultimately depends on sustained adherence.”

For investors, the practical question is positioning. Companies selling hyperpalatable, calorie-dense products face structural headwinds. Companies selling protein-rich, nutrient-dense foods in smaller portions have tailwinds. The data shows users shifting toward yogurt, fresh fruit, and nutrition bars. Package sizes may need to shrink. Marketing strategies may need to pivot from “craveable” to “satisfying.”

The next few quarters of earnings calls will be interesting. At some point, an analyst will ask the GLP-1 question directly. The honest answer from management would be: we don’t know the full impact yet, but 16% of households having a user, 8% declines in fast food spending, and the fastest-growing prescription category in the country is not something we can ignore.

Does AI mean the demand on labor goes up?

me@philippdubach.com (Philipp D. Dubach) — Thu, 15 Jan 2026 00:00:00 +0000

Joe Weisenthal from Bloomberg, this week:

All my shower thoughts now are about designing efficient workflows for synthesizing, collecting, labeling and annotating data.

Same. Since I started building every app and tool I thought would make my life easier, my workflow more efficient, I haven’t stopped. Apparently non-developers are now writing apps instead of buying them. This is the AI productivity paradox in miniature: the tools get better and we do more, not less.

The assumed narrative is still AI displaces jobs, humans collect UBI, society figures out leisure. But the trajectory might be more work, not less. A recent NBER study found that workers in AI-exposed occupations now work roughly 3 extra hours per week—and leisure time has dropped by the same amount. Upwork’s research puts it bluntly: 77% of employees say AI tools have added to their workload.

The Jevons paradox is 160 years old: when James Watt made steam engines more efficient, coal consumption didn’t fall. It exploded. Efficiency made coal useful in new ways. Satya Nadella referenced this for AI after DeepSeek rattled the markets. Erik Brynjolfsson argues it applies to AI-augmented occupations—coders, radiologists, translators. Make something more efficient and you find more things to do with it.

When I can build an app in a weekend that used to take months, I don’t build one. I build six. When I can write a report in an hour, I write five. The friction that once protected us from infinite expectations evaporates. This is the Jevons paradox applied not just to markets or coal, but to our own time and cognitive capacity—a kind of psychological rebound effect where internal expectations outrun what’s actually sustainable.

Keynes predicted a 15-hour work week by now. We got the productivity gains. We work longer hours than ever. Only 21% of employees actually use the time AI saves them for personal life. The rest reinvest it right back into work. When capability expands, so does the definition of “enough.” The bar rises.

If AI makes me 10x more productive, that’s not 10x more free time. That’s 10x more I could be doing. In a competitive environment—founding, climbing, anything with stakes—someone who uses that 10x while I rest will outrun me. The fear was displacement. The reality might be inescapability.

Parkinson’s Law: work expands to fill time available. The AI corollary: work expands to fill capabilities available. More capability means more possibility—and more obligation. We should know where this points.

Agent-based Systems for Modeling Wealth Distribution

me@philippdubach.com (Philipp D. Dubach) — Sat, 30 Aug 2025 00:00:00 +0000

A question Gary Stevenson, the self-proclaimed best trader in the world, has been asking for some time is if a wealth tax can fix Britain’s economy.

[…] he believed the continued parlous state of the economy would halt any interest rate hikes. The reason? Because when ordinary people receive money, they spend it, stimulating the economy, while the wealthy tend to save it. But our economic model promotes the concentration of wealth among a select few at the expense of everybody else’s living standards.

Owen Jones on Gary Stevenson for The Guardian

Something I generally find very useful and appealing is visualizing systems, models and complexities. Any wealth distribution model is by nature a complex system, and agent-based simulation is one of the best ways to make that complexity visible. The Affine Wealth Model

a stochastic, agent-based, binary-transaction Asset-Exchange Model (AEM) for wealth distribution that allows for agents with negative wealth

elegantly demonstrates how random transactions inevitably lead to Pareto distributions without intervention. In this Jupyter Notebook Fabio Manganiello provides great visualizations of the wealth model. He shows how wealth distributes in an open market where a set of agents trades without any mechanisms in place to prevent a situation of extreme inequality.

It can be seen in the left graph that with no tax wealth quickly stashes up in the pockets of a very small group of agents, while most of the other agents end up piling up in the lowest bucket. As we introduce a wealth tax of 25%, then 5% (right graph) and 1% we can see how the distribution becomes more even and therefore more desirable from the perspective of wealth equality, and also very stable over time, with the agents in the highest buckets quickly having at most 3-4x of their initial amount.

As with any model, the paper as well as the simulation have it’s limitations, but again my interest is more in the way a few lines of code can visualize a economic relationships elegantly. It would be interesting to further investigate: (1) How sensitive are these equilibrium distributions to the transaction constraint (max_exchanged_share)? Does allowing larger transfers accelerate concentration or fundamentally alter the steady-state Gini coefficient? (2) The above wealth tax implementation taxes the sender - but what happens if we model progressive taxation on received amounts above median wealth instead? Does the locus of taxation matter for distributional outcomes?

Behavioral Economics & Transit Policy

me@philippdubach.com (Philipp D. Dubach) — Sun, 22 Jun 2025 00:00:00 +0000

Over the weekend a WSJ editorial on the 2025 New York City mayoral election called one of the potential Democratic candidates Zohran Mamdani “a literal socialist” for - among other things - running on the promise of free bus rides for all:

Zohran won New York’s first fare-free bus pilot on five lines across the city. As Mayor, he’ll permanently eliminate the fare on every city bus […] Fast and free buses will not only make buses reliable and accessible but will improve safety for riders and operators – creating the world-class service New Yorkers deserve.

Free public transit seems to be a recurring idea among politicians: For some reason, making it free feels revolutionary in a way that making it cheaper never could. There’s actually some solid behavioral economics behind this intuition: “Zero as a Special Price: The True Value of Free Products.” (Yes, before the fabricated data scandals, Ariely did write research that has replicated consistently.) The basic finding: people don’t treat “free” as just another very low price. When you price something at zero, it gets a special psychological boost that makes people value it way more than they should based on pure cost-benefit analysis: Give people a choice between a Hershey’s Kiss for 1¢ and a Lindt truffle for 15¢. Most people choose the obviously superior Lindt. Now make it Hershey’s for free versus Lindt for 14¢—keeping the price difference exactly the same—and suddenly everyone wants the Hershey’s. Free doesn’t just eliminate cost; it creates additional perceived value. The mechanism is pure affect. “Free” makes people feel good in a way that “1¢” doesn’t, even though the economic difference is trivial. When you force people to think analytically about the trade-offs, the effect disappears. But in normal decision-making, that warm fuzzy feeling of getting something for nothing dominates rational calculation.

The difference between a $2.75 bus fare and $0 isn’t meaningfully different from the difference between $2.75 and $0.75 for most riders’ budgets. But psychologically? Free transit feels like a gift from the city. Cheap transit feels like commerce. The first activates social norms (gratitude, civic participation, shared ownership). The second activates market norms (cost-benefit analysis, value-for-money calculations, consumer complaints when service is bad). On the other side, any positive price, no matter how small, forces people into analytical mode. People start thinking about trade-offs, evaluating whether the service is worth it, considering alternatives. This is why congestion pricing works so well. A $5 charge to drive in Manhattan (or Singapore, London, Stockholm, Milan, Gothenburg) isn’t going to bankrupt anyone who can afford to drive in Manhattan. But it makes people think about each trip in a way they never did when driving felt “free” (ignoring gas, parking, insurance, etc.). Once you’re thinking analytically rather than just following habit, you’re much more likely to take the subway.

But!! Free transit might actually make it easier to cut transit funding, not harder. Right now, when transit agencies face budget cuts, fare-paying riders get angry. They’re customers! They paid for service! They demand value for money! This creates a natural constituency defending transit budgets. Make transit free, and you’ve eliminated that market relationship. Riders become passive beneficiaries rather than paying customers. When service gets worse, they can’t complain about not getting their money’s worth; they’re getting exactly what they paid for. If I were a politician looking to slash subsidies without political blowback, step one would be eliminating fares. Tell everyone it’s about equity and access. Then, once people stop thinking of themselves as customers, start the real cuts. No more late-night service; hey, it’s free! Longer waits, dirtier stations, broken escalators; what did you expect for nothing? The behavioral economics is clear: when something is free, people have lower expectations and less standing to complain. The zero-price effect works both ways. None of this means transit shouldn’t be affordable!